Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Maintenance rules: every pull request that changes user-visible behaviour (the REST surface, AQL, validation, storage/migrations, configuration, CLI, container/Helm artifacts) adds a fragment under changelog.d/ in the same PR, in the format changelog.d/README.md gives, and never edits Unreleased here (a CI guard enforces both). Cutting a release assembles the fragments into the new version’s section (scripts/release/changelog.sh --assemble), moves the link references on, and tags vX.Y.Z; the release workflow refuses a tag that has no matching section here.

Unreleased

5.7.1-rc1 - 2026-11-13

5.7.1-rc1 is the first release candidate of 5.7.1, published for testing. 5.7.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.7.1 itself replaces it. Operational policy moves out of ferroehr.toml into the CDR as openEHR demographic data. Who may do what is coded content of parties written through the Demographic API, read on every request with no restart and capped by the grant ceiling [authz.grants] allowed: system clients with their imports and exports, the clients persons sign in through and the delegates that act for them, Basic users’ roles and scopes, trusted import systems, data origins, change-feed readers, SCIM clients, federation agreements, just-in-time admission and the privileges beyond clinical access. The audit category map, the access delay and the Cedar policy set become admin registers. The configuration shrinks from 920 keys in 5.7.0 to 855, and the reference names the 10 keys a deployment must set, where 5.7.0 named none; a removed key refuses the start as an unknown key, with no removal aid. Before you try it: run ferroehr config check on your file; name the Basic user who writes the first policy parties in [authz.rbac] first_admin, or give an administrator through a token issuer in role_issuers; then write each client, agreement and grant your file held through the Demographic API or ferroehr party-binding grant, and list each operation in [authz.grants] allowed, as the book’s Operational policy page shows. Representatives of a patient sign in through the patient portal under the representations the deployment records. FerroEHR runs on ferrohealth-smart 0.24.1, ferrohealth-authz 0.17.0, ferrohealth-session 0.19.0, ihe-atna and ehds-logging 0.18.0, ihe-iti 0.9.0 and ferrobridge-term 0.10.0. Party migrations 0033 to 0038 add the reverse-relationship validity, the policy party mark, the grant and client-provider rows, the standalone representations and who once represented whom, and clinical migration 0036 the admin registers. Helm chart 32.0.0-rc1 carries appVersion 5.7.1-rc1. Security carries one advisory, GHSA-gwh2-7hjm-4pj4: an issuer written with credentials in its URL rendered them unmasked.

Added

  • FerroEHR ships a complete default set of demographic definitions under deploy/definitions/: the professional PERSON and ROLE, a care-team GROUP, and the membership and representation PARTY_RELATIONSHIP archetypes the party mapping names, each adapted from openEHR CKM archetypes with a PROVENANCE.md naming its sources. The quickstart, its Keycloak overlay, the hosted sandbox, the development lanes and the conformance harness load all five and map them, so SCIM Users, Groups and representation work out of the box; the Helm chart’s default definitions.files (chart 25.1.0) loads them and leaves the party mapping to the operator. Every release attaches each file with a .sha256sum, and the book’s new “Default demographic definitions” page lists them with the keys that name them. A deployment may load its own templates instead (#5428).

  • A system client imports and exports under its own grants (#5432): message_import_ehr needs an import_ehr grant and message_import_ehr_extract an import_ehr_extract grant, and a system client’s message_export_ehrs and message_export_ehr_extracts need an export_ehr_extract grant, each covering the EHR’s subject namespace in its subject_namespace elements. Each commit of a system client, an import included, records a committer whose external_ref names the client’s ACTOR (PARTY_REF type ACTOR), and its access record names that ACTOR and the identifier of the provider it acts for. A ROLE whose grant lacks the terms of its operation (an origin, a device identifier issuer for a device, a facility’s issuer with one identifier or prefix, a subject namespace for an import or an export) is refused with 422 at its commit. The terminology gains the qualifier codes issuer, identifier, identifier_prefix, origin, relay, device and device_identifier_issuer.

  • Operational policy has a home in the party store, with no configuration key. FerroEHR ships example policy definitions: a system’s AGENT (openEHR-DEMOGRAPHIC-AGENT.system.v1), the acts-for PARTY_RELATIONSHIP to the provider it acts for, and the system-client, federation-partner and privileged ROLEs, each with a CAPABILITY whose credentials hold the grant. The binary carries a default template of each, and each boot stores one only when no definition of its id is stored, never overwriting a deployment’s version. A party is a policy party by the codes it holds, in any archetype, as the entry under Changed describes. Every write of a policy party is validated against its own template. Only a caller holding [authz.rbac] admin_role writes or reads a policy party through the Demographic API, a CONTRIBUTION included; any other caller, a SCIM client too, is answered 403. The book’s “Operational policy” page documents the terminology as the contract, with a worked example of each kind (#5434).

  • The default demographic set gains a patient PERSON (deploy/definitions/patient/), adapted from the openEHR CKM person-patient archetype, with its patient identifiers, date of birth and sex in details, its names named by their usage, a postal and an electronic address, and its relationships to personal contacts, payers and providers, and a healthcare provider ORGANISATION (deploy/definitions/provider/), identified by its provider identifier, which the acts-for relationship targets and a federation partner is. The quickstart, the hosted sandbox, the conformance harness, the viewer lane and the Helm chart’s default definitions.files load all seven definitions, and every release attaches each (#5434).

  • Every authenticated request resolves its caller’s policy grants from the party store, read from coded content in parties of any archetype: a capability grants an operation when its credentials hold a DV_CODED_TEXT of the ferroehr-operations terminology (commit_composition, import_ehr, import_ehr_extract, export_ehr_extract, federated_person_access, federated_organisation_access, read_subject_audit, access_service, read_identifiers), and a client is the party carrying a DV_IDENTIFIER of type client_id under the token issuer. A system client resolves by its token’s issuer and client_id, a bound person by their PERSON, each to the operations of the ROLEs it performs that hold on the request’s date by the ROLE’s and the capability’s time_validity and at its instant by the not_before and not_after elements. A grant changed through the Demographic API decides the caller’s next request on every replica with no restart, a second party naming the same client is refused with 409, and a request whose grants cannot be read is 503. Each replica keeps the grants in memory and reads them again when a commit notifies it, so a request adds no database read. ferroehr party-binding refill rebuilds them. See the book’s “Operational policy” page for the terminology contract (#5435).

  • [authz.grants] allowed (FERROEHR__AUTHZ__GRANTS__ALLOWED), the operations the party store may grant. A ROLE committed with a capability coding an operation it does not list is refused with 403, and a stored one is never honoured. The default, ["commit_composition"], admits system client writes only; an import, an export, a federation agreement and each privilege are admitted by name (#5435).

  • A Basic user bound to a PERSON holds the roles and SMART scopes its policy grants name: the new ferroehr-operations codes hold_roles and hold_scopes grant what the capability’s role and scope elements name, under [authz.grants] allowed. A change through the Demographic API applies to the user’s next request with no restart. A token’s roles and scopes still come from the token alone (#5438).

  • A hold_roles grant naming admin_role or emergency_access_role is refused at commit with 403 and never honoured: the administrator authority and break-glass come from the configuration and the token issuers alone. See the book’s “Operational policy” page, “What the grants decide” (#5438).

  • The ferroehr-operations terminology gains the operations data_origin, read_change_feed and scim_provision and the scope elements creating_system, origin, reader, binding_issuer and token_scope. A change an administrator commits through the Demographic API decides the next request with no restart: the outbox is written while a reader is granted and only that reader’s client reads it, a retention period may be keyed on an origin a grant names, and a SCIM request is admitted for a client whose grant is in force. A ROLE that names one of these elements other than as its operation reads it is refused with 400 (#5439).

  • ferroehr party-binding grant commits a client’s first grant on the operator’s authority, such as the scim_provision of the client that provisions the administrators, who cannot log in before their PERSON exists. The compose quickstart, the Keycloak overlay and the hosted sandbox run it in a ferroehr-grant one-shot, and scripts/scim-provision.sh takes the client’s binding issuer and runs it through SCIM_GRANT (#5439).

  • The audit category map, the EHDS Art. 3(3) access delay and the Cedar policy set are admin registers under {ext}/admin/registers/{audit-categories,access-delay,cedar-policies}: admin_role reads one, replaces it whole over the version If-Match names (412 on a stale write, the new version as the ETag) and lists its history. Every version is kept with its committer, every read and write is an access record, and a change applies on the next request on every replica with no restart. A document that does not hold is refused with 400, so a Cedar policy set that does not parse or validate never replaces the set in force (#5442).

  • Representatives in the patient portal (#5446). A parent, a guardian or a proxy signs in through a patients issuer, as the RFC 8693 act of the patient’s token or under their own login naming the patient’s EHR, and reads the patient’s record while their PERSON holds a live representation of the patient: a PARTY_RELATIONSHIP carrying ferroehr-operations::represents, in any archetype. Who represents whom and until when, including the age and legal-capacity rules of national law, is the representations and time_validity the deployment records; FerroEHR builds in no country’s rules. The patient authorises a proxy with POST /demographic/party_relationship and revokes it with a DELETE, effective on the next request, and every access is recorded bound to the representative, naming the subject.

  • [authz.party_binding.patient] names the template of a patient’s PERSON (template_id), the path of its identifiers (identifier) and the path of its name (name), and defaults to the shipped openEHR-DEMOGRAPHIC-PERSON.patient.v1.0.0. Boot judges it against the definition store, and a PERSON written under that template is validated against it (#5462).

  • GET /ferroehr/rest/extensions/v1/demographic/party/{uid_based_id}/reverse_relationships evaluates the RM 1.2.0 function PARTY.reverse_relationships() in either specification profile: a JSON array of the LOCATABLE_REFs a stable-profile party read carries, as they stand now or, for a version id or a version_at_time, as they stood at that version’s commit. No ITS-REST resource serves the function, so this is a FerroEHR extension (#5473).

  • AQL over the demographic store serves every demographic root: GROUP and AGENT as PERSON is, the abstract PARTY and ACTOR over the concrete classes below them, and a standalone PARTY_RELATIONSHIP. The policy filter applies under every root, through CONTAINS and the join to an EHR, so a caller who is no administrator never receives a system AGENT, a policy ROLE, an acts-for relationship or a party holding one. Each reverse_relationships reference now names exactly one relationship, by its uid, by a unique node id and name, or by its position (#5474).

Changed

  • The book’s command-line page shows how to run ferroehr config check, party-binding check and party-binding adopt in each place FerroEHR runs. On a host you call the binary. Under Docker, docker run … config check checks a mounted file before an upgrade, docker compose run --rm --no-deps ferroehr config check checks the configuration of a service that refuses to start, and docker compose exec runs the database commands. Under the Helm chart, kubectl exec deploy/<release>-ferroehr -c ferroehr runs them, and a one-off pod copied from a failing pod runs config check on a configuration the server refuses. The party-binding migration page links each of its steps to these forms. Each example was run against the 5.7.0 image and chart 25.0.1 (#5416).

  • The FerroHEALTH family crates move to their newest releases: ferrohealth-smart 0.18.0 to 0.19.1, ferrohealth-session 0.13.0 to 0.14.0, ferrohealth-authz 0.11.0 to 0.12.0, ferrobridge-term 0.5.0 to 0.6.0, ihe-atna and ehds-logging 0.12.0 to 0.13.0, and ihe-iti 0.4.0 to 0.5.0. A private_key_jwt client assertion FerroEHR signs for a token request or an onward exchange is now valid for 60 seconds, down from 300, so an authorization server that caps the lifetime below five minutes accepts it. No configuration key changes (#5417).

  • The FerroHEALTH family crates move to their newest releases: ferrohealth-smart 0.19.1 to 0.21.0, ferrohealth-session 0.14.0 to 0.16.0, ferrohealth-authz 0.12.0 to 0.14.0, ferrobridge-term 0.6.0 to 0.8.0, ihe-atna and ehds-logging 0.13.0 to 0.15.0, and ihe-iti 0.5.0 to 0.7.0. FerroEHR’s own checks of [authz.rbac] role_issuers and of the issuer the SMART discovery document describes give way to the shared crate’s: a refused role issuer still renders with its userinfo masked, and the refusal now names only the first refused issuer of the list (#5422).

  • Breaking: the SMART discovery document describes exactly one issuer, never a machines issuer and never by table order. With [auth.smart] discovery = true, 5.7.0 described the first issuer whose metadata it read, whatever its kind, so in the quickstart with the Keycloak overlay the document sent user launches to the machines SCIM issuer, which signs no user token (SMART App Launch 2.2.0, Conformance, §Metadata: authorization_endpoint is where a launch starts). The document now describes the issuer the new key [auth] described_issuer names, else the one issuer that is not of kind machines and whose metadata FerroEHR has (read from it, introspected, or supplied by [auth.smart] issuer_metadata_file). With discovery on, naming a machines, unknown or metadata-less issuer refuses the start. Without the key, more than one such issuer refuses the start, naming them and the key, and none refuses it, saying no issuer can be described. The rule is ferrohealth-smart’s, and described_issuer is the same [auth] key in every FerroHEALTH product. Upgrade step: set [auth] described_issuer when you run more than one user-facing issuer (persons, patients or federation) with discovery on (#5425) (#5426).

  • The quickstart, its Keycloak overlay, the hosted sandbox and the development lanes map the care-team GROUP, the membership and the representation of the default demographic definitions, with a group claim on their [[auth.party]] tables, so a fresh start serves SCIM Users, Groups and representation out of the box. The book’s “Default demographic definitions” page and the chart README show the full mapping; the Helm chart’s own party mapping stays the operator’s (#5428).

  • A system caller no longer imports with no grant (#5432): before, any system/ caller could import an EHR or an extract; it now needs a grant of the import operation covering the EHR’s subject namespace.

  • The default professional definitions follow openEHR CKM more closely: the professional ROLE (openEHR-DEMOGRAPHIC-ROLE.healthcare_professional.v1.0.0) records a clinician’s registrations and qualifications as its CAPABILITY, adapted from the CKM individual_credentials archetype, and the professional PERSON names its name by usage and its email address by type and usage, as CKM codes them, with free text still accepted. Every default definition’s derived_from names its CKM citeable id and modification time, and the representation relationship is re-cited to the person-to-person relationship of the CKM PERSON archetypes. The text of each definition changes under the same id, so a definition store that holds the 5.7.0 text of the professional PERSON or ROLE refuses ferroehr definitions load of the new one with 409 (#5434).

  • Breaking: which clients count as backend clients and which reach the DEMOGRAPHIC API is policy the CDR holds, read on every request with no restart, never a configuration key. [auth.smart] backend_clients and [auth.smart] demographic_clients are removed, and a file or FERROEHR__AUTH__SMART__* variable that sets either is refused as an unknown key. A client’s system/aql-* scope counts when the client’s AGENT performs a ROLE whose capability codes ferroehr-operations::query_system_wide, and the DEMOGRAPHIC API admits a client whose ROLE codes ferroehr-operations::access_demographic, beside a caller holding [auth] product_scope. Both are new operations of the terminology, admitted by name in [authz.grants] allowed, whose default admits neither. A grant counts for the client’s own app-only token; a person signed in through the client holds the person’s grants. To move a deployment, write an AGENT carrying the client’s client_id and a ROLE coding the operation through the Demographic API, and list the operation in [authz.grants] allowed; the book’s “Operational policy” page has the worked example. [auth.smart] patient_issuers stays: it is trust, the issuers whose ehrId claim names an EHR of this platform (#5437).

  • Federation agreements are read from the CDR on every request. An agreement is a grant of federated_person_access or federated_organisation_access on the partner’s party, which carries a client_id identifier for each client its federation issuer issues tokens to; its terms are elements of the capability’s credentials named by ferroehr-operations codes (agreement_name, agreement_version, jurisdiction, exchange, purpose, gdpr_art_6_1, gdpr_art_9_2, national_provision, consent, read_scope, template, discard_window, ehdsi_service, requests_per_second, burst, opt_out_override, access_class, requester_log_retention), and its validity is the grant’s own. A verified external caller is admitted only under the agreement its partner holds for its accessor kind, so a new, changed or ended agreement takes effect on the next request with no restart. Every legal check the configuration ran at boot now refuses the ROLE at its commit with 422, naming each rule it breaks. /management/info and the boot line list every agreement in force with its partner’s issuer and client. The legal checks hold no rule of one country: the jurisdiction is the agreement’s own element, and the consent source is registry or local. See the book’s “Operational policy” page, “Federation agreements” (#5440).

  • Just-in-time creation of a provisional PERSON at a first login is admitted by CDR data instead of configuration. An administrator writes a ROLE whose capability grants ferroehr-operations::create_at_login through the Demographic API; its credentials name the token issuer (login_issuer) and the admission conditions (admit_idp, admit_group, admit_role, admit_assurance, and admit_guests for a guest), and any grant in force admits a login that meets every condition it declares. Each first login reads the grants from the replica’s in-memory copy, so a change applies to the next login with no restart. A grant that names no issuer or declares no condition is refused with 422, and [authz.grants] allowed must list create_at_login, which the default does not. See the book’s “Operational policy” page, “Just-in-time admission” (#5441).

  • ferroehr.default.toml opens with a “What you must set” block: the ten keys a deployment that holds real data sets ([deployment] profile, [[auth.issuer]] issuer, kind and audience, [authz.party_binding.provider] issuer and id, [api] system_id, [database] url, [privacy] subject_namespaces and [record] subject_index). Every other key is marked optional with its built-in default, and each key’s note says why it is in the file: infrastructure, trust, or a tuning key. The configuration types carry the classification (x-ferroehr-class), and the reference generator, run by the config-reference guard, fails on a key with no class, an optional key with no default, or a class entry that names no key. The classification adds and removes no key. The shipped configurations (the quickstart, the hosted sandbox, the development, perf and conformance files, the chart’s values.yaml, values-production.yaml and CI overlays) no longer restate a default: deployment.spec_profile, auth.enabled, [[auth.issuer]] profile = "rfc9068", the default [server] listener and body limit, the chart’s api, database, authz.rbac, authz.abac, authz.party_binding.namespace, management.endpoints, record.signing, record.multimedia and audit.store defaults and the like. The server reads the same values as before (#5455).

  • Breaking (chart): the chart moves to 27.0.0. Its values.yaml no longer carries config.deployment.spec_profile, config.api.openapi_ui, admin, terminology, config.database.migrations, acquire_timeout, statement_timeout, config.authz.rbac, config.authz.abac, config.authz.party_binding.namespace, config.management.endpoints, config.record.signing, config.record.multimedia or config.audit.store, each of which only restated the server’s default. A values file that sets them still works; a tool that reads them from the chart’s defaults finds them unset (#5455).

  • ferroehr.default.toml lists 855 keys, down from 920 in 5.7.0, and marks 10 of them as must-set, where 5.7.0 marked none. 72 keys are gone: the operational policy moved into the CDR (#5432, #5436 to #5441, #5521), three tables became admin registers (#5442), the representation mapping became a code the relationship carries (#5446) and the three public URLs became one (#5465). 7 keys are new, each infrastructure or trust: [server] public_base_url (#5465), [auth] described_issuer (#5425), the three keys of [authz.party_binding.patient] (#5462), the grant ceiling [authz.grants] allowed (#5435) and [authz.rbac] first_admin, the one Basic user that holds admin_role from the file and writes the first policy parties (#5438). Both counts are the ones the file’s “What you must set” block states (#5455).

  • The shipped ferroehr.default.toml sets only the keys of its “What you must set” block: [deployment] profile, [api] system_id and [privacy] subject_namespaces are its live lines, and every other key is a #? line that shows its default. A copy of the file sets 3 keys where it set 115 (122 in 5.7.0); this change removes and adds no key. The book’s “Operational policy” page merges “Which checks read a grant” into “What the grants decide” (#5455).

  • The party mapping defaults to FerroEHR’s default demographic definitions. Each boot stores the seven definitions of deploy/definitions/ whose id the definition store does not hold, as it stores the policy definitions, and never overwrites a stored one. With no mapping key set, [authz.party_binding] person, professional_role, group, membership, representation and the new patient section map them, so an empty configuration boots with SCIM Users, Groups and representation working. A section a deployment sets replaces that section’s default, and nothing is stored under a deployment’s own ids. A SCIM client whose [[auth.party]] table sets no group_identifier_issuer writes its Groups’ externalIds under its binding_issuer with /groups appended, where the start was refused before. The quickstart (both its EHR index variants), the hosted sandbox, the development, conformance, performance and viewer lanes and the Helm chart drop the mapping keys that restated the default and their ferroehr definitions load step, and the chart’s definitions.files is empty by default (chart 27.1.0). A deployment that maps its own templates keeps its keys and its definition list (#5462).

  • The FerroHEALTH family crates move to the releases that take the DPoP external URL from the product: ferrohealth-smart 0.21.0 to 0.22.0, ferrohealth-authz 0.14.0 to 0.15.0, ferrohealth-session 0.16.0 to 0.17.0, ferrobridge-term 0.8.0 to 0.9.0, ihe-atna and ehds-logging 0.15.0 to 0.16.0, and ihe-iti 0.7.0 to 0.8.0 (#5465).

  • A policy party is recognised by what it holds, in any archetype or template, and no longer by its archetype id: a CAPABILITY whose credentials hold a DV_CODED_TEXT of FerroEHR’s operation terminology ferroehr-operations, a PARTY_RELATIONSHIP whose name or details hold ferroehr-operations::acts_for, or a DV_IDENTIFIER of type client_id. A deployment’s own ROLE archetype whose capability holds import_ehr is a policy party, and a party of the shipped policy archetypes that holds none of these codes is an ordinary one. The shipped policy definitions are examples that carry the codes: each policy ROLE requires an Operation element (id90) in its credentials, and the acts-for relationship an acts-for code element (id6) in its details. A 403 for a policy party now names what makes it one in place of the archetype. The book’s “Operational policy” page states the rule (#5485).

Removed

  • Breaking: [[authz.system_client]] is removed from ferroehr.toml, with every key it held (issuer, client_id, provider, facilities, writes, origins, device_identifier_issuers); a configuration that still sets it refuses boot as an unknown key. A system client is now granted in the CDR: an AGENT (or any ACTOR) carrying its client_id, acting for its provider ORGANISATION by an acts-for relationship (or, with none, for the deployment’s [authz.party_binding] provider), and performing a ROLE whose capability grants commit_composition with its templates, facilities, origin (relay or device) and device identifier issuers, each read by its ferroehr-operations code; a change applies on the next request with no restart. The Helm chart 31.0.0 drops config.authz.system_client (a major bump: a values file that sets it is refused), and the Keycloak quickstart grants its service client commit_composition beside query_system_wide through the Demographic API. See the book’s “Operational policy” page (#5436).

  • The example ROLE openEHR-DEMOGRAPHIC-ROLE.data_migration.v1 is removed (#5432): a migration is a system client whose ROLE grants the import and export operations, on the system client example ROLE, which now names each credentials element by its ferroehr-operations code (#5436).

  • Breaking: the client lists ferrohealth-smart read while it classifies a token are policy the CDR holds, read on every request with no restart. [[auth.issuer]] first_party_clients, [[auth.issuer.delegates]] (with client, sub and issuer) and [[auth.scopeless_subjects]] (with issuer and sub) are removed, and a file or FERROEHR__AUTH__* variable that sets one is refused as an unknown key. The server answers the crate’s standing questions from the grants of the ACTOR carrying the client identifier under the token’s issuer, through three new operations of ferroehr-operations: sign_in_persons (a client persons sign in through), act_for_persons (a client, or the RFC 8693 actor a person’s token names, that acts for a person) and access_without_scope (a system caller admitted to every SMART on openEHR requirement without a resource scope). [authz.grants] allowed admits each by name, and its default admits none. The answer comes from the copy of the grant rows each replica holds, so a token adds no database transaction, and a party store that cannot answer refuses the token with 503. To move a deployment, write the client’s AGENT and a ROLE coding the operation through the Demographic API (or ferroehr party-binding grant), and list the operation in [authz.grants] allowed. The Helm chart 32.0.0 drops first_party_clients from its example values (a major bump), and the Keycloak quickstart grants its viewer sign_in_persons. ferrohealth-smart moves to 0.24, ferrohealth-authz to 0.17, ferrohealth-session to 0.19, ihe-atna and ehds-logging to 0.18, ihe-iti to 0.9 and ferrobridge-term to 0.10 (#5437).

  • Breaking: a Basic user is a credential alone. [[auth.basic.users]] roles and scopes are removed, and a file that sets either is refused as an unknown key. A Basic user bound to a PERSON holds the roles and SMART scopes the hold_roles and hold_scopes grants of its PERSON name, never a privileged role. The one trust key added, [authz.rbac] first_admin (FERROEHR__AUTHZ__RBAC__FIRST_ADMIN), names the Basic user that holds admin_role and [auth] product_scope from the configuration: the first administrator, who writes the first policy parties; a name that is no Basic user refuses the start. ferroehr party-binding grant --basic-user <user> grants an operation to the PERSON a Basic user binds to, on the operator’s authority. The quickstart, the hosted sandbox, the development lanes and the performance gate make their Basic user the first administrator and grant its scopes and roles this way (scripts/scim-provision.sh does it for the lanes it provisions), and the development configuration no longer declares import_role. The Helm chart 32.0.0 drops roles and scopes from its Basic users (#5438).

  • Breaking: [authz.import] trusted_systems, [[authz.origin]], [record.change_feed] readers and [[auth.scim.client]] are removed from ferroehr.toml. Each is now a grant in the CDR, read on every request, and a configuration that still sets one refuses boot as an unknown key. An import’s creating_system element vouches for the systems whose wrapped committers it may carry, a data_origin grant names an origin, a read_change_feed grant names a reader of the change feed, and a scim_provision grant names a SCIM client’s binding_issuer and token_scope; [authz.grants] allowed must admit each operation. The Helm chart 30.0.0 drops the outbox-readers line of its notes and documents the grants (a major bump: a values file that sets one of these keys is refused) (#5439).

  • Breaking: [authz.federation] and its [[authz.federation.agreement]] tables are removed from ferroehr.toml (issuer, accessor, name, version, jurisdiction, exchange, purpose, legal_ground, consent, reads, writes, discard, discard_window, ehdsi_service, caller_rate, opt_out_override, access_class, chained_log_retention). A configuration that still holds them refuses boot as an unknown key; write each agreement to the CDR as the book’s “Operational policy” page describes. The Helm chart 29.0.0 drops config.authz.federation.agreement (a major bump: a values file that sets it is refused) (#5440).

  • Breaking: the admission keys of [[authz.party_binding.issuer]] are removed from ferroehr.toml: jit, guests_jit, admit_idp, admit_groups, admit_roles and admit_assurance. A configuration that still sets one refuses boot as an unknown key. Admission is a create_at_login grant in the CDR. The table keeps issuer, user_name_claim and idp_claim (#5441).

  • Breaking: [audit.categories], [privacy.access_delay] and [authz.abac.cedar] are removed from ferroehr.toml. Each is now an admin register, and a configuration that still sets one refuses boot as an unknown key. Write the category map, the access delay and the Cedar policy set through {ext}/admin/registers/{register} once the server runs; until then the shipped category map applies, the delay is off, and the Cedar set is empty, so an ABAC deployment on the embedded engine permits nothing until its policies are written. The Helm chart 26.0.0 drops config.audit.categories (a major bump: a values file that sets it is refused), and the server no longer reads Cedar policy files from a directory (#5442).

  • Breaking: [authz.party_binding.representation] is removed, with its archetype_id and relationship_codes (#5446). A representation is recognised by the ferroehr-operations::represents code it carries in any archetype, and names its openEHR subject_relationship code itself; the section is now an unknown key. The refused-patient-representative refusal is gone: a patients token whose act names a representative is admitted while a live representation binds them.

  • The v5.7.0 configuration removal aids are gone. A key a release before v5.7.0 read, in the server’s ferroehr.toml or the viewer’s configuration, is now an unknown key like any other: the start is refused naming it and the keys its table does read, where v5.7.0 named the key that replaced it. The v5.7.0 changelog still maps every old key to its replacement (#5463).

  • Breaking: [auth.smart] public_base_url, [auth.scim] public_base_url and [auth.dpop] external_url are removed from ferroehr.toml, and [server] public_base_url takes their place. A deployment names the URL clients reach it at once. The SMART discovery document’s baseUrl is that URL with server.base_path and api.base_path, every SCIM meta.location hangs off it with server.base_path and /scim/v2, and a DPoP proof’s htu must name it with server.base_path and the request path (RFC 9449 §4.3). The SMART baseUrl and the SCIM locations now carry server.base_path, which they left out before. The key is required with auth.smart.discovery, [auth.scim] or [auth.dpop], and a configuration that still sets a removed key refuses boot as an unknown key: move its value to [server] public_base_url. ferroehr.default.toml lists two keys fewer. The Helm chart 28.0.0 takes config.server.public_base_url in place of config.auth.smart.public_base_url and config.auth.scim.public_base_url (a major bump: a values file that sets either is refused at boot) (#5465).

  • Breaking: four role names are removed from [authz.rbac]: subject_audit_role, access_service_roles, import_role and identifier_reader_role. A configuration that still sets one refuses boot as an unknown key. Each permission is now a grant in the CDR and nothing else, under [authz.grants] allowed: read_subject_audit reads the access log of one subject (GET /fhir/r4/AuditEvent with one patient), access_service makes a caller an EHDS Art. 10 access service held to an opt-out, import_ehr and import_ehr_extract admit the matching import, and read_identifiers admits GET /admin/party/identifiers beside the admin role. A role a token or a Basic user names grants none of them. Recording an access opt-out is refused with 400 while [authz.grants] allowed does not allow access_service. [authz.rbac] role_issuers now governs admin_role and emergency_access_role alone, and the development compose configuration drops import_role. See the book’s “Operational policy” page, “What the grants decide” (#5521).

Fixed

  • A Web Template keeps the time of a single EVENT that the level removal collapses. ITS-REST simplified formats master04, §“Conditionally Collapsed Wrapper Types”, collapses an EVENT with occurrences 0..1 into its OBSERVATION, and the builder dropped the event’s RM-mandatory time with it (RM data structures, EVENT.time 1..1). FLAT and STRUCTURED had no key for the event time, so a written time was lost and read back as the history origin. The OBSERVATION now carries an in-context time leaf at the event’s own path (…/events[at0002]/time). It builds into that event and reads out of it, as a retained event’s time does. In openehr-sdt (#5421).

  • The worked [authz.party_binding] example of the book’s authentication page names only definitions that exist: its PERSON, ROLE, GROUP, membership and representation ids and paths are those of the shipped default set, and a test boots the example from the page against the files (#5428).

  • ferroehr party-binding refill and the first fill rebuild the client identifiers of every kind of ACTOR, a PERSON, an ORGANISATION and a GROUP included; before, only an AGENT’s were rebuilt, so a client carried by another kind of party held no grant after a refill (#5436).

  • A demographic AQL query by a caller without [authz.rbac] admin_role no longer returns policy parties. A query under any demographic root, alone or under VERSION, leaves out every version of a policy party, together with every node it reaches inside them through CONTAINS or a path. The rule is part of the query’s one SQL statement, so COUNT, the other aggregates and paging count only the parties the caller may read, and no extra statement runs. An administrator’s query returns them, as the Demographic API read does. The book’s “Operational policy” page states the rule (#5450).

  • A specialised ADL 2 archetype whose differential path sits on an attribute below the root (ELEMENT[at0002] matches { /value[at9001]/defining_code matches {...} }) is read from the object that owns the attribute, as AOM2 C_ATTRIBUTE.path defines it. The upload no longer refuses such a path with VDIFP when the flat parent carries it, and the flat form overlays the constraint on the node the path names instead of adding it as a new attribute of the owning object (#5452).

  • Archetype validation accepts a term binding keyed by a path that runs through a use_node internal reference, as AOM 2 VTTBK allows (“a path that is valid in the flat archetype”) and ADL 2 §Paths and Proxy Objects forms it, with the target’s node id. The 2013 CKM openEHR-EHR-OBSERVATION.apgar.v1 no longer fails VTTBK; a CDR whose terminology server serves LOINC still refuses it under VETDF, since it binds LOINC answer codes without their check digit (#5457).

  • Breaking: a party’s inline relationship must be sourced at that party on every write. RM demographic PARTY.Relationships_validity (relationships.for_all (r | r.source = self)) was judged only on a PUT whose body carried no uid. A POST /demographic/{kind}, a demographic CONTRIBUTION version, and a PUT whose body uid named another party’s container committed a relationship sourced at another party, which then showed in that party’s reverse_relationships. Each is now refused 422 naming PARTY.Relationships_validity and the relationship’s source, and commits nothing. A new party’s container id is assigned by the commit, so a create that carries a relationship is refused: add the relationship with an update once the party exists. A PUT, or a CONTRIBUTION version that modifies a party, whose body uid names another container than the one it writes is refused 422 at uid (#5471).

  • Breaking: demographic references must name a party’s version container of the right kind. RM demographic master02-demographic_package.adoc §Party Relationships has relationship references denote “the Version container of a Party” by HIER_OBJECT_ID, role.adoc makes performer the “Reference to Version container of Actor playing the role”, and actor.adoc makes roles the “Identifiers of the Version container for each Role”. An inline relationship whose source or target named an OBJECT_VERSION_ID, a ROLE.performer typed ROLE or naming a version or a party this server does not hold, an ACTOR.roles entry typed PERSON or naming a version or a missing ROLE, and a relationship container between parties that do not exist all committed. Each is now refused 422 on the direct routes and in a demographic CONTRIBUTION, naming the attribute and its path; a reference to a party the same CONTRIBUTION creates, or to a logically deleted one, stands. A relationship container now appears in its target’s reverse_relationships (stable profile), as a reference to the container (#5472).

  • A party read of an older version, by its OBJECT_VERSION_ID or at a version_at_time, listed in reverse_relationships the relationships other parties hold today. It now lists them as they stood when that version was committed, and so do a VERSION<PARTY> read and the representation a create or update returns, which carried no reverse_relationships before. The reverse-relationship index keeps when each relationship held: the party migration 0033_relationship_target_validity rebuilds it from the stored party versions, once, at startup (#5473).

  • A party that ever had a policy version is now hidden from a caller without [authz.rbac] admin_role on both read paths, every version of it. Demographic AQL judged each version on its own, so it served the ordinary versions of a party an administrator had turned from a policy party into an ordinary one, while the Demographic API refused them. The commit that stores a policy version now marks the party in the same transaction, nothing clears the mark, and the party migration marks the parties already stored; an import, an archive load and ferroehr party-binding refill mark from the stored versions. The Demographic API read and write rules and demographic AQL all key on that one mark, which AQL reads from the head row each query already joins, so no statement is added (#5485).

  • An import or export grant can reach an EHR whose EHR_STATUS.subject has no external_ref, such as an anonymous record’s PARTY_SELF, which no grant could reach once [authz.rbac] import_role was removed: a subject_namespace element holding the DV_CODED_TEXT no_subject_namespace of ferroehr-operations covers exactly those EHRs, for import_ehr, import_ehr_extract and export_ehr_extract. A grant naming real namespaces still does not reach them, the code reaches no EHR whose subject names a namespace, and a grant naming no subject_namespace is still refused at its commit. See the book’s “Operational policy” page (#5530).

  • A former patient-portal representative is no longer admitted as the patient once their representation ends or is deleted (#5532). On the own-login path, a login whose PERSON holds or has ever held a representation of the EHR’s subject is judged only as a representative: admitted while a representation is live, refused with refused-representation-unrecorded once none is. FerroEHR keeps who once represented whom in the binding side table, and a refill rebuilds it from every version.

Security

Shipped with functional changes because: the fix of [[auth.attributes]], [[auth.federation]], [[auth.party]] and auth.scopeless_subjects comes only with ferrohealth-smart 0.19, which has no security-only release and changes the public API FerroEHR’s authentication is written against, and three of the FerroEHR keys the fix covers ([[authz.system_client]], [[authz.federation.agreement]] and [[auth.scim.client]]) were already replaced on main by grants the CDR holds, so a security-only patch of 5.7.0 would change code that no later release carries.

  • An issuer written with a user name and password in its URL userinfo (RFC 3986 §3.2.1) showed that password wherever 5.7.0 renders its configuration: ferroehr config check, GET /management/env, GET /ferroehr/rest/extensions/v1/admin/config, ferroehr report, the binding posture and the refusals at start-up (GHSA-gwh2-7hjm-4pj4, CWE-532, CWE-200). The keys that held the issuer as a plain string were [[authz.federation.agreement]] issuer, [[authz.party_binding.issuer]] issuer, [[authz.system_client]] issuer, [[auth.scim.client]] issuer, authz.rbac.role_issuers and auth.smart.patient_issuers, and, through ferrohealth-smart 0.18, [[auth.attributes]], [[auth.federation]], [[auth.party]] and auth.scopeless_subjects. Each of these keys that 5.7.1 still reads holds the issuer as a SecretUrl and renders it as https://***@host/..., as [[auth.issuer]] already did; the others are removed with the move of policy into the CDR. Only a deployment that wrote credentials into an issuer identifier is affected: remove them from every key that names the issuer and rotate the credential (#5418).

5.7.0 - 2026-11-11

5.7.0 follows 5.6.3 as the stable release and carries every change of its four release candidates: every person a login names is bound to a demographic PERSON, provisioned over SCIM or created just in time; the configuration is one table per concern with typed durations and sizes; AQL reads the attestations of a version; and FerroEHR runs on the newest FerroHEALTH family releases, with every outbound client built from the family endpoint table and every URL and issuer rendered with its credentials masked. It breaks deployments, each break with its upgrade step under Changed or Removed. Before you upgrade from 5.6.3: run ferroehr config check on your file, which names every key 5.6.3 read with its replacement, all at once, and remove [authz.professional], which refuses the start; give each persons issuer its [[auth.party]] table and each Basic user its [[authz.party_binding.basic_user]] mapping; give each existing PERSON its professional ROLE with ferroehr party-binding adopt before the first SCIM run, since a PERSON that performs no such ROLE binds no login, and list the persons still without one with ferroehr party-binding check; and add the persons issuers of a deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR only after FerroBRIDGE v0.0.23 reaches it and a FerroFED release carries FerroHEALTH/FerroFED#1615, which no release does yet. The steps for party binding are on the book’s bootstrap and migration page. Party migrations 0027 to 0032, clinical migration 0035 and audit migrations 0023 to 0026 add the binding, SCIM, DPoP replay, refusal and access-record tables, columns and codes. The quickstart, the hosted sandbox and the chart move to FerroPIX 0.0.26 and FerroTERM 0.4.5. Helm chart 25.0.1 carries appVersion 5.7.0. Security carries two advisories, GHSA-jhg2-jwjp-hq74 and GHSA-pfgx-xq28-2629, the zlib fixes in the three images, and the not_affected judgements of the MIT Kerberos CVEs in the PostgreSQL image, the thirteen of MITKRB5-SA-2026-001 among them, and of the five OpenSSL 3.5.7 CVEs of 2026-11-10 in the three images.

Added

  • AQL reads the attestations of a version: v/attestations and any ATTESTATION attribute below it (is_pending, reason, proof, items, committer, time_committed) under FROM EHR e CONTAINS VERSION v CONTAINS COMPOSITION c, with predicates on attestations and EXISTS, and v/commit_audit/is_pending and v/commit_audit/reason on a commit audit that is an ATTESTATION. SELECT v/attestations returns the list, and a comparison below attestations holds when any attestation matches. The FerroEHR extension function LAST(v/attestations/<path>) reads the most recent attestation, so WHERE LAST(v/attestations/is_pending) = true lists the versions awaiting signature, including one marked pending again after it was signed (#4989).

  • The party binding’s configuration, each table with its boot refusals: [authz.party_binding] (provider as one table or one table per tenant, namespace, basic_user), [authz.import] trusted_systems, [[authz.system_client]] (the provider, facilities, templates and origins of each client-credentials client that commits) and [[authz.federation.agreement]] (what a federation issuer’s external persons or organisation may do, on which legal ground), with [authz.rbac] role_issuers and import_role. A template an agreement or a system client writes that the store does not hold refuses the start. The boot line and GET /management/info list the issuers with their kind and identifier issuer, the provider, and every agreement with its accessor and legal ground, and an agreement that admits an organisation without a named person is logged at warn (#5003).

  • [auth.idp.entra], the Microsoft Graph access of the idp-entra adapter of ferrohealth-smart, which resolves the groups of an Entra ID groups overage for an issuer whose profile sets adapter = "idp-entra" (#5003).

  • [auth.dpop] and [[auth.issuer]] sender_constraint: tokens bound to a DPoP key (RFC 9449) or a mutual-TLS client certificate (RFC 8705), checked on every request. A DPoP proof is recorded by the replica that accepted it, so the Helm chart refuses a sender-constrained issuer on more than one replica (#5003).

  • The login binding reads the local party source on every request, with no cache: the new party migration 0027_binding_side_table adds the indexed binding_identifier, binding_membership and binding_representation relations, written in each party commit’s transaction from the PERSONs of the mapped templates that perform a professional ROLE or represent another person, their GROUP memberships and their representation relationships, and rebuilt from the party versions alone. A login matching two parties, an inactive, abandoned or deleted party, a token without the key claim, a party without the needed marker and unreadable groups are refused with 403 and their binding code; a party store or identity provider that cannot answer is answered 503 (#5004).

  • ferroehr definitions load <path>... stores OPT 1.4 (.opt) and ADL 2 (.adls) definitions before serve, through the validation and versioning of the REST upload. A file already stored with the same text is left alone and other text under a stored id is refused, so a deployment runs it at every start to give its party mapping the templates it needs at boot. deploy/definitions/professional/ holds a professional PERSON and ROLE pair adapted from openEHR CKM archetypes for the development lanes and the quickstart (#5006).

  • The compose quickstart binds its user to a professional PERSON: the ferroehr-definitions one-shot loads the PERSON and ROLE templates the party mapping names before the server starts, and ferroehr-provision creates the user through /scim/v2/Users with a token of the development issuer, then reports healthy, so docker compose up --wait returns once the login is bound. The Helm chart (23.3.0) takes definitions.files, which an init container stores with ferroehr definitions load before the server container starts. The conformance, browser and Kubernetes lanes load the same templates and provision every login they use through SCIM, with the committed test client of docs/conformance/party/scim/ (#5006).

  • ferroehr party-binding adopt <PARTY_ID>... gives each prepared PERSON of an existing deployment a professional ROLE of the mapping’s role template, with the PERSON as performer, and a new PERSON version listing it in roles, all in one demographic CONTRIBUTION committed by the first administrator named with --admin or --admin-key, and writes one access record per adopted PERSON through the configured audit trail. It refuses the whole adoption, naming them, while any listed PERSON’s key was introduced by a committer that is no [[scim.client]], named administrator or --migration, so a key planted before the upgrade is never adopted. A PERSON that already performs the ROLE is reported and left alone. The report is JSON on stdout (#5009).

  • ferroehr party-binding check is the read-only pre-flight of a migration to party binding, and the one report of the provisional persons just-in-time creation left incomplete. It reads the party versions alone, so it runs before any boot fill, and writes nothing. Per key issuer it reports the PERSONs that carry a binding key, every selector value two PERSONs carry, the inactive and abandoned ones, the provisional ones older than provisional_report_days, the key-bearing ones that perform no professional ROLE, and every mapped Basic user with the PERSON it binds. For each key it names the committer of the version that introduced it, and fails unless that committer is a [[scim.client]], an administrator named with --admin or --admin-key, a migration declared with --migration, or just-in-time creation on an issuer with jit = true. --expected <file> lists every login of the identity provider with no PERSON. The report is JSON on stdout, and the command exits non-zero on a duplicate, a missing person or a key that fails the provenance rule (#5009).

  • At boot, FerroEHR logs a warn naming every Basic user mapped in [[authz.party_binding.basic_user]] while [authz.assurance] minimum is unset: a Basic credential carries no assurance level, so such a person is never refused on patient data (#5009).

  • The book has a “Who is the committer” page under Security. It covers the committer the server writes for each kind of caller, the placeholder a client may still send, the refusals with their status codes, and the deviation from the ITS-REST merge rule. It also shows how a professional’s PERSON gets its login identifier. Enterprise identity providers now opens with the generic standards path and adds the entra preset’s binding and provisioning settings, an Okta example, a Keycloak realm walk-through and a SAML section. The records of processing and the DPIA gain the workforce-identity activity, the federation node page gains its callers and agreements, and the go-live checklist checks the binding and the caller rates (#5010).

  • The access record of a MyHealth@EU contact point’s request carries the request message’s UUID, the WS-Addressing MessageID its token relays as ehdsi.message_id, as the record’s request-message-id. The record then joins the contact point’s Patient Privacy audit entry (eHDSI Audit Trail Profile 10.1.0 §2.3.12 to §2.3.32) (#5011).

  • The one PERSON mapping SCIM provisioning, just-in-time creation and the login binding share: [[authz.party_binding.person.template]] with the binding key, userName and externalId selectors of each template, [[authz.party_binding.person.attribute]] with the SCIM attribute, the OpenID Connect claim and the path of each mapped value, and [authz.party_binding.professional_role], [authz.party_binding.group], [authz.party_binding.membership] and [authz.party_binding.representation]. Boot refuses a mapping that breaks one of its rules, naming the key, including a template that is absent, rooted at another class or lacks a mapped path, and deleting a template the mapping names is refused with 409 (#5013).

  • The new party migration 0028_binding_selectors holds each binding key, userName (compared after RFC 8265 width mapping, lower case and NFC), externalId and group displayName to one live party, so a commit giving one a second party is refused with 409. The first boot fills the side table under the mapping’s digest, a replica with another mapping refuses to start, and ferroehr party-binding refill rebuilds it after a mapping change, refusing while a replica runs and naming the parties when a value is carried twice (#5013).

  • POST /admin/party/binding-key re-keys a PERSON’s binding key: it refuses the administrator’s own login and a key already bound, retires the old key in the same commit and writes one access record carrying both (#5013).

  • [scim] mounts the SCIM 2.0 service provider at /scim/v2/ (off by default): [[scim.client]] declares each provisioning client by issuer, client_id, scope and binding_issuer, and public_base_url names the origin its documents hang off. /scim/v2/ServiceProviderConfig answers without a token; /ResourceTypes and /Schemas answer a declared client acting for nobody that holds its scope, within [server.caller_rate]. Every refusal is an RFC 7644 Error message. The start is refused for a client of a non-machines issuer, one declared twice, two clients sharing a binding_issuer, or a binding_issuer no [[auth.party]] table carries (#5014).

  • The SCIM endpoint answers the Microsoft Entra ID provisioning client as Microsoft’s “Develop a SCIM endpoint” guide documents it: a value-filtered attribute such as emails[type eq "work"].value serves its type and declares primary read-only, so the primary Entra sends is ignored (RFC 7644 §3.3); a PATCH of userName renames the User; and the Test Connection probes are answered with an empty ListResponse and 404. ferroehr config check refuses a [scim] the person mapping cannot serve, as serve does. The Helm chart exposes config.scim and routes /scim/v2 on every Ingress host when it is enabled, and the book walks an operator through provisioning from Entra ID, with the Workload Identity Federation checks the authorization server makes (#5014).

  • /scim/v2/Users/{id} takes PUT, PATCH and DELETE: a write commits the next version of the professional’s PERSON, clearing a mapped attribute a PUT leaves out, keeping the binding key and externalId, and moving the lifecycle state by active (complete or incomplete to inactive or abandoned and back, an abandoned User completed in two versions, an attribute change to an inactive User refused until it is reactivated). DELETE logically deletes the PERSON and its professional ROLE in one CONTRIBUTION and leaves the binding key as a tombstone. An If-Match naming another version is 412, and ServiceProviderConfig declares PATCH. [scim] max_deactivations and deactivation_window_minutes bound the deactivations and deletions of one client, counted in the database so replicas share the bound, with 429 past it (#5014).

  • /scim/v2/Users serves a client’s Users: POST creates the professional’s PERSON and its professional ROLE in one demographic CONTRIBUTION, complete when the template holds in full and incomplete otherwise, GET /Users/{id} answers with the version’s entity tag and 304 to a matching If-None-Match, and GET /Users answers a ListResponse filtered by eq and and over userName, externalId and id, with any other filter 400 invalidFilter. A client sees only the Users whose binding key its binding_issuer issues, every SCIM request is an access record of the client naming the PERSON, and ServiceProviderConfig declares entity tags. The start is refused when [scim] is on and the party mapping has no default person template with a user_name selector or no professional role (#5014).

  • /scim/v2/Groups serves SCIM Groups as demographic GROUPs of [authz.party_binding.group], each member holding an inline PARTY_RELATIONSHIP of [authz.party_binding.membership] from its PERSON to the GROUP, with no ROLE minted: create, read, eq/and filters on displayName, externalId and id, excludedAttributes=members, PATCH (answered 204), PUT and DELETE, with the entity tags of Users. A Group belongs to the client that created it, its members are that client’s Users, its displayName is unique, and its externalId sits under the client’s group_identifier_issuer. A remove of members with a value list removes only those members, an Entra compatibility deviation from RFC 7644 §3.5.2.2. Deleting a User removes its memberships in the same CONTRIBUTION. [scim] max_group_members (default 5000) caps a Group (#5015).

  • Just-in-time creation of a provisional PERSON at a first login, off by default and enabled per persons issuer in a new [[authz.party_binding.issuer]] table (jit, guests_jit, user_name_claim, idp_claim, admit_idp, admit_groups, admit_roles, admit_assurance). A login of a person or a delegated caller that matches no party and meets every declared admission condition commits one 553|incomplete| PERSON holding the binding key, the userName and the name claims of the access token, with its professional ROLE, as a creation by the person, and is served bound to it; any other mapped attribute is written only where its mapping sets jit = true, and externalId never. Concurrent first requests commit one PERSON, a deprovisioned or inactive login is never recreated, a taken userName refuses the login, and a SCIM client of the key’s issuer completes the PERSON. Boot refuses jit without an admission condition or a mapping that can write it. ferroehr party-binding check lists the provisional persons older than [authz.party_binding] provisional_report_days (default 30) (#5016).

  • A party of a mapped template is validated against that template on every demographic write: a PERSON or GROUP whose archetype_details.template_id names a template of the shared party mapping, and a PARTY_RELATIONSHIP of its membership archetype. A 532|complete| version must meet every constraint, and a 553|incomplete| one may miss mandatory nodes but nothing it carries may be wrong (RM common master06 §Incomplete Content); the refusal is a 422 naming each failing path. Every other party keeps the reference-model checks alone, so parties created through the Demographic API under other templates are accepted as before (#5042).

  • The FHIR AuditEvent, the DICOM audit message and the ITI-20 syslog feed carry the binding fields of the access record: the binding outcome, the party and its groups, the provider acted for, the facilities of a system commit, the representation with its relationship, the actor chain, the external accessor’s asserted fields and the agreement with its legal ground, consent and access class, and a chained request’s correlation id. ehds-logging 0.5.1 writes them as the ehds-accessor and ehds-relayed entities, in the DICOM message as participant objects of type Other. The FHIR record also writes the external and system acting modes. FerroEHR moves to ehds-logging and ihe-atna 0.5.1 and ihe-iti 0.0.77 (#5043).

  • A bound representative reaches only the EHR whose EHR_STATUS names the subject it represents. Every composition, EHR_STATUS, directory, contribution and query request is held to that EHR whatever the token’s scopes cover, and a query must name it with ehr_id. A representative whose token names a subject its party has no live representation relationship to, or whose issuer no [[auth.party]] table binds, is refused with 403 and refused-representation-unrecorded (#5086).

  • The Helm chart (23.1.0) accepts bridge.cdr.oauth.tokenEndpoint for an authorization server whose metadata FerroBRIDGE cannot read at bridge.cdr.oauth.issuer. FerroBRIDGE then requests its client credentials token from that endpoint, and issuer still names the machines issuer the CDR trusts (#5093).

  • A MyHealth@EU contact point’s caller (a contact_point = true federation issuer) reaches patient data only for the patient of its treatment relationship confirmation. The token’s ehdsi.patient.id (system|value) is resolved to its EHR through the EHR index, and every patient-data request must name that EHR. A token without it is refused with 403 and ERROR_SEC_DATA_INTEGRITY_NOT_ENSURED, recorded as refused-trc-missing. Another patient’s EHR, an unknown EHR and a request naming no EHR all answer one 404 of the new problem type trc-patient-mismatch, carrying the agreement service’s *_NOT_FOUND code and recorded as a refused-trc-patient-mismatch security event (eHDSI XCA Profile 8.0.1 §4.2.1, §4.2.2). A contact point’s person agreement now names its ehdsi_service; upgrade step: set it, and list the issuer in [smart] patient_issuers (#5102).

  • An external person under an agreement with discard discards a dispensation of a listed template, as a logical delete, within discard_window_hours and from the same point of care: the dispensation’s health_care_facility carries the token’s provider identifier and its eHDSI organization-id. A token carrying permissions records and discards a dispensation only with PPD-046 (eHDSI SAML Profile 10.0.0 §2.5) (#5102).

  • A read under a federation agreement with a consent condition is served when the token carries the consent decision reference for its patient (Wabvpz Art. 1(j) and Art. 15a(1)), in the claim or authorization_details member the issuer’s [[auth.federation]] table maps to consent.reference, and refused with 403 without it. The reference is one string, the token’s patient/ grant holds it to the launch-context patient, and the access record carries it with the other external fields. ferrohealth-smart moves to 0.9.3, which reads the field (#5106).

  • Every access record of an authenticated request carries the outcome of the party binding (binding: bound, provisional, external, organisation, system, self or a refused-* code) with the fields it needs: the bound party, its groups, the provider, the guest mark, a representative’s representation, a delegated token’s actor chain, a verified external accessor’s fields and agreement, and the facilities a system caller’s compositions named. acting_mode gains delegated, external and system. The audit migration 0023_binding_outcome.sql adds the columns and restores a per-row digest version: records written before stay version 1 and keep verifying, new records are version 2, and a version-1 record after a version-2 one is refused (#5118).

  • Every refused request is an access record through the gated path, and a refusal of the login binding or of a caller kind carries its refused-* binding code and the party it names. Refusals of tokens that failed verification are budgeted by source network and those of verified callers by issuer and subject, so neither crowds out the other; past a caller’s limit its refusals become one summary record with the count and the EHR ids they named, and past a budget one summary record per issuer and outcome is written. Security-event refusals are recorded for every answered request, and a refusal record the system log cannot take is counted in atna_audit_refusals_unwritten. The limits, windows and budgets are kept per replica (#5119).

  • A chained organisation request under a federation agreement is served only when its token carries the requester’s correlation id in the claim or attested authorization_details member the issuer’s [[auth.federation]] table maps to correlation_id. Without it the request is refused with 403 and recorded under the binding code refused-correlation-missing (audit migration 0025); a request header never stands in for it. The organisation access record’s agreement carries the id and its source beside the legal ground and the access class. A chained agreement whose issuer maps no correlation_id, or maps it from a constant, refuses the start (#5120).

  • On the MyHealth@EU contact-point route, the committer of an external person names the professional by the HP Identity Assertion’s NameID: DV_IDENTIFIER{issuer = the assertion's issuer, id = the NameID, type = "hp_identifier:<NameID format>"}, beside the voucher’s oauth2 identifier (eHDSI SAML Profile 10.0.0 §2.1), and a dispensation’s discard is a logical delete under the same committer. A contact point’s COMPOSITION must carry the HCPO’s organization-id in EVENT_CONTEXT.health_care_facility.identifiers, under the issuer urn:oasis:names:tc:xspa:1.0:subject:organization-id, beside the provider’s identifier, or it is refused with 422. ferrohealth-smart moves to 0.9.4, which reads ehdsi.assertion.name_id_format and refuses a contact-point token without one of the three formats as ERROR_HPI_INSUFFICIENT_INFORMATION; upgrade step: a contact point’s authorization server writes the NameID format into its tokens (#5144).

  • A federation caller above its agreement’s caller_rate, a MyHealth@EU contact point’s included, is answered 429 with Retry-After right after authentication and before any patient lookup, keyed on the token’s subject per replica, and every such refusal is counted into the caller’s summary record (#5149).

  • A deleted SCIM User’s contacts are erased from every version of its PERSON, keeping the name, the binding key and the identifiers for the record-keeping duties that need them. [scim] contact_erasure_minutes (default 0, at most 43200) sets when: 0 erases them in the DELETE itself, a period leaves them to a sweep that runs every minute. Each erased version’s node rows are derived again from the new body, its server signature is removed, and the erasure is an access record naming party:<id>#contacts. Party migration 0030 adds the scim_contact_erasure table (#5178).

  • openehr-rm types generic RM type names and paths: typing::RmType parses and prints names such as DV_INTERVAL<DV_QUANTITY>, typing::attribute_type gives an attribute of an instantiated class with the actual argument in place of the formal parameter, and typing::path_type types every step of an RmPath with its container. Each model::RmAttribute and model::RmTypeRef keeps the formal parameter it is written as (formal) beside the resolved bound. paths::write::item_at_path_mut and paths::write::set_at_path write at a path into a canonical instance, positional predicates included, creating the members the path names and changing nothing when a write is refused (#5187).

  • openehr-sdt: flat::navigate::Navigator gives every instance of a Web Template path in a composition, with each step’s occurrence, container index, positional path and JSON pointer, reaching a renamed instance where the template’s name is redundant and an OPT2 archetype root through the complete identifier the template recorded. flat::navigate::relative compares two paths step by step. flat::attribute_keys gives the FLAT key of an RM attribute of a structural class (the master05-rm_mapping.adoc tables, checked against them). A WebTemplateNode carries facts (the embedded template id of an OPT 1.4 archetype root and the terms its archetype defines) and every coded option its terminology, so no consumer walks the operational template (#5189).

  • openehr-base gives identifier and time parts as written: ObjectVersionId::object_id_str and ObjectVersionId::versioned_object_id keep the case of the object_id part, Iso8601Time::fraction_digits and Iso8601DateTime::fraction_digits return the digits of the fractional second as text, and Iso8601Duration::parts returns each component’s designator and number verbatim (DurationDesignator, DurationPart), with Iso8601Duration::fraction_digits for the seconds (#5190).

  • openehr-its reads the ITS-REST ETag as a typed entity tag: rest::etag::EntityTag parses and writes the RFC 9110 form, weak or strong, with the weak and strong comparisons. It types the tag as an ObjectVersionId or a HierObjectId. Every generated response-headers struct that carries an ETag gains entity_tag() beside the raw etag field (#5191).

  • openehr-term names the null_flavours and setting groups: codes::NullFlavour and codes::Setting give each concept’s code and English rubric and read a code back, held equal to the bundled terminology by tests. Every typed group’s ALL is now a slice (#5192).

  • openehr-am: HridRef reads an archetype reference with the version precision it was written at (VersionPrecision: major, minor or full, BASE Artefact Identification §Referencing), gives the version as written and matches an ArchetypeHrid by the specification’s rule, the namespace never a wildcard. ArchetypeHrid::local_semantic_id gives the namespace-free interface form a Web Template nodeId carries, which the v2_4 Web Template builder now uses (#5193).

  • openehr-its: rest::stored_query types a stored query’s qualified name (QualifiedQueryName, [{namespace}::]{query-name}, with check_registrable refusing the reserved aql), its version (QueryVersion, major.minor.patch in numeric order) and the version pattern a request addresses (VersionPattern, exact, {major} or {major}.{minor}, selecting the highest match), per ITS-REST 1.1.0. The generated stored-query operations and the StoredQuery and RESULT_SET documents carry them, so a name or version outside the grammar is a 400 where the request is decoded (#5207).

  • openehr_rm::Generation::earliest_for_composition answers the oldest RM generation whose canonical-JSON decode accepts a composition, so a builder can name ARCHETYPED.rm_version without naming a generation module; a value no generation decodes is a typed NoGenerationDecodes error carrying every generation’s refusal (#5242).

  • The access record takes the purposes of use and the professional roles from the access token. An issuer with an [[auth.attributes]] table states them through the IHE IUA purpose_of_use and subject_role claims (IHE ITI TF-2 §3.71.4.2.2.1.1) or the purpose_of_use of RFC 9396 authorization_details. The new token_attributes column of the audit store holds them, a new record is chain digest version 3, and the EHDS access record and the BALP AuditEvent write them with their code systems. A request whose [audit] purpose_header names a purpose the token does not state is refused with 403, and the refusal is recorded. The header alone keeps working for an issuer without the table (#5323).

Changed

  • The published server, viewer and PostgreSQL images take their org.opencontainers.image.base.name and base.digest labels and annotations from the runtime stage’s digest-pinned FROM line at build time, so a base-image bump changes that one line. An image built directly from a Dockerfile no longer carries these two labels (#4953).

  • [auth] takes the ferrohealth-smart 0.9 shape. Every trusted issuer is an [[auth.issuer]] table with a required kind (persons, patients, machines or federation) and a profile (rfc9068 by default, or entra, keycloak, okta, nl), and holds every setting of its tokens itself. A configuration in the old shape refuses the start, naming what replaces it. Upgrade steps: write a flat issuer = "…" as an [[auth.issuer]] table; move audience, algorithms, clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds, jwks_file, timeout_ms, mode, role_claims and [auth.introspection] onto each issuer table ([auth.issuer.introspection]); give each issuer its kind, with patients and professionals at separate issuers (a realm or tenant each), since one issuer is one kind; replace token_profile = "jwt" with token_types = ["JWT"] and the required_claims the issuer’s tokens carry, and drop token_profile = "rfc9068"; and write each scopeless_subjects entry as { issuer = "…", sub = "…" } (#5003).

  • A persons token is admitted only through a client its issuer lists in first_party_clients or delegates. Upgrade step: list the deployment’s own clients, the viewer’s among them, in first_party_clients, and a third-party client that acts for a person in delegates. The other 0.9 profile keys (person_key, tenant_claim, client_claims, app_only, guest, impersonation_marker, expect, user_claims, person_equality, unstable_claims, max_actor_chain, representation, scope_claim, app_only_scope_claim, adapter, sender_constraint and the onward_* keys) are documented in the configuration reference (#5003).

  • An array of tables, such as [[auth.issuer]], is set from the environment by element index (FERROEHR__AUTH__ISSUER__0__ISSUER), through ferrohealth-config 0.2.10 (#5003).

  • [smart] issuer_metadata_file is the metadata of the [auth] issuer it names, wherever that issuer stands in the list, and a file naming no [auth] issuer refuses the start. The SMART discovery document describes the first introspected issuer, else that named issuer, else the first issuer whose metadata is read from it, else the first issuer, so a Basic users’ issuer listed first no longer takes the discovery document (#5003).

  • The family crates move together: ferrohealth-smart 0.9.1, ferrohealth-authz 0.3.0, ferrohealth-session 0.3.1, ihe-atna and ehds-logging 0.4.0, ihe-iti 0.0.76, ferrohealth-config 0.2.10, ferrohealth-server 0.5.1 and ferrohealth-ui 0.6.9. ihe-atna and ehds-logging now escape every control and bidirectional formatting character of a value they write as its code point. With ferrohealth-ui 0.6.9 the viewer’s session watch no longer throws when the party relationship form, which holds a control named target, is submitted (#5003).

  • The terminology client moves to ferrobridge-term 0.1.143 (FerroBRIDGE v0.0.22), which takes the ferrohealth-smart 0.9 token source (#5003).

  • Every person is bound to a demographic party, with no switch. A persons issuer needs its [[auth.party]] table, two issuers of one node may not share an audience, and [authz.party_binding] provider (the healthcare provider the deployment acts for, EHDS Art. 13(4) and Annex II 3.2(a)) is required once [auth] admits a credential; each refuses the start naming what to set. A Basic user is a person: it maps to the binding key of its PERSON in [[authz.party_binding.basic_user]] (user, key_issuer, key), and a Basic user without a mapping refuses the start, naming the user. Upgrade steps: add an [[auth.party]] table to each persons issuer, set provider, and map each Basic user under the identifier_issuer of an [[auth.party]] table, which needs a persons issuer; a deployment of Basic users alone declares one whose jwks_file is a key set that signs nothing, as the quickstart and the sandbox do (#5003).

  • A token grants [authz.rbac] admin_role, emergency_access_role, import_role and identifier_reader_role only when its issuer is listed in the new [authz.rbac] role_issuers, each a persons issuer; any other kind refuses the start. Upgrade step: list the issuers whose role claims grant these roles; a Basic user keeps the roles its own table grants (#5003).

  • The compose quickstart, the dev and conformance stacks, the hosted sandbox and the Helm chart’s Basic overlays bind their Basic users under a persons issuer that signs nothing: its key set, which every FerroEHR image carries at /opt/sandbox-posture/basic-users-jwks.json, is the public half of a key pair whose private key was discarded (#5003).

  • The quickstart has one copy. docker-compose.yml on main pulls the FerroEHR images at :main and carries main’s configuration, which CI checks with the server image the change builds; the copy attached to each release pulls that release’s images, written by the release lane, which refuses a copy that still names :main. Download the release asset (releases/latest/download/docker-compose.yml) to run a release (#5003).

  • The Helm chart moves to 22.0.0. config.auth.issuer is a list of issuer tables, each with its kind, profile and audience, and values.yaml exposes config.auth.party, config.authz.party_binding, config.authz.rbac.role_issuers, config.authz.system_client and config.authz.federation.agreement. The chart refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling. Upgrade steps: rewrite config.auth as the [auth] upgrade above describes, set config.authz.party_binding.provider, and list the issuers whose roles grant the administrator role in config.authz.rbac.role_issuers (#5003).

  • A PERSON binds a login only while it performs a professional ROLE of the party mapping’s role template that names it as performer; a PERSON that carries the binding key and performs no such ROLE is refused like an unbound one (403, refused-unbound). SCIM provisioning and just-in-time creation commit the ROLE with the PERSON. Upgrade: give each existing PERSON its ROLE with ferroehr party-binding adopt before the first SCIM run, and list the persons still without one with ferroehr party-binding check (#5004).

  • A caller bound to a demographic PERSON commits under the committer the server writes: a PARTY_IDENTIFIED with the PERSON’s name, the login identifier (the issuer profile’s person_key value under the token issuer, type oauth2, or the Basic user under ferroehr, type basic), the provider identifier (type healthcare-provider) and an external_ref to the PERSON, on every commit path, attestations and CONTRIBUTIONs included. Upgrade: stop sending a committer, or send the placeholder {"_type": "PARTY_IDENTIFIED", "identifiers": [{"_type": "DV_IDENTIFIER", "issuer": "<iss>", "id": "<sub>"}]}, which must name the authenticated caller; a stated committer with another name, identifier or external_ref, and a stated PARTY_RELATED, is refused with 403 (#5005).

  • A patient’s committer is PARTY_SELF with no external_ref, and a representative commits as a PARTY_RELATED naming their own PERSON and the coded relationship. A patient-context token commits no demographic version. Upgrade: read the person’s own entries by the committer’s PARTY_SELF type, not by its former external_ref (#5005).

  • Importing an EHR Extract needs the role [authz.rbac] import_role names, or a system/ caller, and a wrapped committer naming a party or a login of this deployment is refused unless its version was created by a system [authz.import] trusted_systems lists. Upgrade: set import_role and give it to the accounts that import (#5005).

  • A person of a persons issuer whose login no demographic party carries is refused with 403 and refused-unbound on every request, with no fallback to a login-only committer. The problem names the identifier issuer to provision the person under, a token without the key claim is refused with a problem naming that claim, and each refusal counts into the new binding_refusals metric by issuer and outcome. The Keycloak quickstart overlay provisions its four demo users through SCIM with the ferroehr-provision-keycloak one-shot. Upgrade: provision every professional (through SCIM or the demographic API) before you add a persons issuer to the upgraded deployment (#5006).

  • A deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR adds its persons issuers only after the bridge and the gateway reach it under the caller: FerroBRIDGE v0.0.23, and the FerroFED release that carries FerroHEALTH/FerroFED#1615, planned for v0.0.20. Until then the committer rule refuses their writes. No FerroFED release reaches the CDR under the caller yet, so a deployment with a FerroFED gateway adds its persons issuers only once one does. The book’s new “Party binding: bootstrap and migration” page carries the bootstrap path, the four migration steps and these prerequisites, and “Breaking changes of party binding” lists every incompatible change with its upgrade step (#5009).

  • The hosted sandbox follows the bootstrap path of party binding: its configuration declares the sandbox-scim client ([[scim.client]], with the development issuer trusted as a machines issuer for it), the box’s new ferroehr-provision service provisions the demo PERSON through /scim/v2/ after every reset, and scripts/sandbox/reseed.sh waits for the demo login to bind before it seeds, then checks the committer of the marker’s EHR_STATUS version references a PERSON. The deploy probe’s new bootstrap family (P-BOOTSTRAP-01) shows the compose quickstart does the same from an empty database: the quickstart user is a User of the SCIM client, and a commit of the Basic user names that PERSON (#5009).

  • With a party mapping configured, the parties the login binding relies on are written through the demographic API by an administrator (authz.rbac.admin_role) only: a PERSON of a mapped template carrying a binding key, userName or externalId, its professional ROLE, a representation relationship, and a GROUP of the group template. Any other caller’s write is refused with 403, a version that changes a present binding key is refused with 403, and an administrator’s version that leaves a protected identifier out is refused with 422. Upgrade: write these parties as an administrator, and change a key through POST /admin/party/binding-key (#5013).

  • FerroEHR records every DPoP proof it accepts in its PostgreSQL database until the proof’s window ends, so a proof is accepted once across all replicas, and a proof replayed to another replica is refused there (RFC 9449 §11.1). If the database cannot record a proof, the request is refused with 503. The Helm chart (22.1.0) no longer refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling (#5072).

  • An AQL query scoped to one EHR now takes the streaming plan shape whatever its ORDER BY, DISTINCT or aggregates, whether the EHR comes from the ehr_id request parameter or from an ehr_id/value equality in the query. On a store of 100,000 compositions such a query used to spend 16 to 64 ms in PostgreSQL’s planner for an execution of about 1 ms, on every new connection and every new statement text; it now plans in 1 to 3 ms (#5076).

  • The Helm chart (23.0.0) connects FerroBRIDGE to the CDR as a machine. FerroBRIDGE uses an OAuth 2.0 client credentials client of a machines issuer that the CDR trusts, set in bridge.cdr.oauth (issuer, clientId, and optionally scope, audience and resource). The client secret comes from the client_secret key of bridge.cdr.existingSecret. The chart no longer accepts a Basic user for this connection: from ferrohealth-smart 0.9 a Basic credential belongs to a person and holds no system/ scope. The chart refuses to render when bridge.cdr.oauth.issuer is not listed in config.auth.issuer with kind: machines. To upgrade, add that issuer, create the client, and replace the user and password keys of the Secret with client_secret (#5085).

  • A verified external person’s commit records the professional as its committer: a PARTY_IDENTIFIED with the asserted name, the hp_identifier under the issuer "<country_code>:<issuing_authority_name>" and the token’s iss and sub, with no external_ref. A country code that is no upper-case ISO 3166-1 alpha-2 code, or an issuing authority holding a :, control or bidi character, is refused with 403 naming the field. The COMPOSITION must name the token’s healthcare provider (provider.healthcare_provider_identifier under provider.issuing_authority_name) as its health_care_facility, or it is refused with 422. Upgrade step: map provider.issuing_authority_name in each [[auth.federation]] table whose agreements carry writes, with a constant source where the token does not carry it (#5097).

  • A system caller (an app-only token by its issuer profile’s app_only markers) holds system/ scopes only: a token carrying a user/ or patient/ scope is refused with 403. It commits into an EHR only as a client an [[authz.system_client]] table declares, under the committer PARTY_IDENTIFIED{name = <client>, identifiers = [{issuer = <iss>, id = <client>, type = "client"}]}, and each COMPOSITION it commits must be of a template its writes lists (403), name a facility its facilities lists in context.health_care_facility, or in feeder_audit.originating_system_audit.location when it has no context (422), and be either a device act from a device client or a relayed person act naming the professional by an identifier from a relay client. Upgrade: declare every client that commits in [[authz.system_client]] with its writes, facilities and origins, and give its compositions the facility identifier (#5100).

  • An attestation is accepted only from a caller bound to a professional PERSON; one from a patient, a representative, a system caller or an unbound caller is refused with 403. Upgrade: attest under the professional’s own token (#5100).

  • A guest of a persons issuer, by its profile’s guest markers, is refused with 403 (refused-guest) unless the issuer’s [[auth.party]] table sets guests = true. Upgrade: set guests = true on the issuers whose guests act under your responsibility (#5100).

  • A verified external person or organisation of a federation issuer is admitted only under the [[authz.federation.agreement]] of its issuer and accessor, holding no scope beyond the agreement’s reads, and is an access-service caller by its kind: an opted-out EHR answers 404 to it, and only an external person under an agreement with opt_out_override reaches one under an emergency purpose. It commits only a COMPOSITION of a template the agreement’s writes lists (an organisation nothing), and a read under a consent-conditioned agreement, and every request under an NL agreement, is refused until the consent decision reference and the Wabvpz Art. 15f roles can be read. Upgrade: declare an agreement for every federation issuer and accessor, list the issuer in [smart] patient_issuers, and give its tokens only the patient/ scopes the agreement reads (#5101).

  • Under an NL [[authz.federation.agreement]], Wabvpz Art. 15f now decides admission in place of the refusal of every request. A caller whose nl.role is the UZI role 01.008 (bedrijfsarts) or 01.048 (verzekeringsarts) is refused with 403 as refused-wabvpz-role, and a caller whose nl.organisation_type is missing or is not a care provider type of Nictiz’s OrganisatieTypeCodelijst is refused with 403 as refused-wabvpz-organisation-type, so an insurer is refused. Every NL agreement declares purpose = "care"; one without a purpose, or for a keuring, refuses the start. The code lists are vendored by scripts/vendor/nl-role-codes.sh into docs/specs/nl-role-codes/. Upgrade: add purpose = "care" to every NL agreement and map nl.organisation_type (and nl.role for persons) in its issuer’s [[auth.federation]] table (#5107).

  • A delegated token’s access record now has acting_mode = "delegated" where it had client, and a system caller’s system. The FHIR rendering still names a delegated access as a client one. Upgrade: a review query that selects delegated access by acting_mode = 'client' also selects 'delegated' (#5118).

  • FerroEHR builds on ferrohealth-server 0.6.1, whose public caller-rate take holds a federation caller to its agreement’s caller_rate. Its 0.6.0 break, the revocation feature’s move to der 0.8, touches no feature FerroEHR enables (#5149).

  • The source tree no longer carries the UZI role code CodeSystem (uzi-rolcode) of Nictiz’s zib 2020 package: its codes and titles reproduce the UZI-register CPS, which CIBG does not license for reproduction. scripts/vendor/nl-role-codes.sh writes it into the git-ignored .vendor-cache/nl-role-codes/, as it does the CPS, and --test-inputs fetches only what the Wabvpz tests read. Nictiz’s care provider types stay under docs/specs/nl-role-codes/. Run scripts/vendor/nl-role-codes.sh --test-inputs once before running the ferroehr-rest tests (#5175).

  • A contact point’s patient.id is read through ferrohealth-smart’s TrcPatientId, the family’s one definition of the TRC patient. A value with whitespace in its system, or a value holding a control character, | or a CX delimiter (^, &, ~, \), is refused as a missing binding, as the contact point that writes the claim refuses it (#5195).

  • [auth] takes three boot refusals from ferrohealth-smart: a Basic user’s hash below the Argon2id floor (m=19456, t=2, p=1), an issuer that carries userinfo, and private or symmetric key material in a jwks_file. They are refused as before, with the crate’s messages. A jwks_file that is no JWK Set now refuses the start where the crate reads the key set (#5209).

  • The family crates move together to one ferrohealth-smart: 0.13.1, with ferrohealth-authz 0.7.0, ferrohealth-session 0.7.0, ihe-atna and ehds-logging 0.8.0, ihe-iti 0.0.83, ferrohealth-server 0.6.4 and ferrobridge-term 0.1.148. The 0.13 line refuses a JSON document it cannot read without quoting it (GHSA-mhv7-3975-6xq3), so a jwks_file or [smart] issuer_metadata_file that does not parse is refused at boot with the crate’s message, which names the failure class and position; 0.13.1 keeps a refused key set, introspection or metadata body out of the error (GHSA-5qm7-v4fw-p2v9). The OAuth client credentials tables ([authz.abac.remote.oauth], [ehr_index.oauth], [terminology.oauth] and [audit.fhir_feed.oauth]) read as before (#5209).

  • The API’s response security headers come from the family server stack (ferrohealth-server 0.6.2) with the same six headers and values. They now also reach the management listener on its own port, and Strict-Transport-Security: max-age=31536000 is sent when FerroEHR or the viewer serves TLS itself ([server.tls], RFC 6797 §7.2). The viewer gains Cross-Origin-Resource-Policy: same-site (#5210).

  • The server and the viewer read the parts of an OBJECT_VERSION_ID through openehr-base and its BASE grammar, never by cutting at ::. A value that is no OBJECT_VERSION_ID now names no version: the AQL branch refusal, the version item tags and the viewer’s version numbers leave it alone, a reference to a container is read as the whole value (so uuid::extension no longer names the container uuid), and the viewer’s deletes refuse it before calling the CDR. Every part keeps its case (#5211).

  • ETag and If-Match values are read through openehr-its’s RFC 9110 entity tag on the server and in the viewer. The server accepts the same forms as before. The viewer reads its CDR’s ETag strictly: a value that is no entity tag names no version (#5212).

  • The viewer’s audit screen reads the ITI-81 searchset and each AuditEvent as FHIR R4 resources, and a refusal’s OperationOutcome too. A searchset that is no FHIR R4 Bundle of AuditEvents now shows as a failed read, never as a page of empty rows (#5214).

  • The EHR index’s boot check reads the PIX Manager’s CapabilityStatement through ihe-iti’s PIXm client, with its transport and its client-credentials token, so a Manager behind IUA answers the probe as it answers ITI-83. A failed probe is reported as the client’s refusal: no answer, the status it answered, or no FHIR R4 CapabilityStatement (#5216).

  • [audit.syslog] reads its TLS material from [audit.syslog.tls], the outbound TLS table of every FerroHEALTH product (trust_roots_file, client_identity_file or client_identity, min_version), and hands ihe-atna the rustls configuration ferrohealth-server builds from it. The keys tls_ca_file, tls_identity_cert_file and tls_identity_key_file are retired and refused at boot by name: name the roots file as trust_roots_file, and concatenate the client certificate chain and its private key into the one PEM client_identity_file names. The client speaks TLS 1.3 unless min_version = "1.2" lowers it (#5217).

  • A SCIM Group membership is now held by the member: an inline PARTY_RELATIONSHIP from the member’s PERSON to the GROUP, stored in the PERSON’s version (RM demographic §Party Relationships). A membership change versions only that PERSON, so its cost no longer grows with the Group’s size. The GROUP is versioned only by a rename, an externalId change or its deletion. The Group’s members is read from the membership index (#5220).

  • A Group’s ETag is now its GROUP version plus a membership revision, so it still changes with every membership change. The member User’s ETag now changes too, while its body stays the same (#5220).

  • Deactivating a User (active = false) drops all of its memberships in the same version, because an inactive PERSON cannot be versioned again. A reactivated User belongs to no Group until its client adds it again (#5220).

  • Microsoft Entra ID’s disable-then-remove order now succeeds, since the later remove finds nothing to remove. An add naming an inactive User is 400 invalidValue (“reactivate the User first”) (#5220).

  • A refused SCIM write now names the rule it breaks in detail (#5220).

  • scripts/perf/scim-groups.sh keeps the database log, df -h and free -m of a failed run (#5220).

  • A party commit no longer re-derives the login binding’s side table, the reverse-relationship index and the protection check from the whole stored body. Each hook reads what the committed body holds, derived once in memory, and writes only the rows that changed, and ferroehr party-binding refill derives the rows with the same code. A party commit without a server signature no longer reassembles its body. A membership change at the cap of a 5 000-member SCIM Group went from 810 ms to 358 ms at p99 on a developer laptop (#5220).

  • openehr-rm writes the RFC 8785 canonical form a VERSION signature covers in one pass over the version, sorting each object’s members by reference. A 3.3 MB GROUP version canonicalises about seven times faster than through serde_jcs, with byte-identical output, so signed versions verify unchanged. serde_jcs remains the test oracle (#5220).

  • The stored-query service takes the typed name, version and version pattern of openehr_its::rest::stored_query end to end, so it reads one grammar with the REST layer: the three-part {namespace}::{formalism}::{query-name} name, which ITS-REST 1.1.0 Qualified_query_name.md does not define, is no longer accepted anywhere (the formalism is the query_type parameter), and DELETE /admin/query/{qualified_query_name}/{version} answers 400 for a name outside [{namespace}::]{query-name} or a version that is no exact major.minor.patch, where it answered 404 (#5223).

  • Boot and ferroehr config check refuse an [[auth.party]] table whose identifier_type is the type of no [[authz.party_binding.person.template]] key, naming the issuer, the type and the types the templates write: no login of such an issuer could ever match a party, and every one would be refused. Upgrade step: give the deployment a person template whose key carries the table’s identifier_type (or set the table to the type the template writes). The Helm CI and production values and the hosted sandbox now carry the professional PERSON and ROLE mapping, and the sandbox loads the two templates before the CDR starts (#5231).

  • openehr_query::printer::to_aql writes each WHERE comparison with one space on each side of the operator (c/name/value = $nameValue), as the QUERY 1.1.0 AQL syntax examples do; path predicates keep their unspaced form ([ehr_id/value=$ehrUid]). The query builder’s generated AQL follows (#5241).

  • openehr_sdt::flat::navigate::Navigator owns what it copies from its Web Template (the name-fallback table and the OPT2 archetype references) and carries no template lifetime, so a consumer can keep it beside the template it was built from; it is Debug, Clone, Send and Sync (#5243).

  • Breaking: the configuration tree is one table per concern. ferroehr.toml moves to [deployment], [api], [database], [auth] (with [auth.smart] and [auth.scim]), [authz] (with [[authz.origin]]), [management], [audit], [privacy] (with identifier_protection, cohort and access_delay), [record] (with subject_index, versioning, signing, query, change_feed and multimedia), [terminology] and [pixm]. FerroEHR’s own keys take one duration type ("500ms", "30s", "10min", "12h", "7d", UCUM unit codes, so a bare number and m are refused), one size type ("256KiB", "64MiB", IEC binary units, decimal units refused) and one calendar retention period ("P5Y", "P1830D", unset keeps forever). An optional feature is on when its table is present: [terminology], [pixm], [auth.scim], [audit.syslog], [audit.fhir_feed], [privacy.identifier_protection] and [privacy.access_delay] carry no enabled; [authz.abac], [record.multimedia] and [auth.smart] keep a switch because some of their keys apply while off. record.subject_index has no default, so the server refuses to start until "pixm" or "local" is chosen. [api.cors] replaces cors_permissive with the Fetch Standard response headers (allow_origins, allow_methods, allow_headers, expose_headers, allow_credentials, max_age). The environment follows the tree (FERROEHR__DB__URL is FERROEHR__DATABASE__URL, FERROEHR__EHR_INDEX__PROVIDER is FERROEHR__RECORD__SUBJECT_INDEX). The Helm chart moves to its 25 line with config: on the same tree. Every key a release before v5.7.0 read is refused at load with its replacement named, all at once, so ferroehr config check --config <old file> prints the whole move list; the refusals ship in v5.7.0 only and the next release removes them. To upgrade, run config check on the old file and move each key it names: spec_profile, deployment_profile, deployment_accepts and licence.file are deployment.spec_profile, deployment.profile, deployment.accepts and deployment.licence_file; [rest] is [api] (swagger_ui is openapi_ui, body_limit_bytes is body_limit, cors_permissive = true is [api.cors] allow_origins = ["*"], [rest.identity] is [api.options] with solution and vendor; the version keys are facts of the build), admin.enabled is api.admin and terminology.api_enabled is api.terminology; [db] is [database] (acquire_timeout_secs is acquire_timeout, statement_timeout_ms is statement_timeout, migrate is migrations), [storage.<domain>] is [database.<domain>] and [storage.migrate.<domain>] url is [database.<domain>] migrate_url; [smart] is [auth.smart] (enabled is discovery, episode.enabled is episode_launch, launch_base64_json = true is "launch-base64-json" in auth.capabilities); [scim] is [auth.scim] (max_results is filter_max_results, deactivation_window_minutes is deactivation_window, contact_erasure_minutes is contact_erasure_delay); [authz.abac.cedar] reload_secs is reload_interval, [authz.abac.remote] server, connect_timeout_ms and request_timeout_ms are url, connect_timeout and timeout, authz.party_binding.provisional_report_days is provisional_report_after, and an agreement’s discard_window_hours and chained_log_retention_years are discard_window and chained_log_retention; [[origins.registered]] is [[authz.origin]]; management.port is management.listen (an address:port) and management.profiling.max_seconds is max_duration; audit.store.retention_days and retention_years are audit.store.retention, verify_interval_seconds and full_verify_interval_seconds are verify_interval and full_verify_interval; [audit.syslog] host and port are url = "tls://host:port", spool_max_messages is spool_max_events, spool_max_bytes is spool_max_size, retry_max_ms is retry.max_backoff, and [audit.fhir_feed] poll_interval_ms and retry_max_ms are poll_interval and retry.max_backoff; [demographic.identifier_protection] is [privacy.identifier_protection], [cohort] is [privacy.cohort], and [access_delay] is [privacy.access_delay] with hours as max_delay; ehr_index.provider is record.subject_index ("ihe" is "pixm") and the rest of [ehr_index] is [pixm] (pixm_url is url, invocation is method, request_timeout_ms is timeout); versioning.skip_identical_resend is record.versioning.idempotent_resend; [signing] is [record.signing] (key_path is key_file, retired_key_paths is retired_key_files), [query] is [record.query] (timeout_ms is timeout), [outbox] is [record.change_feed] (retention_days is retention, prune_interval_secs is prune_interval), and [multimedia] is [record.multimedia] (endpoint is url, threshold_bytes is threshold, and allow_http is retired: an http:// url is plain HTTP); [terminology.server] is [terminology] (fail_on_error is on_error = "accept" | "refuse", request_timeout_ms is timeout). The family sections take the same types, from ferrohealth-server 0.8, ferrohealth-telemetry 0.4, ferrohealth-smart 0.15 and ferrohealth-session 0.9: [server] request_timeout_ms, header_read_timeout_ms, idle_timeout_ms, drain_delay_ms, shutdown_timeout_ms and retry_after_s are request_timeout, header_read_timeout, idle_timeout, drain_delay, shutdown_timeout and retry_after, body_limit_bytes and compression_min_bytes are body_limit and compression_min_size, and [server.tls] reload_interval_s is reload_interval; [telemetry.otlp] trust_roots_file and client_identity_file move into [telemetry.otlp.tls], and an https:// collector is dialed over it; [auth] ca_file is [auth.tls] trust_roots_file, an [[auth.issuer]] table’s clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds and timeout_ms are clock_skew, jwks_max_age, jwks_refresh_interval and timeout, its introspection cache_max_age_seconds is cache_max_age, [auth.basic] cache_ttl_seconds is cache_ttl, [auth.dpop] max_age_seconds and clock_skew_seconds are max_age and clock_skew, [auth.idp.entra] timeout_ms and max_response_bytes are timeout and max_response_size, and an oauth grant’s ca_file, client_certificate_file and client_key_file are its tls table’s trust_roots_file and client_identity_file (one PEM), with refresh_margin_ms, timeout_ms and previous_key_overlap_ms as refresh_margin, timeout and previous_key_overlap. The [auth] keys v5.6 read for every issuer (audience, algorithms, token_profile, jwks_file, mode, introspection, role_claims and the rest) live on each [[auth.issuer]] table, and scopeless_subjects entries are { issuer, sub } tables. In the viewer, cdr.request_timeout_secs is cdr.timeout, [oidc] ca_file is [oidc.tls] trust_roots_file, [oidc.tls] client_certificate and client_key are client_identity_file, and [session] idle_timeout_s, absolute_timeout_s and sign_in_timeout_s are idle_timeout, absolute_timeout and sign_in_timeout. Each of these is refused in v5.7.0 naming its replacement, in the CDR and the viewer alike (#5259).

  • Breaking: a patients token that names an actor is refused. With ferrohealth-smart 0.15, a patients issuer reads a token whose top-level act (RFC 8693 §4.1) names someone other than the patient only when its [[auth.issuer]] table sets act_represents_patient = true; without it the token is refused 401, where v5.6 admitted it as the patient’s own access. With it, FerroEHR refuses the token 403 (refused-patient-representative), because it binds no party to someone acting for a patient, and the access log records the refusal (#5259).

  • SECURITY.md, security.txt (its Contact: mailto: field), the post-market record, the CRA technical documentation and the book’s compliance pages name security@cadasto.com as the email route for a vulnerability report, in place of info@cadasto.com, which stays the single point of contact for complaints and incidents, the manufacturer’s address on the running server and the address for licensing and other business questions (#5272).

  • The shipped ferroehr.toml template (app/ferroehr/assets/ferroehr.default.toml) is generated from the configuration types with tools/config-reference: every key, each under its description and set to its default, a secret or a key with no default as a #? placeholder line. The book’s new page “Every configuration key” includes it whole. CI fails when it drifts from the types. No key changes: a ferroehr.toml that loads today loads unchanged (#5316).

  • The family crates move to their newest releases: ferrohealth-smart 0.18.0, ferrohealth-session 0.12.0, ferrohealth-authz 0.11.0, ferrohealth-server 0.8.4 (its TLS certificate verifier is built once per process), ferrohealth-telemetry 0.4.1, ihe-atna and ehds-logging 0.12.0, ihe-iti 0.4.0, ferrobridge-term 0.5.0, and fhir-types and fhir-search 0.1.131. Every [[auth.issuer]], onward and OAuth issuer renders with its credentials masked. A system caller, a patient and a patient’s representative carry the purpose of use their token states when their issuer has an [[auth.attributes]] table. An IUA purpose_of_use of empty codes states no purpose (#5320).

  • ferroehr and ferroehr-viewer run through the FerroHEALTH family’s run path. A configuration either binary refuses (an unknown or invalid key, a broken semantic rule, an [auth] table with no mechanism, a SCIM mapping it cannot serve, or for the viewer a TLS file that does not read or an unusable [oidc] or [session] table) now exits 78 (EX_CONFIG of sysexits(3)) where it exited 1, so an orchestrator can tell a configuration to fix from a failure to retry. Every other failure still exits 1, and the failure is written to stderr as one line with its whole cause, prefixed with the binary’s name. Upgrade: a script or probe that expects exactly 1 from config check on a refused file expects 78 (#5321).

  • The configuration is redacted by its secret types alone. GET /management/env, GET /ferroehr/rest/extensions/v1/admin/config and ferroehr config check show the tree as its Secret and SecretUrl values serialize, and the second masking by key name and by URL shape is gone. [audit.syslog] url, [auth.smart] platform_base_url and public_base_url, and [auth.scim] public_base_url are secret URLs now, as are the OAuth, onward and introspection endpoints of ferrohealth-smart 0.17 and the viewer’s [oidc] endpoints of ferrohealth-session 0.11, so each renders with its credentials and query masked. A platform_base_url given as a bare path renders as *** (#5322).

  • Every outbound integration (the terminology server, the PIXm identity source, the ITI-20 FHIR Feed, the remote policy decision point and the viewer’s CDR client) is built by ferrohealth-smart’s OutboundClient from its endpoint table, and FerroEHR’s own client builder is gone. A PIXm exchange’s timeout now bounds the exchange with its tries and the waits between them, and the viewer refuses a [cdr] timeout of zero at start (#5324).

  • The family crates move to their newest releases: ferrohealth-smart 0.17.0, ferrohealth-session 0.11.0, ferrohealth-authz 0.10.0, ferrohealth-config 0.4.4, ferrohealth-server 0.8.3, ferrohealth-secret 0.2.1, ferrohealth-usage-report 0.4.1, ferrobridge-term 0.4.0, ihe-atna and ehds-logging 0.11.0, and ihe-iti 0.3.0. Every OAuth token endpoint and introspection endpoint renders with its credentials masked, and a TLS refusal of an outbound client names its <key>.tls table (#5324).

  • [auth.idp.entra] is read by ferrohealth-smart’s [auth] section, so an unknown key under it is refused with the crate’s field list. The key path is unchanged. The effective configuration now shows [auth.idp] as an empty table when no adapter is configured (#5325).

  • The build compiles FHIR R4’s resources alone: fhir-types is required with its r4-resources feature in place of the crate-wide resources, and the FerroBRIDGE crates (fhir-types, fhir-search, ferrobridge-term) are required by caret ranges in place of exact versions (#5327).

  • Breaking: the outbound integrations take the family’s endpoint table. [terminology], [pixm], [audit.fhir_feed] and [authz.abac.remote] read where they call and how from one [<x>.endpoint] table, the same keys for every outbound integration of every FerroHEALTH product: url, timeout, connect_timeout, retry (max_attempts, initial_backoff, max_backoff), tls (trust_roots_file, client_identity_file, min_version) and credentials (HTTP Basic user with password or password_file, or the client credentials grant [...credentials.oauth]). terminology.url, timeout and [terminology.oauth] are terminology.endpoint.url, timeout and [terminology.endpoint.credentials.oauth]; pixm.url, timeout, [pixm.tls] and [pixm.oauth] are pixm.endpoint.url, timeout, [pixm.endpoint.tls] and [pixm.endpoint.credentials.oauth]; audit.fhir_feed.url, [audit.fhir_feed.retry] and [audit.fhir_feed.oauth] are audit.fhir_feed.endpoint.url, [audit.fhir_feed.endpoint.retry] and [audit.fhir_feed.endpoint.credentials.oauth]; authz.abac.remote.url, connect_timeout, timeout and [authz.abac.remote.oauth] are authz.abac.remote.endpoint.url, connect_timeout, timeout and [authz.abac.remote.endpoint.credentials.oauth]. The v5.6 keys are refused naming their new home. Every client now honours retry: the terminology, PIXm and remote policy clients try a call that got no answer or a 5xx up to max_attempts times (three by default; each tried once before), so set max_attempts = 1 to keep one attempt, and the FHIR feed’s delivery rounds take max_attempts tries each, from initial_backoff, before the round waits max_backoff; the feed still never gives a record up. The environment follows the tree (FERROEHR__TERMINOLOGY__ENDPOINT__URL, FERROEHR__AUDIT__FHIR_FEED__ENDPOINT__URL), and so do the Helm chart’s values and its client secret routing, in the chart’s 25 line (#5331).

  • The FerroEHR PostgreSQL image is built on the 2026-11-09 respin of postgres:18.6 (sha256:e3d24f69…), a same-version respin with rebuilt base packages. A full scan of the rebuilt image on amd64 and arm64 finds the same findings as the image of the previous base, each judged in security/vex/postgres-os.openvex.json, and a gate scan finds no fixable HIGH or CRITICAL finding (#5332).

  • The development and Keycloak example compose files run Keycloak 26.8.2 (quay.io/keycloak/keycloak:26.8.2@sha256:723b1d6e…). Of the 26.8.2 upgrading notes, the shipped realm meets the ones that touch it: the Admin REST API’s new audience check passes for the master realm’s admin-cli token the browser-test harness uses, and the realm uses no token exchange, identity broker, SAML client, CIBA or dynamic client registration. The ferroehr-service client-credentials token carries the same audience and scopes as under 26.8.1 (#5333).

  • The compose quickstart, the hosted sandbox and the Helm chart deploy FerroPIX 0.0.26 (ghcr.io/ferrohealth/ferropix:0.0.26@sha256:85d40e6fdcdb39f782007eff91eb059861a4e039ca3f8fb56e37f2fa0f28d78a, the FerroPIX chart 0.3.19) and FerroTERM 0.4.5 (ghcr.io/ferrohealth/ferroterm:0.4.5@sha256:7c20d39e1989fce1dfb92e6ddfd853dfa9a52bb9b7b973fbba42430963ef75bd), and the development issuer is mock-oauth2-server 6.1.5. FerroPIX’s configuration moves to the [[auth.issuer]] shape it now requires: the issuer is kind = "machines" and carries its own audience, and [auth] audience is gone. A chart deployment that sets ferropix.config.auth moves its audience onto each issuer the same way (#5351).

  • The nine openehr-* crates follow The Cargo Book’s SemVer rules, judged by cargo-semver-checks 0.51.0 against their newest release on crates.io on every pull request, in the merge queue and in the release plan. Their next publish moves the line from 0.0.x to 0.1.0. From then on a compatible change ships as a patch and only a breaking change moves the minor, so a caret requirement such as openehr-rm = "0.1" takes every compatible release. A public dependency moved to another compatibility line counts as a breaking change (#5353).

  • openehr-rm writes the ECMAScript number form of the RFC 8785 canonical form a VERSION signature covers with ryu-js 1.0, which carries the upstream fix for a panic in its fixed-point formatter, and is held byte-equal to serde_json_canonicalizer 0.4 as its test oracle in place of serde_jcs, so the lock holds one ryu-js line. Its output is unchanged (#5372).

  • ferroehr.eu takes the FerroHEALTH family layout: the “Oxide & Iron” brand tokens light and dark, a hero badge rendered from the workspace version at build time, a content hash on the stylesheet link, linked capability cards, the conformance record, the viewer screenshot, deployment cards, the family and the four parts of the book, and a 404 page in the same style. The book groups its pages under Evaluate, Integrate, Operate and Contribute, keeping every page’s address, opens in the recoloured light and navy themes and gains a Changelog page. The README’s quick start is the anchored steps of docker/smoke-test.sh, held by scripts/checks/readme-quickstart.sh (#5382).

  • The quickstart docker-compose.yml each release attaches now carries its .sha256sum; the README and the book download it with curl -fsSLO from the latest release and start it with docker compose up --wait, and the README check refuses a host path the file would read from the repository beyond the backup directories (#5387).

  • The viewer moves to ferrohealth-session 0.13.0, which types the OpenID Provider issuer, redirect_uri and post_logout_redirect_uri of [oidc] as secret URLs, so a password written into one of them renders masked in the viewer’s config check, its refusals and its configuration page (#5394).

Removed

  • [authz.professional] (claim, client_tokens, acting_for_claim) and its 403 for a token that names no natural person. The issuer’s kind and profile classify the caller, and the access record’s acting mode and professional come from that classification: a person’s own token is the person, and a delegated token is the client acting for the person it names. Upgrade step: remove the table, which refuses the start, and list a client that acts for a person in its issuer’s delegates (#5003).

  • The named refusal of authz.rbac.user_role, which v5.6 introduced, is gone under the one-release rule for removal aids: the key is now refused as an unknown key when the configuration loads (#5259).

Fixed

  • A 553|incomplete| commit no longer refuses an empty mandatory archetype slot: RM common master06 §Incomplete Content sets every occurrences lower limit to zero, a slot’s included, while too many fillers are still refused. The OPT 1.4 writer now closes a mandatory slot once the fillers placed in it carry its lower bound, as ADL 2 §Slot Filling and Redefinition allows, and still refuses to close one they leave unmet (#5042).

  • The example of a party template whose identity the template marks optional is a party of its class carrying one identity at every detail level, where the required level answered 500: RM demographic PARTY.identities admits no party without one. A party template that constrains no identity at all answers 400 Bad Request naming its class (#5057).

  • An AQL VERSION standard predicate on commit_audit/time_committed applies its operator: VERSION v[commit_audit/time_committed/value > $t] returns every version committed after $t, where it returned the trunk version current at $t whatever the operator (QUERY 1.1.0 master03 §Standard predicate). A number, boolean or NULL literal as the bound, a string bound no ISO 8601 form reads, and a time_committed path other than the whole value or /value are refused with 400 (#5062).

  • AQL refuses a number, boolean or null parameter or literal compared with a date/time path, in WHERE and in a VERSION predicate, with the same 400 a non-ISO 8601 string gets; such a bound used to become NULL and match nothing (#5064).

  • The served OpenAPI document now lists, on both template example operations, the 400 answered for a party template that constrains no PARTY_IDENTITY (#5066).

  • The compose quickstart, the hosted sandbox and the Helm chart run FerroPIX 0.0.18 (ghcr.io/ferrohealth/ferropix:0.0.18@sha256:6fb622b5b9ade34c8a4053112c8f166ff495d8b32bf99f35529d01e512b7cc42), with the FerroPIX chart 0.3.10 built for it. The 0.0.16 they pinned had an outbox sender that never slept while its outbox was empty, so an idle install kept about 1.2 CPU cores busy between FerroPIX and its PostgreSQL at about 8,000 transactions a second (FerroHEALTH/FerroPIX#490). 0.0.18 also makes FerroPIX’s [auth] scopeless_subjects and role_claims hold per issuer, which changes nothing for a FerroPIX that trusts one issuer, as every configuration FerroEHR ships does. The chart moves to 21.0.5 (#5073).

  • An OPT2 built from a template that specialises a published archetype now carries that archetype, by its full id, as parent_archetype_id, as the OPERATIONAL_TEMPLATE invariant is_specialised requires (AOM2 §3.6.6), so a reader of the OPT finds the archetype its root specialises. The ADL form of an OPT still prints no specialize section (OPT2 §Artefact Structure), so the parent travels in the canonical JSON form (#5075).

  • A refused-key-claim-missing access record names the key claim the issuer’s binding reads (its person_key) in binding_detail.claim, so an operator sees which claim the token lacked. It records the name only, never a claim value (#5139).

  • A 403 for a caller kind is recorded with its own binding code: refused-system-personal-scope for a system caller holding a user/ or patient/ scope, refused-no-agreement for an external caller whose issuer has no agreement for its accessor kind, and refused-scope-beyond-agreement for an external caller holding a scope its agreement does not read. Audit migration 0024 widens the binding constraint; the problem text of each refusal is unchanged (#5140).

  • openehr-sdt builds no /context node, and no start_time or setting leaf, in the Web Template of an OPT 1.4 whose COMPOSITION prohibits context (existence 0..0, or an EVENT_CONTEXT of occurrences 0..0), and the template example of such a template carries no context. A template that leaves context open or optional keeps the node as before (#5147).

  • openehr-sdt gives the /context node it synthesizes in a Web Template the min of the template’s existence on COMPOSITION.context, and 0 where the template leaves context unconstrained, since the RM makes context optional (0..1). A client following the Web Template no longer treats the context as mandatory when the template does not require it, and the required template example of such a template carries no context (#5154).

  • An ADL2 template that deletes an attribute (existence matches {0}), for example COMPOSITION.context, keeps the prohibition after it compiles to its operational template: openehr-adl’s create_opt now returns a RawOpt that names each removed attribute beside the OPT, and build_web_template_v2_4 takes that list. A template that deletes context gets no /context node, and a commit that carries one is refused, as it is for an OPT 1.4 template that prohibits context. An OPT2 uploaded already compiled carries no such list and reads a deleted attribute at its reference-model existence. Breaking for crate users: create_opt returns RawOpt and build_web_template_v2_4 takes the prohibited attributes (#5154).

  • A composition commit whose template_id names a stored ADL2 archetype rather than a template is refused with 422 as an unknown operational template; it failed at the insert with a server error before (#5177).

  • openehr-rm types REFERENCE_RANGE.range as DV_INTERVAL<T>: ReferenceRange is now generic over its limit type (ReferenceRange<DvQuantity> on a DV_QUANTITY, ReferenceRange<DvOrdered> on the other ordered types), its is_in_range and Range_is_simple work at that type, and the static RM model gives REFERENCE_RANGE.range and IMPORTED_VERSION.item their argument. Each model::RmAttribute and model::RmTypeRef now keeps the formal parameter it is written as (formal, Some("T") for DV_INTERVAL.lower) beside the resolved bound. The RM class table declares REFERENCE_RANGE<T> and leaves T unused; the generator restores it from a cited table and refuses a schema that no longer matches (#5196) (#5188).

  • An access record names the client the token’s issuer profile reads from client_claims (azp under the entra and keycloak profiles), the one the caller was admitted as. A token carrying a different client_id claim beside it was recorded under that claim. The [origins] register and the SCIM client lookup read the same client (#5194).

  • An Interval or DV_INTERVAL read from canonical JSON or XML that omits a boundedness or inclusion flag now takes it from its bound. An absent bound reads as unbounded and not included, and a present bound as bounded and included. A flag the wire states is kept as written. Before, an omitted flag always read as bounded and included, so an interval with only an upper came out with a bounded, included lower side that had no value. ITS-JSON requires all four flags, and BASE ties none of them to its bound, so reading the omission is our own design (#5203).

  • The viewer’s New contribution commits as the signed-in login. Each audit carries the placeholder naming that login, and the server records the committer it derives from the login. The free-text committer field is gone. For a login bound to a demographic PERSON, a stated name that differs from the PERSON’s name was refused with 403 (#5222).

  • The management surface answers a caller the login binding refuses with 403 and its binding code, such as refused-unbound, as every other route does, in place of 401: the credential authenticated, so new credentials cannot help (RFC 9110 §15.5.4). An absent or invalid credential still answers 401 with its challenge (#5250).

  • The party binding posture in ferroehr report and the boot log masks the user name and password of an issuer URL, as every other rendering of the configuration does (#5259).

  • GET /definition/query/{namespace}:: lists every stored query in the namespace again, where it answered 400 since the typed stored-query name: ITS-REST 1.1.0 definition_query_list reads its name as a pattern in which an empty part is a wildcard, so ::{name} also lists the names that start with it in any namespace. openehr-its gains rest::stored_query::QueryNamePattern, the type the generated DefinitionQueryListParams carries. The store and the single-query reads still refuse an empty name or namespace (#5268).

  • The identifier scan ([privacy.identifier_scan]) reads the name of the ELEMENT that holds a DV_IDENTIFIER when the identifier’s type is absent or names neither a person nor an item, with the same word rules. An untyped laboratory order number in the Receiver order identifier or Requester order identifier element of laboratory_test_result.v1 that passes the NHS Number check by chance is now stored under the default rule set. A person word still wins, so an untyped value in an element named NHS number or Patient identifier stays scanned, and a type that names a kind takes precedence over the element name (#5291).

  • The Helm chart’s description and README name config.deployment.profile, the key the chart refuses to render without since 5.7.0, and the README’s install command sets it; the command it showed set the old config.deployment_profile and did not render (chart 25.0.1) (#5341).

  • A representative’s write is held to the EHDS Art. 5 rules exactly as the person’s. A representative may no longer alter an entry a health professional inserted or name a professional as composer. A version committed as PARTY_RELATED counts as inserted by the person, so the person and the representative may change it, and the contribution list reports inserted_by_person = true for it (#5359).

  • The hosted sandbox’s configuration (deploy/hosted/ferroehr.sandbox.toml) declares its SCIM service provider under [auth.scim] and [[auth.scim.client]], the tables 5.7.0 reads, so the sandbox starts on 5.7.0. Every server configuration the repository ships beside the quickstart and the chart (the hosted sandbox’s and each compose overlay’s) and every viewer configuration (the hosted sandbox’s, the quickstart’s with each overlay, and the chart’s viewer Deployment) is now read by the program built from the change under test in CI, so a shipped file the release refuses fails its pull request (#5396).

Security

Shipped with functional changes because: the advisory reaches only a deployment that trusts several issuers, which no configuration FerroEHR ships does, and its fix changes what such a configuration must state at boot, so it lands on main beside the AQL version attestations already merged for this release; a security-only patch of the last release remains open to the release cut.

  • A Basic user never carries a system/ scope, through ferrohealth-smart 0.9.1 (GHSA-jhg2-jwjp-hq74, CWE-269, CWE-863). A system/ scope granted to a Basic user gave that person every resource of a type, a backend service’s reach. Such a user now refuses the start, naming the user and the scope. Upgrade step: give each Basic user the user/ scopes a person holds, and move each machine on Basic to a client of the authorization server under the client credentials grant (#5003).

  • [auth] scopeless_subjects and role_claims hold per issuer, through ferrohealth-smart 0.8.9 (GHSA-pfgx-xq28-2629, CWE-863, CWE-639). Before it, beside several [[auth.issuer]] tables both lists applied to every trusted issuer’s tokens: a token from any of them whose sub was listed met every SMART on openEHR requirement without a scope, and any of them could assert every role the RBAC checks, ADMIN included. An entry is now written <issuer>#<value> and holds for that issuer alone; beside several issuers a bare entry, or one naming an issuer the section does not trust, refuses the boot. Upgrade step for a deployment with several issuers: write each scopeless_subjects entry as <issuer>#<sub>, and write each role_claims path as <issuer>#<path> for the issuers whose roles count (or set role_claims = [] to read none), since the default role_claims is bare and refused. A deployment with one issuer changes nothing (#5047).

  • The ferroehr-postgres image carries zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 from trixie-security, which fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). The image of 5.6.3 and earlier releases carries 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action (security/vex/postgres-os.openvex.json). The ferroehr and ferroehr-viewer images keep the distroless base’s zlib, which neither program loads (security/vex/ferroehr-os.openvex.json) (#5137).

  • The ferroehr and ferroehr-viewer images move to the gcr.io/distroless/cc-debian13:nonroot build of 2026-11-08 (sha256:cc66d222…), whose zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). Neither program loads zlib. The images of 5.6.3 and earlier releases carry 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action, as the ferroehr-postgres image of those releases is (security/vex/ferroehr-os.openvex.json) (#5152, #5329).

  • CVE-2026-112321, CVE-2026-112322 and CVE-2026-112323 (memory and CPU exhaustion in kadmind through the AUTH_GSSAPI handler) and CVE-2026-119079 (CPU exhaustion in the KDC decoding an AS-REQ or TGS-REQ) in MIT Kerberos 5 have no Debian fix for trixie and are judged not_affected in ferroehr-postgres. The image carries neither kadmind, a KDC nor libgssrpc, and the PostgreSQL server never decodes either request; the configuration FerroEHR ships enables no GSS-API authentication (security/vex/postgres-os.openvex.json) (#5344).

  • Thirteen MIT Kerberos 5 CVEs of MITKRB5-SA-2026-001 in the ferroehr-postgres image’s libgssapi-krb5-2, libk5crypto3, libkrb5-3 and libkrb5support0 1.21.3-5+deb13u1 have no Debian fix for trixie and are judged not_affected. The image keeps the four libraries because the postgres binary links libgssapi_krb5.so.2. Eleven sit in code the image does not carry: the KDC (CVE-2026-107364), its PKINIT plugin (CVE-2026-71200), kadmind (CVE-2026-107884, CVE-2026-112318), libgssrpc (CVE-2026-112314, CVE-2026-112315, CVE-2026-112320, CVE-2026-118380), the kadm5 libraries (CVE-2026-112316, CVE-2026-112319) and libkdb5 (CVE-2026-112317). The PAC checksum flaw (CVE-2026-111423) and the CAMMAC checksum flaw (CVE-2026-71199) sit in libkrb5, which the server reaches only when it accepts a Kerberos ticket. The configuration FerroEHR ships has no gss or hostgssenc line in pg_hba.conf and no server keytab, and the server never reads a PAC or an authentication indicator (security/vex/postgres-os.openvex.json) (#5374).

  • The five OpenSSL 3.5.7 CVEs of the OpenSSL advisory of 2026-11-10 have no Debian fix for trixie and are judged not_affected in the three images: the AES-GCM-SIV flaws CVE-2026-88823 and CVE-2026-95692, the QUIC flaws CVE-2026-88822 and CVE-2026-95691, and the CMS_encrypt() loop CVE-2026-107496. The ferroehr and ferroehr-viewer programs never map libssl.so.3 or libcrypto.so.3: each links only libc, libm and libgcc_s, imports no dlopen, and speaks TLS through rustls (security/vex/ferroehr-os.openvex.json). The PostgreSQL server maps both for TLS over TCP, which never selects AES-GCM-SIV (no TLS cipher suite is one), never opens a QUIC connection and never writes CMS, and the image never runs the openssl program or loads the legacy provider (security/vex/postgres-os.openvex.json) (#5393).

5.7.0-rc4 - 2026-11-11

5.7.0-rc4 is the fourth release candidate of 5.7.0, published for testing. It carries every change since 5.6.3, those of the 5.7.0-rc3, 5.7.0-rc2 and 5.7.0-rc1 sections below included. 5.6.3 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.7.0 itself replaces it. 5.7.0-rc3 was published, and its sandbox check failed: the hosted sandbox’s configuration still declared [scim], which 5.7.0 reads as [auth.scim], so the server refused it. The sandbox file now declares [auth.scim], and CI reads every server and viewer configuration the repository ships through the program built from the change, so a shipped file the release refuses fails its pull request. The viewer moves to ferrohealth-session 0.13.0, so its OpenID Provider issuer and redirect URIs render with their credentials masked. Before you try it, run ferroehr config check on your file: it names every key a release before 5.7.0 read with its replacement, all at once. The steps for party binding are on the book’s bootstrap and migration page. Helm chart 25.0.1-rc5 carries appVersion 5.7.0-rc4. Security carries the two advisories of 5.7.0-rc2, GHSA-jhg2-jwjp-hq74 and GHSA-pfgx-xq28-2629, the zlib fixes in the three images, and the not_affected judgements of the MIT Kerberos CVEs in the PostgreSQL image, the thirteen of MITKRB5-SA-2026-001 among them, and of the five OpenSSL 3.5.7 CVEs of 2026-11-10 in the three images.

Added

  • AQL reads the attestations of a version: v/attestations and any ATTESTATION attribute below it (is_pending, reason, proof, items, committer, time_committed) under FROM EHR e CONTAINS VERSION v CONTAINS COMPOSITION c, with predicates on attestations and EXISTS, and v/commit_audit/is_pending and v/commit_audit/reason on a commit audit that is an ATTESTATION. SELECT v/attestations returns the list, and a comparison below attestations holds when any attestation matches. The FerroEHR extension function LAST(v/attestations/<path>) reads the most recent attestation, so WHERE LAST(v/attestations/is_pending) = true lists the versions awaiting signature, including one marked pending again after it was signed (#4989).

  • The party binding’s configuration, each table with its boot refusals: [authz.party_binding] (provider as one table or one table per tenant, namespace, basic_user), [authz.import] trusted_systems, [[authz.system_client]] (the provider, facilities, templates and origins of each client-credentials client that commits) and [[authz.federation.agreement]] (what a federation issuer’s external persons or organisation may do, on which legal ground), with [authz.rbac] role_issuers and import_role. A template an agreement or a system client writes that the store does not hold refuses the start. The boot line and GET /management/info list the issuers with their kind and identifier issuer, the provider, and every agreement with its accessor and legal ground, and an agreement that admits an organisation without a named person is logged at warn (#5003).

  • [auth.idp.entra], the Microsoft Graph access of the idp-entra adapter of ferrohealth-smart, which resolves the groups of an Entra ID groups overage for an issuer whose profile sets adapter = "idp-entra" (#5003).

  • [auth.dpop] and [[auth.issuer]] sender_constraint: tokens bound to a DPoP key (RFC 9449) or a mutual-TLS client certificate (RFC 8705), checked on every request. A DPoP proof is recorded by the replica that accepted it, so the Helm chart refuses a sender-constrained issuer on more than one replica (#5003).

  • The login binding reads the local party source on every request, with no cache: the new party migration 0027_binding_side_table adds the indexed binding_identifier, binding_membership and binding_representation relations, written in each party commit’s transaction from the PERSONs of the mapped templates that perform a professional ROLE or represent another person, their GROUP memberships and their representation relationships, and rebuilt from the party versions alone. A login matching two parties, an inactive, abandoned or deleted party, a token without the key claim, a party without the needed marker and unreadable groups are refused with 403 and their binding code; a party store or identity provider that cannot answer is answered 503 (#5004).

  • ferroehr definitions load <path>... stores OPT 1.4 (.opt) and ADL 2 (.adls) definitions before serve, through the validation and versioning of the REST upload. A file already stored with the same text is left alone and other text under a stored id is refused, so a deployment runs it at every start to give its party mapping the templates it needs at boot. deploy/definitions/professional/ holds a professional PERSON and ROLE pair adapted from openEHR CKM archetypes for the development lanes and the quickstart (#5006).

  • The compose quickstart binds its user to a professional PERSON: the ferroehr-definitions one-shot loads the PERSON and ROLE templates the party mapping names before the server starts, and ferroehr-provision creates the user through /scim/v2/Users with a token of the development issuer, then reports healthy, so docker compose up --wait returns once the login is bound. The Helm chart (23.3.0) takes definitions.files, which an init container stores with ferroehr definitions load before the server container starts. The conformance, browser and Kubernetes lanes load the same templates and provision every login they use through SCIM, with the committed test client of docs/conformance/party/scim/ (#5006).

  • ferroehr party-binding adopt <PARTY_ID>... gives each prepared PERSON of an existing deployment a professional ROLE of the mapping’s role template, with the PERSON as performer, and a new PERSON version listing it in roles, all in one demographic CONTRIBUTION committed by the first administrator named with --admin or --admin-key, and writes one access record per adopted PERSON through the configured audit trail. It refuses the whole adoption, naming them, while any listed PERSON’s key was introduced by a committer that is no [[scim.client]], named administrator or --migration, so a key planted before the upgrade is never adopted. A PERSON that already performs the ROLE is reported and left alone. The report is JSON on stdout (#5009).

  • ferroehr party-binding check is the read-only pre-flight of a migration to party binding, and the one report of the provisional persons just-in-time creation left incomplete. It reads the party versions alone, so it runs before any boot fill, and writes nothing. Per key issuer it reports the PERSONs that carry a binding key, every selector value two PERSONs carry, the inactive and abandoned ones, the provisional ones older than provisional_report_days, the key-bearing ones that perform no professional ROLE, and every mapped Basic user with the PERSON it binds. For each key it names the committer of the version that introduced it, and fails unless that committer is a [[scim.client]], an administrator named with --admin or --admin-key, a migration declared with --migration, or just-in-time creation on an issuer with jit = true. --expected <file> lists every login of the identity provider with no PERSON. The report is JSON on stdout, and the command exits non-zero on a duplicate, a missing person or a key that fails the provenance rule (#5009).

  • At boot, FerroEHR logs a warn naming every Basic user mapped in [[authz.party_binding.basic_user]] while [authz.assurance] minimum is unset: a Basic credential carries no assurance level, so such a person is never refused on patient data (#5009).

  • The book has a “Who is the committer” page under Security. It covers the committer the server writes for each kind of caller, the placeholder a client may still send, the refusals with their status codes, and the deviation from the ITS-REST merge rule. It also shows how a professional’s PERSON gets its login identifier. Enterprise identity providers now opens with the generic standards path and adds the entra preset’s binding and provisioning settings, an Okta example, a Keycloak realm walk-through and a SAML section. The records of processing and the DPIA gain the workforce-identity activity, the federation node page gains its callers and agreements, and the go-live checklist checks the binding and the caller rates (#5010).

  • The access record of a MyHealth@EU contact point’s request carries the request message’s UUID, the WS-Addressing MessageID its token relays as ehdsi.message_id, as the record’s request-message-id. The record then joins the contact point’s Patient Privacy audit entry (eHDSI Audit Trail Profile 10.1.0 §2.3.12 to §2.3.32) (#5011).

  • The one PERSON mapping SCIM provisioning, just-in-time creation and the login binding share: [[authz.party_binding.person.template]] with the binding key, userName and externalId selectors of each template, [[authz.party_binding.person.attribute]] with the SCIM attribute, the OpenID Connect claim and the path of each mapped value, and [authz.party_binding.professional_role], [authz.party_binding.group], [authz.party_binding.membership] and [authz.party_binding.representation]. Boot refuses a mapping that breaks one of its rules, naming the key, including a template that is absent, rooted at another class or lacks a mapped path, and deleting a template the mapping names is refused with 409 (#5013).

  • The new party migration 0028_binding_selectors holds each binding key, userName (compared after RFC 8265 width mapping, lower case and NFC), externalId and group displayName to one live party, so a commit giving one a second party is refused with 409. The first boot fills the side table under the mapping’s digest, a replica with another mapping refuses to start, and ferroehr party-binding refill rebuilds it after a mapping change, refusing while a replica runs and naming the parties when a value is carried twice (#5013).

  • POST /admin/party/binding-key re-keys a PERSON’s binding key: it refuses the administrator’s own login and a key already bound, retires the old key in the same commit and writes one access record carrying both (#5013).

  • [scim] mounts the SCIM 2.0 service provider at /scim/v2/ (off by default): [[scim.client]] declares each provisioning client by issuer, client_id, scope and binding_issuer, and public_base_url names the origin its documents hang off. /scim/v2/ServiceProviderConfig answers without a token; /ResourceTypes and /Schemas answer a declared client acting for nobody that holds its scope, within [server.caller_rate]. Every refusal is an RFC 7644 Error message. The start is refused for a client of a non-machines issuer, one declared twice, two clients sharing a binding_issuer, or a binding_issuer no [[auth.party]] table carries (#5014).

  • The SCIM endpoint answers the Microsoft Entra ID provisioning client as Microsoft’s “Develop a SCIM endpoint” guide documents it: a value-filtered attribute such as emails[type eq "work"].value serves its type and declares primary read-only, so the primary Entra sends is ignored (RFC 7644 §3.3); a PATCH of userName renames the User; and the Test Connection probes are answered with an empty ListResponse and 404. ferroehr config check refuses a [scim] the person mapping cannot serve, as serve does. The Helm chart exposes config.scim and routes /scim/v2 on every Ingress host when it is enabled, and the book walks an operator through provisioning from Entra ID, with the Workload Identity Federation checks the authorization server makes (#5014).

  • /scim/v2/Users/{id} takes PUT, PATCH and DELETE: a write commits the next version of the professional’s PERSON, clearing a mapped attribute a PUT leaves out, keeping the binding key and externalId, and moving the lifecycle state by active (complete or incomplete to inactive or abandoned and back, an abandoned User completed in two versions, an attribute change to an inactive User refused until it is reactivated). DELETE logically deletes the PERSON and its professional ROLE in one CONTRIBUTION and leaves the binding key as a tombstone. An If-Match naming another version is 412, and ServiceProviderConfig declares PATCH. [scim] max_deactivations and deactivation_window_minutes bound the deactivations and deletions of one client, counted in the database so replicas share the bound, with 429 past it (#5014).

  • /scim/v2/Users serves a client’s Users: POST creates the professional’s PERSON and its professional ROLE in one demographic CONTRIBUTION, complete when the template holds in full and incomplete otherwise, GET /Users/{id} answers with the version’s entity tag and 304 to a matching If-None-Match, and GET /Users answers a ListResponse filtered by eq and and over userName, externalId and id, with any other filter 400 invalidFilter. A client sees only the Users whose binding key its binding_issuer issues, every SCIM request is an access record of the client naming the PERSON, and ServiceProviderConfig declares entity tags. The start is refused when [scim] is on and the party mapping has no default person template with a user_name selector or no professional role (#5014).

  • /scim/v2/Groups serves SCIM Groups as demographic GROUPs of [authz.party_binding.group], each member holding an inline PARTY_RELATIONSHIP of [authz.party_binding.membership] from its PERSON to the GROUP, with no ROLE minted: create, read, eq/and filters on displayName, externalId and id, excludedAttributes=members, PATCH (answered 204), PUT and DELETE, with the entity tags of Users. A Group belongs to the client that created it, its members are that client’s Users, its displayName is unique, and its externalId sits under the client’s group_identifier_issuer. A remove of members with a value list removes only those members, an Entra compatibility deviation from RFC 7644 §3.5.2.2. Deleting a User removes its memberships in the same CONTRIBUTION. [scim] max_group_members (default 5000) caps a Group (#5015).

  • Just-in-time creation of a provisional PERSON at a first login, off by default and enabled per persons issuer in a new [[authz.party_binding.issuer]] table (jit, guests_jit, user_name_claim, idp_claim, admit_idp, admit_groups, admit_roles, admit_assurance). A login of a person or a delegated caller that matches no party and meets every declared admission condition commits one 553|incomplete| PERSON holding the binding key, the userName and the name claims of the access token, with its professional ROLE, as a creation by the person, and is served bound to it; any other mapped attribute is written only where its mapping sets jit = true, and externalId never. Concurrent first requests commit one PERSON, a deprovisioned or inactive login is never recreated, a taken userName refuses the login, and a SCIM client of the key’s issuer completes the PERSON. Boot refuses jit without an admission condition or a mapping that can write it. ferroehr party-binding check lists the provisional persons older than [authz.party_binding] provisional_report_days (default 30) (#5016).

  • A party of a mapped template is validated against that template on every demographic write: a PERSON or GROUP whose archetype_details.template_id names a template of the shared party mapping, and a PARTY_RELATIONSHIP of its membership archetype. A 532|complete| version must meet every constraint, and a 553|incomplete| one may miss mandatory nodes but nothing it carries may be wrong (RM common master06 §Incomplete Content); the refusal is a 422 naming each failing path. Every other party keeps the reference-model checks alone, so parties created through the Demographic API under other templates are accepted as before (#5042).

  • The FHIR AuditEvent, the DICOM audit message and the ITI-20 syslog feed carry the binding fields of the access record: the binding outcome, the party and its groups, the provider acted for, the facilities of a system commit, the representation with its relationship, the actor chain, the external accessor’s asserted fields and the agreement with its legal ground, consent and access class, and a chained request’s correlation id. ehds-logging 0.5.1 writes them as the ehds-accessor and ehds-relayed entities, in the DICOM message as participant objects of type Other. The FHIR record also writes the external and system acting modes. FerroEHR moves to ehds-logging and ihe-atna 0.5.1 and ihe-iti 0.0.77 (#5043).

  • A bound representative reaches only the EHR whose EHR_STATUS names the subject it represents. Every composition, EHR_STATUS, directory, contribution and query request is held to that EHR whatever the token’s scopes cover, and a query must name it with ehr_id. A representative whose token names a subject its party has no live representation relationship to, or whose issuer no [[auth.party]] table binds, is refused with 403 and refused-representation-unrecorded (#5086).

  • The Helm chart (23.1.0) accepts bridge.cdr.oauth.tokenEndpoint for an authorization server whose metadata FerroBRIDGE cannot read at bridge.cdr.oauth.issuer. FerroBRIDGE then requests its client credentials token from that endpoint, and issuer still names the machines issuer the CDR trusts (#5093).

  • A MyHealth@EU contact point’s caller (a contact_point = true federation issuer) reaches patient data only for the patient of its treatment relationship confirmation. The token’s ehdsi.patient.id (system|value) is resolved to its EHR through the EHR index, and every patient-data request must name that EHR. A token without it is refused with 403 and ERROR_SEC_DATA_INTEGRITY_NOT_ENSURED, recorded as refused-trc-missing. Another patient’s EHR, an unknown EHR and a request naming no EHR all answer one 404 of the new problem type trc-patient-mismatch, carrying the agreement service’s *_NOT_FOUND code and recorded as a refused-trc-patient-mismatch security event (eHDSI XCA Profile 8.0.1 §4.2.1, §4.2.2). A contact point’s person agreement now names its ehdsi_service; upgrade step: set it, and list the issuer in [smart] patient_issuers (#5102).

  • An external person under an agreement with discard discards a dispensation of a listed template, as a logical delete, within discard_window_hours and from the same point of care: the dispensation’s health_care_facility carries the token’s provider identifier and its eHDSI organization-id. A token carrying permissions records and discards a dispensation only with PPD-046 (eHDSI SAML Profile 10.0.0 §2.5) (#5102).

  • A read under a federation agreement with a consent condition is served when the token carries the consent decision reference for its patient (Wabvpz Art. 1(j) and Art. 15a(1)), in the claim or authorization_details member the issuer’s [[auth.federation]] table maps to consent.reference, and refused with 403 without it. The reference is one string, the token’s patient/ grant holds it to the launch-context patient, and the access record carries it with the other external fields. ferrohealth-smart moves to 0.9.3, which reads the field (#5106).

  • Every access record of an authenticated request carries the outcome of the party binding (binding: bound, provisional, external, organisation, system, self or a refused-* code) with the fields it needs: the bound party, its groups, the provider, the guest mark, a representative’s representation, a delegated token’s actor chain, a verified external accessor’s fields and agreement, and the facilities a system caller’s compositions named. acting_mode gains delegated, external and system. The audit migration 0023_binding_outcome.sql adds the columns and restores a per-row digest version: records written before stay version 1 and keep verifying, new records are version 2, and a version-1 record after a version-2 one is refused (#5118).

  • Every refused request is an access record through the gated path, and a refusal of the login binding or of a caller kind carries its refused-* binding code and the party it names. Refusals of tokens that failed verification are budgeted by source network and those of verified callers by issuer and subject, so neither crowds out the other; past a caller’s limit its refusals become one summary record with the count and the EHR ids they named, and past a budget one summary record per issuer and outcome is written. Security-event refusals are recorded for every answered request, and a refusal record the system log cannot take is counted in atna_audit_refusals_unwritten. The limits, windows and budgets are kept per replica (#5119).

  • A chained organisation request under a federation agreement is served only when its token carries the requester’s correlation id in the claim or attested authorization_details member the issuer’s [[auth.federation]] table maps to correlation_id. Without it the request is refused with 403 and recorded under the binding code refused-correlation-missing (audit migration 0025); a request header never stands in for it. The organisation access record’s agreement carries the id and its source beside the legal ground and the access class. A chained agreement whose issuer maps no correlation_id, or maps it from a constant, refuses the start (#5120).

  • On the MyHealth@EU contact-point route, the committer of an external person names the professional by the HP Identity Assertion’s NameID: DV_IDENTIFIER{issuer = the assertion's issuer, id = the NameID, type = "hp_identifier:<NameID format>"}, beside the voucher’s oauth2 identifier (eHDSI SAML Profile 10.0.0 §2.1), and a dispensation’s discard is a logical delete under the same committer. A contact point’s COMPOSITION must carry the HCPO’s organization-id in EVENT_CONTEXT.health_care_facility.identifiers, under the issuer urn:oasis:names:tc:xspa:1.0:subject:organization-id, beside the provider’s identifier, or it is refused with 422. ferrohealth-smart moves to 0.9.4, which reads ehdsi.assertion.name_id_format and refuses a contact-point token without one of the three formats as ERROR_HPI_INSUFFICIENT_INFORMATION; upgrade step: a contact point’s authorization server writes the NameID format into its tokens (#5144).

  • A federation caller above its agreement’s caller_rate, a MyHealth@EU contact point’s included, is answered 429 with Retry-After right after authentication and before any patient lookup, keyed on the token’s subject per replica, and every such refusal is counted into the caller’s summary record (#5149).

  • A deleted SCIM User’s contacts are erased from every version of its PERSON, keeping the name, the binding key and the identifiers for the record-keeping duties that need them. [scim] contact_erasure_minutes (default 0, at most 43200) sets when: 0 erases them in the DELETE itself, a period leaves them to a sweep that runs every minute. Each erased version’s node rows are derived again from the new body, its server signature is removed, and the erasure is an access record naming party:<id>#contacts. Party migration 0030 adds the scim_contact_erasure table (#5178).

  • openehr-rm types generic RM type names and paths: typing::RmType parses and prints names such as DV_INTERVAL<DV_QUANTITY>, typing::attribute_type gives an attribute of an instantiated class with the actual argument in place of the formal parameter, and typing::path_type types every step of an RmPath with its container. Each model::RmAttribute and model::RmTypeRef keeps the formal parameter it is written as (formal) beside the resolved bound. paths::write::item_at_path_mut and paths::write::set_at_path write at a path into a canonical instance, positional predicates included, creating the members the path names and changing nothing when a write is refused (#5187).

  • openehr-sdt: flat::navigate::Navigator gives every instance of a Web Template path in a composition, with each step’s occurrence, container index, positional path and JSON pointer, reaching a renamed instance where the template’s name is redundant and an OPT2 archetype root through the complete identifier the template recorded. flat::navigate::relative compares two paths step by step. flat::attribute_keys gives the FLAT key of an RM attribute of a structural class (the master05-rm_mapping.adoc tables, checked against them). A WebTemplateNode carries facts (the embedded template id of an OPT 1.4 archetype root and the terms its archetype defines) and every coded option its terminology, so no consumer walks the operational template (#5189).

  • openehr-base gives identifier and time parts as written: ObjectVersionId::object_id_str and ObjectVersionId::versioned_object_id keep the case of the object_id part, Iso8601Time::fraction_digits and Iso8601DateTime::fraction_digits return the digits of the fractional second as text, and Iso8601Duration::parts returns each component’s designator and number verbatim (DurationDesignator, DurationPart), with Iso8601Duration::fraction_digits for the seconds (#5190).

  • openehr-its reads the ITS-REST ETag as a typed entity tag: rest::etag::EntityTag parses and writes the RFC 9110 form, weak or strong, with the weak and strong comparisons. It types the tag as an ObjectVersionId or a HierObjectId. Every generated response-headers struct that carries an ETag gains entity_tag() beside the raw etag field (#5191).

  • openehr-term names the null_flavours and setting groups: codes::NullFlavour and codes::Setting give each concept’s code and English rubric and read a code back, held equal to the bundled terminology by tests. Every typed group’s ALL is now a slice (#5192).

  • openehr-am: HridRef reads an archetype reference with the version precision it was written at (VersionPrecision: major, minor or full, BASE Artefact Identification §Referencing), gives the version as written and matches an ArchetypeHrid by the specification’s rule, the namespace never a wildcard. ArchetypeHrid::local_semantic_id gives the namespace-free interface form a Web Template nodeId carries, which the v2_4 Web Template builder now uses (#5193).

  • openehr-its: rest::stored_query types a stored query’s qualified name (QualifiedQueryName, [{namespace}::]{query-name}, with check_registrable refusing the reserved aql), its version (QueryVersion, major.minor.patch in numeric order) and the version pattern a request addresses (VersionPattern, exact, {major} or {major}.{minor}, selecting the highest match), per ITS-REST 1.1.0. The generated stored-query operations and the StoredQuery and RESULT_SET documents carry them, so a name or version outside the grammar is a 400 where the request is decoded (#5207).

  • openehr_rm::Generation::earliest_for_composition answers the oldest RM generation whose canonical-JSON decode accepts a composition, so a builder can name ARCHETYPED.rm_version without naming a generation module; a value no generation decodes is a typed NoGenerationDecodes error carrying every generation’s refusal (#5242).

  • The access record takes the purposes of use and the professional roles from the access token. An issuer with an [[auth.attributes]] table states them through the IHE IUA purpose_of_use and subject_role claims (IHE ITI TF-2 §3.71.4.2.2.1.1) or the purpose_of_use of RFC 9396 authorization_details. The new token_attributes column of the audit store holds them, a new record is chain digest version 3, and the EHDS access record and the BALP AuditEvent write them with their code systems. A request whose [audit] purpose_header names a purpose the token does not state is refused with 403, and the refusal is recorded. The header alone keeps working for an issuer without the table (#5323).

Changed

  • The published server, viewer and PostgreSQL images take their org.opencontainers.image.base.name and base.digest labels and annotations from the runtime stage’s digest-pinned FROM line at build time, so a base-image bump changes that one line. An image built directly from a Dockerfile no longer carries these two labels (#4953).

  • [auth] takes the ferrohealth-smart 0.9 shape. Every trusted issuer is an [[auth.issuer]] table with a required kind (persons, patients, machines or federation) and a profile (rfc9068 by default, or entra, keycloak, okta, nl), and holds every setting of its tokens itself. A configuration in the old shape refuses the start, naming what replaces it. Upgrade steps: write a flat issuer = "…" as an [[auth.issuer]] table; move audience, algorithms, clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds, jwks_file, timeout_ms, mode, role_claims and [auth.introspection] onto each issuer table ([auth.issuer.introspection]); give each issuer its kind, with patients and professionals at separate issuers (a realm or tenant each), since one issuer is one kind; replace token_profile = "jwt" with token_types = ["JWT"] and the required_claims the issuer’s tokens carry, and drop token_profile = "rfc9068"; and write each scopeless_subjects entry as { issuer = "…", sub = "…" } (#5003).

  • A persons token is admitted only through a client its issuer lists in first_party_clients or delegates. Upgrade step: list the deployment’s own clients, the viewer’s among them, in first_party_clients, and a third-party client that acts for a person in delegates. The other 0.9 profile keys (person_key, tenant_claim, client_claims, app_only, guest, impersonation_marker, expect, user_claims, person_equality, unstable_claims, max_actor_chain, representation, scope_claim, app_only_scope_claim, adapter, sender_constraint and the onward_* keys) are documented in the configuration reference (#5003).

  • An array of tables, such as [[auth.issuer]], is set from the environment by element index (FERROEHR__AUTH__ISSUER__0__ISSUER), through ferrohealth-config 0.2.10 (#5003).

  • [smart] issuer_metadata_file is the metadata of the [auth] issuer it names, wherever that issuer stands in the list, and a file naming no [auth] issuer refuses the start. The SMART discovery document describes the first introspected issuer, else that named issuer, else the first issuer whose metadata is read from it, else the first issuer, so a Basic users’ issuer listed first no longer takes the discovery document (#5003).

  • The family crates move together: ferrohealth-smart 0.9.1, ferrohealth-authz 0.3.0, ferrohealth-session 0.3.1, ihe-atna and ehds-logging 0.4.0, ihe-iti 0.0.76, ferrohealth-config 0.2.10, ferrohealth-server 0.5.1 and ferrohealth-ui 0.6.9. ihe-atna and ehds-logging now escape every control and bidirectional formatting character of a value they write as its code point. With ferrohealth-ui 0.6.9 the viewer’s session watch no longer throws when the party relationship form, which holds a control named target, is submitted (#5003).

  • The terminology client moves to ferrobridge-term 0.1.143 (FerroBRIDGE v0.0.22), which takes the ferrohealth-smart 0.9 token source (#5003).

  • Every person is bound to a demographic party, with no switch. A persons issuer needs its [[auth.party]] table, two issuers of one node may not share an audience, and [authz.party_binding] provider (the healthcare provider the deployment acts for, EHDS Art. 13(4) and Annex II 3.2(a)) is required once [auth] admits a credential; each refuses the start naming what to set. A Basic user is a person: it maps to the binding key of its PERSON in [[authz.party_binding.basic_user]] (user, key_issuer, key), and a Basic user without a mapping refuses the start, naming the user. Upgrade steps: add an [[auth.party]] table to each persons issuer, set provider, and map each Basic user under the identifier_issuer of an [[auth.party]] table, which needs a persons issuer; a deployment of Basic users alone declares one whose jwks_file is a key set that signs nothing, as the quickstart and the sandbox do (#5003).

  • A token grants [authz.rbac] admin_role, emergency_access_role, import_role and identifier_reader_role only when its issuer is listed in the new [authz.rbac] role_issuers, each a persons issuer; any other kind refuses the start. Upgrade step: list the issuers whose role claims grant these roles; a Basic user keeps the roles its own table grants (#5003).

  • The compose quickstart, the dev and conformance stacks, the hosted sandbox and the Helm chart’s Basic overlays bind their Basic users under a persons issuer that signs nothing: its key set, which every FerroEHR image carries at /opt/sandbox-posture/basic-users-jwks.json, is the public half of a key pair whose private key was discarded (#5003).

  • The quickstart has one copy. docker-compose.yml on main pulls the FerroEHR images at :main and carries main’s configuration, which CI checks with the server image the change builds; the copy attached to each release pulls that release’s images, written by the release lane, which refuses a copy that still names :main. Download the release asset (releases/latest/download/docker-compose.yml) to run a release (#5003).

  • The Helm chart moves to 22.0.0. config.auth.issuer is a list of issuer tables, each with its kind, profile and audience, and values.yaml exposes config.auth.party, config.authz.party_binding, config.authz.rbac.role_issuers, config.authz.system_client and config.authz.federation.agreement. The chart refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling. Upgrade steps: rewrite config.auth as the [auth] upgrade above describes, set config.authz.party_binding.provider, and list the issuers whose roles grant the administrator role in config.authz.rbac.role_issuers (#5003).

  • A caller bound to a demographic PERSON commits under the committer the server writes: a PARTY_IDENTIFIED with the PERSON’s name, the login identifier (the issuer profile’s person_key value under the token issuer, type oauth2, or the Basic user under ferroehr, type basic), the provider identifier (type healthcare-provider) and an external_ref to the PERSON, on every commit path, attestations and CONTRIBUTIONs included. Upgrade: stop sending a committer, or send the placeholder {"_type": "PARTY_IDENTIFIED", "identifiers": [{"_type": "DV_IDENTIFIER", "issuer": "<iss>", "id": "<sub>"}]}, which must name the authenticated caller; a stated committer with another name, identifier or external_ref, and a stated PARTY_RELATED, is refused with 403 (#5005).

  • A patient’s committer is PARTY_SELF with no external_ref, and a representative commits as a PARTY_RELATED naming their own PERSON and the coded relationship. A patient-context token commits no demographic version. Upgrade: read the person’s own entries by the committer’s PARTY_SELF type, not by its former external_ref (#5005).

  • Importing an EHR Extract needs the role [authz.rbac] import_role names, or a system/ caller, and a wrapped committer naming a party or a login of this deployment is refused unless its version was created by a system [authz.import] trusted_systems lists. Upgrade: set import_role and give it to the accounts that import (#5005).

  • A person of a persons issuer whose login no demographic party carries is refused with 403 and refused-unbound on every request, with no fallback to a login-only committer. The problem names the identifier issuer to provision the person under, a token without the key claim is refused with a problem naming that claim, and each refusal counts into the new binding_refusals metric by issuer and outcome. The Keycloak quickstart overlay provisions its four demo users through SCIM with the ferroehr-provision-keycloak one-shot. Upgrade: provision every professional (through SCIM or the demographic API) before you add a persons issuer to the upgraded deployment (#5006).

  • A deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR adds its persons issuers only after the bridge and the gateway reach it under the caller: FerroBRIDGE v0.0.23 and FerroFED v0.0.17. Until then the committer rule refuses their writes. FerroFED v0.0.17 follows this release, so a deployment with a FerroFED gateway adds its persons issuers only once that release is out. The book’s new “Party binding: bootstrap and migration” page carries the bootstrap path, the four migration steps and these prerequisites, and “Breaking changes of party binding” lists every incompatible change with its upgrade step (#5009).

  • The hosted sandbox follows the bootstrap path of party binding: its configuration declares the sandbox-scim client ([[scim.client]], with the development issuer trusted as a machines issuer for it), the box’s new ferroehr-provision service provisions the demo PERSON through /scim/v2/ after every reset, and scripts/sandbox/reseed.sh waits for the demo login to bind before it seeds, then checks the committer of the marker’s EHR_STATUS version references a PERSON. The deploy probe’s new bootstrap family (P-BOOTSTRAP-01) shows the compose quickstart does the same from an empty database: the quickstart user is a User of the SCIM client, and a commit of the Basic user names that PERSON (#5009).

  • With a party mapping configured, the parties the login binding relies on are written through the demographic API by an administrator (authz.rbac.admin_role) only: a PERSON of a mapped template carrying a binding key, userName or externalId, its professional ROLE, a representation relationship, and a GROUP of the group template. Any other caller’s write is refused with 403, a version that changes a present binding key is refused with 403, and an administrator’s version that leaves a protected identifier out is refused with 422. Upgrade: write these parties as an administrator, and change a key through POST /admin/party/binding-key (#5013).

  • FerroEHR records every DPoP proof it accepts in its PostgreSQL database until the proof’s window ends, so a proof is accepted once across all replicas, and a proof replayed to another replica is refused there (RFC 9449 §11.1). If the database cannot record a proof, the request is refused with 503. The Helm chart (22.1.0) no longer refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling (#5072).

  • An AQL query scoped to one EHR now takes the streaming plan shape whatever its ORDER BY, DISTINCT or aggregates, whether the EHR comes from the ehr_id request parameter or from an ehr_id/value equality in the query. On a store of 100,000 compositions such a query used to spend 16 to 64 ms in PostgreSQL’s planner for an execution of about 1 ms, on every new connection and every new statement text; it now plans in 1 to 3 ms (#5076).

  • The Helm chart (23.0.0) connects FerroBRIDGE to the CDR as a machine. FerroBRIDGE uses an OAuth 2.0 client credentials client of a machines issuer that the CDR trusts, set in bridge.cdr.oauth (issuer, clientId, and optionally scope, audience and resource). The client secret comes from the client_secret key of bridge.cdr.existingSecret. The chart no longer accepts a Basic user for this connection: from ferrohealth-smart 0.9 a Basic credential belongs to a person and holds no system/ scope. The chart refuses to render when bridge.cdr.oauth.issuer is not listed in config.auth.issuer with kind: machines. To upgrade, add that issuer, create the client, and replace the user and password keys of the Secret with client_secret (#5085).

  • A verified external person’s commit records the professional as its committer: a PARTY_IDENTIFIED with the asserted name, the hp_identifier under the issuer "<country_code>:<issuing_authority_name>" and the token’s iss and sub, with no external_ref. A country code that is no upper-case ISO 3166-1 alpha-2 code, or an issuing authority holding a :, control or bidi character, is refused with 403 naming the field. The COMPOSITION must name the token’s healthcare provider (provider.healthcare_provider_identifier under provider.issuing_authority_name) as its health_care_facility, or it is refused with 422. Upgrade step: map provider.issuing_authority_name in each [[auth.federation]] table whose agreements carry writes, with a constant source where the token does not carry it (#5097).

  • A system caller (an app-only token by its issuer profile’s app_only markers) holds system/ scopes only: a token carrying a user/ or patient/ scope is refused with 403. It commits into an EHR only as a client an [[authz.system_client]] table declares, under the committer PARTY_IDENTIFIED{name = <client>, identifiers = [{issuer = <iss>, id = <client>, type = "client"}]}, and each COMPOSITION it commits must be of a template its writes lists (403), name a facility its facilities lists in context.health_care_facility, or in feeder_audit.originating_system_audit.location when it has no context (422), and be either a device act from a device client or a relayed person act naming the professional by an identifier from a relay client. Upgrade: declare every client that commits in [[authz.system_client]] with its writes, facilities and origins, and give its compositions the facility identifier (#5100).

  • An attestation is accepted only from a caller bound to a professional PERSON; one from a patient, a representative, a system caller or an unbound caller is refused with 403. Upgrade: attest under the professional’s own token (#5100).

  • A guest of a persons issuer, by its profile’s guest markers, is refused with 403 (refused-guest) unless the issuer’s [[auth.party]] table sets guests = true. Upgrade: set guests = true on the issuers whose guests act under your responsibility (#5100).

  • A verified external person or organisation of a federation issuer is admitted only under the [[authz.federation.agreement]] of its issuer and accessor, holding no scope beyond the agreement’s reads, and is an access-service caller by its kind: an opted-out EHR answers 404 to it, and only an external person under an agreement with opt_out_override reaches one under an emergency purpose. It commits only a COMPOSITION of a template the agreement’s writes lists (an organisation nothing), and a read under a consent-conditioned agreement, and every request under an NL agreement, is refused until the consent decision reference and the Wabvpz Art. 15f roles can be read. Upgrade: declare an agreement for every federation issuer and accessor, list the issuer in [smart] patient_issuers, and give its tokens only the patient/ scopes the agreement reads (#5101).

  • Under an NL [[authz.federation.agreement]], Wabvpz Art. 15f now decides admission in place of the refusal of every request. A caller whose nl.role is the UZI role 01.008 (bedrijfsarts) or 01.048 (verzekeringsarts) is refused with 403 as refused-wabvpz-role, and a caller whose nl.organisation_type is missing or is not a care provider type of Nictiz’s OrganisatieTypeCodelijst is refused with 403 as refused-wabvpz-organisation-type, so an insurer is refused. Every NL agreement declares purpose = "care"; one without a purpose, or for a keuring, refuses the start. The code lists are vendored by scripts/vendor/nl-role-codes.sh into docs/specs/nl-role-codes/. Upgrade: add purpose = "care" to every NL agreement and map nl.organisation_type (and nl.role for persons) in its issuer’s [[auth.federation]] table (#5107).

  • A delegated token’s access record now has acting_mode = "delegated" where it had client, and a system caller’s system. The FHIR rendering still names a delegated access as a client one. Upgrade: a review query that selects delegated access by acting_mode = 'client' also selects 'delegated' (#5118).

  • FerroEHR builds on ferrohealth-server 0.6.1, whose public caller-rate take holds a federation caller to its agreement’s caller_rate. Its 0.6.0 break, the revocation feature’s move to der 0.8, touches no feature FerroEHR enables (#5149).

  • The source tree no longer carries the UZI role code CodeSystem (uzi-rolcode) of Nictiz’s zib 2020 package: its codes and titles reproduce the UZI-register CPS, which CIBG does not license for reproduction. scripts/vendor/nl-role-codes.sh writes it into the git-ignored .vendor-cache/nl-role-codes/, as it does the CPS, and --test-inputs fetches only what the Wabvpz tests read. Nictiz’s care provider types stay under docs/specs/nl-role-codes/. Run scripts/vendor/nl-role-codes.sh --test-inputs once before running the ferroehr-rest tests (#5175).

  • A contact point’s patient.id is read through ferrohealth-smart’s TrcPatientId, the family’s one definition of the TRC patient. A value with whitespace in its system, or a value holding a control character, | or a CX delimiter (^, &, ~, \), is refused as a missing binding, as the contact point that writes the claim refuses it (#5195).

  • [auth] takes three boot refusals from ferrohealth-smart: a Basic user’s hash below the Argon2id floor (m=19456, t=2, p=1), an issuer that carries userinfo, and private or symmetric key material in a jwks_file. They are refused as before, with the crate’s messages. A jwks_file that is no JWK Set now refuses the start where the crate reads the key set (#5209).

  • The family crates move together to one ferrohealth-smart: 0.13.1, with ferrohealth-authz 0.7.0, ferrohealth-session 0.7.0, ihe-atna and ehds-logging 0.8.0, ihe-iti 0.0.83, ferrohealth-server 0.6.4 and ferrobridge-term 0.1.148. The 0.13 line refuses a JSON document it cannot read without quoting it (GHSA-mhv7-3975-6xq3), so a jwks_file or [smart] issuer_metadata_file that does not parse is refused at boot with the crate’s message, which names the failure class and position; 0.13.1 keeps a refused key set, introspection or metadata body out of the error (GHSA-5qm7-v4fw-p2v9). The OAuth client credentials tables ([authz.abac.remote.oauth], [ehr_index.oauth], [terminology.oauth] and [audit.fhir_feed.oauth]) read as before (#5209).

  • The API’s response security headers come from the family server stack (ferrohealth-server 0.6.2) with the same six headers and values. They now also reach the management listener on its own port, and Strict-Transport-Security: max-age=31536000 is sent when FerroEHR or the viewer serves TLS itself ([server.tls], RFC 6797 §7.2). The viewer gains Cross-Origin-Resource-Policy: same-site (#5210).

  • The server and the viewer read the parts of an OBJECT_VERSION_ID through openehr-base and its BASE grammar, never by cutting at ::. A value that is no OBJECT_VERSION_ID now names no version: the AQL branch refusal, the version item tags and the viewer’s version numbers leave it alone, a reference to a container is read as the whole value (so uuid::extension no longer names the container uuid), and the viewer’s deletes refuse it before calling the CDR. Every part keeps its case (#5211).

  • ETag and If-Match values are read through openehr-its’s RFC 9110 entity tag on the server and in the viewer. The server accepts the same forms as before. The viewer reads its CDR’s ETag strictly: a value that is no entity tag names no version (#5212).

  • The viewer’s audit screen reads the ITI-81 searchset and each AuditEvent as FHIR R4 resources, and a refusal’s OperationOutcome too. A searchset that is no FHIR R4 Bundle of AuditEvents now shows as a failed read, never as a page of empty rows (#5214).

  • The EHR index’s boot check reads the PIX Manager’s CapabilityStatement through ihe-iti’s PIXm client, with its transport and its client-credentials token, so a Manager behind IUA answers the probe as it answers ITI-83. A failed probe is reported as the client’s refusal: no answer, the status it answered, or no FHIR R4 CapabilityStatement (#5216).

  • [audit.syslog] reads its TLS material from [audit.syslog.tls], the outbound TLS table of every FerroHEALTH product (trust_roots_file, client_identity_file or client_identity, min_version), and hands ihe-atna the rustls configuration ferrohealth-server builds from it. The keys tls_ca_file, tls_identity_cert_file and tls_identity_key_file are retired and refused at boot by name: name the roots file as trust_roots_file, and concatenate the client certificate chain and its private key into the one PEM client_identity_file names. The client speaks TLS 1.3 unless min_version = "1.2" lowers it (#5217).

  • A SCIM Group membership is now held by the member: an inline PARTY_RELATIONSHIP from the member’s PERSON to the GROUP, stored in the PERSON’s version (RM demographic §Party Relationships). A membership change versions only that PERSON, so its cost no longer grows with the Group’s size. The GROUP is versioned only by a rename, an externalId change or its deletion. The Group’s members is read from the membership index (#5220).

  • A Group’s ETag is now its GROUP version plus a membership revision, so it still changes with every membership change. The member User’s ETag now changes too, while its body stays the same (#5220).

  • Deactivating a User (active = false) drops all of its memberships in the same version, because an inactive PERSON cannot be versioned again. A reactivated User belongs to no Group until its client adds it again (#5220).

  • Microsoft Entra ID’s disable-then-remove order now succeeds, since the later remove finds nothing to remove. An add naming an inactive User is 400 invalidValue (“reactivate the User first”) (#5220).

  • A refused SCIM write now names the rule it breaks in detail (#5220).

  • scripts/perf/scim-groups.sh keeps the database log, df -h and free -m of a failed run (#5220).

  • A party commit no longer re-derives the login binding’s side table, the reverse-relationship index and the protection check from the whole stored body. Each hook reads what the committed body holds, derived once in memory, and writes only the rows that changed, and ferroehr party-binding refill derives the rows with the same code. A party commit without a server signature no longer reassembles its body. A membership change at the cap of a 5 000-member SCIM Group went from 810 ms to 358 ms at p99 on a developer laptop (#5220).

  • openehr-rm writes the RFC 8785 canonical form a VERSION signature covers in one pass over the version, sorting each object’s members by reference. A 3.3 MB GROUP version canonicalises about seven times faster than through serde_jcs, with byte-identical output, so signed versions verify unchanged. serde_jcs remains the test oracle (#5220).

  • The stored-query service takes the typed name, version and version pattern of openehr_its::rest::stored_query end to end, so it reads one grammar with the REST layer: the three-part {namespace}::{formalism}::{query-name} name, which ITS-REST 1.1.0 Qualified_query_name.md does not define, is no longer accepted anywhere (the formalism is the query_type parameter), and DELETE /admin/query/{qualified_query_name}/{version} answers 400 for a name outside [{namespace}::]{query-name} or a version that is no exact major.minor.patch, where it answered 404 (#5223).

  • Boot and ferroehr config check refuse an [[auth.party]] table whose identifier_type is the type of no [[authz.party_binding.person.template]] key, naming the issuer, the type and the types the templates write: no login of such an issuer could ever match a party, and every one would be refused. Upgrade step: give the deployment a person template whose key carries the table’s identifier_type (or set the table to the type the template writes). The Helm CI and production values and the hosted sandbox now carry the professional PERSON and ROLE mapping, and the sandbox loads the two templates before the CDR starts (#5231).

  • openehr_query::printer::to_aql writes each WHERE comparison with one space on each side of the operator (c/name/value = $nameValue), as the QUERY 1.1.0 AQL syntax examples do; path predicates keep their unspaced form ([ehr_id/value=$ehrUid]). The query builder’s generated AQL follows (#5241).

  • openehr_sdt::flat::navigate::Navigator owns what it copies from its Web Template (the name-fallback table and the OPT2 archetype references) and carries no template lifetime, so a consumer can keep it beside the template it was built from; it is Debug, Clone, Send and Sync (#5243).

  • Breaking: the configuration tree is one table per concern. ferroehr.toml moves to [deployment], [api], [database], [auth] (with [auth.smart] and [auth.scim]), [authz] (with [[authz.origin]]), [management], [audit], [privacy] (with identifier_protection, cohort and access_delay), [record] (with subject_index, versioning, signing, query, change_feed and multimedia), [terminology] and [pixm]. FerroEHR’s own keys take one duration type ("500ms", "30s", "10min", "12h", "7d", UCUM unit codes, so a bare number and m are refused), one size type ("256KiB", "64MiB", IEC binary units, decimal units refused) and one calendar retention period ("P5Y", "P1830D", unset keeps forever). An optional feature is on when its table is present: [terminology], [pixm], [auth.scim], [audit.syslog], [audit.fhir_feed], [privacy.identifier_protection] and [privacy.access_delay] carry no enabled; [authz.abac], [record.multimedia] and [auth.smart] keep a switch because some of their keys apply while off. record.subject_index has no default, so the server refuses to start until "pixm" or "local" is chosen. [api.cors] replaces cors_permissive with the Fetch Standard response headers (allow_origins, allow_methods, allow_headers, expose_headers, allow_credentials, max_age). The environment follows the tree (FERROEHR__DB__URL is FERROEHR__DATABASE__URL, FERROEHR__EHR_INDEX__PROVIDER is FERROEHR__RECORD__SUBJECT_INDEX). The Helm chart moves to its 25 line with config: on the same tree. Every key a release before v5.7.0 read is refused at load with its replacement named, all at once, so ferroehr config check --config <old file> prints the whole move list; the refusals ship in v5.7.0 only and the next release removes them. To upgrade, run config check on the old file and move each key it names: spec_profile, deployment_profile, deployment_accepts and licence.file are deployment.spec_profile, deployment.profile, deployment.accepts and deployment.licence_file; [rest] is [api] (swagger_ui is openapi_ui, body_limit_bytes is body_limit, cors_permissive = true is [api.cors] allow_origins = ["*"], [rest.identity] is [api.options] with solution and vendor; the version keys are facts of the build), admin.enabled is api.admin and terminology.api_enabled is api.terminology; [db] is [database] (acquire_timeout_secs is acquire_timeout, statement_timeout_ms is statement_timeout, migrate is migrations), [storage.<domain>] is [database.<domain>] and [storage.migrate.<domain>] url is [database.<domain>] migrate_url; [smart] is [auth.smart] (enabled is discovery, episode.enabled is episode_launch, launch_base64_json = true is "launch-base64-json" in auth.capabilities); [scim] is [auth.scim] (max_results is filter_max_results, deactivation_window_minutes is deactivation_window, contact_erasure_minutes is contact_erasure_delay); [authz.abac.cedar] reload_secs is reload_interval, [authz.abac.remote] server, connect_timeout_ms and request_timeout_ms are url, connect_timeout and timeout, authz.party_binding.provisional_report_days is provisional_report_after, and an agreement’s discard_window_hours and chained_log_retention_years are discard_window and chained_log_retention; [[origins.registered]] is [[authz.origin]]; management.port is management.listen (an address:port) and management.profiling.max_seconds is max_duration; audit.store.retention_days and retention_years are audit.store.retention, verify_interval_seconds and full_verify_interval_seconds are verify_interval and full_verify_interval; [audit.syslog] host and port are url = "tls://host:port", spool_max_messages is spool_max_events, spool_max_bytes is spool_max_size, retry_max_ms is retry.max_backoff, and [audit.fhir_feed] poll_interval_ms and retry_max_ms are poll_interval and retry.max_backoff; [demographic.identifier_protection] is [privacy.identifier_protection], [cohort] is [privacy.cohort], and [access_delay] is [privacy.access_delay] with hours as max_delay; ehr_index.provider is record.subject_index ("ihe" is "pixm") and the rest of [ehr_index] is [pixm] (pixm_url is url, invocation is method, request_timeout_ms is timeout); versioning.skip_identical_resend is record.versioning.idempotent_resend; [signing] is [record.signing] (key_path is key_file, retired_key_paths is retired_key_files), [query] is [record.query] (timeout_ms is timeout), [outbox] is [record.change_feed] (retention_days is retention, prune_interval_secs is prune_interval), and [multimedia] is [record.multimedia] (endpoint is url, threshold_bytes is threshold, and allow_http is retired: an http:// url is plain HTTP); [terminology.server] is [terminology] (fail_on_error is on_error = "accept" | "refuse", request_timeout_ms is timeout). The family sections take the same types, from ferrohealth-server 0.8, ferrohealth-telemetry 0.4, ferrohealth-smart 0.15 and ferrohealth-session 0.9: [server] request_timeout_ms, header_read_timeout_ms, idle_timeout_ms, drain_delay_ms, shutdown_timeout_ms and retry_after_s are request_timeout, header_read_timeout, idle_timeout, drain_delay, shutdown_timeout and retry_after, body_limit_bytes and compression_min_bytes are body_limit and compression_min_size, and [server.tls] reload_interval_s is reload_interval; [telemetry.otlp] trust_roots_file and client_identity_file move into [telemetry.otlp.tls], and an https:// collector is dialed over it; [auth] ca_file is [auth.tls] trust_roots_file, an [[auth.issuer]] table’s clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds and timeout_ms are clock_skew, jwks_max_age, jwks_refresh_interval and timeout, its introspection cache_max_age_seconds is cache_max_age, [auth.basic] cache_ttl_seconds is cache_ttl, [auth.dpop] max_age_seconds and clock_skew_seconds are max_age and clock_skew, [auth.idp.entra] timeout_ms and max_response_bytes are timeout and max_response_size, and an oauth grant’s ca_file, client_certificate_file and client_key_file are its tls table’s trust_roots_file and client_identity_file (one PEM), with refresh_margin_ms, timeout_ms and previous_key_overlap_ms as refresh_margin, timeout and previous_key_overlap. The [auth] keys v5.6 read for every issuer (audience, algorithms, token_profile, jwks_file, mode, introspection, role_claims and the rest) live on each [[auth.issuer]] table, and scopeless_subjects entries are { issuer, sub } tables. In the viewer, cdr.request_timeout_secs is cdr.timeout, [oidc] ca_file is [oidc.tls] trust_roots_file, [oidc.tls] client_certificate and client_key are client_identity_file, and [session] idle_timeout_s, absolute_timeout_s and sign_in_timeout_s are idle_timeout, absolute_timeout and sign_in_timeout. Each of these is refused in v5.7.0 naming its replacement, in the CDR and the viewer alike (#5259).

  • Breaking: a patients token that names an actor is refused. With ferrohealth-smart 0.15, a patients issuer reads a token whose top-level act (RFC 8693 §4.1) names someone other than the patient only when its [[auth.issuer]] table sets act_represents_patient = true; without it the token is refused 401, where v5.6 admitted it as the patient’s own access. With it, FerroEHR refuses the token 403 (refused-patient-representative), because it binds no party to someone acting for a patient, and the access log records the refusal (#5259).

  • SECURITY.md, security.txt (its Contact: mailto: field), the post-market record, the CRA technical documentation and the book’s compliance pages name security@cadasto.com as the email route for a vulnerability report, in place of info@cadasto.com, which stays the single point of contact for complaints and incidents, the manufacturer’s address on the running server and the address for licensing and other business questions (#5272).

  • The shipped ferroehr.toml template (app/ferroehr/assets/ferroehr.default.toml) is generated from the configuration types with tools/config-reference: every key, each under its description and set to its default, a secret or a key with no default as a #? placeholder line. The book’s new page “Every configuration key” includes it whole. CI fails when it drifts from the types. No key changes: a ferroehr.toml that loads today loads unchanged (#5316).

  • The family crates move to their newest releases: ferrohealth-smart 0.18.0, ferrohealth-session 0.12.0, ferrohealth-authz 0.11.0, ferrohealth-server 0.8.4 (its TLS certificate verifier is built once per process), ferrohealth-telemetry 0.4.1, ihe-atna and ehds-logging 0.12.0, ihe-iti 0.4.0, ferrobridge-term 0.5.0, and fhir-types and fhir-search 0.1.131. Every [[auth.issuer]], onward and OAuth issuer renders with its credentials masked. A system caller, a patient and a patient’s representative carry the purpose of use their token states when their issuer has an [[auth.attributes]] table. An IUA purpose_of_use of empty codes states no purpose (#5320).

  • ferroehr and ferroehr-viewer run through the FerroHEALTH family’s run path. A configuration either binary refuses (an unknown or invalid key, a broken semantic rule, an [auth] table with no mechanism, a SCIM mapping it cannot serve, or for the viewer a TLS file that does not read or an unusable [oidc] or [session] table) now exits 78 (EX_CONFIG of sysexits(3)) where it exited 1, so an orchestrator can tell a configuration to fix from a failure to retry. Every other failure still exits 1, and the failure is written to stderr as one line with its whole cause, prefixed with the binary’s name. Upgrade: a script or probe that expects exactly 1 from config check on a refused file expects 78 (#5321).

  • The configuration is redacted by its secret types alone. GET /management/env, GET /ferroehr/rest/extensions/v1/admin/config and ferroehr config check show the tree as its Secret and SecretUrl values serialize, and the second masking by key name and by URL shape is gone. [audit.syslog] url, [auth.smart] platform_base_url and public_base_url, and [auth.scim] public_base_url are secret URLs now, as are the OAuth, onward and introspection endpoints of ferrohealth-smart 0.17 and the viewer’s [oidc] endpoints of ferrohealth-session 0.11, so each renders with its credentials and query masked. A platform_base_url given as a bare path renders as *** (#5322).

  • Every outbound integration (the terminology server, the PIXm identity source, the ITI-20 FHIR Feed, the remote policy decision point and the viewer’s CDR client) is built by ferrohealth-smart’s OutboundClient from its endpoint table, and FerroEHR’s own client builder is gone. A PIXm exchange’s timeout now bounds the exchange with its tries and the waits between them, and the viewer refuses a [cdr] timeout of zero at start (#5324).

  • The family crates move to their newest releases: ferrohealth-smart 0.17.0, ferrohealth-session 0.11.0, ferrohealth-authz 0.10.0, ferrohealth-config 0.4.4, ferrohealth-server 0.8.3, ferrohealth-secret 0.2.1, ferrohealth-usage-report 0.4.1, ferrobridge-term 0.4.0, ihe-atna and ehds-logging 0.11.0, and ihe-iti 0.3.0. Every OAuth token endpoint and introspection endpoint renders with its credentials masked, and a TLS refusal of an outbound client names its <key>.tls table (#5324).

  • [auth.idp.entra] is read by ferrohealth-smart’s [auth] section, so an unknown key under it is refused with the crate’s field list. The key path is unchanged. The effective configuration now shows [auth.idp] as an empty table when no adapter is configured (#5325).

  • The build compiles FHIR R4’s resources alone: fhir-types is required with its r4-resources feature in place of the crate-wide resources, and the FerroBRIDGE crates (fhir-types, fhir-search, ferrobridge-term) are required by caret ranges in place of exact versions (#5327).

  • Breaking: the outbound integrations take the family’s endpoint table. [terminology], [pixm], [audit.fhir_feed] and [authz.abac.remote] read where they call and how from one [<x>.endpoint] table, the same keys for every outbound integration of every FerroHEALTH product: url, timeout, connect_timeout, retry (max_attempts, initial_backoff, max_backoff), tls (trust_roots_file, client_identity_file, min_version) and credentials (HTTP Basic user with password or password_file, or the client credentials grant [...credentials.oauth]). terminology.url, timeout and [terminology.oauth] are terminology.endpoint.url, timeout and [terminology.endpoint.credentials.oauth]; pixm.url, timeout, [pixm.tls] and [pixm.oauth] are pixm.endpoint.url, timeout, [pixm.endpoint.tls] and [pixm.endpoint.credentials.oauth]; audit.fhir_feed.url, [audit.fhir_feed.retry] and [audit.fhir_feed.oauth] are audit.fhir_feed.endpoint.url, [audit.fhir_feed.endpoint.retry] and [audit.fhir_feed.endpoint.credentials.oauth]; authz.abac.remote.url, connect_timeout, timeout and [authz.abac.remote.oauth] are authz.abac.remote.endpoint.url, connect_timeout, timeout and [authz.abac.remote.endpoint.credentials.oauth]. The v5.6 keys are refused naming their new home. Every client now honours retry: the terminology, PIXm and remote policy clients try a call that got no answer or a 5xx up to max_attempts times (three by default; each tried once before), so set max_attempts = 1 to keep one attempt, and the FHIR feed’s delivery rounds take max_attempts tries each, from initial_backoff, before the round waits max_backoff; the feed still never gives a record up. The environment follows the tree (FERROEHR__TERMINOLOGY__ENDPOINT__URL, FERROEHR__AUDIT__FHIR_FEED__ENDPOINT__URL), and so do the Helm chart’s values and its client secret routing, in the chart’s 25 line (#5331).

  • The FerroEHR PostgreSQL image is built on the 2026-11-09 respin of postgres:18.6 (sha256:e3d24f69…), a same-version respin with rebuilt base packages. A full scan of the rebuilt image on amd64 and arm64 finds the same findings as the image of the previous base, each judged in security/vex/postgres-os.openvex.json, and a gate scan finds no fixable HIGH or CRITICAL finding (#5332).

  • The development and Keycloak example compose files run Keycloak 26.8.2 (quay.io/keycloak/keycloak:26.8.2@sha256:723b1d6e…). Of the 26.8.2 upgrading notes, the shipped realm meets the ones that touch it: the Admin REST API’s new audience check passes for the master realm’s admin-cli token the browser-test harness uses, and the realm uses no token exchange, identity broker, SAML client, CIBA or dynamic client registration. The ferroehr-service client-credentials token carries the same audience and scopes as under 26.8.1 (#5333).

  • The compose quickstart, the hosted sandbox and the Helm chart deploy FerroPIX 0.0.26 (ghcr.io/ferrohealth/ferropix:0.0.26@sha256:85d40e6fdcdb39f782007eff91eb059861a4e039ca3f8fb56e37f2fa0f28d78a, the FerroPIX chart 0.3.19) and FerroTERM 0.4.5 (ghcr.io/ferrohealth/ferroterm:0.4.5@sha256:7c20d39e1989fce1dfb92e6ddfd853dfa9a52bb9b7b973fbba42430963ef75bd), and the development issuer is mock-oauth2-server 6.1.5. FerroPIX’s configuration moves to the [[auth.issuer]] shape it now requires: the issuer is kind = "machines" and carries its own audience, and [auth] audience is gone. A chart deployment that sets ferropix.config.auth moves its audience onto each issuer the same way (#5351).

  • The nine openehr-* crates follow The Cargo Book’s SemVer rules, judged by cargo-semver-checks 0.51.0 against their newest release on crates.io on every pull request, in the merge queue and in the release plan. Their next publish moves the line from 0.0.x to 0.1.0. From then on a compatible change ships as a patch and only a breaking change moves the minor, so a caret requirement such as openehr-rm = "0.1" takes every compatible release. A public dependency moved to another compatibility line counts as a breaking change (#5353).

  • openehr-rm writes the ECMAScript number form of the RFC 8785 canonical form a VERSION signature covers with ryu-js 1.0, which carries the upstream fix for a panic in its fixed-point formatter, and is held byte-equal to serde_json_canonicalizer 0.4 as its test oracle in place of serde_jcs, so the lock holds one ryu-js line. Its output is unchanged (#5372).

  • ferroehr.eu takes the FerroHEALTH family layout: the “Oxide & Iron” brand tokens light and dark, a hero badge rendered from the workspace version at build time, a content hash on the stylesheet link, linked capability cards, the conformance record, the viewer screenshot, deployment cards, the family and the four parts of the book, and a 404 page in the same style. The book groups its pages under Evaluate, Integrate, Operate and Contribute, keeping every page’s address, opens in the recoloured light and navy themes and gains a Changelog page. The README’s quick start is the anchored steps of docker/smoke-test.sh, held by scripts/checks/readme-quickstart.sh (#5382).

  • The quickstart docker-compose.yml each release attaches now carries its .sha256sum; the README and the book download it with curl -fsSLO from the latest release and start it with docker compose up --wait, and the README check refuses a host path the file would read from the repository beyond the backup directories (#5387).

  • The viewer moves to ferrohealth-session 0.13.0, which types the OpenID Provider issuer, redirect_uri and post_logout_redirect_uri of [oidc] as secret URLs, so a password written into one of them renders masked in the viewer’s config check, its refusals and its configuration page (#5394).

Removed

  • [authz.professional] (claim, client_tokens, acting_for_claim) and its 403 for a token that names no natural person. The issuer’s kind and profile classify the caller, and the access record’s acting mode and professional come from that classification: a person’s own token is the person, and a delegated token is the client acting for the person it names. Upgrade step: remove the table, which refuses the start, and list a client that acts for a person in its issuer’s delegates (#5003).

  • The named refusal of authz.rbac.user_role, which v5.6 introduced, is gone under the one-release rule for removal aids: the key is now refused as an unknown key when the configuration loads (#5259).

Fixed

  • A 553|incomplete| commit no longer refuses an empty mandatory archetype slot: RM common master06 §Incomplete Content sets every occurrences lower limit to zero, a slot’s included, while too many fillers are still refused. The OPT 1.4 writer now closes a mandatory slot once the fillers placed in it carry its lower bound, as ADL 2 §Slot Filling and Redefinition allows, and still refuses to close one they leave unmet (#5042).

  • The example of a party template whose identity the template marks optional is a party of its class carrying one identity at every detail level, where the required level answered 500: RM demographic PARTY.identities admits no party without one. A party template that constrains no identity at all answers 400 Bad Request naming its class (#5057).

  • An AQL VERSION standard predicate on commit_audit/time_committed applies its operator: VERSION v[commit_audit/time_committed/value > $t] returns every version committed after $t, where it returned the trunk version current at $t whatever the operator (QUERY 1.1.0 master03 §Standard predicate). A number, boolean or NULL literal as the bound, a string bound no ISO 8601 form reads, and a time_committed path other than the whole value or /value are refused with 400 (#5062).

  • AQL refuses a number, boolean or null parameter or literal compared with a date/time path, in WHERE and in a VERSION predicate, with the same 400 a non-ISO 8601 string gets; such a bound used to become NULL and match nothing (#5064).

  • The served OpenAPI document now lists, on both template example operations, the 400 answered for a party template that constrains no PARTY_IDENTITY (#5066).

  • The compose quickstart, the hosted sandbox and the Helm chart run FerroPIX 0.0.18 (ghcr.io/ferrohealth/ferropix:0.0.18@sha256:6fb622b5b9ade34c8a4053112c8f166ff495d8b32bf99f35529d01e512b7cc42), with the FerroPIX chart 0.3.10 built for it. The 0.0.16 they pinned had an outbox sender that never slept while its outbox was empty, so an idle install kept about 1.2 CPU cores busy between FerroPIX and its PostgreSQL at about 8,000 transactions a second (FerroHEALTH/FerroPIX#490). 0.0.18 also makes FerroPIX’s [auth] scopeless_subjects and role_claims hold per issuer, which changes nothing for a FerroPIX that trusts one issuer, as every configuration FerroEHR ships does. The chart moves to 21.0.5 (#5073).

  • An OPT2 built from a template that specialises a published archetype now carries that archetype, by its full id, as parent_archetype_id, as the OPERATIONAL_TEMPLATE invariant is_specialised requires (AOM2 §3.6.6), so a reader of the OPT finds the archetype its root specialises. The ADL form of an OPT still prints no specialize section (OPT2 §Artefact Structure), so the parent travels in the canonical JSON form (#5075).

  • A refused-key-claim-missing access record names the key claim the issuer’s binding reads (its person_key) in binding_detail.claim, so an operator sees which claim the token lacked. It records the name only, never a claim value (#5139).

  • A 403 for a caller kind is recorded with its own binding code: refused-system-personal-scope for a system caller holding a user/ or patient/ scope, refused-no-agreement for an external caller whose issuer has no agreement for its accessor kind, and refused-scope-beyond-agreement for an external caller holding a scope its agreement does not read. Audit migration 0024 widens the binding constraint; the problem text of each refusal is unchanged (#5140).

  • openehr-sdt builds no /context node, and no start_time or setting leaf, in the Web Template of an OPT 1.4 whose COMPOSITION prohibits context (existence 0..0, or an EVENT_CONTEXT of occurrences 0..0), and the template example of such a template carries no context. A template that leaves context open or optional keeps the node as before (#5147).

  • openehr-sdt gives the /context node it synthesizes in a Web Template the min of the template’s existence on COMPOSITION.context, and 0 where the template leaves context unconstrained, since the RM makes context optional (0..1). A client following the Web Template no longer treats the context as mandatory when the template does not require it, and the required template example of such a template carries no context (#5154).

  • An ADL2 template that deletes an attribute (existence matches {0}), for example COMPOSITION.context, keeps the prohibition after it compiles to its operational template: openehr-adl’s create_opt now returns a RawOpt that names each removed attribute beside the OPT, and build_web_template_v2_4 takes that list. A template that deletes context gets no /context node, and a commit that carries one is refused, as it is for an OPT 1.4 template that prohibits context. An OPT2 uploaded already compiled carries no such list and reads a deleted attribute at its reference-model existence. Breaking for crate users: create_opt returns RawOpt and build_web_template_v2_4 takes the prohibited attributes (#5154).

  • A composition commit whose template_id names a stored ADL2 archetype rather than a template is refused with 422 as an unknown operational template; it failed at the insert with a server error before (#5177).

  • openehr-rm types REFERENCE_RANGE.range as DV_INTERVAL<T>: ReferenceRange is now generic over its limit type (ReferenceRange<DvQuantity> on a DV_QUANTITY, ReferenceRange<DvOrdered> on the other ordered types), its is_in_range and Range_is_simple work at that type, and the static RM model gives REFERENCE_RANGE.range and IMPORTED_VERSION.item their argument. Each model::RmAttribute and model::RmTypeRef now keeps the formal parameter it is written as (formal, Some("T") for DV_INTERVAL.lower) beside the resolved bound. The RM class table declares REFERENCE_RANGE<T> and leaves T unused; the generator restores it from a cited table and refuses a schema that no longer matches (#5196) (#5188).

  • An access record names the client the token’s issuer profile reads from client_claims (azp under the entra and keycloak profiles), the one the caller was admitted as. A token carrying a different client_id claim beside it was recorded under that claim. The [origins] register and the SCIM client lookup read the same client (#5194).

  • An Interval or DV_INTERVAL read from canonical JSON or XML that omits a boundedness or inclusion flag now takes it from its bound. An absent bound reads as unbounded and not included, and a present bound as bounded and included. A flag the wire states is kept as written. Before, an omitted flag always read as bounded and included, so an interval with only an upper came out with a bounded, included lower side that had no value. ITS-JSON requires all four flags, and BASE ties none of them to its bound, so reading the omission is our own design (#5203).

  • The viewer’s New contribution commits as the signed-in login. Each audit carries the placeholder naming that login, and the server records the committer it derives from the login. The free-text committer field is gone. For a login bound to a demographic PERSON, a stated name that differs from the PERSON’s name was refused with 403 (#5222).

  • The management surface answers a caller the login binding refuses with 403 and its binding code, such as refused-unbound, as every other route does, in place of 401: the credential authenticated, so new credentials cannot help (RFC 9110 §15.5.4). An absent or invalid credential still answers 401 with its challenge (#5250).

  • The party binding posture in ferroehr report and the boot log masks the user name and password of an issuer URL, as every other rendering of the configuration does (#5259).

  • GET /definition/query/{namespace}:: lists every stored query in the namespace again, where it answered 400 since the typed stored-query name: ITS-REST 1.1.0 definition_query_list reads its name as a pattern in which an empty part is a wildcard, so ::{name} also lists the names that start with it in any namespace. openehr-its gains rest::stored_query::QueryNamePattern, the type the generated DefinitionQueryListParams carries. The store and the single-query reads still refuse an empty name or namespace (#5268).

  • The identifier scan ([privacy.identifier_scan]) reads the name of the ELEMENT that holds a DV_IDENTIFIER when the identifier’s type is absent or names neither a person nor an item, with the same word rules. An untyped laboratory order number in the Receiver order identifier or Requester order identifier element of laboratory_test_result.v1 that passes the NHS Number check by chance is now stored under the default rule set. A person word still wins, so an untyped value in an element named NHS number or Patient identifier stays scanned, and a type that names a kind takes precedence over the element name (#5291).

  • The Helm chart’s description and README name config.deployment.profile, the key the chart refuses to render without since 5.7.0, and the README’s install command sets it; the command it showed set the old config.deployment_profile and did not render (chart 25.0.1) (#5341).

  • A representative’s write is held to the EHDS Art. 5 rules exactly as the person’s. A representative may no longer alter an entry a health professional inserted or name a professional as composer. A version committed as PARTY_RELATED counts as inserted by the person, so the person and the representative may change it, and the contribution list reports inserted_by_person = true for it (#5359).

  • The hosted sandbox’s configuration (deploy/hosted/ferroehr.sandbox.toml) declares its SCIM service provider under [auth.scim] and [[auth.scim.client]], the tables 5.7.0 reads, so the sandbox starts on 5.7.0. Every server configuration the repository ships beside the quickstart and the chart (the hosted sandbox’s and each compose overlay’s) and every viewer configuration (the hosted sandbox’s, the quickstart’s with each overlay, and the chart’s viewer Deployment) is now read by the program built from the change under test in CI, so a shipped file the release refuses fails its pull request (#5396).

Security

Shipped with functional changes because: the advisory reaches only a deployment that trusts several issuers, which no configuration FerroEHR ships does, and its fix changes what such a configuration must state at boot, so it lands on main beside the AQL version attestations already merged for this release; a security-only patch of the last release remains open to the release cut.

  • A Basic user never carries a system/ scope, through ferrohealth-smart 0.9.1 (GHSA-jhg2-jwjp-hq74, CWE-269, CWE-863). A system/ scope granted to a Basic user gave that person every resource of a type, a backend service’s reach. Such a user now refuses the start, naming the user and the scope. Upgrade step: give each Basic user the user/ scopes a person holds, and move each machine on Basic to a client of the authorization server under the client credentials grant (#5003).

  • [auth] scopeless_subjects and role_claims hold per issuer, through ferrohealth-smart 0.8.9 (GHSA-pfgx-xq28-2629, CWE-863, CWE-639). Before it, beside several [[auth.issuer]] tables both lists applied to every trusted issuer’s tokens: a token from any of them whose sub was listed met every SMART on openEHR requirement without a scope, and any of them could assert every role the RBAC checks, ADMIN included. An entry is now written <issuer>#<value> and holds for that issuer alone; beside several issuers a bare entry, or one naming an issuer the section does not trust, refuses the boot. Upgrade step for a deployment with several issuers: write each scopeless_subjects entry as <issuer>#<sub>, and write each role_claims path as <issuer>#<path> for the issuers whose roles count (or set role_claims = [] to read none), since the default role_claims is bare and refused. A deployment with one issuer changes nothing (#5047).

  • The ferroehr-postgres image carries zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 from trixie-security, which fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). The image of 5.6.3 and earlier releases carries 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action (security/vex/postgres-os.openvex.json). The ferroehr and ferroehr-viewer images keep the distroless base’s zlib, which neither program loads (security/vex/ferroehr-os.openvex.json) (#5137).

  • The ferroehr and ferroehr-viewer images move to the gcr.io/distroless/cc-debian13:nonroot build of 2026-11-08 (sha256:cc66d222…), whose zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). Neither program loads zlib. The images of 5.6.3 and earlier releases carry 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action, as the ferroehr-postgres image of those releases is (security/vex/ferroehr-os.openvex.json) (#5152, #5329).

  • CVE-2026-112321, CVE-2026-112322 and CVE-2026-112323 (memory and CPU exhaustion in kadmind through the AUTH_GSSAPI handler) and CVE-2026-119079 (CPU exhaustion in the KDC decoding an AS-REQ or TGS-REQ) in MIT Kerberos 5 have no Debian fix for trixie and are judged not_affected in ferroehr-postgres. The image carries neither kadmind, a KDC nor libgssrpc, and the PostgreSQL server never decodes either request; the configuration FerroEHR ships enables no GSS-API authentication (security/vex/postgres-os.openvex.json) (#5344).

  • Thirteen MIT Kerberos 5 CVEs of MITKRB5-SA-2026-001 in the ferroehr-postgres image’s libgssapi-krb5-2, libk5crypto3, libkrb5-3 and libkrb5support0 1.21.3-5+deb13u1 have no Debian fix for trixie and are judged not_affected. The image keeps the four libraries because the postgres binary links libgssapi_krb5.so.2. Eleven sit in code the image does not carry: the KDC (CVE-2026-107364), its PKINIT plugin (CVE-2026-71200), kadmind (CVE-2026-107884, CVE-2026-112318), libgssrpc (CVE-2026-112314, CVE-2026-112315, CVE-2026-112320, CVE-2026-118380), the kadm5 libraries (CVE-2026-112316, CVE-2026-112319) and libkdb5 (CVE-2026-112317). The PAC checksum flaw (CVE-2026-111423) and the CAMMAC checksum flaw (CVE-2026-71199) sit in libkrb5, which the server reaches only when it accepts a Kerberos ticket. The configuration FerroEHR ships has no gss or hostgssenc line in pg_hba.conf and no server keytab, and the server never reads a PAC or an authentication indicator (security/vex/postgres-os.openvex.json) (#5374).

  • The five OpenSSL 3.5.7 CVEs of the OpenSSL advisory of 2026-11-10 have no Debian fix for trixie and are judged not_affected in the three images: the AES-GCM-SIV flaws CVE-2026-88823 and CVE-2026-95692, the QUIC flaws CVE-2026-88822 and CVE-2026-95691, and the CMS_encrypt() loop CVE-2026-107496. The ferroehr and ferroehr-viewer programs never map libssl.so.3 or libcrypto.so.3: each links only libc, libm and libgcc_s, imports no dlopen, and speaks TLS through rustls (security/vex/ferroehr-os.openvex.json). The PostgreSQL server maps both for TLS over TCP, which never selects AES-GCM-SIV (no TLS cipher suite is one), never opens a QUIC connection and never writes CMS, and the image never runs the openssl program or loads the legacy provider (security/vex/postgres-os.openvex.json) (#5393).

5.7.0-rc3 - 2026-11-11

5.7.0-rc3 is the third release candidate of 5.7.0, published for testing. It carries every change since 5.6.3, those of the 5.7.0-rc2 and 5.7.0-rc1 sections below included. 5.6.3 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.7.0 itself replaces it. Since 5.7.0-rc2, FerroEHR runs on the newest FerroHEALTH family releases (ferrohealth-smart 0.18.0, ferrohealth-session 0.12.0, ferrohealth-authz 0.11.0, ihe-atna and ehds-logging 0.12.0, ihe-iti 0.4.0, ferrobridge-term 0.5.0), every outbound client is built from the family endpoint table, and the configuration is redacted by its secret types alone, so every URL and issuer renders with its credentials masked. The quickstart, the hosted sandbox and the chart move to FerroPIX 0.0.26 and FerroTERM 0.4.5. The audit record takes the purpose of use and the professional role from an IHE IUA token and records a contact point’s request message id; audit migration 0026 adds their columns. The openehr-* crates are judged by Cargo SemVer, and their next publish moves the line to 0.1.0. Before you try it, run ferroehr config check on your file: it names every key a release before 5.7.0 read with its replacement, all at once. The steps for party binding are on the book’s bootstrap and migration page. Helm chart 25.0.1-rc4 carries appVersion 5.7.0-rc3. Security carries the two advisories of 5.7.0-rc2, GHSA-jhg2-jwjp-hq74 and GHSA-pfgx-xq28-2629, the zlib fixes in the three images, and the not_affected judgements of the MIT Kerberos CVEs in the PostgreSQL image, the thirteen of MITKRB5-SA-2026-001 among them, and of the five OpenSSL 3.5.7 CVEs of 2026-11-10 in the three images.

Added

  • AQL reads the attestations of a version: v/attestations and any ATTESTATION attribute below it (is_pending, reason, proof, items, committer, time_committed) under FROM EHR e CONTAINS VERSION v CONTAINS COMPOSITION c, with predicates on attestations and EXISTS, and v/commit_audit/is_pending and v/commit_audit/reason on a commit audit that is an ATTESTATION. SELECT v/attestations returns the list, and a comparison below attestations holds when any attestation matches. The FerroEHR extension function LAST(v/attestations/<path>) reads the most recent attestation, so WHERE LAST(v/attestations/is_pending) = true lists the versions awaiting signature, including one marked pending again after it was signed (#4989).

  • The party binding’s configuration, each table with its boot refusals: [authz.party_binding] (provider as one table or one table per tenant, namespace, basic_user), [authz.import] trusted_systems, [[authz.system_client]] (the provider, facilities, templates and origins of each client-credentials client that commits) and [[authz.federation.agreement]] (what a federation issuer’s external persons or organisation may do, on which legal ground), with [authz.rbac] role_issuers and import_role. A template an agreement or a system client writes that the store does not hold refuses the start. The boot line and GET /management/info list the issuers with their kind and identifier issuer, the provider, and every agreement with its accessor and legal ground, and an agreement that admits an organisation without a named person is logged at warn (#5003).

  • [auth.idp.entra], the Microsoft Graph access of the idp-entra adapter of ferrohealth-smart, which resolves the groups of an Entra ID groups overage for an issuer whose profile sets adapter = "idp-entra" (#5003).

  • [auth.dpop] and [[auth.issuer]] sender_constraint: tokens bound to a DPoP key (RFC 9449) or a mutual-TLS client certificate (RFC 8705), checked on every request. A DPoP proof is recorded by the replica that accepted it, so the Helm chart refuses a sender-constrained issuer on more than one replica (#5003).

  • The login binding reads the local party source on every request, with no cache: the new party migration 0027_binding_side_table adds the indexed binding_identifier, binding_membership and binding_representation relations, written in each party commit’s transaction from the PERSONs of the mapped templates that perform a professional ROLE or represent another person, their GROUP memberships and their representation relationships, and rebuilt from the party versions alone. A login matching two parties, an inactive, abandoned or deleted party, a token without the key claim, a party without the needed marker and unreadable groups are refused with 403 and their binding code; a party store or identity provider that cannot answer is answered 503 (#5004).

  • ferroehr definitions load <path>... stores OPT 1.4 (.opt) and ADL 2 (.adls) definitions before serve, through the validation and versioning of the REST upload. A file already stored with the same text is left alone and other text under a stored id is refused, so a deployment runs it at every start to give its party mapping the templates it needs at boot. deploy/definitions/professional/ holds a professional PERSON and ROLE pair adapted from openEHR CKM archetypes for the development lanes and the quickstart (#5006).

  • The compose quickstart binds its user to a professional PERSON: the ferroehr-definitions one-shot loads the PERSON and ROLE templates the party mapping names before the server starts, and ferroehr-provision creates the user through /scim/v2/Users with a token of the development issuer, then reports healthy, so docker compose up --wait returns once the login is bound. The Helm chart (23.3.0) takes definitions.files, which an init container stores with ferroehr definitions load before the server container starts. The conformance, browser and Kubernetes lanes load the same templates and provision every login they use through SCIM, with the committed test client of docs/conformance/party/scim/ (#5006).

  • ferroehr party-binding adopt <PARTY_ID>... gives each prepared PERSON of an existing deployment a professional ROLE of the mapping’s role template, with the PERSON as performer, and a new PERSON version listing it in roles, all in one demographic CONTRIBUTION committed by the first administrator named with --admin or --admin-key, and writes one access record per adopted PERSON through the configured audit trail. It refuses the whole adoption, naming them, while any listed PERSON’s key was introduced by a committer that is no [[scim.client]], named administrator or --migration, so a key planted before the upgrade is never adopted. A PERSON that already performs the ROLE is reported and left alone. The report is JSON on stdout (#5009).

  • ferroehr party-binding check is the read-only pre-flight of a migration to party binding, and the one report of the provisional persons just-in-time creation left incomplete. It reads the party versions alone, so it runs before any boot fill, and writes nothing. Per key issuer it reports the PERSONs that carry a binding key, every selector value two PERSONs carry, the inactive and abandoned ones, the provisional ones older than provisional_report_days, the key-bearing ones that perform no professional ROLE, and every mapped Basic user with the PERSON it binds. For each key it names the committer of the version that introduced it, and fails unless that committer is a [[scim.client]], an administrator named with --admin or --admin-key, a migration declared with --migration, or just-in-time creation on an issuer with jit = true. --expected <file> lists every login of the identity provider with no PERSON. The report is JSON on stdout, and the command exits non-zero on a duplicate, a missing person or a key that fails the provenance rule (#5009).

  • At boot, FerroEHR logs a warn naming every Basic user mapped in [[authz.party_binding.basic_user]] while [authz.assurance] minimum is unset: a Basic credential carries no assurance level, so such a person is never refused on patient data (#5009).

  • The book has a “Who is the committer” page under Security. It covers the committer the server writes for each kind of caller, the placeholder a client may still send, the refusals with their status codes, and the deviation from the ITS-REST merge rule. It also shows how a professional’s PERSON gets its login identifier. Enterprise identity providers now opens with the generic standards path and adds the entra preset’s binding and provisioning settings, an Okta example, a Keycloak realm walk-through and a SAML section. The records of processing and the DPIA gain the workforce-identity activity, the federation node page gains its callers and agreements, and the go-live checklist checks the binding and the caller rates (#5010).

  • The access record of a MyHealth@EU contact point’s request carries the request message’s UUID, the WS-Addressing MessageID its token relays as ehdsi.message_id, as the record’s request-message-id. The record then joins the contact point’s Patient Privacy audit entry (eHDSI Audit Trail Profile 10.1.0 §2.3.12 to §2.3.32) (#5011).

  • The one PERSON mapping SCIM provisioning, just-in-time creation and the login binding share: [[authz.party_binding.person.template]] with the binding key, userName and externalId selectors of each template, [[authz.party_binding.person.attribute]] with the SCIM attribute, the OpenID Connect claim and the path of each mapped value, and [authz.party_binding.professional_role], [authz.party_binding.group], [authz.party_binding.membership] and [authz.party_binding.representation]. Boot refuses a mapping that breaks one of its rules, naming the key, including a template that is absent, rooted at another class or lacks a mapped path, and deleting a template the mapping names is refused with 409 (#5013).

  • The new party migration 0028_binding_selectors holds each binding key, userName (compared after RFC 8265 width mapping, lower case and NFC), externalId and group displayName to one live party, so a commit giving one a second party is refused with 409. The first boot fills the side table under the mapping’s digest, a replica with another mapping refuses to start, and ferroehr party-binding refill rebuilds it after a mapping change, refusing while a replica runs and naming the parties when a value is carried twice (#5013).

  • POST /admin/party/binding-key re-keys a PERSON’s binding key: it refuses the administrator’s own login and a key already bound, retires the old key in the same commit and writes one access record carrying both (#5013).

  • [scim] mounts the SCIM 2.0 service provider at /scim/v2/ (off by default): [[scim.client]] declares each provisioning client by issuer, client_id, scope and binding_issuer, and public_base_url names the origin its documents hang off. /scim/v2/ServiceProviderConfig answers without a token; /ResourceTypes and /Schemas answer a declared client acting for nobody that holds its scope, within [server.caller_rate]. Every refusal is an RFC 7644 Error message. The start is refused for a client of a non-machines issuer, one declared twice, two clients sharing a binding_issuer, or a binding_issuer no [[auth.party]] table carries (#5014).

  • The SCIM endpoint answers the Microsoft Entra ID provisioning client as Microsoft’s “Develop a SCIM endpoint” guide documents it: a value-filtered attribute such as emails[type eq "work"].value serves its type and declares primary read-only, so the primary Entra sends is ignored (RFC 7644 §3.3); a PATCH of userName renames the User; and the Test Connection probes are answered with an empty ListResponse and 404. ferroehr config check refuses a [scim] the person mapping cannot serve, as serve does. The Helm chart exposes config.scim and routes /scim/v2 on every Ingress host when it is enabled, and the book walks an operator through provisioning from Entra ID, with the Workload Identity Federation checks the authorization server makes (#5014).

  • /scim/v2/Users/{id} takes PUT, PATCH and DELETE: a write commits the next version of the professional’s PERSON, clearing a mapped attribute a PUT leaves out, keeping the binding key and externalId, and moving the lifecycle state by active (complete or incomplete to inactive or abandoned and back, an abandoned User completed in two versions, an attribute change to an inactive User refused until it is reactivated). DELETE logically deletes the PERSON and its professional ROLE in one CONTRIBUTION and leaves the binding key as a tombstone. An If-Match naming another version is 412, and ServiceProviderConfig declares PATCH. [scim] max_deactivations and deactivation_window_minutes bound the deactivations and deletions of one client, counted in the database so replicas share the bound, with 429 past it (#5014).

  • /scim/v2/Users serves a client’s Users: POST creates the professional’s PERSON and its professional ROLE in one demographic CONTRIBUTION, complete when the template holds in full and incomplete otherwise, GET /Users/{id} answers with the version’s entity tag and 304 to a matching If-None-Match, and GET /Users answers a ListResponse filtered by eq and and over userName, externalId and id, with any other filter 400 invalidFilter. A client sees only the Users whose binding key its binding_issuer issues, every SCIM request is an access record of the client naming the PERSON, and ServiceProviderConfig declares entity tags. The start is refused when [scim] is on and the party mapping has no default person template with a user_name selector or no professional role (#5014).

  • /scim/v2/Groups serves SCIM Groups as demographic GROUPs of [authz.party_binding.group], each member holding an inline PARTY_RELATIONSHIP of [authz.party_binding.membership] from its PERSON to the GROUP, with no ROLE minted: create, read, eq/and filters on displayName, externalId and id, excludedAttributes=members, PATCH (answered 204), PUT and DELETE, with the entity tags of Users. A Group belongs to the client that created it, its members are that client’s Users, its displayName is unique, and its externalId sits under the client’s group_identifier_issuer. A remove of members with a value list removes only those members, an Entra compatibility deviation from RFC 7644 §3.5.2.2. Deleting a User removes its memberships in the same CONTRIBUTION. [scim] max_group_members (default 5000) caps a Group (#5015).

  • Just-in-time creation of a provisional PERSON at a first login, off by default and enabled per persons issuer in a new [[authz.party_binding.issuer]] table (jit, guests_jit, user_name_claim, idp_claim, admit_idp, admit_groups, admit_roles, admit_assurance). A login of a person or a delegated caller that matches no party and meets every declared admission condition commits one 553|incomplete| PERSON holding the binding key, the userName and the name claims of the access token, with its professional ROLE, as a creation by the person, and is served bound to it; any other mapped attribute is written only where its mapping sets jit = true, and externalId never. Concurrent first requests commit one PERSON, a deprovisioned or inactive login is never recreated, a taken userName refuses the login, and a SCIM client of the key’s issuer completes the PERSON. Boot refuses jit without an admission condition or a mapping that can write it. ferroehr party-binding check lists the provisional persons older than [authz.party_binding] provisional_report_days (default 30) (#5016).

  • A party of a mapped template is validated against that template on every demographic write: a PERSON or GROUP whose archetype_details.template_id names a template of the shared party mapping, and a PARTY_RELATIONSHIP of its membership archetype. A 532|complete| version must meet every constraint, and a 553|incomplete| one may miss mandatory nodes but nothing it carries may be wrong (RM common master06 §Incomplete Content); the refusal is a 422 naming each failing path. Every other party keeps the reference-model checks alone, so parties created through the Demographic API under other templates are accepted as before (#5042).

  • The FHIR AuditEvent, the DICOM audit message and the ITI-20 syslog feed carry the binding fields of the access record: the binding outcome, the party and its groups, the provider acted for, the facilities of a system commit, the representation with its relationship, the actor chain, the external accessor’s asserted fields and the agreement with its legal ground, consent and access class, and a chained request’s correlation id. ehds-logging 0.5.1 writes them as the ehds-accessor and ehds-relayed entities, in the DICOM message as participant objects of type Other. The FHIR record also writes the external and system acting modes. FerroEHR moves to ehds-logging and ihe-atna 0.5.1 and ihe-iti 0.0.77 (#5043).

  • A bound representative reaches only the EHR whose EHR_STATUS names the subject it represents. Every composition, EHR_STATUS, directory, contribution and query request is held to that EHR whatever the token’s scopes cover, and a query must name it with ehr_id. A representative whose token names a subject its party has no live representation relationship to, or whose issuer no [[auth.party]] table binds, is refused with 403 and refused-representation-unrecorded (#5086).

  • The Helm chart (23.1.0) accepts bridge.cdr.oauth.tokenEndpoint for an authorization server whose metadata FerroBRIDGE cannot read at bridge.cdr.oauth.issuer. FerroBRIDGE then requests its client credentials token from that endpoint, and issuer still names the machines issuer the CDR trusts (#5093).

  • A MyHealth@EU contact point’s caller (a contact_point = true federation issuer) reaches patient data only for the patient of its treatment relationship confirmation. The token’s ehdsi.patient.id (system|value) is resolved to its EHR through the EHR index, and every patient-data request must name that EHR. A token without it is refused with 403 and ERROR_SEC_DATA_INTEGRITY_NOT_ENSURED, recorded as refused-trc-missing. Another patient’s EHR, an unknown EHR and a request naming no EHR all answer one 404 of the new problem type trc-patient-mismatch, carrying the agreement service’s *_NOT_FOUND code and recorded as a refused-trc-patient-mismatch security event (eHDSI XCA Profile 8.0.1 §4.2.1, §4.2.2). A contact point’s person agreement now names its ehdsi_service; upgrade step: set it, and list the issuer in [smart] patient_issuers (#5102).

  • An external person under an agreement with discard discards a dispensation of a listed template, as a logical delete, within discard_window_hours and from the same point of care: the dispensation’s health_care_facility carries the token’s provider identifier and its eHDSI organization-id. A token carrying permissions records and discards a dispensation only with PPD-046 (eHDSI SAML Profile 10.0.0 §2.5) (#5102).

  • A read under a federation agreement with a consent condition is served when the token carries the consent decision reference for its patient (Wabvpz Art. 1(j) and Art. 15a(1)), in the claim or authorization_details member the issuer’s [[auth.federation]] table maps to consent.reference, and refused with 403 without it. The reference is one string, the token’s patient/ grant holds it to the launch-context patient, and the access record carries it with the other external fields. ferrohealth-smart moves to 0.9.3, which reads the field (#5106).

  • Every access record of an authenticated request carries the outcome of the party binding (binding: bound, provisional, external, organisation, system, self or a refused-* code) with the fields it needs: the bound party, its groups, the provider, the guest mark, a representative’s representation, a delegated token’s actor chain, a verified external accessor’s fields and agreement, and the facilities a system caller’s compositions named. acting_mode gains delegated, external and system. The audit migration 0023_binding_outcome.sql adds the columns and restores a per-row digest version: records written before stay version 1 and keep verifying, new records are version 2, and a version-1 record after a version-2 one is refused (#5118).

  • Every refused request is an access record through the gated path, and a refusal of the login binding or of a caller kind carries its refused-* binding code and the party it names. Refusals of tokens that failed verification are budgeted by source network and those of verified callers by issuer and subject, so neither crowds out the other; past a caller’s limit its refusals become one summary record with the count and the EHR ids they named, and past a budget one summary record per issuer and outcome is written. Security-event refusals are recorded for every answered request, and a refusal record the system log cannot take is counted in atna_audit_refusals_unwritten. The limits, windows and budgets are kept per replica (#5119).

  • A chained organisation request under a federation agreement is served only when its token carries the requester’s correlation id in the claim or attested authorization_details member the issuer’s [[auth.federation]] table maps to correlation_id. Without it the request is refused with 403 and recorded under the binding code refused-correlation-missing (audit migration 0025); a request header never stands in for it. The organisation access record’s agreement carries the id and its source beside the legal ground and the access class. A chained agreement whose issuer maps no correlation_id, or maps it from a constant, refuses the start (#5120).

  • On the MyHealth@EU contact-point route, the committer of an external person names the professional by the HP Identity Assertion’s NameID: DV_IDENTIFIER{issuer = the assertion's issuer, id = the NameID, type = "hp_identifier:<NameID format>"}, beside the voucher’s oauth2 identifier (eHDSI SAML Profile 10.0.0 §2.1), and a dispensation’s discard is a logical delete under the same committer. A contact point’s COMPOSITION must carry the HCPO’s organization-id in EVENT_CONTEXT.health_care_facility.identifiers, under the issuer urn:oasis:names:tc:xspa:1.0:subject:organization-id, beside the provider’s identifier, or it is refused with 422. ferrohealth-smart moves to 0.9.4, which reads ehdsi.assertion.name_id_format and refuses a contact-point token without one of the three formats as ERROR_HPI_INSUFFICIENT_INFORMATION; upgrade step: a contact point’s authorization server writes the NameID format into its tokens (#5144).

  • A federation caller above its agreement’s caller_rate, a MyHealth@EU contact point’s included, is answered 429 with Retry-After right after authentication and before any patient lookup, keyed on the token’s subject per replica, and every such refusal is counted into the caller’s summary record (#5149).

  • A deleted SCIM User’s contacts are erased from every version of its PERSON, keeping the name, the binding key and the identifiers for the record-keeping duties that need them. [scim] contact_erasure_minutes (default 0, at most 43200) sets when: 0 erases them in the DELETE itself, a period leaves them to a sweep that runs every minute. Each erased version’s node rows are derived again from the new body, its server signature is removed, and the erasure is an access record naming party:<id>#contacts. Party migration 0030 adds the scim_contact_erasure table (#5178).

  • openehr-rm types generic RM type names and paths: typing::RmType parses and prints names such as DV_INTERVAL<DV_QUANTITY>, typing::attribute_type gives an attribute of an instantiated class with the actual argument in place of the formal parameter, and typing::path_type types every step of an RmPath with its container. Each model::RmAttribute and model::RmTypeRef keeps the formal parameter it is written as (formal) beside the resolved bound. paths::write::item_at_path_mut and paths::write::set_at_path write at a path into a canonical instance, positional predicates included, creating the members the path names and changing nothing when a write is refused (#5187).

  • openehr-sdt: flat::navigate::Navigator gives every instance of a Web Template path in a composition, with each step’s occurrence, container index, positional path and JSON pointer, reaching a renamed instance where the template’s name is redundant and an OPT2 archetype root through the complete identifier the template recorded. flat::navigate::relative compares two paths step by step. flat::attribute_keys gives the FLAT key of an RM attribute of a structural class (the master05-rm_mapping.adoc tables, checked against them). A WebTemplateNode carries facts (the embedded template id of an OPT 1.4 archetype root and the terms its archetype defines) and every coded option its terminology, so no consumer walks the operational template (#5189).

  • openehr-base gives identifier and time parts as written: ObjectVersionId::object_id_str and ObjectVersionId::versioned_object_id keep the case of the object_id part, Iso8601Time::fraction_digits and Iso8601DateTime::fraction_digits return the digits of the fractional second as text, and Iso8601Duration::parts returns each component’s designator and number verbatim (DurationDesignator, DurationPart), with Iso8601Duration::fraction_digits for the seconds (#5190).

  • openehr-its reads the ITS-REST ETag as a typed entity tag: rest::etag::EntityTag parses and writes the RFC 9110 form, weak or strong, with the weak and strong comparisons. It types the tag as an ObjectVersionId or a HierObjectId. Every generated response-headers struct that carries an ETag gains entity_tag() beside the raw etag field (#5191).

  • openehr-term names the null_flavours and setting groups: codes::NullFlavour and codes::Setting give each concept’s code and English rubric and read a code back, held equal to the bundled terminology by tests. Every typed group’s ALL is now a slice (#5192).

  • openehr-am: HridRef reads an archetype reference with the version precision it was written at (VersionPrecision: major, minor or full, BASE Artefact Identification §Referencing), gives the version as written and matches an ArchetypeHrid by the specification’s rule, the namespace never a wildcard. ArchetypeHrid::local_semantic_id gives the namespace-free interface form a Web Template nodeId carries, which the v2_4 Web Template builder now uses (#5193).

  • openehr-its: rest::stored_query types a stored query’s qualified name (QualifiedQueryName, [{namespace}::]{query-name}, with check_registrable refusing the reserved aql), its version (QueryVersion, major.minor.patch in numeric order) and the version pattern a request addresses (VersionPattern, exact, {major} or {major}.{minor}, selecting the highest match), per ITS-REST 1.1.0. The generated stored-query operations and the StoredQuery and RESULT_SET documents carry them, so a name or version outside the grammar is a 400 where the request is decoded (#5207).

  • openehr_rm::Generation::earliest_for_composition answers the oldest RM generation whose canonical-JSON decode accepts a composition, so a builder can name ARCHETYPED.rm_version without naming a generation module; a value no generation decodes is a typed NoGenerationDecodes error carrying every generation’s refusal (#5242).

  • The access record takes the purposes of use and the professional roles from the access token. An issuer with an [[auth.attributes]] table states them through the IHE IUA purpose_of_use and subject_role claims (IHE ITI TF-2 §3.71.4.2.2.1.1) or the purpose_of_use of RFC 9396 authorization_details. The new token_attributes column of the audit store holds them, a new record is chain digest version 3, and the EHDS access record and the BALP AuditEvent write them with their code systems. A request whose [audit] purpose_header names a purpose the token does not state is refused with 403, and the refusal is recorded. The header alone keeps working for an issuer without the table (#5323).

Changed

  • The published server, viewer and PostgreSQL images take their org.opencontainers.image.base.name and base.digest labels and annotations from the runtime stage’s digest-pinned FROM line at build time, so a base-image bump changes that one line. An image built directly from a Dockerfile no longer carries these two labels (#4953).

  • [auth] takes the ferrohealth-smart 0.9 shape. Every trusted issuer is an [[auth.issuer]] table with a required kind (persons, patients, machines or federation) and a profile (rfc9068 by default, or entra, keycloak, okta, nl), and holds every setting of its tokens itself. A configuration in the old shape refuses the start, naming what replaces it. Upgrade steps: write a flat issuer = "…" as an [[auth.issuer]] table; move audience, algorithms, clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds, jwks_file, timeout_ms, mode, role_claims and [auth.introspection] onto each issuer table ([auth.issuer.introspection]); give each issuer its kind, with patients and professionals at separate issuers (a realm or tenant each), since one issuer is one kind; replace token_profile = "jwt" with token_types = ["JWT"] and the required_claims the issuer’s tokens carry, and drop token_profile = "rfc9068"; and write each scopeless_subjects entry as { issuer = "…", sub = "…" } (#5003).

  • A persons token is admitted only through a client its issuer lists in first_party_clients or delegates. Upgrade step: list the deployment’s own clients, the viewer’s among them, in first_party_clients, and a third-party client that acts for a person in delegates. The other 0.9 profile keys (person_key, tenant_claim, client_claims, app_only, guest, impersonation_marker, expect, user_claims, person_equality, unstable_claims, max_actor_chain, representation, scope_claim, app_only_scope_claim, adapter, sender_constraint and the onward_* keys) are documented in the configuration reference (#5003).

  • An array of tables, such as [[auth.issuer]], is set from the environment by element index (FERROEHR__AUTH__ISSUER__0__ISSUER), through ferrohealth-config 0.2.10 (#5003).

  • [smart] issuer_metadata_file is the metadata of the [auth] issuer it names, wherever that issuer stands in the list, and a file naming no [auth] issuer refuses the start. The SMART discovery document describes the first introspected issuer, else that named issuer, else the first issuer whose metadata is read from it, else the first issuer, so a Basic users’ issuer listed first no longer takes the discovery document (#5003).

  • The family crates move together: ferrohealth-smart 0.9.1, ferrohealth-authz 0.3.0, ferrohealth-session 0.3.1, ihe-atna and ehds-logging 0.4.0, ihe-iti 0.0.76, ferrohealth-config 0.2.10, ferrohealth-server 0.5.1 and ferrohealth-ui 0.6.9. ihe-atna and ehds-logging now escape every control and bidirectional formatting character of a value they write as its code point. With ferrohealth-ui 0.6.9 the viewer’s session watch no longer throws when the party relationship form, which holds a control named target, is submitted (#5003).

  • The terminology client moves to ferrobridge-term 0.1.143 (FerroBRIDGE v0.0.22), which takes the ferrohealth-smart 0.9 token source (#5003).

  • Every person is bound to a demographic party, with no switch. A persons issuer needs its [[auth.party]] table, two issuers of one node may not share an audience, and [authz.party_binding] provider (the healthcare provider the deployment acts for, EHDS Art. 13(4) and Annex II 3.2(a)) is required once [auth] admits a credential; each refuses the start naming what to set. A Basic user is a person: it maps to the binding key of its PERSON in [[authz.party_binding.basic_user]] (user, key_issuer, key), and a Basic user without a mapping refuses the start, naming the user. Upgrade steps: add an [[auth.party]] table to each persons issuer, set provider, and map each Basic user under the identifier_issuer of an [[auth.party]] table, which needs a persons issuer; a deployment of Basic users alone declares one whose jwks_file is a key set that signs nothing, as the quickstart and the sandbox do (#5003).

  • A token grants [authz.rbac] admin_role, emergency_access_role, import_role and identifier_reader_role only when its issuer is listed in the new [authz.rbac] role_issuers, each a persons issuer; any other kind refuses the start. Upgrade step: list the issuers whose role claims grant these roles; a Basic user keeps the roles its own table grants (#5003).

  • The compose quickstart, the dev and conformance stacks, the hosted sandbox and the Helm chart’s Basic overlays bind their Basic users under a persons issuer that signs nothing: its key set, which every FerroEHR image carries at /opt/sandbox-posture/basic-users-jwks.json, is the public half of a key pair whose private key was discarded (#5003).

  • The quickstart has one copy. docker-compose.yml on main pulls the FerroEHR images at :main and carries main’s configuration, which CI checks with the server image the change builds; the copy attached to each release pulls that release’s images, written by the release lane, which refuses a copy that still names :main. Download the release asset (releases/latest/download/docker-compose.yml) to run a release (#5003).

  • The Helm chart moves to 22.0.0. config.auth.issuer is a list of issuer tables, each with its kind, profile and audience, and values.yaml exposes config.auth.party, config.authz.party_binding, config.authz.rbac.role_issuers, config.authz.system_client and config.authz.federation.agreement. The chart refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling. Upgrade steps: rewrite config.auth as the [auth] upgrade above describes, set config.authz.party_binding.provider, and list the issuers whose roles grant the administrator role in config.authz.rbac.role_issuers (#5003).

  • A caller bound to a demographic PERSON commits under the committer the server writes: a PARTY_IDENTIFIED with the PERSON’s name, the login identifier (the issuer profile’s person_key value under the token issuer, type oauth2, or the Basic user under ferroehr, type basic), the provider identifier (type healthcare-provider) and an external_ref to the PERSON, on every commit path, attestations and CONTRIBUTIONs included. Upgrade: stop sending a committer, or send the placeholder {"_type": "PARTY_IDENTIFIED", "identifiers": [{"_type": "DV_IDENTIFIER", "issuer": "<iss>", "id": "<sub>"}]}, which must name the authenticated caller; a stated committer with another name, identifier or external_ref, and a stated PARTY_RELATED, is refused with 403 (#5005).

  • A patient’s committer is PARTY_SELF with no external_ref, and a representative commits as a PARTY_RELATED naming their own PERSON and the coded relationship. A patient-context token commits no demographic version. Upgrade: read the person’s own entries by the committer’s PARTY_SELF type, not by its former external_ref (#5005).

  • Importing an EHR Extract needs the role [authz.rbac] import_role names, or a system/ caller, and a wrapped committer naming a party or a login of this deployment is refused unless its version was created by a system [authz.import] trusted_systems lists. Upgrade: set import_role and give it to the accounts that import (#5005).

  • A person of a persons issuer whose login no demographic party carries is refused with 403 and refused-unbound on every request, with no fallback to a login-only committer. The problem names the identifier issuer to provision the person under, a token without the key claim is refused with a problem naming that claim, and each refusal counts into the new binding_refusals metric by issuer and outcome. The Keycloak quickstart overlay provisions its four demo users through SCIM with the ferroehr-provision-keycloak one-shot. Upgrade: provision every professional (through SCIM or the demographic API) before you add a persons issuer to the upgraded deployment (#5006).

  • A deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR adds its persons issuers only after the bridge and the gateway reach it under the caller: FerroBRIDGE v0.0.23 and FerroFED v0.0.17. Until then the committer rule refuses their writes. FerroFED v0.0.17 follows this release, so a deployment with a FerroFED gateway adds its persons issuers only once that release is out. The book’s new “Party binding: bootstrap and migration” page carries the bootstrap path, the four migration steps and these prerequisites, and “Breaking changes of party binding” lists every incompatible change with its upgrade step (#5009).

  • The hosted sandbox follows the bootstrap path of party binding: its configuration declares the sandbox-scim client ([[scim.client]], with the development issuer trusted as a machines issuer for it), the box’s new ferroehr-provision service provisions the demo PERSON through /scim/v2/ after every reset, and scripts/sandbox/reseed.sh waits for the demo login to bind before it seeds, then checks the committer of the marker’s EHR_STATUS version references a PERSON. The deploy probe’s new bootstrap family (P-BOOTSTRAP-01) shows the compose quickstart does the same from an empty database: the quickstart user is a User of the SCIM client, and a commit of the Basic user names that PERSON (#5009).

  • With a party mapping configured, the parties the login binding relies on are written through the demographic API by an administrator (authz.rbac.admin_role) only: a PERSON of a mapped template carrying a binding key, userName or externalId, its professional ROLE, a representation relationship, and a GROUP of the group template. Any other caller’s write is refused with 403, a version that changes a present binding key is refused with 403, and an administrator’s version that leaves a protected identifier out is refused with 422. Upgrade: write these parties as an administrator, and change a key through POST /admin/party/binding-key (#5013).

  • FerroEHR records every DPoP proof it accepts in its PostgreSQL database until the proof’s window ends, so a proof is accepted once across all replicas, and a proof replayed to another replica is refused there (RFC 9449 §11.1). If the database cannot record a proof, the request is refused with 503. The Helm chart (22.1.0) no longer refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling (#5072).

  • An AQL query scoped to one EHR now takes the streaming plan shape whatever its ORDER BY, DISTINCT or aggregates, whether the EHR comes from the ehr_id request parameter or from an ehr_id/value equality in the query. On a store of 100,000 compositions such a query used to spend 16 to 64 ms in PostgreSQL’s planner for an execution of about 1 ms, on every new connection and every new statement text; it now plans in 1 to 3 ms (#5076).

  • The Helm chart (23.0.0) connects FerroBRIDGE to the CDR as a machine. FerroBRIDGE uses an OAuth 2.0 client credentials client of a machines issuer that the CDR trusts, set in bridge.cdr.oauth (issuer, clientId, and optionally scope, audience and resource). The client secret comes from the client_secret key of bridge.cdr.existingSecret. The chart no longer accepts a Basic user for this connection: from ferrohealth-smart 0.9 a Basic credential belongs to a person and holds no system/ scope. The chart refuses to render when bridge.cdr.oauth.issuer is not listed in config.auth.issuer with kind: machines. To upgrade, add that issuer, create the client, and replace the user and password keys of the Secret with client_secret (#5085).

  • A verified external person’s commit records the professional as its committer: a PARTY_IDENTIFIED with the asserted name, the hp_identifier under the issuer "<country_code>:<issuing_authority_name>" and the token’s iss and sub, with no external_ref. A country code that is no upper-case ISO 3166-1 alpha-2 code, or an issuing authority holding a :, control or bidi character, is refused with 403 naming the field. The COMPOSITION must name the token’s healthcare provider (provider.healthcare_provider_identifier under provider.issuing_authority_name) as its health_care_facility, or it is refused with 422. Upgrade step: map provider.issuing_authority_name in each [[auth.federation]] table whose agreements carry writes, with a constant source where the token does not carry it (#5097).

  • A system caller (an app-only token by its issuer profile’s app_only markers) holds system/ scopes only: a token carrying a user/ or patient/ scope is refused with 403. It commits into an EHR only as a client an [[authz.system_client]] table declares, under the committer PARTY_IDENTIFIED{name = <client>, identifiers = [{issuer = <iss>, id = <client>, type = "client"}]}, and each COMPOSITION it commits must be of a template its writes lists (403), name a facility its facilities lists in context.health_care_facility, or in feeder_audit.originating_system_audit.location when it has no context (422), and be either a device act from a device client or a relayed person act naming the professional by an identifier from a relay client. Upgrade: declare every client that commits in [[authz.system_client]] with its writes, facilities and origins, and give its compositions the facility identifier (#5100).

  • An attestation is accepted only from a caller bound to a professional PERSON; one from a patient, a representative, a system caller or an unbound caller is refused with 403. Upgrade: attest under the professional’s own token (#5100).

  • A guest of a persons issuer, by its profile’s guest markers, is refused with 403 (refused-guest) unless the issuer’s [[auth.party]] table sets guests = true. Upgrade: set guests = true on the issuers whose guests act under your responsibility (#5100).

  • A verified external person or organisation of a federation issuer is admitted only under the [[authz.federation.agreement]] of its issuer and accessor, holding no scope beyond the agreement’s reads, and is an access-service caller by its kind: an opted-out EHR answers 404 to it, and only an external person under an agreement with opt_out_override reaches one under an emergency purpose. It commits only a COMPOSITION of a template the agreement’s writes lists (an organisation nothing), and a read under a consent-conditioned agreement, and every request under an NL agreement, is refused until the consent decision reference and the Wabvpz Art. 15f roles can be read. Upgrade: declare an agreement for every federation issuer and accessor, list the issuer in [smart] patient_issuers, and give its tokens only the patient/ scopes the agreement reads (#5101).

  • Under an NL [[authz.federation.agreement]], Wabvpz Art. 15f now decides admission in place of the refusal of every request. A caller whose nl.role is the UZI role 01.008 (bedrijfsarts) or 01.048 (verzekeringsarts) is refused with 403 as refused-wabvpz-role, and a caller whose nl.organisation_type is missing or is not a care provider type of Nictiz’s OrganisatieTypeCodelijst is refused with 403 as refused-wabvpz-organisation-type, so an insurer is refused. Every NL agreement declares purpose = "care"; one without a purpose, or for a keuring, refuses the start. The code lists are vendored by scripts/vendor/nl-role-codes.sh into docs/specs/nl-role-codes/. Upgrade: add purpose = "care" to every NL agreement and map nl.organisation_type (and nl.role for persons) in its issuer’s [[auth.federation]] table (#5107).

  • A delegated token’s access record now has acting_mode = "delegated" where it had client, and a system caller’s system. The FHIR rendering still names a delegated access as a client one. Upgrade: a review query that selects delegated access by acting_mode = 'client' also selects 'delegated' (#5118).

  • FerroEHR builds on ferrohealth-server 0.6.1, whose public caller-rate take holds a federation caller to its agreement’s caller_rate. Its 0.6.0 break, the revocation feature’s move to der 0.8, touches no feature FerroEHR enables (#5149).

  • The source tree no longer carries the UZI role code CodeSystem (uzi-rolcode) of Nictiz’s zib 2020 package: its codes and titles reproduce the UZI-register CPS, which CIBG does not license for reproduction. scripts/vendor/nl-role-codes.sh writes it into the git-ignored .vendor-cache/nl-role-codes/, as it does the CPS, and --test-inputs fetches only what the Wabvpz tests read. Nictiz’s care provider types stay under docs/specs/nl-role-codes/. Run scripts/vendor/nl-role-codes.sh --test-inputs once before running the ferroehr-rest tests (#5175).

  • A contact point’s patient.id is read through ferrohealth-smart’s TrcPatientId, the family’s one definition of the TRC patient. A value with whitespace in its system, or a value holding a control character, | or a CX delimiter (^, &, ~, \), is refused as a missing binding, as the contact point that writes the claim refuses it (#5195).

  • [auth] takes three boot refusals from ferrohealth-smart: a Basic user’s hash below the Argon2id floor (m=19456, t=2, p=1), an issuer that carries userinfo, and private or symmetric key material in a jwks_file. They are refused as before, with the crate’s messages. A jwks_file that is no JWK Set now refuses the start where the crate reads the key set (#5209).

  • The family crates move together to one ferrohealth-smart: 0.13.1, with ferrohealth-authz 0.7.0, ferrohealth-session 0.7.0, ihe-atna and ehds-logging 0.8.0, ihe-iti 0.0.83, ferrohealth-server 0.6.4 and ferrobridge-term 0.1.148. The 0.13 line refuses a JSON document it cannot read without quoting it (GHSA-mhv7-3975-6xq3), so a jwks_file or [smart] issuer_metadata_file that does not parse is refused at boot with the crate’s message, which names the failure class and position; 0.13.1 keeps a refused key set, introspection or metadata body out of the error (GHSA-5qm7-v4fw-p2v9). The OAuth client credentials tables ([authz.abac.remote.oauth], [ehr_index.oauth], [terminology.oauth] and [audit.fhir_feed.oauth]) read as before (#5209).

  • The API’s response security headers come from the family server stack (ferrohealth-server 0.6.2) with the same six headers and values. They now also reach the management listener on its own port, and Strict-Transport-Security: max-age=31536000 is sent when FerroEHR or the viewer serves TLS itself ([server.tls], RFC 6797 §7.2). The viewer gains Cross-Origin-Resource-Policy: same-site (#5210).

  • The server and the viewer read the parts of an OBJECT_VERSION_ID through openehr-base and its BASE grammar, never by cutting at ::. A value that is no OBJECT_VERSION_ID now names no version: the AQL branch refusal, the version item tags and the viewer’s version numbers leave it alone, a reference to a container is read as the whole value (so uuid::extension no longer names the container uuid), and the viewer’s deletes refuse it before calling the CDR. Every part keeps its case (#5211).

  • ETag and If-Match values are read through openehr-its’s RFC 9110 entity tag on the server and in the viewer. The server accepts the same forms as before. The viewer reads its CDR’s ETag strictly: a value that is no entity tag names no version (#5212).

  • The viewer’s audit screen reads the ITI-81 searchset and each AuditEvent as FHIR R4 resources, and a refusal’s OperationOutcome too. A searchset that is no FHIR R4 Bundle of AuditEvents now shows as a failed read, never as a page of empty rows (#5214).

  • The EHR index’s boot check reads the PIX Manager’s CapabilityStatement through ihe-iti’s PIXm client, with its transport and its client-credentials token, so a Manager behind IUA answers the probe as it answers ITI-83. A failed probe is reported as the client’s refusal: no answer, the status it answered, or no FHIR R4 CapabilityStatement (#5216).

  • [audit.syslog] reads its TLS material from [audit.syslog.tls], the outbound TLS table of every FerroHEALTH product (trust_roots_file, client_identity_file or client_identity, min_version), and hands ihe-atna the rustls configuration ferrohealth-server builds from it. The keys tls_ca_file, tls_identity_cert_file and tls_identity_key_file are retired and refused at boot by name: name the roots file as trust_roots_file, and concatenate the client certificate chain and its private key into the one PEM client_identity_file names. The client speaks TLS 1.3 unless min_version = "1.2" lowers it (#5217).

  • A SCIM Group membership is now held by the member: an inline PARTY_RELATIONSHIP from the member’s PERSON to the GROUP, stored in the PERSON’s version (RM demographic §Party Relationships). A membership change versions only that PERSON, so its cost no longer grows with the Group’s size. The GROUP is versioned only by a rename, an externalId change or its deletion. The Group’s members is read from the membership index (#5220).

  • A Group’s ETag is now its GROUP version plus a membership revision, so it still changes with every membership change. The member User’s ETag now changes too, while its body stays the same (#5220).

  • Deactivating a User (active = false) drops all of its memberships in the same version, because an inactive PERSON cannot be versioned again. A reactivated User belongs to no Group until its client adds it again (#5220).

  • Microsoft Entra ID’s disable-then-remove order now succeeds, since the later remove finds nothing to remove. An add naming an inactive User is 400 invalidValue (“reactivate the User first”) (#5220).

  • A refused SCIM write now names the rule it breaks in detail (#5220).

  • scripts/perf/scim-groups.sh keeps the database log, df -h and free -m of a failed run (#5220).

  • A party commit no longer re-derives the login binding’s side table, the reverse-relationship index and the protection check from the whole stored body. Each hook reads what the committed body holds, derived once in memory, and writes only the rows that changed, and ferroehr party-binding refill derives the rows with the same code. A party commit without a server signature no longer reassembles its body. A membership change at the cap of a 5 000-member SCIM Group went from 810 ms to 358 ms at p99 on a developer laptop (#5220).

  • openehr-rm writes the RFC 8785 canonical form a VERSION signature covers in one pass over the version, sorting each object’s members by reference. A 3.3 MB GROUP version canonicalises about seven times faster than through serde_jcs, with byte-identical output, so signed versions verify unchanged. serde_jcs remains the test oracle (#5220).

  • The stored-query service takes the typed name, version and version pattern of openehr_its::rest::stored_query end to end, so it reads one grammar with the REST layer: the three-part {namespace}::{formalism}::{query-name} name, which ITS-REST 1.1.0 Qualified_query_name.md does not define, is no longer accepted anywhere (the formalism is the query_type parameter), and DELETE /admin/query/{qualified_query_name}/{version} answers 400 for a name outside [{namespace}::]{query-name} or a version that is no exact major.minor.patch, where it answered 404 (#5223).

  • Boot and ferroehr config check refuse an [[auth.party]] table whose identifier_type is the type of no [[authz.party_binding.person.template]] key, naming the issuer, the type and the types the templates write: no login of such an issuer could ever match a party, and every one would be refused. Upgrade step: give the deployment a person template whose key carries the table’s identifier_type (or set the table to the type the template writes). The Helm CI and production values and the hosted sandbox now carry the professional PERSON and ROLE mapping, and the sandbox loads the two templates before the CDR starts (#5231).

  • openehr_query::printer::to_aql writes each WHERE comparison with one space on each side of the operator (c/name/value = $nameValue), as the QUERY 1.1.0 AQL syntax examples do; path predicates keep their unspaced form ([ehr_id/value=$ehrUid]). The query builder’s generated AQL follows (#5241).

  • openehr_sdt::flat::navigate::Navigator owns what it copies from its Web Template (the name-fallback table and the OPT2 archetype references) and carries no template lifetime, so a consumer can keep it beside the template it was built from; it is Debug, Clone, Send and Sync (#5243).

  • Breaking: the configuration tree is one table per concern. ferroehr.toml moves to [deployment], [api], [database], [auth] (with [auth.smart] and [auth.scim]), [authz] (with [[authz.origin]]), [management], [audit], [privacy] (with identifier_protection, cohort and access_delay), [record] (with subject_index, versioning, signing, query, change_feed and multimedia), [terminology] and [pixm]. FerroEHR’s own keys take one duration type ("500ms", "30s", "10min", "12h", "7d", UCUM unit codes, so a bare number and m are refused), one size type ("256KiB", "64MiB", IEC binary units, decimal units refused) and one calendar retention period ("P5Y", "P1830D", unset keeps forever). An optional feature is on when its table is present: [terminology], [pixm], [auth.scim], [audit.syslog], [audit.fhir_feed], [privacy.identifier_protection] and [privacy.access_delay] carry no enabled; [authz.abac], [record.multimedia] and [auth.smart] keep a switch because some of their keys apply while off. record.subject_index has no default, so the server refuses to start until "pixm" or "local" is chosen. [api.cors] replaces cors_permissive with the Fetch Standard response headers (allow_origins, allow_methods, allow_headers, expose_headers, allow_credentials, max_age). The environment follows the tree (FERROEHR__DB__URL is FERROEHR__DATABASE__URL, FERROEHR__EHR_INDEX__PROVIDER is FERROEHR__RECORD__SUBJECT_INDEX). The Helm chart moves to its 25 line with config: on the same tree. Every key a release before v5.7.0 read is refused at load with its replacement named, all at once, so ferroehr config check --config <old file> prints the whole move list; the refusals ship in v5.7.0 only and the next release removes them. To upgrade, run config check on the old file and move each key it names: spec_profile, deployment_profile, deployment_accepts and licence.file are deployment.spec_profile, deployment.profile, deployment.accepts and deployment.licence_file; [rest] is [api] (swagger_ui is openapi_ui, body_limit_bytes is body_limit, cors_permissive = true is [api.cors] allow_origins = ["*"], [rest.identity] is [api.options] with solution and vendor; the version keys are facts of the build), admin.enabled is api.admin and terminology.api_enabled is api.terminology; [db] is [database] (acquire_timeout_secs is acquire_timeout, statement_timeout_ms is statement_timeout, migrate is migrations), [storage.<domain>] is [database.<domain>] and [storage.migrate.<domain>] url is [database.<domain>] migrate_url; [smart] is [auth.smart] (enabled is discovery, episode.enabled is episode_launch, launch_base64_json = true is "launch-base64-json" in auth.capabilities); [scim] is [auth.scim] (max_results is filter_max_results, deactivation_window_minutes is deactivation_window, contact_erasure_minutes is contact_erasure_delay); [authz.abac.cedar] reload_secs is reload_interval, [authz.abac.remote] server, connect_timeout_ms and request_timeout_ms are url, connect_timeout and timeout, authz.party_binding.provisional_report_days is provisional_report_after, and an agreement’s discard_window_hours and chained_log_retention_years are discard_window and chained_log_retention; [[origins.registered]] is [[authz.origin]]; management.port is management.listen (an address:port) and management.profiling.max_seconds is max_duration; audit.store.retention_days and retention_years are audit.store.retention, verify_interval_seconds and full_verify_interval_seconds are verify_interval and full_verify_interval; [audit.syslog] host and port are url = "tls://host:port", spool_max_messages is spool_max_events, spool_max_bytes is spool_max_size, retry_max_ms is retry.max_backoff, and [audit.fhir_feed] poll_interval_ms and retry_max_ms are poll_interval and retry.max_backoff; [demographic.identifier_protection] is [privacy.identifier_protection], [cohort] is [privacy.cohort], and [access_delay] is [privacy.access_delay] with hours as max_delay; ehr_index.provider is record.subject_index ("ihe" is "pixm") and the rest of [ehr_index] is [pixm] (pixm_url is url, invocation is method, request_timeout_ms is timeout); versioning.skip_identical_resend is record.versioning.idempotent_resend; [signing] is [record.signing] (key_path is key_file, retired_key_paths is retired_key_files), [query] is [record.query] (timeout_ms is timeout), [outbox] is [record.change_feed] (retention_days is retention, prune_interval_secs is prune_interval), and [multimedia] is [record.multimedia] (endpoint is url, threshold_bytes is threshold, and allow_http is retired: an http:// url is plain HTTP); [terminology.server] is [terminology] (fail_on_error is on_error = "accept" | "refuse", request_timeout_ms is timeout). The family sections take the same types, from ferrohealth-server 0.8, ferrohealth-telemetry 0.4, ferrohealth-smart 0.15 and ferrohealth-session 0.9: [server] request_timeout_ms, header_read_timeout_ms, idle_timeout_ms, drain_delay_ms, shutdown_timeout_ms and retry_after_s are request_timeout, header_read_timeout, idle_timeout, drain_delay, shutdown_timeout and retry_after, body_limit_bytes and compression_min_bytes are body_limit and compression_min_size, and [server.tls] reload_interval_s is reload_interval; [telemetry.otlp] trust_roots_file and client_identity_file move into [telemetry.otlp.tls], and an https:// collector is dialed over it; [auth] ca_file is [auth.tls] trust_roots_file, an [[auth.issuer]] table’s clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds and timeout_ms are clock_skew, jwks_max_age, jwks_refresh_interval and timeout, its introspection cache_max_age_seconds is cache_max_age, [auth.basic] cache_ttl_seconds is cache_ttl, [auth.dpop] max_age_seconds and clock_skew_seconds are max_age and clock_skew, [auth.idp.entra] timeout_ms and max_response_bytes are timeout and max_response_size, and an oauth grant’s ca_file, client_certificate_file and client_key_file are its tls table’s trust_roots_file and client_identity_file (one PEM), with refresh_margin_ms, timeout_ms and previous_key_overlap_ms as refresh_margin, timeout and previous_key_overlap. The [auth] keys v5.6 read for every issuer (audience, algorithms, token_profile, jwks_file, mode, introspection, role_claims and the rest) live on each [[auth.issuer]] table, and scopeless_subjects entries are { issuer, sub } tables. In the viewer, cdr.request_timeout_secs is cdr.timeout, [oidc] ca_file is [oidc.tls] trust_roots_file, [oidc.tls] client_certificate and client_key are client_identity_file, and [session] idle_timeout_s, absolute_timeout_s and sign_in_timeout_s are idle_timeout, absolute_timeout and sign_in_timeout. Each of these is refused in v5.7.0 naming its replacement, in the CDR and the viewer alike (#5259).

  • Breaking: a patients token that names an actor is refused. With ferrohealth-smart 0.15, a patients issuer reads a token whose top-level act (RFC 8693 §4.1) names someone other than the patient only when its [[auth.issuer]] table sets act_represents_patient = true; without it the token is refused 401, where v5.6 admitted it as the patient’s own access. With it, FerroEHR refuses the token 403 (refused-patient-representative), because it binds no party to someone acting for a patient, and the access log records the refusal (#5259).

  • SECURITY.md, security.txt (its Contact: mailto: field), the post-market record, the CRA technical documentation and the book’s compliance pages name security@cadasto.com as the email route for a vulnerability report, in place of info@cadasto.com, which stays the single point of contact for complaints and incidents, the manufacturer’s address on the running server and the address for licensing and other business questions (#5272).

  • The shipped ferroehr.toml template (app/ferroehr/assets/ferroehr.default.toml) is generated from the configuration types with tools/config-reference: every key, each under its description and set to its default, a secret or a key with no default as a #? placeholder line. The book’s new page “Every configuration key” includes it whole. CI fails when it drifts from the types. No key changes: a ferroehr.toml that loads today loads unchanged (#5316).

  • The family crates move to their newest releases: ferrohealth-smart 0.18.0, ferrohealth-session 0.12.0, ferrohealth-authz 0.11.0, ferrohealth-server 0.8.4 (its TLS certificate verifier is built once per process), ferrohealth-telemetry 0.4.1, ihe-atna and ehds-logging 0.12.0, ihe-iti 0.4.0, ferrobridge-term 0.5.0, and fhir-types and fhir-search 0.1.131. Every [[auth.issuer]], onward and OAuth issuer renders with its credentials masked. A system caller, a patient and a patient’s representative carry the purpose of use their token states when their issuer has an [[auth.attributes]] table. An IUA purpose_of_use of empty codes states no purpose (#5320).

  • ferroehr and ferroehr-viewer run through the FerroHEALTH family’s run path. A configuration either binary refuses (an unknown or invalid key, a broken semantic rule, an [auth] table with no mechanism, a SCIM mapping it cannot serve, or for the viewer a TLS file that does not read or an unusable [oidc] or [session] table) now exits 78 (EX_CONFIG of sysexits(3)) where it exited 1, so an orchestrator can tell a configuration to fix from a failure to retry. Every other failure still exits 1, and the failure is written to stderr as one line with its whole cause, prefixed with the binary’s name. Upgrade: a script or probe that expects exactly 1 from config check on a refused file expects 78 (#5321).

  • The configuration is redacted by its secret types alone. GET /management/env, GET /ferroehr/rest/extensions/v1/admin/config and ferroehr config check show the tree as its Secret and SecretUrl values serialize, and the second masking by key name and by URL shape is gone. [audit.syslog] url, [auth.smart] platform_base_url and public_base_url, and [auth.scim] public_base_url are secret URLs now, as are the OAuth, onward and introspection endpoints of ferrohealth-smart 0.17 and the viewer’s [oidc] endpoints of ferrohealth-session 0.11, so each renders with its credentials and query masked. A platform_base_url given as a bare path renders as *** (#5322).

  • Every outbound integration (the terminology server, the PIXm identity source, the ITI-20 FHIR Feed, the remote policy decision point and the viewer’s CDR client) is built by ferrohealth-smart’s OutboundClient from its endpoint table, and FerroEHR’s own client builder is gone. A PIXm exchange’s timeout now bounds the exchange with its tries and the waits between them, and the viewer refuses a [cdr] timeout of zero at start (#5324).

  • The family crates move to their newest releases: ferrohealth-smart 0.17.0, ferrohealth-session 0.11.0, ferrohealth-authz 0.10.0, ferrohealth-config 0.4.4, ferrohealth-server 0.8.3, ferrohealth-secret 0.2.1, ferrohealth-usage-report 0.4.1, ferrobridge-term 0.4.0, ihe-atna and ehds-logging 0.11.0, and ihe-iti 0.3.0. Every OAuth token endpoint and introspection endpoint renders with its credentials masked, and a TLS refusal of an outbound client names its <key>.tls table (#5324).

  • [auth.idp.entra] is read by ferrohealth-smart’s [auth] section, so an unknown key under it is refused with the crate’s field list. The key path is unchanged. The effective configuration now shows [auth.idp] as an empty table when no adapter is configured (#5325).

  • The build compiles FHIR R4’s resources alone: fhir-types is required with its r4-resources feature in place of the crate-wide resources, and the FerroBRIDGE crates (fhir-types, fhir-search, ferrobridge-term) are required by caret ranges in place of exact versions (#5327).

  • Breaking: the outbound integrations take the family’s endpoint table. [terminology], [pixm], [audit.fhir_feed] and [authz.abac.remote] read where they call and how from one [<x>.endpoint] table, the same keys for every outbound integration of every FerroHEALTH product: url, timeout, connect_timeout, retry (max_attempts, initial_backoff, max_backoff), tls (trust_roots_file, client_identity_file, min_version) and credentials (HTTP Basic user with password or password_file, or the client credentials grant [...credentials.oauth]). terminology.url, timeout and [terminology.oauth] are terminology.endpoint.url, timeout and [terminology.endpoint.credentials.oauth]; pixm.url, timeout, [pixm.tls] and [pixm.oauth] are pixm.endpoint.url, timeout, [pixm.endpoint.tls] and [pixm.endpoint.credentials.oauth]; audit.fhir_feed.url, [audit.fhir_feed.retry] and [audit.fhir_feed.oauth] are audit.fhir_feed.endpoint.url, [audit.fhir_feed.endpoint.retry] and [audit.fhir_feed.endpoint.credentials.oauth]; authz.abac.remote.url, connect_timeout, timeout and [authz.abac.remote.oauth] are authz.abac.remote.endpoint.url, connect_timeout, timeout and [authz.abac.remote.endpoint.credentials.oauth]. The v5.6 keys are refused naming their new home. Every client now honours retry: the terminology, PIXm and remote policy clients try a call that got no answer or a 5xx up to max_attempts times (three by default; each tried once before), so set max_attempts = 1 to keep one attempt, and the FHIR feed’s delivery rounds take max_attempts tries each, from initial_backoff, before the round waits max_backoff; the feed still never gives a record up. The environment follows the tree (FERROEHR__TERMINOLOGY__ENDPOINT__URL, FERROEHR__AUDIT__FHIR_FEED__ENDPOINT__URL), and so do the Helm chart’s values and its client secret routing, in the chart’s 25 line (#5331).

  • The FerroEHR PostgreSQL image is built on the 2026-11-09 respin of postgres:18.6 (sha256:e3d24f69…), a same-version respin with rebuilt base packages. A full scan of the rebuilt image on amd64 and arm64 finds the same findings as the image of the previous base, each judged in security/vex/postgres-os.openvex.json, and a gate scan finds no fixable HIGH or CRITICAL finding (#5332).

  • The development and Keycloak example compose files run Keycloak 26.8.2 (quay.io/keycloak/keycloak:26.8.2@sha256:723b1d6e…). Of the 26.8.2 upgrading notes, the shipped realm meets the ones that touch it: the Admin REST API’s new audience check passes for the master realm’s admin-cli token the browser-test harness uses, and the realm uses no token exchange, identity broker, SAML client, CIBA or dynamic client registration. The ferroehr-service client-credentials token carries the same audience and scopes as under 26.8.1 (#5333).

  • The compose quickstart, the hosted sandbox and the Helm chart deploy FerroPIX 0.0.26 (ghcr.io/ferrohealth/ferropix:0.0.26@sha256:85d40e6fdcdb39f782007eff91eb059861a4e039ca3f8fb56e37f2fa0f28d78a, the FerroPIX chart 0.3.19) and FerroTERM 0.4.5 (ghcr.io/ferrohealth/ferroterm:0.4.5@sha256:7c20d39e1989fce1dfb92e6ddfd853dfa9a52bb9b7b973fbba42430963ef75bd), and the development issuer is mock-oauth2-server 6.1.5. FerroPIX’s configuration moves to the [[auth.issuer]] shape it now requires: the issuer is kind = "machines" and carries its own audience, and [auth] audience is gone. A chart deployment that sets ferropix.config.auth moves its audience onto each issuer the same way (#5351).

  • The nine openehr-* crates follow The Cargo Book’s SemVer rules, judged by cargo-semver-checks 0.51.0 against their newest release on crates.io on every pull request, in the merge queue and in the release plan. Their next publish moves the line from 0.0.x to 0.1.0. From then on a compatible change ships as a patch and only a breaking change moves the minor, so a caret requirement such as openehr-rm = "0.1" takes every compatible release. A public dependency moved to another compatibility line counts as a breaking change (#5353).

  • openehr-rm writes the ECMAScript number form of the RFC 8785 canonical form a VERSION signature covers with ryu-js 1.0, which carries the upstream fix for a panic in its fixed-point formatter, and is held byte-equal to serde_json_canonicalizer 0.4 as its test oracle in place of serde_jcs, so the lock holds one ryu-js line. Its output is unchanged (#5372).

  • ferroehr.eu takes the FerroHEALTH family layout: the “Oxide & Iron” brand tokens light and dark, a hero badge rendered from the workspace version at build time, a content hash on the stylesheet link, linked capability cards, the conformance record, the viewer screenshot, deployment cards, the family and the four parts of the book, and a 404 page in the same style. The book groups its pages under Evaluate, Integrate, Operate and Contribute, keeping every page’s address, opens in the recoloured light and navy themes and gains a Changelog page. The README’s quick start is the anchored steps of docker/smoke-test.sh, held by scripts/checks/readme-quickstart.sh (#5382).

  • The quickstart docker-compose.yml each release attaches now carries its .sha256sum; the README and the book download it with curl -fsSLO from the latest release and start it with docker compose up --wait, and the README check refuses a host path the file would read from the repository beyond the backup directories (#5387).

Removed

  • [authz.professional] (claim, client_tokens, acting_for_claim) and its 403 for a token that names no natural person. The issuer’s kind and profile classify the caller, and the access record’s acting mode and professional come from that classification: a person’s own token is the person, and a delegated token is the client acting for the person it names. Upgrade step: remove the table, which refuses the start, and list a client that acts for a person in its issuer’s delegates (#5003).

  • The named refusal of authz.rbac.user_role, which v5.6 introduced, is gone under the one-release rule for removal aids: the key is now refused as an unknown key when the configuration loads (#5259).

Fixed

  • A 553|incomplete| commit no longer refuses an empty mandatory archetype slot: RM common master06 §Incomplete Content sets every occurrences lower limit to zero, a slot’s included, while too many fillers are still refused. The OPT 1.4 writer now closes a mandatory slot once the fillers placed in it carry its lower bound, as ADL 2 §Slot Filling and Redefinition allows, and still refuses to close one they leave unmet (#5042).

  • The example of a party template whose identity the template marks optional is a party of its class carrying one identity at every detail level, where the required level answered 500: RM demographic PARTY.identities admits no party without one. A party template that constrains no identity at all answers 400 Bad Request naming its class (#5057).

  • An AQL VERSION standard predicate on commit_audit/time_committed applies its operator: VERSION v[commit_audit/time_committed/value > $t] returns every version committed after $t, where it returned the trunk version current at $t whatever the operator (QUERY 1.1.0 master03 §Standard predicate). A number, boolean or NULL literal as the bound, a string bound no ISO 8601 form reads, and a time_committed path other than the whole value or /value are refused with 400 (#5062).

  • AQL refuses a number, boolean or null parameter or literal compared with a date/time path, in WHERE and in a VERSION predicate, with the same 400 a non-ISO 8601 string gets; such a bound used to become NULL and match nothing (#5064).

  • The served OpenAPI document now lists, on both template example operations, the 400 answered for a party template that constrains no PARTY_IDENTITY (#5066).

  • The compose quickstart, the hosted sandbox and the Helm chart run FerroPIX 0.0.18 (ghcr.io/ferrohealth/ferropix:0.0.18@sha256:6fb622b5b9ade34c8a4053112c8f166ff495d8b32bf99f35529d01e512b7cc42), with the FerroPIX chart 0.3.10 built for it. The 0.0.16 they pinned had an outbox sender that never slept while its outbox was empty, so an idle install kept about 1.2 CPU cores busy between FerroPIX and its PostgreSQL at about 8,000 transactions a second (FerroHEALTH/FerroPIX#490). 0.0.18 also makes FerroPIX’s [auth] scopeless_subjects and role_claims hold per issuer, which changes nothing for a FerroPIX that trusts one issuer, as every configuration FerroEHR ships does. The chart moves to 21.0.5 (#5073).

  • An OPT2 built from a template that specialises a published archetype now carries that archetype, by its full id, as parent_archetype_id, as the OPERATIONAL_TEMPLATE invariant is_specialised requires (AOM2 §3.6.6), so a reader of the OPT finds the archetype its root specialises. The ADL form of an OPT still prints no specialize section (OPT2 §Artefact Structure), so the parent travels in the canonical JSON form (#5075).

  • A refused-key-claim-missing access record names the key claim the issuer’s binding reads (its person_key) in binding_detail.claim, so an operator sees which claim the token lacked. It records the name only, never a claim value (#5139).

  • A 403 for a caller kind is recorded with its own binding code: refused-system-personal-scope for a system caller holding a user/ or patient/ scope, refused-no-agreement for an external caller whose issuer has no agreement for its accessor kind, and refused-scope-beyond-agreement for an external caller holding a scope its agreement does not read. Audit migration 0024 widens the binding constraint; the problem text of each refusal is unchanged (#5140).

  • openehr-sdt builds no /context node, and no start_time or setting leaf, in the Web Template of an OPT 1.4 whose COMPOSITION prohibits context (existence 0..0, or an EVENT_CONTEXT of occurrences 0..0), and the template example of such a template carries no context. A template that leaves context open or optional keeps the node as before (#5147).

  • openehr-sdt gives the /context node it synthesizes in a Web Template the min of the template’s existence on COMPOSITION.context, and 0 where the template leaves context unconstrained, since the RM makes context optional (0..1). A client following the Web Template no longer treats the context as mandatory when the template does not require it, and the required template example of such a template carries no context (#5154).

  • An ADL2 template that deletes an attribute (existence matches {0}), for example COMPOSITION.context, keeps the prohibition after it compiles to its operational template: openehr-adl’s create_opt now returns a RawOpt that names each removed attribute beside the OPT, and build_web_template_v2_4 takes that list. A template that deletes context gets no /context node, and a commit that carries one is refused, as it is for an OPT 1.4 template that prohibits context. An OPT2 uploaded already compiled carries no such list and reads a deleted attribute at its reference-model existence. Breaking for crate users: create_opt returns RawOpt and build_web_template_v2_4 takes the prohibited attributes (#5154).

  • A composition commit whose template_id names a stored ADL2 archetype rather than a template is refused with 422 as an unknown operational template; it failed at the insert with a server error before (#5177).

  • openehr-rm types REFERENCE_RANGE.range as DV_INTERVAL<T>: ReferenceRange is now generic over its limit type (ReferenceRange<DvQuantity> on a DV_QUANTITY, ReferenceRange<DvOrdered> on the other ordered types), its is_in_range and Range_is_simple work at that type, and the static RM model gives REFERENCE_RANGE.range and IMPORTED_VERSION.item their argument. Each model::RmAttribute and model::RmTypeRef now keeps the formal parameter it is written as (formal, Some("T") for DV_INTERVAL.lower) beside the resolved bound. The RM class table declares REFERENCE_RANGE<T> and leaves T unused; the generator restores it from a cited table and refuses a schema that no longer matches (#5196) (#5188).

  • An access record names the client the token’s issuer profile reads from client_claims (azp under the entra and keycloak profiles), the one the caller was admitted as. A token carrying a different client_id claim beside it was recorded under that claim. The [origins] register and the SCIM client lookup read the same client (#5194).

  • An Interval or DV_INTERVAL read from canonical JSON or XML that omits a boundedness or inclusion flag now takes it from its bound. An absent bound reads as unbounded and not included, and a present bound as bounded and included. A flag the wire states is kept as written. Before, an omitted flag always read as bounded and included, so an interval with only an upper came out with a bounded, included lower side that had no value. ITS-JSON requires all four flags, and BASE ties none of them to its bound, so reading the omission is our own design (#5203).

  • The viewer’s New contribution commits as the signed-in login. Each audit carries the placeholder naming that login, and the server records the committer it derives from the login. The free-text committer field is gone. For a login bound to a demographic PERSON, a stated name that differs from the PERSON’s name was refused with 403 (#5222).

  • The management surface answers a caller the login binding refuses with 403 and its binding code, such as refused-unbound, as every other route does, in place of 401: the credential authenticated, so new credentials cannot help (RFC 9110 §15.5.4). An absent or invalid credential still answers 401 with its challenge (#5250).

  • The party binding posture in ferroehr report and the boot log masks the user name and password of an issuer URL, as every other rendering of the configuration does (#5259).

  • GET /definition/query/{namespace}:: lists every stored query in the namespace again, where it answered 400 since the typed stored-query name: ITS-REST 1.1.0 definition_query_list reads its name as a pattern in which an empty part is a wildcard, so ::{name} also lists the names that start with it in any namespace. openehr-its gains rest::stored_query::QueryNamePattern, the type the generated DefinitionQueryListParams carries. The store and the single-query reads still refuse an empty name or namespace (#5268).

  • The identifier scan ([privacy.identifier_scan]) reads the name of the ELEMENT that holds a DV_IDENTIFIER when the identifier’s type is absent or names neither a person nor an item, with the same word rules. An untyped laboratory order number in the Receiver order identifier or Requester order identifier element of laboratory_test_result.v1 that passes the NHS Number check by chance is now stored under the default rule set. A person word still wins, so an untyped value in an element named NHS number or Patient identifier stays scanned, and a type that names a kind takes precedence over the element name (#5291).

  • The Helm chart’s description and README name config.deployment.profile, the key the chart refuses to render without since 5.7.0, and the README’s install command sets it; the command it showed set the old config.deployment_profile and did not render (chart 25.0.1) (#5341).

  • A representative’s write is held to the EHDS Art. 5 rules exactly as the person’s. A representative may no longer alter an entry a health professional inserted or name a professional as composer. A version committed as PARTY_RELATED counts as inserted by the person, so the person and the representative may change it, and the contribution list reports inserted_by_person = true for it (#5359).

Security

Shipped with functional changes because: the advisory reaches only a deployment that trusts several issuers, which no configuration FerroEHR ships does, and its fix changes what such a configuration must state at boot, so it lands on main beside the AQL version attestations already merged for this release; a security-only patch of the last release remains open to the release cut.

  • A Basic user never carries a system/ scope, through ferrohealth-smart 0.9.1 (GHSA-jhg2-jwjp-hq74, CWE-269, CWE-863). A system/ scope granted to a Basic user gave that person every resource of a type, a backend service’s reach. Such a user now refuses the start, naming the user and the scope. Upgrade step: give each Basic user the user/ scopes a person holds, and move each machine on Basic to a client of the authorization server under the client credentials grant (#5003).

  • [auth] scopeless_subjects and role_claims hold per issuer, through ferrohealth-smart 0.8.9 (GHSA-pfgx-xq28-2629, CWE-863, CWE-639). Before it, beside several [[auth.issuer]] tables both lists applied to every trusted issuer’s tokens: a token from any of them whose sub was listed met every SMART on openEHR requirement without a scope, and any of them could assert every role the RBAC checks, ADMIN included. An entry is now written <issuer>#<value> and holds for that issuer alone; beside several issuers a bare entry, or one naming an issuer the section does not trust, refuses the boot. Upgrade step for a deployment with several issuers: write each scopeless_subjects entry as <issuer>#<sub>, and write each role_claims path as <issuer>#<path> for the issuers whose roles count (or set role_claims = [] to read none), since the default role_claims is bare and refused. A deployment with one issuer changes nothing (#5047).

  • The ferroehr-postgres image carries zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 from trixie-security, which fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). The image of 5.6.3 and earlier releases carries 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action (security/vex/postgres-os.openvex.json). The ferroehr and ferroehr-viewer images keep the distroless base’s zlib, which neither program loads (security/vex/ferroehr-os.openvex.json) (#5137).

  • The ferroehr and ferroehr-viewer images move to the gcr.io/distroless/cc-debian13:nonroot build of 2026-11-08 (sha256:cc66d222…), whose zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). Neither program loads zlib. The images of 5.6.3 and earlier releases carry 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action, as the ferroehr-postgres image of those releases is (security/vex/ferroehr-os.openvex.json) (#5152, #5329).

  • CVE-2026-112321, CVE-2026-112322 and CVE-2026-112323 (memory and CPU exhaustion in kadmind through the AUTH_GSSAPI handler) and CVE-2026-119079 (CPU exhaustion in the KDC decoding an AS-REQ or TGS-REQ) in MIT Kerberos 5 have no Debian fix for trixie and are judged not_affected in ferroehr-postgres. The image carries neither kadmind, a KDC nor libgssrpc, and the PostgreSQL server never decodes either request; the configuration FerroEHR ships enables no GSS-API authentication (security/vex/postgres-os.openvex.json) (#5344).

  • Thirteen MIT Kerberos 5 CVEs of MITKRB5-SA-2026-001 in the ferroehr-postgres image’s libgssapi-krb5-2, libk5crypto3, libkrb5-3 and libkrb5support0 1.21.3-5+deb13u1 have no Debian fix for trixie and are judged not_affected. The image keeps the four libraries because the postgres binary links libgssapi_krb5.so.2. Eleven sit in code the image does not carry: the KDC (CVE-2026-107364), its PKINIT plugin (CVE-2026-71200), kadmind (CVE-2026-107884, CVE-2026-112318), libgssrpc (CVE-2026-112314, CVE-2026-112315, CVE-2026-112320, CVE-2026-118380), the kadm5 libraries (CVE-2026-112316, CVE-2026-112319) and libkdb5 (CVE-2026-112317). The PAC checksum flaw (CVE-2026-111423) and the CAMMAC checksum flaw (CVE-2026-71199) sit in libkrb5, which the server reaches only when it accepts a Kerberos ticket. The configuration FerroEHR ships has no gss or hostgssenc line in pg_hba.conf and no server keytab, and the server never reads a PAC or an authentication indicator (security/vex/postgres-os.openvex.json) (#5374).

  • The five OpenSSL 3.5.7 CVEs of the OpenSSL advisory of 2026-11-10 have no Debian fix for trixie and are judged not_affected in the three images: the AES-GCM-SIV flaws CVE-2026-88823 and CVE-2026-95692, the QUIC flaws CVE-2026-88822 and CVE-2026-95691, and the CMS_encrypt() loop CVE-2026-107496. The ferroehr and ferroehr-viewer programs never map libssl.so.3 or libcrypto.so.3: each links only libc, libm and libgcc_s, imports no dlopen, and speaks TLS through rustls (security/vex/ferroehr-os.openvex.json). The PostgreSQL server maps both for TLS over TCP, which never selects AES-GCM-SIV (no TLS cipher suite is one), never opens a QUIC connection and never writes CMS, and the image never runs the openssl program or loads the legacy provider (security/vex/postgres-os.openvex.json) (#5393).

5.7.0-rc2 - 2026-11-10

5.7.0-rc2 is the second release candidate of 5.7.0, published for testing. 5.7.0-rc1 was tagged but not published: the release’s performance guard booted the candidate without the login provisioner its quickstart ships, so its load test was refused as an unbound login. The guard now starts the provisioner (#5346), and the server is unchanged. These notes carry every change since 5.6.3, those of the 5.7.0-rc1 section below included. 5.6.3 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.7.0 itself replaces it. Before you try it: the configuration is one table per concern with typed durations and sizes, so run ferroehr config check on your file first, which names every key a release before 5.7.0 read with its replacement, all at once; every person is bound to a demographic PERSON, so each persons issuer needs its [[auth.party]] table and each Basic user its [[authz.party_binding.basic_user]] mapping, with the steps on the book’s bootstrap and migration page; and a deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR adds its persons issuers only after FerroBRIDGE v0.0.23 and FerroFED v0.0.17 reach it. Party migrations 0027 to 0032, clinical migration 0035 and audit migrations 0023 to 0025 add the binding, SCIM, DPoP replay and refusal tables, columns and codes. Helm chart 25.0.1-rc2 carries appVersion 5.7.0-rc2. Security carries two advisories, GHSA-jhg2-jwjp-hq74 and GHSA-pfgx-xq28-2629, the zlib fixes of CVE-2026-76844, CVE-2026-85091 and CVE-2026-27171 in the three images, and the not_affected judgements of four krb5 CVEs in the PostgreSQL image.

Added

  • AQL reads the attestations of a version: v/attestations and any ATTESTATION attribute below it (is_pending, reason, proof, items, committer, time_committed) under FROM EHR e CONTAINS VERSION v CONTAINS COMPOSITION c, with predicates on attestations and EXISTS, and v/commit_audit/is_pending and v/commit_audit/reason on a commit audit that is an ATTESTATION. SELECT v/attestations returns the list, and a comparison below attestations holds when any attestation matches. The FerroEHR extension function LAST(v/attestations/<path>) reads the most recent attestation, so WHERE LAST(v/attestations/is_pending) = true lists the versions awaiting signature, including one marked pending again after it was signed (#4989).

  • The party binding’s configuration, each table with its boot refusals: [authz.party_binding] (provider as one table or one table per tenant, namespace, basic_user), [authz.import] trusted_systems, [[authz.system_client]] (the provider, facilities, templates and origins of each client-credentials client that commits) and [[authz.federation.agreement]] (what a federation issuer’s external persons or organisation may do, on which legal ground), with [authz.rbac] role_issuers and import_role. A template an agreement or a system client writes that the store does not hold refuses the start. The boot line and GET /management/info list the issuers with their kind and identifier issuer, the provider, and every agreement with its accessor and legal ground, and an agreement that admits an organisation without a named person is logged at warn (#5003).

  • [auth.idp.entra], the Microsoft Graph access of the idp-entra adapter of ferrohealth-smart, which resolves the groups of an Entra ID groups overage for an issuer whose profile sets adapter = "idp-entra" (#5003).

  • [auth.dpop] and [[auth.issuer]] sender_constraint: tokens bound to a DPoP key (RFC 9449) or a mutual-TLS client certificate (RFC 8705), checked on every request. A DPoP proof is recorded by the replica that accepted it, so the Helm chart refuses a sender-constrained issuer on more than one replica (#5003).

  • The login binding reads the local party source on every request, with no cache: the new party migration 0027_binding_side_table adds the indexed binding_identifier, binding_membership and binding_representation relations, written in each party commit’s transaction from the PERSONs of the mapped templates that perform a professional ROLE or represent another person, their GROUP memberships and their representation relationships, and rebuilt from the party versions alone. A login matching two parties, an inactive, abandoned or deleted party, a token without the key claim, a party without the needed marker and unreadable groups are refused with 403 and their binding code; a party store or identity provider that cannot answer is answered 503 (#5004).

  • ferroehr definitions load <path>... stores OPT 1.4 (.opt) and ADL 2 (.adls) definitions before serve, through the validation and versioning of the REST upload. A file already stored with the same text is left alone and other text under a stored id is refused, so a deployment runs it at every start to give its party mapping the templates it needs at boot. deploy/definitions/professional/ holds a professional PERSON and ROLE pair adapted from openEHR CKM archetypes for the development lanes and the quickstart (#5006).

  • The compose quickstart binds its user to a professional PERSON: the ferroehr-definitions one-shot loads the PERSON and ROLE templates the party mapping names before the server starts, and ferroehr-provision creates the user through /scim/v2/Users with a token of the development issuer, then reports healthy, so docker compose up --wait returns once the login is bound. The Helm chart (23.3.0) takes definitions.files, which an init container stores with ferroehr definitions load before the server container starts. The conformance, browser and Kubernetes lanes load the same templates and provision every login they use through SCIM, with the committed test client of docs/conformance/party/scim/ (#5006).

  • ferroehr party-binding adopt <PARTY_ID>... gives each prepared PERSON of an existing deployment a professional ROLE of the mapping’s role template, with the PERSON as performer, and a new PERSON version listing it in roles, all in one demographic CONTRIBUTION committed by the first administrator named with --admin or --admin-key, and writes one access record per adopted PERSON through the configured audit trail. It refuses the whole adoption, naming them, while any listed PERSON’s key was introduced by a committer that is no [[scim.client]], named administrator or --migration, so a key planted before the upgrade is never adopted. A PERSON that already performs the ROLE is reported and left alone. The report is JSON on stdout (#5009).

  • ferroehr party-binding check is the read-only pre-flight of a migration to party binding, and the one report of the provisional persons just-in-time creation left incomplete. It reads the party versions alone, so it runs before any boot fill, and writes nothing. Per key issuer it reports the PERSONs that carry a binding key, every selector value two PERSONs carry, the inactive and abandoned ones, the provisional ones older than provisional_report_days, the key-bearing ones that perform no professional ROLE, and every mapped Basic user with the PERSON it binds. For each key it names the committer of the version that introduced it, and fails unless that committer is a [[scim.client]], an administrator named with --admin or --admin-key, a migration declared with --migration, or just-in-time creation on an issuer with jit = true. --expected <file> lists every login of the identity provider with no PERSON. The report is JSON on stdout, and the command exits non-zero on a duplicate, a missing person or a key that fails the provenance rule (#5009).

  • At boot, FerroEHR logs a warn naming every Basic user mapped in [[authz.party_binding.basic_user]] while [authz.assurance] minimum is unset: a Basic credential carries no assurance level, so such a person is never refused on patient data (#5009).

  • The one PERSON mapping SCIM provisioning, just-in-time creation and the login binding share: [[authz.party_binding.person.template]] with the binding key, userName and externalId selectors of each template, [[authz.party_binding.person.attribute]] with the SCIM attribute, the OpenID Connect claim and the path of each mapped value, and [authz.party_binding.professional_role], [authz.party_binding.group], [authz.party_binding.membership] and [authz.party_binding.representation]. Boot refuses a mapping that breaks one of its rules, naming the key, including a template that is absent, rooted at another class or lacks a mapped path, and deleting a template the mapping names is refused with 409 (#5013).

  • The new party migration 0028_binding_selectors holds each binding key, userName (compared after RFC 8265 width mapping, lower case and NFC), externalId and group displayName to one live party, so a commit giving one a second party is refused with 409. The first boot fills the side table under the mapping’s digest, a replica with another mapping refuses to start, and ferroehr party-binding refill rebuilds it after a mapping change, refusing while a replica runs and naming the parties when a value is carried twice (#5013).

  • POST /admin/party/binding-key re-keys a PERSON’s binding key: it refuses the administrator’s own login and a key already bound, retires the old key in the same commit and writes one access record carrying both (#5013).

  • [scim] mounts the SCIM 2.0 service provider at /scim/v2/ (off by default): [[scim.client]] declares each provisioning client by issuer, client_id, scope and binding_issuer, and public_base_url names the origin its documents hang off. /scim/v2/ServiceProviderConfig answers without a token; /ResourceTypes and /Schemas answer a declared client acting for nobody that holds its scope, within [server.caller_rate]. Every refusal is an RFC 7644 Error message. The start is refused for a client of a non-machines issuer, one declared twice, two clients sharing a binding_issuer, or a binding_issuer no [[auth.party]] table carries (#5014).

  • The SCIM endpoint answers the Microsoft Entra ID provisioning client as Microsoft’s “Develop a SCIM endpoint” guide documents it: a value-filtered attribute such as emails[type eq "work"].value serves its type and declares primary read-only, so the primary Entra sends is ignored (RFC 7644 §3.3); a PATCH of userName renames the User; and the Test Connection probes are answered with an empty ListResponse and 404. ferroehr config check refuses a [scim] the person mapping cannot serve, as serve does. The Helm chart exposes config.scim and routes /scim/v2 on every Ingress host when it is enabled, and the book walks an operator through provisioning from Entra ID, with the Workload Identity Federation checks the authorization server makes (#5014).

  • /scim/v2/Users/{id} takes PUT, PATCH and DELETE: a write commits the next version of the professional’s PERSON, clearing a mapped attribute a PUT leaves out, keeping the binding key and externalId, and moving the lifecycle state by active (complete or incomplete to inactive or abandoned and back, an abandoned User completed in two versions, an attribute change to an inactive User refused until it is reactivated). DELETE logically deletes the PERSON and its professional ROLE in one CONTRIBUTION and leaves the binding key as a tombstone. An If-Match naming another version is 412, and ServiceProviderConfig declares PATCH. [scim] max_deactivations and deactivation_window_minutes bound the deactivations and deletions of one client, counted in the database so replicas share the bound, with 429 past it (#5014).

  • /scim/v2/Users serves a client’s Users: POST creates the professional’s PERSON and its professional ROLE in one demographic CONTRIBUTION, complete when the template holds in full and incomplete otherwise, GET /Users/{id} answers with the version’s entity tag and 304 to a matching If-None-Match, and GET /Users answers a ListResponse filtered by eq and and over userName, externalId and id, with any other filter 400 invalidFilter. A client sees only the Users whose binding key its binding_issuer issues, every SCIM request is an access record of the client naming the PERSON, and ServiceProviderConfig declares entity tags. The start is refused when [scim] is on and the party mapping has no default person template with a user_name selector or no professional role (#5014).

  • /scim/v2/Groups serves SCIM Groups as demographic GROUPs of [authz.party_binding.group], each member holding an inline PARTY_RELATIONSHIP of [authz.party_binding.membership] from its PERSON to the GROUP, with no ROLE minted: create, read, eq/and filters on displayName, externalId and id, excludedAttributes=members, PATCH (answered 204), PUT and DELETE, with the entity tags of Users. A Group belongs to the client that created it, its members are that client’s Users, its displayName is unique, and its externalId sits under the client’s group_identifier_issuer. A remove of members with a value list removes only those members, an Entra compatibility deviation from RFC 7644 §3.5.2.2. Deleting a User removes its memberships in the same CONTRIBUTION. [scim] max_group_members (default 5000) caps a Group (#5015).

  • Just-in-time creation of a provisional PERSON at a first login, off by default and enabled per persons issuer in a new [[authz.party_binding.issuer]] table (jit, guests_jit, user_name_claim, idp_claim, admit_idp, admit_groups, admit_roles, admit_assurance). A login of a person or a delegated caller that matches no party and meets every declared admission condition commits one 553|incomplete| PERSON holding the binding key, the userName and the name claims of the access token, with its professional ROLE, as a creation by the person, and is served bound to it; any other mapped attribute is written only where its mapping sets jit = true, and externalId never. Concurrent first requests commit one PERSON, a deprovisioned or inactive login is never recreated, a taken userName refuses the login, and a SCIM client of the key’s issuer completes the PERSON. Boot refuses jit without an admission condition or a mapping that can write it. ferroehr party-binding check lists the provisional persons older than [authz.party_binding] provisional_report_days (default 30) (#5016).

  • A party of a mapped template is validated against that template on every demographic write: a PERSON or GROUP whose archetype_details.template_id names a template of the shared party mapping, and a PARTY_RELATIONSHIP of its membership archetype. A 532|complete| version must meet every constraint, and a 553|incomplete| one may miss mandatory nodes but nothing it carries may be wrong (RM common master06 §Incomplete Content); the refusal is a 422 naming each failing path. Every other party keeps the reference-model checks alone, so parties created through the Demographic API under other templates are accepted as before (#5042).

  • The FHIR AuditEvent, the DICOM audit message and the ITI-20 syslog feed carry the binding fields of the access record: the binding outcome, the party and its groups, the provider acted for, the facilities of a system commit, the representation with its relationship, the actor chain, the external accessor’s asserted fields and the agreement with its legal ground, consent and access class, and a chained request’s correlation id. ehds-logging 0.5.1 writes them as the ehds-accessor and ehds-relayed entities, in the DICOM message as participant objects of type Other. The FHIR record also writes the external and system acting modes. FerroEHR moves to ehds-logging and ihe-atna 0.5.1 and ihe-iti 0.0.77 (#5043).

  • A bound representative reaches only the EHR whose EHR_STATUS names the subject it represents. Every composition, EHR_STATUS, directory, contribution and query request is held to that EHR whatever the token’s scopes cover, and a query must name it with ehr_id. A representative whose token names a subject its party has no live representation relationship to, or whose issuer no [[auth.party]] table binds, is refused with 403 and refused-representation-unrecorded (#5086).

  • The Helm chart (23.1.0) accepts bridge.cdr.oauth.tokenEndpoint for an authorization server whose metadata FerroBRIDGE cannot read at bridge.cdr.oauth.issuer. FerroBRIDGE then requests its client credentials token from that endpoint, and issuer still names the machines issuer the CDR trusts (#5093).

  • A MyHealth@EU contact point’s caller (a contact_point = true federation issuer) reaches patient data only for the patient of its treatment relationship confirmation. The token’s ehdsi.patient.id (system|value) is resolved to its EHR through the EHR index, and every patient-data request must name that EHR. A token without it is refused with 403 and ERROR_SEC_DATA_INTEGRITY_NOT_ENSURED, recorded as refused-trc-missing. Another patient’s EHR, an unknown EHR and a request naming no EHR all answer one 404 of the new problem type trc-patient-mismatch, carrying the agreement service’s *_NOT_FOUND code and recorded as a refused-trc-patient-mismatch security event (eHDSI XCA Profile 8.0.1 §4.2.1, §4.2.2). A contact point’s person agreement now names its ehdsi_service; upgrade step: set it, and list the issuer in [smart] patient_issuers (#5102).

  • An external person under an agreement with discard discards a dispensation of a listed template, as a logical delete, within discard_window_hours and from the same point of care: the dispensation’s health_care_facility carries the token’s provider identifier and its eHDSI organization-id. A token carrying permissions records and discards a dispensation only with PPD-046 (eHDSI SAML Profile 10.0.0 §2.5) (#5102).

  • A read under a federation agreement with a consent condition is served when the token carries the consent decision reference for its patient (Wabvpz Art. 1(j) and Art. 15a(1)), in the claim or authorization_details member the issuer’s [[auth.federation]] table maps to consent.reference, and refused with 403 without it. The reference is one string, the token’s patient/ grant holds it to the launch-context patient, and the access record carries it with the other external fields. ferrohealth-smart moves to 0.9.3, which reads the field (#5106).

  • Every access record of an authenticated request carries the outcome of the party binding (binding: bound, provisional, external, organisation, system, self or a refused-* code) with the fields it needs: the bound party, its groups, the provider, the guest mark, a representative’s representation, a delegated token’s actor chain, a verified external accessor’s fields and agreement, and the facilities a system caller’s compositions named. acting_mode gains delegated, external and system. The audit migration 0023_binding_outcome.sql adds the columns and restores a per-row digest version: records written before stay version 1 and keep verifying, new records are version 2, and a version-1 record after a version-2 one is refused (#5118).

  • Every refused request is an access record through the gated path, and a refusal of the login binding or of a caller kind carries its refused-* binding code and the party it names. Refusals of tokens that failed verification are budgeted by source network and those of verified callers by issuer and subject, so neither crowds out the other; past a caller’s limit its refusals become one summary record with the count and the EHR ids they named, and past a budget one summary record per issuer and outcome is written. Security-event refusals are recorded for every answered request, and a refusal record the system log cannot take is counted in atna_audit_refusals_unwritten. The limits, windows and budgets are kept per replica (#5119).

  • A chained organisation request under a federation agreement is served only when its token carries the requester’s correlation id in the claim or attested authorization_details member the issuer’s [[auth.federation]] table maps to correlation_id. Without it the request is refused with 403 and recorded under the binding code refused-correlation-missing (audit migration 0025); a request header never stands in for it. The organisation access record’s agreement carries the id and its source beside the legal ground and the access class. A chained agreement whose issuer maps no correlation_id, or maps it from a constant, refuses the start (#5120).

  • On the MyHealth@EU contact-point route, the committer of an external person names the professional by the HP Identity Assertion’s NameID: DV_IDENTIFIER{issuer = the assertion's issuer, id = the NameID, type = "hp_identifier:<NameID format>"}, beside the voucher’s oauth2 identifier (eHDSI SAML Profile 10.0.0 §2.1), and a dispensation’s discard is a logical delete under the same committer. A contact point’s COMPOSITION must carry the HCPO’s organization-id in EVENT_CONTEXT.health_care_facility.identifiers, under the issuer urn:oasis:names:tc:xspa:1.0:subject:organization-id, beside the provider’s identifier, or it is refused with 422. ferrohealth-smart moves to 0.9.4, which reads ehdsi.assertion.name_id_format and refuses a contact-point token without one of the three formats as ERROR_HPI_INSUFFICIENT_INFORMATION; upgrade step: a contact point’s authorization server writes the NameID format into its tokens (#5144).

  • A federation caller above its agreement’s caller_rate, a MyHealth@EU contact point’s included, is answered 429 with Retry-After right after authentication and before any patient lookup, keyed on the token’s subject per replica, and every such refusal is counted into the caller’s summary record (#5149).

  • A deleted SCIM User’s contacts are erased from every version of its PERSON, keeping the name, the binding key and the identifiers for the record-keeping duties that need them. [scim] contact_erasure_minutes (default 0, at most 43200) sets when: 0 erases them in the DELETE itself, a period leaves them to a sweep that runs every minute. Each erased version’s node rows are derived again from the new body, its server signature is removed, and the erasure is an access record naming party:<id>#contacts. Party migration 0030 adds the scim_contact_erasure table (#5178).

  • openehr-rm types generic RM type names and paths: typing::RmType parses and prints names such as DV_INTERVAL<DV_QUANTITY>, typing::attribute_type gives an attribute of an instantiated class with the actual argument in place of the formal parameter, and typing::path_type types every step of an RmPath with its container. Each model::RmAttribute and model::RmTypeRef keeps the formal parameter it is written as (formal) beside the resolved bound. paths::write::item_at_path_mut and paths::write::set_at_path write at a path into a canonical instance, positional predicates included, creating the members the path names and changing nothing when a write is refused (#5187).

  • openehr-sdt: flat::navigate::Navigator gives every instance of a Web Template path in a composition, with each step’s occurrence, container index, positional path and JSON pointer, reaching a renamed instance where the template’s name is redundant and an OPT2 archetype root through the complete identifier the template recorded. flat::navigate::relative compares two paths step by step. flat::attribute_keys gives the FLAT key of an RM attribute of a structural class (the master05-rm_mapping.adoc tables, checked against them). A WebTemplateNode carries facts (the embedded template id of an OPT 1.4 archetype root and the terms its archetype defines) and every coded option its terminology, so no consumer walks the operational template (#5189).

  • openehr-base gives identifier and time parts as written: ObjectVersionId::object_id_str and ObjectVersionId::versioned_object_id keep the case of the object_id part, Iso8601Time::fraction_digits and Iso8601DateTime::fraction_digits return the digits of the fractional second as text, and Iso8601Duration::parts returns each component’s designator and number verbatim (DurationDesignator, DurationPart), with Iso8601Duration::fraction_digits for the seconds (#5190).

  • openehr-its reads the ITS-REST ETag as a typed entity tag: rest::etag::EntityTag parses and writes the RFC 9110 form, weak or strong, with the weak and strong comparisons. It types the tag as an ObjectVersionId or a HierObjectId. Every generated response-headers struct that carries an ETag gains entity_tag() beside the raw etag field (#5191).

  • openehr-term names the null_flavours and setting groups: codes::NullFlavour and codes::Setting give each concept’s code and English rubric and read a code back, held equal to the bundled terminology by tests. Every typed group’s ALL is now a slice (#5192).

  • openehr-am: HridRef reads an archetype reference with the version precision it was written at (VersionPrecision: major, minor or full, BASE Artefact Identification §Referencing), gives the version as written and matches an ArchetypeHrid by the specification’s rule, the namespace never a wildcard. ArchetypeHrid::local_semantic_id gives the namespace-free interface form a Web Template nodeId carries, which the v2_4 Web Template builder now uses (#5193).

  • openehr-its: rest::stored_query types a stored query’s qualified name (QualifiedQueryName, [{namespace}::]{query-name}, with check_registrable refusing the reserved aql), its version (QueryVersion, major.minor.patch in numeric order) and the version pattern a request addresses (VersionPattern, exact, {major} or {major}.{minor}, selecting the highest match), per ITS-REST 1.1.0. The generated stored-query operations and the StoredQuery and RESULT_SET documents carry them, so a name or version outside the grammar is a 400 where the request is decoded (#5207).

  • openehr_rm::Generation::earliest_for_composition answers the oldest RM generation whose canonical-JSON decode accepts a composition, so a builder can name ARCHETYPED.rm_version without naming a generation module; a value no generation decodes is a typed NoGenerationDecodes error carrying every generation’s refusal (#5242).

Changed

  • The published server, viewer and PostgreSQL images take their org.opencontainers.image.base.name and base.digest labels and annotations from the runtime stage’s digest-pinned FROM line at build time, so a base-image bump changes that one line. An image built directly from a Dockerfile no longer carries these two labels (#4953).

  • [auth] takes the ferrohealth-smart 0.9 shape. Every trusted issuer is an [[auth.issuer]] table with a required kind (persons, patients, machines or federation) and a profile (rfc9068 by default, or entra, keycloak, okta, nl), and holds every setting of its tokens itself. A configuration in the old shape refuses the start, naming what replaces it. Upgrade steps: write a flat issuer = "…" as an [[auth.issuer]] table; move audience, algorithms, clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds, jwks_file, timeout_ms, mode, role_claims and [auth.introspection] onto each issuer table ([auth.issuer.introspection]); give each issuer its kind, with patients and professionals at separate issuers (a realm or tenant each), since one issuer is one kind; replace token_profile = "jwt" with token_types = ["JWT"] and the required_claims the issuer’s tokens carry, and drop token_profile = "rfc9068"; and write each scopeless_subjects entry as { issuer = "…", sub = "…" } (#5003).

  • A persons token is admitted only through a client its issuer lists in first_party_clients or delegates. Upgrade step: list the deployment’s own clients, the viewer’s among them, in first_party_clients, and a third-party client that acts for a person in delegates. The other 0.9 profile keys (person_key, tenant_claim, client_claims, app_only, guest, impersonation_marker, expect, user_claims, person_equality, unstable_claims, max_actor_chain, representation, scope_claim, app_only_scope_claim, adapter, sender_constraint and the onward_* keys) are documented in the configuration reference (#5003).

  • An array of tables, such as [[auth.issuer]], is set from the environment by element index (FERROEHR__AUTH__ISSUER__0__ISSUER), through ferrohealth-config 0.2.10 (#5003).

  • [smart] issuer_metadata_file is the metadata of the [auth] issuer it names, wherever that issuer stands in the list, and a file naming no [auth] issuer refuses the start. The SMART discovery document describes the first introspected issuer, else that named issuer, else the first issuer whose metadata is read from it, else the first issuer, so a Basic users’ issuer listed first no longer takes the discovery document (#5003).

  • The family crates move together: ferrohealth-smart 0.9.1, ferrohealth-authz 0.3.0, ferrohealth-session 0.3.1, ihe-atna and ehds-logging 0.4.0, ihe-iti 0.0.76, ferrohealth-config 0.2.10, ferrohealth-server 0.5.1 and ferrohealth-ui 0.6.9. ihe-atna and ehds-logging now escape every control and bidirectional formatting character of a value they write as its code point. With ferrohealth-ui 0.6.9 the viewer’s session watch no longer throws when the party relationship form, which holds a control named target, is submitted (#5003).

  • The terminology client moves to ferrobridge-term 0.1.143 (FerroBRIDGE v0.0.22), which takes the ferrohealth-smart 0.9 token source (#5003).

  • Every person is bound to a demographic party, with no switch. A persons issuer needs its [[auth.party]] table, two issuers of one node may not share an audience, and [authz.party_binding] provider (the healthcare provider the deployment acts for, EHDS Art. 13(4) and Annex II 3.2(a)) is required once [auth] admits a credential; each refuses the start naming what to set. A Basic user is a person: it maps to the binding key of its PERSON in [[authz.party_binding.basic_user]] (user, key_issuer, key), and a Basic user without a mapping refuses the start, naming the user. Upgrade steps: add an [[auth.party]] table to each persons issuer, set provider, and map each Basic user under the identifier_issuer of an [[auth.party]] table, which needs a persons issuer; a deployment of Basic users alone declares one whose jwks_file is a key set that signs nothing, as the quickstart and the sandbox do (#5003).

  • A token grants [authz.rbac] admin_role, emergency_access_role, import_role and identifier_reader_role only when its issuer is listed in the new [authz.rbac] role_issuers, each a persons issuer; any other kind refuses the start. Upgrade step: list the issuers whose role claims grant these roles; a Basic user keeps the roles its own table grants (#5003).

  • The compose quickstart, the dev and conformance stacks, the hosted sandbox and the Helm chart’s Basic overlays bind their Basic users under a persons issuer that signs nothing: its key set, which every FerroEHR image carries at /opt/sandbox-posture/basic-users-jwks.json, is the public half of a key pair whose private key was discarded (#5003).

  • The quickstart has one copy. docker-compose.yml on main pulls the FerroEHR images at :main and carries main’s configuration, which CI checks with the server image the change builds; the copy attached to each release pulls that release’s images, written by the release lane, which refuses a copy that still names :main. Download the release asset (releases/latest/download/docker-compose.yml) to run a release (#5003).

  • The Helm chart moves to 22.0.0. config.auth.issuer is a list of issuer tables, each with its kind, profile and audience, and values.yaml exposes config.auth.party, config.authz.party_binding, config.authz.rbac.role_issuers, config.authz.system_client and config.authz.federation.agreement. The chart refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling. Upgrade steps: rewrite config.auth as the [auth] upgrade above describes, set config.authz.party_binding.provider, and list the issuers whose roles grant the administrator role in config.authz.rbac.role_issuers (#5003).

  • A caller bound to a demographic PERSON commits under the committer the server writes: a PARTY_IDENTIFIED with the PERSON’s name, the login identifier (the issuer profile’s person_key value under the token issuer, type oauth2, or the Basic user under ferroehr, type basic), the provider identifier (type healthcare-provider) and an external_ref to the PERSON, on every commit path, attestations and CONTRIBUTIONs included. Upgrade: stop sending a committer, or send the placeholder {"_type": "PARTY_IDENTIFIED", "identifiers": [{"_type": "DV_IDENTIFIER", "issuer": "<iss>", "id": "<sub>"}]}, which must name the authenticated caller; a stated committer with another name, identifier or external_ref, and a stated PARTY_RELATED, is refused with 403 (#5005).

  • A patient’s committer is PARTY_SELF with no external_ref, and a representative commits as a PARTY_RELATED naming their own PERSON and the coded relationship. A patient-context token commits no demographic version. Upgrade: read the person’s own entries by the committer’s PARTY_SELF type, not by its former external_ref (#5005).

  • Importing an EHR Extract needs the role [authz.rbac] import_role names, or a system/ caller, and a wrapped committer naming a party or a login of this deployment is refused unless its version was created by a system [authz.import] trusted_systems lists. Upgrade: set import_role and give it to the accounts that import (#5005).

  • A person of a persons issuer whose login no demographic party carries is refused with 403 and refused-unbound on every request, with no fallback to a login-only committer. The problem names the identifier issuer to provision the person under, a token without the key claim is refused with a problem naming that claim, and each refusal counts into the new binding_refusals metric by issuer and outcome. The Keycloak quickstart overlay provisions its four demo users through SCIM with the ferroehr-provision-keycloak one-shot. Upgrade: provision every professional (through SCIM or the demographic API) before you add a persons issuer to the upgraded deployment (#5006).

  • A deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR adds its persons issuers only after the bridge and the gateway reach it under the caller: FerroBRIDGE v0.0.23 and FerroFED v0.0.17. Until then the committer rule refuses their writes. FerroFED v0.0.17 follows this release, so a deployment with a FerroFED gateway adds its persons issuers only once that release is out. The book’s new “Party binding: bootstrap and migration” page carries the bootstrap path, the four migration steps and these prerequisites, and “Breaking changes of party binding” lists every incompatible change with its upgrade step (#5009).

  • The hosted sandbox follows the bootstrap path of party binding: its configuration declares the sandbox-scim client ([[scim.client]], with the development issuer trusted as a machines issuer for it), the box’s new ferroehr-provision service provisions the demo PERSON through /scim/v2/ after every reset, and scripts/sandbox/reseed.sh waits for the demo login to bind before it seeds, then checks the committer of the marker’s EHR_STATUS version references a PERSON. The deploy probe’s new bootstrap family (P-BOOTSTRAP-01) shows the compose quickstart does the same from an empty database: the quickstart user is a User of the SCIM client, and a commit of the Basic user names that PERSON (#5009).

  • With a party mapping configured, the parties the login binding relies on are written through the demographic API by an administrator (authz.rbac.admin_role) only: a PERSON of a mapped template carrying a binding key, userName or externalId, its professional ROLE, a representation relationship, and a GROUP of the group template. Any other caller’s write is refused with 403, a version that changes a present binding key is refused with 403, and an administrator’s version that leaves a protected identifier out is refused with 422. Upgrade: write these parties as an administrator, and change a key through POST /admin/party/binding-key (#5013).

  • FerroEHR records every DPoP proof it accepts in its PostgreSQL database until the proof’s window ends, so a proof is accepted once across all replicas, and a proof replayed to another replica is refused there (RFC 9449 §11.1). If the database cannot record a proof, the request is refused with 503. The Helm chart (22.1.0) no longer refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling (#5072).

  • An AQL query scoped to one EHR now takes the streaming plan shape whatever its ORDER BY, DISTINCT or aggregates, whether the EHR comes from the ehr_id request parameter or from an ehr_id/value equality in the query. On a store of 100,000 compositions such a query used to spend 16 to 64 ms in PostgreSQL’s planner for an execution of about 1 ms, on every new connection and every new statement text; it now plans in 1 to 3 ms (#5076).

  • The Helm chart (23.0.0) connects FerroBRIDGE to the CDR as a machine. FerroBRIDGE uses an OAuth 2.0 client credentials client of a machines issuer that the CDR trusts, set in bridge.cdr.oauth (issuer, clientId, and optionally scope, audience and resource). The client secret comes from the client_secret key of bridge.cdr.existingSecret. The chart no longer accepts a Basic user for this connection: from ferrohealth-smart 0.9 a Basic credential belongs to a person and holds no system/ scope. The chart refuses to render when bridge.cdr.oauth.issuer is not listed in config.auth.issuer with kind: machines. To upgrade, add that issuer, create the client, and replace the user and password keys of the Secret with client_secret (#5085).

  • A verified external person’s commit records the professional as its committer: a PARTY_IDENTIFIED with the asserted name, the hp_identifier under the issuer "<country_code>:<issuing_authority_name>" and the token’s iss and sub, with no external_ref. A country code that is no upper-case ISO 3166-1 alpha-2 code, or an issuing authority holding a :, control or bidi character, is refused with 403 naming the field. The COMPOSITION must name the token’s healthcare provider (provider.healthcare_provider_identifier under provider.issuing_authority_name) as its health_care_facility, or it is refused with 422. Upgrade step: map provider.issuing_authority_name in each [[auth.federation]] table whose agreements carry writes, with a constant source where the token does not carry it (#5097).

  • A system caller (an app-only token by its issuer profile’s app_only markers) holds system/ scopes only: a token carrying a user/ or patient/ scope is refused with 403. It commits into an EHR only as a client an [[authz.system_client]] table declares, under the committer PARTY_IDENTIFIED{name = <client>, identifiers = [{issuer = <iss>, id = <client>, type = "client"}]}, and each COMPOSITION it commits must be of a template its writes lists (403), name a facility its facilities lists in context.health_care_facility, or in feeder_audit.originating_system_audit.location when it has no context (422), and be either a device act from a device client or a relayed person act naming the professional by an identifier from a relay client. Upgrade: declare every client that commits in [[authz.system_client]] with its writes, facilities and origins, and give its compositions the facility identifier (#5100).

  • An attestation is accepted only from a caller bound to a professional PERSON; one from a patient, a representative, a system caller or an unbound caller is refused with 403. Upgrade: attest under the professional’s own token (#5100).

  • A guest of a persons issuer, by its profile’s guest markers, is refused with 403 (refused-guest) unless the issuer’s [[auth.party]] table sets guests = true. Upgrade: set guests = true on the issuers whose guests act under your responsibility (#5100).

  • A verified external person or organisation of a federation issuer is admitted only under the [[authz.federation.agreement]] of its issuer and accessor, holding no scope beyond the agreement’s reads, and is an access-service caller by its kind: an opted-out EHR answers 404 to it, and only an external person under an agreement with opt_out_override reaches one under an emergency purpose. It commits only a COMPOSITION of a template the agreement’s writes lists (an organisation nothing), and a read under a consent-conditioned agreement, and every request under an NL agreement, is refused until the consent decision reference and the Wabvpz Art. 15f roles can be read. Upgrade: declare an agreement for every federation issuer and accessor, list the issuer in [smart] patient_issuers, and give its tokens only the patient/ scopes the agreement reads (#5101).

  • Under an NL [[authz.federation.agreement]], Wabvpz Art. 15f now decides admission in place of the refusal of every request. A caller whose nl.role is the UZI role 01.008 (bedrijfsarts) or 01.048 (verzekeringsarts) is refused with 403 as refused-wabvpz-role, and a caller whose nl.organisation_type is missing or is not a care provider type of Nictiz’s OrganisatieTypeCodelijst is refused with 403 as refused-wabvpz-organisation-type, so an insurer is refused. Every NL agreement declares purpose = "care"; one without a purpose, or for a keuring, refuses the start. The code lists are vendored by scripts/vendor/nl-role-codes.sh into docs/specs/nl-role-codes/. Upgrade: add purpose = "care" to every NL agreement and map nl.organisation_type (and nl.role for persons) in its issuer’s [[auth.federation]] table (#5107).

  • A delegated token’s access record now has acting_mode = "delegated" where it had client, and a system caller’s system. The FHIR rendering still names a delegated access as a client one. Upgrade: a review query that selects delegated access by acting_mode = 'client' also selects 'delegated' (#5118).

  • FerroEHR builds on ferrohealth-server 0.6.1, whose public caller-rate take holds a federation caller to its agreement’s caller_rate. Its 0.6.0 break, the revocation feature’s move to der 0.8, touches no feature FerroEHR enables (#5149).

  • The source tree no longer carries the UZI role code CodeSystem (uzi-rolcode) of Nictiz’s zib 2020 package: its codes and titles reproduce the UZI-register CPS, which CIBG does not license for reproduction. scripts/vendor/nl-role-codes.sh writes it into the git-ignored .vendor-cache/nl-role-codes/, as it does the CPS, and --test-inputs fetches only what the Wabvpz tests read. Nictiz’s care provider types stay under docs/specs/nl-role-codes/. Run scripts/vendor/nl-role-codes.sh --test-inputs once before running the ferroehr-rest tests (#5175).

  • [auth] takes three boot refusals from ferrohealth-smart: a Basic user’s hash below the Argon2id floor (m=19456, t=2, p=1), an issuer that carries userinfo, and private or symmetric key material in a jwks_file. They are refused as before, with the crate’s messages. A jwks_file that is no JWK Set now refuses the start where the crate reads the key set (#5209).

  • The family crates move together to one ferrohealth-smart: 0.13.1, with ferrohealth-authz 0.7.0, ferrohealth-session 0.7.0, ihe-atna and ehds-logging 0.8.0, ihe-iti 0.0.83, ferrohealth-server 0.6.4 and ferrobridge-term 0.1.148. The 0.13 line refuses a JSON document it cannot read without quoting it (GHSA-mhv7-3975-6xq3), so a jwks_file or [smart] issuer_metadata_file that does not parse is refused at boot with the crate’s message, which names the failure class and position; 0.13.1 keeps a refused key set, introspection or metadata body out of the error (GHSA-5qm7-v4fw-p2v9). The OAuth client credentials tables ([authz.abac.remote.oauth], [ehr_index.oauth], [terminology.oauth] and [audit.fhir_feed.oauth]) read as before (#5209).

  • The API’s response security headers come from the family server stack (ferrohealth-server 0.6.2) with the same six headers and values. They now also reach the management listener on its own port, and Strict-Transport-Security: max-age=31536000 is sent when FerroEHR or the viewer serves TLS itself ([server.tls], RFC 6797 §7.2). The viewer gains Cross-Origin-Resource-Policy: same-site (#5210).

  • The server and the viewer read the parts of an OBJECT_VERSION_ID through openehr-base and its BASE grammar, never by cutting at ::. A value that is no OBJECT_VERSION_ID now names no version: the AQL branch refusal, the version item tags and the viewer’s version numbers leave it alone, a reference to a container is read as the whole value (so uuid::extension no longer names the container uuid), and the viewer’s deletes refuse it before calling the CDR. Every part keeps its case (#5211).

  • ETag and If-Match values are read through openehr-its’s RFC 9110 entity tag on the server and in the viewer. The server accepts the same forms as before. The viewer reads its CDR’s ETag strictly: a value that is no entity tag names no version (#5212).

  • The viewer’s audit screen reads the ITI-81 searchset and each AuditEvent as FHIR R4 resources, and a refusal’s OperationOutcome too. A searchset that is no FHIR R4 Bundle of AuditEvents now shows as a failed read, never as a page of empty rows (#5214).

  • The EHR index’s boot check reads the PIX Manager’s CapabilityStatement through ihe-iti’s PIXm client, with its transport and its client-credentials token, so a Manager behind IUA answers the probe as it answers ITI-83. A failed probe is reported as the client’s refusal: no answer, the status it answered, or no FHIR R4 CapabilityStatement (#5216).

  • [audit.syslog] reads its TLS material from [audit.syslog.tls], the outbound TLS table of every FerroHEALTH product (trust_roots_file, client_identity_file or client_identity, min_version), and hands ihe-atna the rustls configuration ferrohealth-server builds from it. The keys tls_ca_file, tls_identity_cert_file and tls_identity_key_file are retired and refused at boot by name: name the roots file as trust_roots_file, and concatenate the client certificate chain and its private key into the one PEM client_identity_file names. The client speaks TLS 1.3 unless min_version = "1.2" lowers it (#5217).

  • A SCIM Group membership is now held by the member: an inline PARTY_RELATIONSHIP from the member’s PERSON to the GROUP, stored in the PERSON’s version (RM demographic §Party Relationships). A membership change versions only that PERSON, so its cost no longer grows with the Group’s size. The GROUP is versioned only by a rename, an externalId change or its deletion. The Group’s members is read from the membership index (#5220).

  • A Group’s ETag is now its GROUP version plus a membership revision, so it still changes with every membership change. The member User’s ETag now changes too, while its body stays the same (#5220).

  • Deactivating a User (active = false) drops all of its memberships in the same version, because an inactive PERSON cannot be versioned again. A reactivated User belongs to no Group until its client adds it again (#5220).

  • Microsoft Entra ID’s disable-then-remove order now succeeds, since the later remove finds nothing to remove. An add naming an inactive User is 400 invalidValue (“reactivate the User first”) (#5220).

  • A refused SCIM write now names the rule it breaks in detail (#5220).

  • scripts/perf/scim-groups.sh keeps the database log, df -h and free -m of a failed run (#5220).

  • A party commit no longer re-derives the login binding’s side table, the reverse-relationship index and the protection check from the whole stored body. Each hook reads what the committed body holds, derived once in memory, and writes only the rows that changed, and ferroehr party-binding refill derives the rows with the same code. A party commit without a server signature no longer reassembles its body. A membership change at the cap of a 5 000-member SCIM Group went from 810 ms to 358 ms at p99 on a developer laptop (#5220).

  • openehr-rm writes the RFC 8785 canonical form a VERSION signature covers in one pass over the version, sorting each object’s members by reference. A 3.3 MB GROUP version canonicalises about seven times faster than through serde_jcs, with byte-identical output, so signed versions verify unchanged. serde_jcs remains the test oracle (#5220).

  • The stored-query service takes the typed name, version and version pattern of openehr_its::rest::stored_query end to end, so it reads one grammar with the REST layer: the three-part {namespace}::{formalism}::{query-name} name, which ITS-REST 1.1.0 Qualified_query_name.md does not define, is no longer accepted anywhere (the formalism is the query_type parameter), and DELETE /admin/query/{qualified_query_name}/{version} answers 400 for a name outside [{namespace}::]{query-name} or a version that is no exact major.minor.patch, where it answered 404 (#5223).

  • Boot and ferroehr config check refuse an [[auth.party]] table whose identifier_type is the type of no [[authz.party_binding.person.template]] key, naming the issuer, the type and the types the templates write: no login of such an issuer could ever match a party, and every one would be refused. Upgrade step: give the deployment a person template whose key carries the table’s identifier_type (or set the table to the type the template writes). The Helm CI and production values and the hosted sandbox now carry the professional PERSON and ROLE mapping, and the sandbox loads the two templates before the CDR starts (#5231).

  • openehr_query::printer::to_aql writes each WHERE comparison with one space on each side of the operator (c/name/value = $nameValue), as the QUERY 1.1.0 AQL syntax examples do; path predicates keep their unspaced form ([ehr_id/value=$ehrUid]). The query builder’s generated AQL follows (#5241).

  • openehr_sdt::flat::navigate::Navigator owns what it copies from its Web Template (the name-fallback table and the OPT2 archetype references) and carries no template lifetime, so a consumer can keep it beside the template it was built from; it is Debug, Clone, Send and Sync (#5243).

  • Breaking: the configuration tree is one table per concern. ferroehr.toml moves to [deployment], [api], [database], [auth] (with [auth.smart] and [auth.scim]), [authz] (with [[authz.origin]]), [management], [audit], [privacy] (with identifier_protection, cohort and access_delay), [record] (with subject_index, versioning, signing, query, change_feed and multimedia), [terminology] and [pixm]. FerroEHR’s own keys take one duration type ("500ms", "30s", "10min", "12h", "7d", UCUM unit codes, so a bare number and m are refused), one size type ("256KiB", "64MiB", IEC binary units, decimal units refused) and one calendar retention period ("P5Y", "P1830D", unset keeps forever). An optional feature is on when its table is present: [terminology], [pixm], [auth.scim], [audit.syslog], [audit.fhir_feed], [privacy.identifier_protection] and [privacy.access_delay] carry no enabled; [authz.abac], [record.multimedia] and [auth.smart] keep a switch because some of their keys apply while off. record.subject_index has no default, so the server refuses to start until "pixm" or "local" is chosen. [api.cors] replaces cors_permissive with the Fetch Standard response headers (allow_origins, allow_methods, allow_headers, expose_headers, allow_credentials, max_age). The environment follows the tree (FERROEHR__DB__URL is FERROEHR__DATABASE__URL, FERROEHR__EHR_INDEX__PROVIDER is FERROEHR__RECORD__SUBJECT_INDEX). The Helm chart moves to its 25 line with config: on the same tree. Every key a release before v5.7.0 read is refused at load with its replacement named, all at once, so ferroehr config check --config <old file> prints the whole move list; the refusals ship in v5.7.0 only and the next release removes them. To upgrade, run config check on the old file and move each key it names: spec_profile, deployment_profile, deployment_accepts and licence.file are deployment.spec_profile, deployment.profile, deployment.accepts and deployment.licence_file; [rest] is [api] (swagger_ui is openapi_ui, body_limit_bytes is body_limit, cors_permissive = true is [api.cors] allow_origins = ["*"], [rest.identity] is [api.options] with solution and vendor; the version keys are facts of the build), admin.enabled is api.admin and terminology.api_enabled is api.terminology; [db] is [database] (acquire_timeout_secs is acquire_timeout, statement_timeout_ms is statement_timeout, migrate is migrations), [storage.<domain>] is [database.<domain>] and [storage.migrate.<domain>] url is [database.<domain>] migrate_url; [smart] is [auth.smart] (enabled is discovery, episode.enabled is episode_launch, launch_base64_json = true is "launch-base64-json" in auth.capabilities); [scim] is [auth.scim] (max_results is filter_max_results, deactivation_window_minutes is deactivation_window, contact_erasure_minutes is contact_erasure_delay); [authz.abac.cedar] reload_secs is reload_interval, [authz.abac.remote] server, connect_timeout_ms and request_timeout_ms are url, connect_timeout and timeout, authz.party_binding.provisional_report_days is provisional_report_after, and an agreement’s discard_window_hours and chained_log_retention_years are discard_window and chained_log_retention; [[origins.registered]] is [[authz.origin]]; management.port is management.listen (an address:port) and management.profiling.max_seconds is max_duration; audit.store.retention_days and retention_years are audit.store.retention, verify_interval_seconds and full_verify_interval_seconds are verify_interval and full_verify_interval; [audit.syslog] host and port are url = "tls://host:port", spool_max_messages is spool_max_events, spool_max_bytes is spool_max_size, retry_max_ms is retry.max_backoff, and [audit.fhir_feed] poll_interval_ms and retry_max_ms are poll_interval and retry.max_backoff; [demographic.identifier_protection] is [privacy.identifier_protection], [cohort] is [privacy.cohort], and [access_delay] is [privacy.access_delay] with hours as max_delay; ehr_index.provider is record.subject_index ("ihe" is "pixm") and the rest of [ehr_index] is [pixm] (pixm_url is url, invocation is method, request_timeout_ms is timeout); versioning.skip_identical_resend is record.versioning.idempotent_resend; [signing] is [record.signing] (key_path is key_file, retired_key_paths is retired_key_files), [query] is [record.query] (timeout_ms is timeout), [outbox] is [record.change_feed] (retention_days is retention, prune_interval_secs is prune_interval), and [multimedia] is [record.multimedia] (endpoint is url, threshold_bytes is threshold, and allow_http is retired: an http:// url is plain HTTP); [terminology.server] is [terminology] (fail_on_error is on_error = "accept" | "refuse", request_timeout_ms is timeout). The family sections take the same types, from ferrohealth-server 0.8, ferrohealth-telemetry 0.4, ferrohealth-smart 0.15 and ferrohealth-session 0.9: [server] request_timeout_ms, header_read_timeout_ms, idle_timeout_ms, drain_delay_ms, shutdown_timeout_ms and retry_after_s are request_timeout, header_read_timeout, idle_timeout, drain_delay, shutdown_timeout and retry_after, body_limit_bytes and compression_min_bytes are body_limit and compression_min_size, and [server.tls] reload_interval_s is reload_interval; [telemetry.otlp] trust_roots_file and client_identity_file move into [telemetry.otlp.tls], and an https:// collector is dialed over it; [auth] ca_file is [auth.tls] trust_roots_file, an [[auth.issuer]] table’s clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds and timeout_ms are clock_skew, jwks_max_age, jwks_refresh_interval and timeout, its introspection cache_max_age_seconds is cache_max_age, [auth.basic] cache_ttl_seconds is cache_ttl, [auth.dpop] max_age_seconds and clock_skew_seconds are max_age and clock_skew, [auth.idp.entra] timeout_ms and max_response_bytes are timeout and max_response_size, and an oauth grant’s ca_file, client_certificate_file and client_key_file are its tls table’s trust_roots_file and client_identity_file (one PEM), with refresh_margin_ms, timeout_ms and previous_key_overlap_ms as refresh_margin, timeout and previous_key_overlap. The [auth] keys v5.6 read for every issuer (audience, algorithms, token_profile, jwks_file, mode, introspection, role_claims and the rest) live on each [[auth.issuer]] table, and scopeless_subjects entries are { issuer, sub } tables. In the viewer, cdr.request_timeout_secs is cdr.timeout, [oidc] ca_file is [oidc.tls] trust_roots_file, [oidc.tls] client_certificate and client_key are client_identity_file, and [session] idle_timeout_s, absolute_timeout_s and sign_in_timeout_s are idle_timeout, absolute_timeout and sign_in_timeout. Each of these is refused in v5.7.0 naming its replacement, in the CDR and the viewer alike (#5259).

  • Breaking: a patients token that names an actor is refused. With ferrohealth-smart 0.15, a patients issuer reads a token whose top-level act (RFC 8693 §4.1) names someone other than the patient only when its [[auth.issuer]] table sets act_represents_patient = true; without it the token is refused 401, where v5.6 admitted it as the patient’s own access. With it, FerroEHR refuses the token 403 (refused-patient-representative), because it binds no party to someone acting for a patient, and the access log records the refusal (#5259).

  • SECURITY.md, security.txt (its Contact: mailto: field), the post-market record, the CRA technical documentation and the book’s compliance pages name security@cadasto.com as the email route for a vulnerability report, in place of info@cadasto.com, which stays the single point of contact for complaints and incidents, the manufacturer’s address on the running server and the address for licensing and other business questions (#5272).

  • ferroehr and ferroehr-viewer run through the FerroHEALTH family’s run path. A configuration either binary refuses (an unknown or invalid key, a broken semantic rule, an [auth] table with no mechanism, a SCIM mapping it cannot serve, or for the viewer a TLS file that does not read or an unusable [oidc] or [session] table) now exits 78 (EX_CONFIG of sysexits(3)) where it exited 1, so an orchestrator can tell a configuration to fix from a failure to retry. Every other failure still exits 1, and the failure is written to stderr as one line with its whole cause, prefixed with the binary’s name. Upgrade: a script or probe that expects exactly 1 from config check on a refused file expects 78 (#5321).

  • Breaking: the outbound integrations take the family’s endpoint table. [terminology], [pixm], [audit.fhir_feed] and [authz.abac.remote] read where they call and how from one [<x>.endpoint] table, the same keys for every outbound integration of every FerroHEALTH product: url, timeout, connect_timeout, retry (max_attempts, initial_backoff, max_backoff), tls (trust_roots_file, client_identity_file, min_version) and credentials (HTTP Basic user with password or password_file, or the client credentials grant [...credentials.oauth]). terminology.url, timeout and [terminology.oauth] are terminology.endpoint.url, timeout and [terminology.endpoint.credentials.oauth]; pixm.url, timeout, [pixm.tls] and [pixm.oauth] are pixm.endpoint.url, timeout, [pixm.endpoint.tls] and [pixm.endpoint.credentials.oauth]; audit.fhir_feed.url, [audit.fhir_feed.retry] and [audit.fhir_feed.oauth] are audit.fhir_feed.endpoint.url, [audit.fhir_feed.endpoint.retry] and [audit.fhir_feed.endpoint.credentials.oauth]; authz.abac.remote.url, connect_timeout, timeout and [authz.abac.remote.oauth] are authz.abac.remote.endpoint.url, connect_timeout, timeout and [authz.abac.remote.endpoint.credentials.oauth]. The v5.6 keys are refused naming their new home. Every client now honours retry: the terminology, PIXm and remote policy clients try a call that got no answer or a 5xx up to max_attempts times (three by default; each tried once before), so set max_attempts = 1 to keep one attempt, and the FHIR feed’s delivery rounds take max_attempts tries each, from initial_backoff, before the round waits max_backoff; the feed still never gives a record up. The environment follows the tree (FERROEHR__TERMINOLOGY__ENDPOINT__URL, FERROEHR__AUDIT__FHIR_FEED__ENDPOINT__URL), and so do the Helm chart’s values and its client secret routing, in the chart’s 25 line (#5331).

  • The FerroEHR PostgreSQL image is built on the 2026-11-09 respin of postgres:18.6 (sha256:e3d24f69…), a same-version respin with rebuilt base packages. A full scan of the rebuilt image on amd64 and arm64 finds the same findings as the image of the previous base, each judged in security/vex/postgres-os.openvex.json, and a gate scan finds no fixable HIGH or CRITICAL finding (#5332).

  • The development and Keycloak example compose files run Keycloak 26.8.2 (quay.io/keycloak/keycloak:26.8.2@sha256:723b1d6e…). Of the 26.8.2 upgrading notes, the shipped realm meets the ones that touch it: the Admin REST API’s new audience check passes for the master realm’s admin-cli token the browser-test harness uses, and the realm uses no token exchange, identity broker, SAML client, CIBA or dynamic client registration. The ferroehr-service client-credentials token carries the same audience and scopes as under 26.8.1 (#5333).

Removed

  • [authz.professional] (claim, client_tokens, acting_for_claim) and its 403 for a token that names no natural person. The issuer’s kind and profile classify the caller, and the access record’s acting mode and professional come from that classification: a person’s own token is the person, and a delegated token is the client acting for the person it names. Upgrade step: remove the table, which refuses the start, and list a client that acts for a person in its issuer’s delegates (#5003).

  • The named refusal of authz.rbac.user_role, which v5.6 introduced, is gone under the one-release rule for removal aids: the key is now refused as an unknown key when the configuration loads (#5259).

Fixed

  • A 553|incomplete| commit no longer refuses an empty mandatory archetype slot: RM common master06 §Incomplete Content sets every occurrences lower limit to zero, a slot’s included, while too many fillers are still refused. The OPT 1.4 writer now closes a mandatory slot once the fillers placed in it carry its lower bound, as ADL 2 §Slot Filling and Redefinition allows, and still refuses to close one they leave unmet (#5042).

  • The example of a party template whose identity the template marks optional is a party of its class carrying one identity at every detail level, where the required level answered 500: RM demographic PARTY.identities admits no party without one. A party template that constrains no identity at all answers 400 Bad Request naming its class (#5057).

  • An AQL VERSION standard predicate on commit_audit/time_committed applies its operator: VERSION v[commit_audit/time_committed/value > $t] returns every version committed after $t, where it returned the trunk version current at $t whatever the operator (QUERY 1.1.0 master03 §Standard predicate). A number, boolean or NULL literal as the bound, a string bound no ISO 8601 form reads, and a time_committed path other than the whole value or /value are refused with 400 (#5062).

  • AQL refuses a number, boolean or null parameter or literal compared with a date/time path, in WHERE and in a VERSION predicate, with the same 400 a non-ISO 8601 string gets; such a bound used to become NULL and match nothing (#5064).

  • The served OpenAPI document now lists, on both template example operations, the 400 answered for a party template that constrains no PARTY_IDENTITY (#5066).

  • The compose quickstart, the hosted sandbox and the Helm chart run FerroPIX 0.0.18 (ghcr.io/ferrohealth/ferropix:0.0.18@sha256:6fb622b5b9ade34c8a4053112c8f166ff495d8b32bf99f35529d01e512b7cc42), with the FerroPIX chart 0.3.10 built for it. The 0.0.16 they pinned had an outbox sender that never slept while its outbox was empty, so an idle install kept about 1.2 CPU cores busy between FerroPIX and its PostgreSQL at about 8,000 transactions a second (FerroHEALTH/FerroPIX#490). 0.0.18 also makes FerroPIX’s [auth] scopeless_subjects and role_claims hold per issuer, which changes nothing for a FerroPIX that trusts one issuer, as every configuration FerroEHR ships does. The chart moves to 21.0.5 (#5073).

  • An OPT2 built from a template that specialises a published archetype now carries that archetype, by its full id, as parent_archetype_id, as the OPERATIONAL_TEMPLATE invariant is_specialised requires (AOM2 §3.6.6), so a reader of the OPT finds the archetype its root specialises. The ADL form of an OPT still prints no specialize section (OPT2 §Artefact Structure), so the parent travels in the canonical JSON form (#5075).

  • A refused-key-claim-missing access record names the key claim the issuer’s binding reads (its person_key) in binding_detail.claim, so an operator sees which claim the token lacked. It records the name only, never a claim value (#5139).

  • A 403 for a caller kind is recorded with its own binding code: refused-system-personal-scope for a system caller holding a user/ or patient/ scope, refused-no-agreement for an external caller whose issuer has no agreement for its accessor kind, and refused-scope-beyond-agreement for an external caller holding a scope its agreement does not read. Audit migration 0024 widens the binding constraint; the problem text of each refusal is unchanged (#5140).

  • openehr-sdt builds no /context node, and no start_time or setting leaf, in the Web Template of an OPT 1.4 whose COMPOSITION prohibits context (existence 0..0, or an EVENT_CONTEXT of occurrences 0..0), and the template example of such a template carries no context. A template that leaves context open or optional keeps the node as before (#5147).

  • openehr-sdt gives the /context node it synthesizes in a Web Template the min of the template’s existence on COMPOSITION.context, and 0 where the template leaves context unconstrained, since the RM makes context optional (0..1). A client following the Web Template no longer treats the context as mandatory when the template does not require it, and the required template example of such a template carries no context (#5154).

  • An ADL2 template that deletes an attribute (existence matches {0}), for example COMPOSITION.context, keeps the prohibition after it compiles to its operational template: openehr-adl’s create_opt now returns a RawOpt that names each removed attribute beside the OPT, and build_web_template_v2_4 takes that list. A template that deletes context gets no /context node, and a commit that carries one is refused, as it is for an OPT 1.4 template that prohibits context. An OPT2 uploaded already compiled carries no such list and reads a deleted attribute at its reference-model existence. Breaking for crate users: create_opt returns RawOpt and build_web_template_v2_4 takes the prohibited attributes (#5154).

  • A composition commit whose template_id names a stored ADL2 archetype rather than a template is refused with 422 as an unknown operational template; it failed at the insert with a server error before (#5177).

  • openehr-rm types REFERENCE_RANGE.range as DV_INTERVAL<T>: ReferenceRange is now generic over its limit type (ReferenceRange<DvQuantity> on a DV_QUANTITY, ReferenceRange<DvOrdered> on the other ordered types), its is_in_range and Range_is_simple work at that type, and the static RM model gives REFERENCE_RANGE.range and IMPORTED_VERSION.item their argument. Each model::RmAttribute and model::RmTypeRef now keeps the formal parameter it is written as (formal, Some("T") for DV_INTERVAL.lower) beside the resolved bound. The RM class table declares REFERENCE_RANGE<T> and leaves T unused; the generator restores it from a cited table and refuses a schema that no longer matches (#5196) (#5188).

  • An access record names the client the token’s issuer profile reads from client_claims (azp under the entra and keycloak profiles), the one the caller was admitted as. A token carrying a different client_id claim beside it was recorded under that claim. The [origins] register and the SCIM client lookup read the same client (#5194).

  • An Interval or DV_INTERVAL read from canonical JSON or XML that omits a boundedness or inclusion flag now takes it from its bound. An absent bound reads as unbounded and not included, and a present bound as bounded and included. A flag the wire states is kept as written. Before, an omitted flag always read as bounded and included, so an interval with only an upper came out with a bounded, included lower side that had no value. ITS-JSON requires all four flags, and BASE ties none of them to its bound, so reading the omission is our own design (#5203).

  • The viewer’s New contribution commits as the signed-in login. Each audit carries the placeholder naming that login, and the server records the committer it derives from the login. The free-text committer field is gone. For a login bound to a demographic PERSON, a stated name that differs from the PERSON’s name was refused with 403 (#5222).

  • The management surface answers a caller the login binding refuses with 403 and its binding code, such as refused-unbound, as every other route does, in place of 401: the credential authenticated, so new credentials cannot help (RFC 9110 §15.5.4). An absent or invalid credential still answers 401 with its challenge (#5250).

  • The party binding posture in ferroehr report and the boot log masks the user name and password of an issuer URL, as every other rendering of the configuration does (#5259).

  • GET /definition/query/{namespace}:: lists every stored query in the namespace again, where it answered 400 since the typed stored-query name: ITS-REST 1.1.0 definition_query_list reads its name as a pattern in which an empty part is a wildcard, so ::{name} also lists the names that start with it in any namespace. openehr-its gains rest::stored_query::QueryNamePattern, the type the generated DefinitionQueryListParams carries. The store and the single-query reads still refuse an empty name or namespace (#5268).

  • The identifier scan ([privacy.identifier_scan]) reads the name of the ELEMENT that holds a DV_IDENTIFIER when the identifier’s type is absent or names neither a person nor an item, with the same word rules. An untyped laboratory order number in the Receiver order identifier or Requester order identifier element of laboratory_test_result.v1 that passes the NHS Number check by chance is now stored under the default rule set. A person word still wins, so an untyped value in an element named NHS number or Patient identifier stays scanned, and a type that names a kind takes precedence over the element name (#5291).

  • The Helm chart’s description and README name config.deployment.profile, the key the chart refuses to render without since 5.7.0, and the README’s install command sets it; the command it showed set the old config.deployment_profile and did not render (chart 25.0.1) (#5341).

Security

Shipped with functional changes because: the advisory reaches only a deployment that trusts several issuers, which no configuration FerroEHR ships does, and its fix changes what such a configuration must state at boot, so it lands on main beside the AQL version attestations already merged for this release; a security-only patch of the last release remains open to the release cut.

  • A Basic user never carries a system/ scope, through ferrohealth-smart 0.9.1 (GHSA-jhg2-jwjp-hq74, CWE-269, CWE-863). A system/ scope granted to a Basic user gave that person every resource of a type, a backend service’s reach. Such a user now refuses the start, naming the user and the scope. Upgrade step: give each Basic user the user/ scopes a person holds, and move each machine on Basic to a client of the authorization server under the client credentials grant (#5003).

  • [auth] scopeless_subjects and role_claims hold per issuer, through ferrohealth-smart 0.8.9 (GHSA-pfgx-xq28-2629, CWE-863, CWE-639). Before it, beside several [[auth.issuer]] tables both lists applied to every trusted issuer’s tokens: a token from any of them whose sub was listed met every SMART on openEHR requirement without a scope, and any of them could assert every role the RBAC checks, ADMIN included. An entry is now written <issuer>#<value> and holds for that issuer alone; beside several issuers a bare entry, or one naming an issuer the section does not trust, refuses the boot. Upgrade step for a deployment with several issuers: write each scopeless_subjects entry as <issuer>#<sub>, and write each role_claims path as <issuer>#<path> for the issuers whose roles count (or set role_claims = [] to read none), since the default role_claims is bare and refused. A deployment with one issuer changes nothing (#5047).

  • The ferroehr-postgres image carries zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 from trixie-security, which fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). The image of 5.6.3 and earlier releases carries 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action (security/vex/postgres-os.openvex.json). The ferroehr and ferroehr-viewer images keep the distroless base’s zlib, which neither program loads (security/vex/ferroehr-os.openvex.json) (#5137).

  • The ferroehr and ferroehr-viewer images move to the gcr.io/distroless/cc-debian13:nonroot build of 2026-11-08 (sha256:cc66d222…), whose zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). Neither program loads zlib. The images of 5.6.3 and earlier releases carry 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action, as the ferroehr-postgres image of those releases is (security/vex/ferroehr-os.openvex.json) (#5152, #5329).

  • CVE-2026-112321, CVE-2026-112322 and CVE-2026-112323 (memory and CPU exhaustion in kadmind through the AUTH_GSSAPI handler) and CVE-2026-119079 (CPU exhaustion in the KDC decoding an AS-REQ or TGS-REQ) in MIT Kerberos 5 have no Debian fix for trixie and are judged not_affected in ferroehr-postgres. The image carries neither kadmind, a KDC nor libgssrpc, and the PostgreSQL server never decodes either request; the configuration FerroEHR ships enables no GSS-API authentication (security/vex/postgres-os.openvex.json) (#5344).

5.7.0-rc1 - 2026-11-10

5.7.0-rc1 is the first release candidate of 5.7.0, published for testing. 5.6.3 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.7.0 itself replaces it. Before you try it: the configuration is one table per concern with typed durations and sizes, so run ferroehr config check on your file first, which names every key a release before 5.7.0 read with its replacement, all at once; every person is bound to a demographic PERSON, so each persons issuer needs its [[auth.party]] table and each Basic user its [[authz.party_binding.basic_user]] mapping, with the steps on the book’s bootstrap and migration page; and a deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR adds its persons issuers only after FerroBRIDGE v0.0.23 and FerroFED v0.0.17 reach it. Party migrations 0027 to 0032, clinical migration 0035 and audit migrations 0023 to 0025 add the binding, SCIM, DPoP replay and refusal tables, columns and codes. Helm chart 25.0.1-rc1 carries appVersion 5.7.0-rc1. Security carries two advisories, GHSA-jhg2-jwjp-hq74 and GHSA-pfgx-xq28-2629, the zlib fixes of CVE-2026-76844, CVE-2026-85091 and CVE-2026-27171 in the three images, and the not_affected judgements of four krb5 CVEs in the PostgreSQL image.

Added

  • AQL reads the attestations of a version: v/attestations and any ATTESTATION attribute below it (is_pending, reason, proof, items, committer, time_committed) under FROM EHR e CONTAINS VERSION v CONTAINS COMPOSITION c, with predicates on attestations and EXISTS, and v/commit_audit/is_pending and v/commit_audit/reason on a commit audit that is an ATTESTATION. SELECT v/attestations returns the list, and a comparison below attestations holds when any attestation matches. The FerroEHR extension function LAST(v/attestations/<path>) reads the most recent attestation, so WHERE LAST(v/attestations/is_pending) = true lists the versions awaiting signature, including one marked pending again after it was signed (#4989).

  • The party binding’s configuration, each table with its boot refusals: [authz.party_binding] (provider as one table or one table per tenant, namespace, basic_user), [authz.import] trusted_systems, [[authz.system_client]] (the provider, facilities, templates and origins of each client-credentials client that commits) and [[authz.federation.agreement]] (what a federation issuer’s external persons or organisation may do, on which legal ground), with [authz.rbac] role_issuers and import_role. A template an agreement or a system client writes that the store does not hold refuses the start. The boot line and GET /management/info list the issuers with their kind and identifier issuer, the provider, and every agreement with its accessor and legal ground, and an agreement that admits an organisation without a named person is logged at warn (#5003).

  • [auth.idp.entra], the Microsoft Graph access of the idp-entra adapter of ferrohealth-smart, which resolves the groups of an Entra ID groups overage for an issuer whose profile sets adapter = "idp-entra" (#5003).

  • [auth.dpop] and [[auth.issuer]] sender_constraint: tokens bound to a DPoP key (RFC 9449) or a mutual-TLS client certificate (RFC 8705), checked on every request. A DPoP proof is recorded by the replica that accepted it, so the Helm chart refuses a sender-constrained issuer on more than one replica (#5003).

  • The login binding reads the local party source on every request, with no cache: the new party migration 0027_binding_side_table adds the indexed binding_identifier, binding_membership and binding_representation relations, written in each party commit’s transaction from the PERSONs of the mapped templates that perform a professional ROLE or represent another person, their GROUP memberships and their representation relationships, and rebuilt from the party versions alone. A login matching two parties, an inactive, abandoned or deleted party, a token without the key claim, a party without the needed marker and unreadable groups are refused with 403 and their binding code; a party store or identity provider that cannot answer is answered 503 (#5004).

  • ferroehr definitions load <path>... stores OPT 1.4 (.opt) and ADL 2 (.adls) definitions before serve, through the validation and versioning of the REST upload. A file already stored with the same text is left alone and other text under a stored id is refused, so a deployment runs it at every start to give its party mapping the templates it needs at boot. deploy/definitions/professional/ holds a professional PERSON and ROLE pair adapted from openEHR CKM archetypes for the development lanes and the quickstart (#5006).

  • The compose quickstart binds its user to a professional PERSON: the ferroehr-definitions one-shot loads the PERSON and ROLE templates the party mapping names before the server starts, and ferroehr-provision creates the user through /scim/v2/Users with a token of the development issuer, then reports healthy, so docker compose up --wait returns once the login is bound. The Helm chart (23.3.0) takes definitions.files, which an init container stores with ferroehr definitions load before the server container starts. The conformance, browser and Kubernetes lanes load the same templates and provision every login they use through SCIM, with the committed test client of docs/conformance/party/scim/ (#5006).

  • ferroehr party-binding adopt <PARTY_ID>... gives each prepared PERSON of an existing deployment a professional ROLE of the mapping’s role template, with the PERSON as performer, and a new PERSON version listing it in roles, all in one demographic CONTRIBUTION committed by the first administrator named with --admin or --admin-key, and writes one access record per adopted PERSON through the configured audit trail. It refuses the whole adoption, naming them, while any listed PERSON’s key was introduced by a committer that is no [[scim.client]], named administrator or --migration, so a key planted before the upgrade is never adopted. A PERSON that already performs the ROLE is reported and left alone. The report is JSON on stdout (#5009).

  • ferroehr party-binding check is the read-only pre-flight of a migration to party binding, and the one report of the provisional persons just-in-time creation left incomplete. It reads the party versions alone, so it runs before any boot fill, and writes nothing. Per key issuer it reports the PERSONs that carry a binding key, every selector value two PERSONs carry, the inactive and abandoned ones, the provisional ones older than provisional_report_days, the key-bearing ones that perform no professional ROLE, and every mapped Basic user with the PERSON it binds. For each key it names the committer of the version that introduced it, and fails unless that committer is a [[scim.client]], an administrator named with --admin or --admin-key, a migration declared with --migration, or just-in-time creation on an issuer with jit = true. --expected <file> lists every login of the identity provider with no PERSON. The report is JSON on stdout, and the command exits non-zero on a duplicate, a missing person or a key that fails the provenance rule (#5009).

  • At boot, FerroEHR logs a warn naming every Basic user mapped in [[authz.party_binding.basic_user]] while [authz.assurance] minimum is unset: a Basic credential carries no assurance level, so such a person is never refused on patient data (#5009).

  • The one PERSON mapping SCIM provisioning, just-in-time creation and the login binding share: [[authz.party_binding.person.template]] with the binding key, userName and externalId selectors of each template, [[authz.party_binding.person.attribute]] with the SCIM attribute, the OpenID Connect claim and the path of each mapped value, and [authz.party_binding.professional_role], [authz.party_binding.group], [authz.party_binding.membership] and [authz.party_binding.representation]. Boot refuses a mapping that breaks one of its rules, naming the key, including a template that is absent, rooted at another class or lacks a mapped path, and deleting a template the mapping names is refused with 409 (#5013).

  • The new party migration 0028_binding_selectors holds each binding key, userName (compared after RFC 8265 width mapping, lower case and NFC), externalId and group displayName to one live party, so a commit giving one a second party is refused with 409. The first boot fills the side table under the mapping’s digest, a replica with another mapping refuses to start, and ferroehr party-binding refill rebuilds it after a mapping change, refusing while a replica runs and naming the parties when a value is carried twice (#5013).

  • POST /admin/party/binding-key re-keys a PERSON’s binding key: it refuses the administrator’s own login and a key already bound, retires the old key in the same commit and writes one access record carrying both (#5013).

  • [scim] mounts the SCIM 2.0 service provider at /scim/v2/ (off by default): [[scim.client]] declares each provisioning client by issuer, client_id, scope and binding_issuer, and public_base_url names the origin its documents hang off. /scim/v2/ServiceProviderConfig answers without a token; /ResourceTypes and /Schemas answer a declared client acting for nobody that holds its scope, within [server.caller_rate]. Every refusal is an RFC 7644 Error message. The start is refused for a client of a non-machines issuer, one declared twice, two clients sharing a binding_issuer, or a binding_issuer no [[auth.party]] table carries (#5014).

  • The SCIM endpoint answers the Microsoft Entra ID provisioning client as Microsoft’s “Develop a SCIM endpoint” guide documents it: a value-filtered attribute such as emails[type eq "work"].value serves its type and declares primary read-only, so the primary Entra sends is ignored (RFC 7644 §3.3); a PATCH of userName renames the User; and the Test Connection probes are answered with an empty ListResponse and 404. ferroehr config check refuses a [scim] the person mapping cannot serve, as serve does. The Helm chart exposes config.scim and routes /scim/v2 on every Ingress host when it is enabled, and the book walks an operator through provisioning from Entra ID, with the Workload Identity Federation checks the authorization server makes (#5014).

  • /scim/v2/Users/{id} takes PUT, PATCH and DELETE: a write commits the next version of the professional’s PERSON, clearing a mapped attribute a PUT leaves out, keeping the binding key and externalId, and moving the lifecycle state by active (complete or incomplete to inactive or abandoned and back, an abandoned User completed in two versions, an attribute change to an inactive User refused until it is reactivated). DELETE logically deletes the PERSON and its professional ROLE in one CONTRIBUTION and leaves the binding key as a tombstone. An If-Match naming another version is 412, and ServiceProviderConfig declares PATCH. [scim] max_deactivations and deactivation_window_minutes bound the deactivations and deletions of one client, counted in the database so replicas share the bound, with 429 past it (#5014).

  • /scim/v2/Users serves a client’s Users: POST creates the professional’s PERSON and its professional ROLE in one demographic CONTRIBUTION, complete when the template holds in full and incomplete otherwise, GET /Users/{id} answers with the version’s entity tag and 304 to a matching If-None-Match, and GET /Users answers a ListResponse filtered by eq and and over userName, externalId and id, with any other filter 400 invalidFilter. A client sees only the Users whose binding key its binding_issuer issues, every SCIM request is an access record of the client naming the PERSON, and ServiceProviderConfig declares entity tags. The start is refused when [scim] is on and the party mapping has no default person template with a user_name selector or no professional role (#5014).

  • /scim/v2/Groups serves SCIM Groups as demographic GROUPs of [authz.party_binding.group], each member holding an inline PARTY_RELATIONSHIP of [authz.party_binding.membership] from its PERSON to the GROUP, with no ROLE minted: create, read, eq/and filters on displayName, externalId and id, excludedAttributes=members, PATCH (answered 204), PUT and DELETE, with the entity tags of Users. A Group belongs to the client that created it, its members are that client’s Users, its displayName is unique, and its externalId sits under the client’s group_identifier_issuer. A remove of members with a value list removes only those members, an Entra compatibility deviation from RFC 7644 §3.5.2.2. Deleting a User removes its memberships in the same CONTRIBUTION. [scim] max_group_members (default 5000) caps a Group (#5015).

  • Just-in-time creation of a provisional PERSON at a first login, off by default and enabled per persons issuer in a new [[authz.party_binding.issuer]] table (jit, guests_jit, user_name_claim, idp_claim, admit_idp, admit_groups, admit_roles, admit_assurance). A login of a person or a delegated caller that matches no party and meets every declared admission condition commits one 553|incomplete| PERSON holding the binding key, the userName and the name claims of the access token, with its professional ROLE, as a creation by the person, and is served bound to it; any other mapped attribute is written only where its mapping sets jit = true, and externalId never. Concurrent first requests commit one PERSON, a deprovisioned or inactive login is never recreated, a taken userName refuses the login, and a SCIM client of the key’s issuer completes the PERSON. Boot refuses jit without an admission condition or a mapping that can write it. ferroehr party-binding check lists the provisional persons older than [authz.party_binding] provisional_report_days (default 30) (#5016).

  • A party of a mapped template is validated against that template on every demographic write: a PERSON or GROUP whose archetype_details.template_id names a template of the shared party mapping, and a PARTY_RELATIONSHIP of its membership archetype. A 532|complete| version must meet every constraint, and a 553|incomplete| one may miss mandatory nodes but nothing it carries may be wrong (RM common master06 §Incomplete Content); the refusal is a 422 naming each failing path. Every other party keeps the reference-model checks alone, so parties created through the Demographic API under other templates are accepted as before (#5042).

  • The FHIR AuditEvent, the DICOM audit message and the ITI-20 syslog feed carry the binding fields of the access record: the binding outcome, the party and its groups, the provider acted for, the facilities of a system commit, the representation with its relationship, the actor chain, the external accessor’s asserted fields and the agreement with its legal ground, consent and access class, and a chained request’s correlation id. ehds-logging 0.5.1 writes them as the ehds-accessor and ehds-relayed entities, in the DICOM message as participant objects of type Other. The FHIR record also writes the external and system acting modes. FerroEHR moves to ehds-logging and ihe-atna 0.5.1 and ihe-iti 0.0.77 (#5043).

  • A bound representative reaches only the EHR whose EHR_STATUS names the subject it represents. Every composition, EHR_STATUS, directory, contribution and query request is held to that EHR whatever the token’s scopes cover, and a query must name it with ehr_id. A representative whose token names a subject its party has no live representation relationship to, or whose issuer no [[auth.party]] table binds, is refused with 403 and refused-representation-unrecorded (#5086).

  • The Helm chart (23.1.0) accepts bridge.cdr.oauth.tokenEndpoint for an authorization server whose metadata FerroBRIDGE cannot read at bridge.cdr.oauth.issuer. FerroBRIDGE then requests its client credentials token from that endpoint, and issuer still names the machines issuer the CDR trusts (#5093).

  • A MyHealth@EU contact point’s caller (a contact_point = true federation issuer) reaches patient data only for the patient of its treatment relationship confirmation. The token’s ehdsi.patient.id (system|value) is resolved to its EHR through the EHR index, and every patient-data request must name that EHR. A token without it is refused with 403 and ERROR_SEC_DATA_INTEGRITY_NOT_ENSURED, recorded as refused-trc-missing. Another patient’s EHR, an unknown EHR and a request naming no EHR all answer one 404 of the new problem type trc-patient-mismatch, carrying the agreement service’s *_NOT_FOUND code and recorded as a refused-trc-patient-mismatch security event (eHDSI XCA Profile 8.0.1 §4.2.1, §4.2.2). A contact point’s person agreement now names its ehdsi_service; upgrade step: set it, and list the issuer in [smart] patient_issuers (#5102).

  • An external person under an agreement with discard discards a dispensation of a listed template, as a logical delete, within discard_window_hours and from the same point of care: the dispensation’s health_care_facility carries the token’s provider identifier and its eHDSI organization-id. A token carrying permissions records and discards a dispensation only with PPD-046 (eHDSI SAML Profile 10.0.0 §2.5) (#5102).

  • A read under a federation agreement with a consent condition is served when the token carries the consent decision reference for its patient (Wabvpz Art. 1(j) and Art. 15a(1)), in the claim or authorization_details member the issuer’s [[auth.federation]] table maps to consent.reference, and refused with 403 without it. The reference is one string, the token’s patient/ grant holds it to the launch-context patient, and the access record carries it with the other external fields. ferrohealth-smart moves to 0.9.3, which reads the field (#5106).

  • Every access record of an authenticated request carries the outcome of the party binding (binding: bound, provisional, external, organisation, system, self or a refused-* code) with the fields it needs: the bound party, its groups, the provider, the guest mark, a representative’s representation, a delegated token’s actor chain, a verified external accessor’s fields and agreement, and the facilities a system caller’s compositions named. acting_mode gains delegated, external and system. The audit migration 0023_binding_outcome.sql adds the columns and restores a per-row digest version: records written before stay version 1 and keep verifying, new records are version 2, and a version-1 record after a version-2 one is refused (#5118).

  • Every refused request is an access record through the gated path, and a refusal of the login binding or of a caller kind carries its refused-* binding code and the party it names. Refusals of tokens that failed verification are budgeted by source network and those of verified callers by issuer and subject, so neither crowds out the other; past a caller’s limit its refusals become one summary record with the count and the EHR ids they named, and past a budget one summary record per issuer and outcome is written. Security-event refusals are recorded for every answered request, and a refusal record the system log cannot take is counted in atna_audit_refusals_unwritten. The limits, windows and budgets are kept per replica (#5119).

  • A chained organisation request under a federation agreement is served only when its token carries the requester’s correlation id in the claim or attested authorization_details member the issuer’s [[auth.federation]] table maps to correlation_id. Without it the request is refused with 403 and recorded under the binding code refused-correlation-missing (audit migration 0025); a request header never stands in for it. The organisation access record’s agreement carries the id and its source beside the legal ground and the access class. A chained agreement whose issuer maps no correlation_id, or maps it from a constant, refuses the start (#5120).

  • On the MyHealth@EU contact-point route, the committer of an external person names the professional by the HP Identity Assertion’s NameID: DV_IDENTIFIER{issuer = the assertion's issuer, id = the NameID, type = "hp_identifier:<NameID format>"}, beside the voucher’s oauth2 identifier (eHDSI SAML Profile 10.0.0 §2.1), and a dispensation’s discard is a logical delete under the same committer. A contact point’s COMPOSITION must carry the HCPO’s organization-id in EVENT_CONTEXT.health_care_facility.identifiers, under the issuer urn:oasis:names:tc:xspa:1.0:subject:organization-id, beside the provider’s identifier, or it is refused with 422. ferrohealth-smart moves to 0.9.4, which reads ehdsi.assertion.name_id_format and refuses a contact-point token without one of the three formats as ERROR_HPI_INSUFFICIENT_INFORMATION; upgrade step: a contact point’s authorization server writes the NameID format into its tokens (#5144).

  • A federation caller above its agreement’s caller_rate, a MyHealth@EU contact point’s included, is answered 429 with Retry-After right after authentication and before any patient lookup, keyed on the token’s subject per replica, and every such refusal is counted into the caller’s summary record (#5149).

  • A deleted SCIM User’s contacts are erased from every version of its PERSON, keeping the name, the binding key and the identifiers for the record-keeping duties that need them. [scim] contact_erasure_minutes (default 0, at most 43200) sets when: 0 erases them in the DELETE itself, a period leaves them to a sweep that runs every minute. Each erased version’s node rows are derived again from the new body, its server signature is removed, and the erasure is an access record naming party:<id>#contacts. Party migration 0030 adds the scim_contact_erasure table (#5178).

  • openehr-rm types generic RM type names and paths: typing::RmType parses and prints names such as DV_INTERVAL<DV_QUANTITY>, typing::attribute_type gives an attribute of an instantiated class with the actual argument in place of the formal parameter, and typing::path_type types every step of an RmPath with its container. Each model::RmAttribute and model::RmTypeRef keeps the formal parameter it is written as (formal) beside the resolved bound. paths::write::item_at_path_mut and paths::write::set_at_path write at a path into a canonical instance, positional predicates included, creating the members the path names and changing nothing when a write is refused (#5187).

  • openehr-sdt: flat::navigate::Navigator gives every instance of a Web Template path in a composition, with each step’s occurrence, container index, positional path and JSON pointer, reaching a renamed instance where the template’s name is redundant and an OPT2 archetype root through the complete identifier the template recorded. flat::navigate::relative compares two paths step by step. flat::attribute_keys gives the FLAT key of an RM attribute of a structural class (the master05-rm_mapping.adoc tables, checked against them). A WebTemplateNode carries facts (the embedded template id of an OPT 1.4 archetype root and the terms its archetype defines) and every coded option its terminology, so no consumer walks the operational template (#5189).

  • openehr-base gives identifier and time parts as written: ObjectVersionId::object_id_str and ObjectVersionId::versioned_object_id keep the case of the object_id part, Iso8601Time::fraction_digits and Iso8601DateTime::fraction_digits return the digits of the fractional second as text, and Iso8601Duration::parts returns each component’s designator and number verbatim (DurationDesignator, DurationPart), with Iso8601Duration::fraction_digits for the seconds (#5190).

  • openehr-its reads the ITS-REST ETag as a typed entity tag: rest::etag::EntityTag parses and writes the RFC 9110 form, weak or strong, with the weak and strong comparisons. It types the tag as an ObjectVersionId or a HierObjectId. Every generated response-headers struct that carries an ETag gains entity_tag() beside the raw etag field (#5191).

  • openehr-term names the null_flavours and setting groups: codes::NullFlavour and codes::Setting give each concept’s code and English rubric and read a code back, held equal to the bundled terminology by tests. Every typed group’s ALL is now a slice (#5192).

  • openehr-am: HridRef reads an archetype reference with the version precision it was written at (VersionPrecision: major, minor or full, BASE Artefact Identification §Referencing), gives the version as written and matches an ArchetypeHrid by the specification’s rule, the namespace never a wildcard. ArchetypeHrid::local_semantic_id gives the namespace-free interface form a Web Template nodeId carries, which the v2_4 Web Template builder now uses (#5193).

  • openehr-its: rest::stored_query types a stored query’s qualified name (QualifiedQueryName, [{namespace}::]{query-name}, with check_registrable refusing the reserved aql), its version (QueryVersion, major.minor.patch in numeric order) and the version pattern a request addresses (VersionPattern, exact, {major} or {major}.{minor}, selecting the highest match), per ITS-REST 1.1.0. The generated stored-query operations and the StoredQuery and RESULT_SET documents carry them, so a name or version outside the grammar is a 400 where the request is decoded (#5207).

  • openehr_rm::Generation::earliest_for_composition answers the oldest RM generation whose canonical-JSON decode accepts a composition, so a builder can name ARCHETYPED.rm_version without naming a generation module; a value no generation decodes is a typed NoGenerationDecodes error carrying every generation’s refusal (#5242).

Changed

  • The published server, viewer and PostgreSQL images take their org.opencontainers.image.base.name and base.digest labels and annotations from the runtime stage’s digest-pinned FROM line at build time, so a base-image bump changes that one line. An image built directly from a Dockerfile no longer carries these two labels (#4953).

  • [auth] takes the ferrohealth-smart 0.9 shape. Every trusted issuer is an [[auth.issuer]] table with a required kind (persons, patients, machines or federation) and a profile (rfc9068 by default, or entra, keycloak, okta, nl), and holds every setting of its tokens itself. A configuration in the old shape refuses the start, naming what replaces it. Upgrade steps: write a flat issuer = "…" as an [[auth.issuer]] table; move audience, algorithms, clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds, jwks_file, timeout_ms, mode, role_claims and [auth.introspection] onto each issuer table ([auth.issuer.introspection]); give each issuer its kind, with patients and professionals at separate issuers (a realm or tenant each), since one issuer is one kind; replace token_profile = "jwt" with token_types = ["JWT"] and the required_claims the issuer’s tokens carry, and drop token_profile = "rfc9068"; and write each scopeless_subjects entry as { issuer = "…", sub = "…" } (#5003).

  • A persons token is admitted only through a client its issuer lists in first_party_clients or delegates. Upgrade step: list the deployment’s own clients, the viewer’s among them, in first_party_clients, and a third-party client that acts for a person in delegates. The other 0.9 profile keys (person_key, tenant_claim, client_claims, app_only, guest, impersonation_marker, expect, user_claims, person_equality, unstable_claims, max_actor_chain, representation, scope_claim, app_only_scope_claim, adapter, sender_constraint and the onward_* keys) are documented in the configuration reference (#5003).

  • An array of tables, such as [[auth.issuer]], is set from the environment by element index (FERROEHR__AUTH__ISSUER__0__ISSUER), through ferrohealth-config 0.2.10 (#5003).

  • [smart] issuer_metadata_file is the metadata of the [auth] issuer it names, wherever that issuer stands in the list, and a file naming no [auth] issuer refuses the start. The SMART discovery document describes the first introspected issuer, else that named issuer, else the first issuer whose metadata is read from it, else the first issuer, so a Basic users’ issuer listed first no longer takes the discovery document (#5003).

  • The family crates move together: ferrohealth-smart 0.9.1, ferrohealth-authz 0.3.0, ferrohealth-session 0.3.1, ihe-atna and ehds-logging 0.4.0, ihe-iti 0.0.76, ferrohealth-config 0.2.10, ferrohealth-server 0.5.1 and ferrohealth-ui 0.6.9. ihe-atna and ehds-logging now escape every control and bidirectional formatting character of a value they write as its code point. With ferrohealth-ui 0.6.9 the viewer’s session watch no longer throws when the party relationship form, which holds a control named target, is submitted (#5003).

  • The terminology client moves to ferrobridge-term 0.1.143 (FerroBRIDGE v0.0.22), which takes the ferrohealth-smart 0.9 token source (#5003).

  • Every person is bound to a demographic party, with no switch. A persons issuer needs its [[auth.party]] table, two issuers of one node may not share an audience, and [authz.party_binding] provider (the healthcare provider the deployment acts for, EHDS Art. 13(4) and Annex II 3.2(a)) is required once [auth] admits a credential; each refuses the start naming what to set. A Basic user is a person: it maps to the binding key of its PERSON in [[authz.party_binding.basic_user]] (user, key_issuer, key), and a Basic user without a mapping refuses the start, naming the user. Upgrade steps: add an [[auth.party]] table to each persons issuer, set provider, and map each Basic user under the identifier_issuer of an [[auth.party]] table, which needs a persons issuer; a deployment of Basic users alone declares one whose jwks_file is a key set that signs nothing, as the quickstart and the sandbox do (#5003).

  • A token grants [authz.rbac] admin_role, emergency_access_role, import_role and identifier_reader_role only when its issuer is listed in the new [authz.rbac] role_issuers, each a persons issuer; any other kind refuses the start. Upgrade step: list the issuers whose role claims grant these roles; a Basic user keeps the roles its own table grants (#5003).

  • The compose quickstart, the dev and conformance stacks, the hosted sandbox and the Helm chart’s Basic overlays bind their Basic users under a persons issuer that signs nothing: its key set, which every FerroEHR image carries at /opt/sandbox-posture/basic-users-jwks.json, is the public half of a key pair whose private key was discarded (#5003).

  • The quickstart has one copy. docker-compose.yml on main pulls the FerroEHR images at :main and carries main’s configuration, which CI checks with the server image the change builds; the copy attached to each release pulls that release’s images, written by the release lane, which refuses a copy that still names :main. Download the release asset (releases/latest/download/docker-compose.yml) to run a release (#5003).

  • The Helm chart moves to 22.0.0. config.auth.issuer is a list of issuer tables, each with its kind, profile and audience, and values.yaml exposes config.auth.party, config.authz.party_binding, config.authz.rbac.role_issuers, config.authz.system_client and config.authz.federation.agreement. The chart refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling. Upgrade steps: rewrite config.auth as the [auth] upgrade above describes, set config.authz.party_binding.provider, and list the issuers whose roles grant the administrator role in config.authz.rbac.role_issuers (#5003).

  • A caller bound to a demographic PERSON commits under the committer the server writes: a PARTY_IDENTIFIED with the PERSON’s name, the login identifier (the issuer profile’s person_key value under the token issuer, type oauth2, or the Basic user under ferroehr, type basic), the provider identifier (type healthcare-provider) and an external_ref to the PERSON, on every commit path, attestations and CONTRIBUTIONs included. Upgrade: stop sending a committer, or send the placeholder {"_type": "PARTY_IDENTIFIED", "identifiers": [{"_type": "DV_IDENTIFIER", "issuer": "<iss>", "id": "<sub>"}]}, which must name the authenticated caller; a stated committer with another name, identifier or external_ref, and a stated PARTY_RELATED, is refused with 403 (#5005).

  • A patient’s committer is PARTY_SELF with no external_ref, and a representative commits as a PARTY_RELATED naming their own PERSON and the coded relationship. A patient-context token commits no demographic version. Upgrade: read the person’s own entries by the committer’s PARTY_SELF type, not by its former external_ref (#5005).

  • Importing an EHR Extract needs the role [authz.rbac] import_role names, or a system/ caller, and a wrapped committer naming a party or a login of this deployment is refused unless its version was created by a system [authz.import] trusted_systems lists. Upgrade: set import_role and give it to the accounts that import (#5005).

  • A person of a persons issuer whose login no demographic party carries is refused with 403 and refused-unbound on every request, with no fallback to a login-only committer. The problem names the identifier issuer to provision the person under, a token without the key claim is refused with a problem naming that claim, and each refusal counts into the new binding_refusals metric by issuer and outcome. The Keycloak quickstart overlay provisions its four demo users through SCIM with the ferroehr-provision-keycloak one-shot. Upgrade: provision every professional (through SCIM or the demographic API) before you add a persons issuer to the upgraded deployment (#5006).

  • A deployment whose FerroBRIDGE facade or FerroFED gateway writes to FerroEHR adds its persons issuers only after the bridge and the gateway reach it under the caller: FerroBRIDGE v0.0.23 and FerroFED v0.0.17. Until then the committer rule refuses their writes. FerroFED v0.0.17 follows this release, so a deployment with a FerroFED gateway adds its persons issuers only once that release is out. The book’s new “Party binding: bootstrap and migration” page carries the bootstrap path, the four migration steps and these prerequisites, and “Breaking changes of party binding” lists every incompatible change with its upgrade step (#5009).

  • The hosted sandbox follows the bootstrap path of party binding: its configuration declares the sandbox-scim client ([[scim.client]], with the development issuer trusted as a machines issuer for it), the box’s new ferroehr-provision service provisions the demo PERSON through /scim/v2/ after every reset, and scripts/sandbox/reseed.sh waits for the demo login to bind before it seeds, then checks the committer of the marker’s EHR_STATUS version references a PERSON. The deploy probe’s new bootstrap family (P-BOOTSTRAP-01) shows the compose quickstart does the same from an empty database: the quickstart user is a User of the SCIM client, and a commit of the Basic user names that PERSON (#5009).

  • With a party mapping configured, the parties the login binding relies on are written through the demographic API by an administrator (authz.rbac.admin_role) only: a PERSON of a mapped template carrying a binding key, userName or externalId, its professional ROLE, a representation relationship, and a GROUP of the group template. Any other caller’s write is refused with 403, a version that changes a present binding key is refused with 403, and an administrator’s version that leaves a protected identifier out is refused with 422. Upgrade: write these parties as an administrator, and change a key through POST /admin/party/binding-key (#5013).

  • FerroEHR records every DPoP proof it accepts in its PostgreSQL database until the proof’s window ends, so a proof is accepted once across all replicas, and a proof replayed to another replica is refused there (RFC 9449 §11.1). If the database cannot record a proof, the request is refused with 503. The Helm chart (22.1.0) no longer refuses an issuer with sender_constraint: required beside replicaCount above 1 or autoscaling (#5072).

  • An AQL query scoped to one EHR now takes the streaming plan shape whatever its ORDER BY, DISTINCT or aggregates, whether the EHR comes from the ehr_id request parameter or from an ehr_id/value equality in the query. On a store of 100,000 compositions such a query used to spend 16 to 64 ms in PostgreSQL’s planner for an execution of about 1 ms, on every new connection and every new statement text; it now plans in 1 to 3 ms (#5076).

  • The Helm chart (23.0.0) connects FerroBRIDGE to the CDR as a machine. FerroBRIDGE uses an OAuth 2.0 client credentials client of a machines issuer that the CDR trusts, set in bridge.cdr.oauth (issuer, clientId, and optionally scope, audience and resource). The client secret comes from the client_secret key of bridge.cdr.existingSecret. The chart no longer accepts a Basic user for this connection: from ferrohealth-smart 0.9 a Basic credential belongs to a person and holds no system/ scope. The chart refuses to render when bridge.cdr.oauth.issuer is not listed in config.auth.issuer with kind: machines. To upgrade, add that issuer, create the client, and replace the user and password keys of the Secret with client_secret (#5085).

  • A verified external person’s commit records the professional as its committer: a PARTY_IDENTIFIED with the asserted name, the hp_identifier under the issuer "<country_code>:<issuing_authority_name>" and the token’s iss and sub, with no external_ref. A country code that is no upper-case ISO 3166-1 alpha-2 code, or an issuing authority holding a :, control or bidi character, is refused with 403 naming the field. The COMPOSITION must name the token’s healthcare provider (provider.healthcare_provider_identifier under provider.issuing_authority_name) as its health_care_facility, or it is refused with 422. Upgrade step: map provider.issuing_authority_name in each [[auth.federation]] table whose agreements carry writes, with a constant source where the token does not carry it (#5097).

  • A system caller (an app-only token by its issuer profile’s app_only markers) holds system/ scopes only: a token carrying a user/ or patient/ scope is refused with 403. It commits into an EHR only as a client an [[authz.system_client]] table declares, under the committer PARTY_IDENTIFIED{name = <client>, identifiers = [{issuer = <iss>, id = <client>, type = "client"}]}, and each COMPOSITION it commits must be of a template its writes lists (403), name a facility its facilities lists in context.health_care_facility, or in feeder_audit.originating_system_audit.location when it has no context (422), and be either a device act from a device client or a relayed person act naming the professional by an identifier from a relay client. Upgrade: declare every client that commits in [[authz.system_client]] with its writes, facilities and origins, and give its compositions the facility identifier (#5100).

  • An attestation is accepted only from a caller bound to a professional PERSON; one from a patient, a representative, a system caller or an unbound caller is refused with 403. Upgrade: attest under the professional’s own token (#5100).

  • A guest of a persons issuer, by its profile’s guest markers, is refused with 403 (refused-guest) unless the issuer’s [[auth.party]] table sets guests = true. Upgrade: set guests = true on the issuers whose guests act under your responsibility (#5100).

  • A verified external person or organisation of a federation issuer is admitted only under the [[authz.federation.agreement]] of its issuer and accessor, holding no scope beyond the agreement’s reads, and is an access-service caller by its kind: an opted-out EHR answers 404 to it, and only an external person under an agreement with opt_out_override reaches one under an emergency purpose. It commits only a COMPOSITION of a template the agreement’s writes lists (an organisation nothing), and a read under a consent-conditioned agreement, and every request under an NL agreement, is refused until the consent decision reference and the Wabvpz Art. 15f roles can be read. Upgrade: declare an agreement for every federation issuer and accessor, list the issuer in [smart] patient_issuers, and give its tokens only the patient/ scopes the agreement reads (#5101).

  • Under an NL [[authz.federation.agreement]], Wabvpz Art. 15f now decides admission in place of the refusal of every request. A caller whose nl.role is the UZI role 01.008 (bedrijfsarts) or 01.048 (verzekeringsarts) is refused with 403 as refused-wabvpz-role, and a caller whose nl.organisation_type is missing or is not a care provider type of Nictiz’s OrganisatieTypeCodelijst is refused with 403 as refused-wabvpz-organisation-type, so an insurer is refused. Every NL agreement declares purpose = "care"; one without a purpose, or for a keuring, refuses the start. The code lists are vendored by scripts/vendor/nl-role-codes.sh into docs/specs/nl-role-codes/. Upgrade: add purpose = "care" to every NL agreement and map nl.organisation_type (and nl.role for persons) in its issuer’s [[auth.federation]] table (#5107).

  • A delegated token’s access record now has acting_mode = "delegated" where it had client, and a system caller’s system. The FHIR rendering still names a delegated access as a client one. Upgrade: a review query that selects delegated access by acting_mode = 'client' also selects 'delegated' (#5118).

  • FerroEHR builds on ferrohealth-server 0.6.1, whose public caller-rate take holds a federation caller to its agreement’s caller_rate. Its 0.6.0 break, the revocation feature’s move to der 0.8, touches no feature FerroEHR enables (#5149).

  • The source tree no longer carries the UZI role code CodeSystem (uzi-rolcode) of Nictiz’s zib 2020 package: its codes and titles reproduce the UZI-register CPS, which CIBG does not license for reproduction. scripts/vendor/nl-role-codes.sh writes it into the git-ignored .vendor-cache/nl-role-codes/, as it does the CPS, and --test-inputs fetches only what the Wabvpz tests read. Nictiz’s care provider types stay under docs/specs/nl-role-codes/. Run scripts/vendor/nl-role-codes.sh --test-inputs once before running the ferroehr-rest tests (#5175).

  • [auth] takes three boot refusals from ferrohealth-smart: a Basic user’s hash below the Argon2id floor (m=19456, t=2, p=1), an issuer that carries userinfo, and private or symmetric key material in a jwks_file. They are refused as before, with the crate’s messages. A jwks_file that is no JWK Set now refuses the start where the crate reads the key set (#5209).

  • The family crates move together to one ferrohealth-smart: 0.13.1, with ferrohealth-authz 0.7.0, ferrohealth-session 0.7.0, ihe-atna and ehds-logging 0.8.0, ihe-iti 0.0.83, ferrohealth-server 0.6.4 and ferrobridge-term 0.1.148. The 0.13 line refuses a JSON document it cannot read without quoting it (GHSA-mhv7-3975-6xq3), so a jwks_file or [smart] issuer_metadata_file that does not parse is refused at boot with the crate’s message, which names the failure class and position; 0.13.1 keeps a refused key set, introspection or metadata body out of the error (GHSA-5qm7-v4fw-p2v9). The OAuth client credentials tables ([authz.abac.remote.oauth], [ehr_index.oauth], [terminology.oauth] and [audit.fhir_feed.oauth]) read as before (#5209).

  • The API’s response security headers come from the family server stack (ferrohealth-server 0.6.2) with the same six headers and values. They now also reach the management listener on its own port, and Strict-Transport-Security: max-age=31536000 is sent when FerroEHR or the viewer serves TLS itself ([server.tls], RFC 6797 §7.2). The viewer gains Cross-Origin-Resource-Policy: same-site (#5210).

  • The server and the viewer read the parts of an OBJECT_VERSION_ID through openehr-base and its BASE grammar, never by cutting at ::. A value that is no OBJECT_VERSION_ID now names no version: the AQL branch refusal, the version item tags and the viewer’s version numbers leave it alone, a reference to a container is read as the whole value (so uuid::extension no longer names the container uuid), and the viewer’s deletes refuse it before calling the CDR. Every part keeps its case (#5211).

  • ETag and If-Match values are read through openehr-its’s RFC 9110 entity tag on the server and in the viewer. The server accepts the same forms as before. The viewer reads its CDR’s ETag strictly: a value that is no entity tag names no version (#5212).

  • The viewer’s audit screen reads the ITI-81 searchset and each AuditEvent as FHIR R4 resources, and a refusal’s OperationOutcome too. A searchset that is no FHIR R4 Bundle of AuditEvents now shows as a failed read, never as a page of empty rows (#5214).

  • The EHR index’s boot check reads the PIX Manager’s CapabilityStatement through ihe-iti’s PIXm client, with its transport and its client-credentials token, so a Manager behind IUA answers the probe as it answers ITI-83. A failed probe is reported as the client’s refusal: no answer, the status it answered, or no FHIR R4 CapabilityStatement (#5216).

  • [audit.syslog] reads its TLS material from [audit.syslog.tls], the outbound TLS table of every FerroHEALTH product (trust_roots_file, client_identity_file or client_identity, min_version), and hands ihe-atna the rustls configuration ferrohealth-server builds from it. The keys tls_ca_file, tls_identity_cert_file and tls_identity_key_file are retired and refused at boot by name: name the roots file as trust_roots_file, and concatenate the client certificate chain and its private key into the one PEM client_identity_file names. The client speaks TLS 1.3 unless min_version = "1.2" lowers it (#5217).

  • A SCIM Group membership is now held by the member: an inline PARTY_RELATIONSHIP from the member’s PERSON to the GROUP, stored in the PERSON’s version (RM demographic §Party Relationships). A membership change versions only that PERSON, so its cost no longer grows with the Group’s size. The GROUP is versioned only by a rename, an externalId change or its deletion. The Group’s members is read from the membership index (#5220).

  • A Group’s ETag is now its GROUP version plus a membership revision, so it still changes with every membership change. The member User’s ETag now changes too, while its body stays the same (#5220).

  • Deactivating a User (active = false) drops all of its memberships in the same version, because an inactive PERSON cannot be versioned again. A reactivated User belongs to no Group until its client adds it again (#5220).

  • Microsoft Entra ID’s disable-then-remove order now succeeds, since the later remove finds nothing to remove. An add naming an inactive User is 400 invalidValue (“reactivate the User first”) (#5220).

  • A refused SCIM write now names the rule it breaks in detail (#5220).

  • scripts/perf/scim-groups.sh keeps the database log, df -h and free -m of a failed run (#5220).

  • A party commit no longer re-derives the login binding’s side table, the reverse-relationship index and the protection check from the whole stored body. Each hook reads what the committed body holds, derived once in memory, and writes only the rows that changed, and ferroehr party-binding refill derives the rows with the same code. A party commit without a server signature no longer reassembles its body. A membership change at the cap of a 5 000-member SCIM Group went from 810 ms to 358 ms at p99 on a developer laptop (#5220).

  • openehr-rm writes the RFC 8785 canonical form a VERSION signature covers in one pass over the version, sorting each object’s members by reference. A 3.3 MB GROUP version canonicalises about seven times faster than through serde_jcs, with byte-identical output, so signed versions verify unchanged. serde_jcs remains the test oracle (#5220).

  • The stored-query service takes the typed name, version and version pattern of openehr_its::rest::stored_query end to end, so it reads one grammar with the REST layer: the three-part {namespace}::{formalism}::{query-name} name, which ITS-REST 1.1.0 Qualified_query_name.md does not define, is no longer accepted anywhere (the formalism is the query_type parameter), and DELETE /admin/query/{qualified_query_name}/{version} answers 400 for a name outside [{namespace}::]{query-name} or a version that is no exact major.minor.patch, where it answered 404 (#5223).

  • Boot and ferroehr config check refuse an [[auth.party]] table whose identifier_type is the type of no [[authz.party_binding.person.template]] key, naming the issuer, the type and the types the templates write: no login of such an issuer could ever match a party, and every one would be refused. Upgrade step: give the deployment a person template whose key carries the table’s identifier_type (or set the table to the type the template writes). The Helm CI and production values and the hosted sandbox now carry the professional PERSON and ROLE mapping, and the sandbox loads the two templates before the CDR starts (#5231).

  • openehr_query::printer::to_aql writes each WHERE comparison with one space on each side of the operator (c/name/value = $nameValue), as the QUERY 1.1.0 AQL syntax examples do; path predicates keep their unspaced form ([ehr_id/value=$ehrUid]). The query builder’s generated AQL follows (#5241).

  • openehr_sdt::flat::navigate::Navigator owns what it copies from its Web Template (the name-fallback table and the OPT2 archetype references) and carries no template lifetime, so a consumer can keep it beside the template it was built from; it is Debug, Clone, Send and Sync (#5243).

  • Breaking: the configuration tree is one table per concern. ferroehr.toml moves to [deployment], [api], [database], [auth] (with [auth.smart] and [auth.scim]), [authz] (with [[authz.origin]]), [management], [audit], [privacy] (with identifier_protection, cohort and access_delay), [record] (with subject_index, versioning, signing, query, change_feed and multimedia), [terminology] and [pixm]. FerroEHR’s own keys take one duration type ("500ms", "30s", "10min", "12h", "7d", UCUM unit codes, so a bare number and m are refused), one size type ("256KiB", "64MiB", IEC binary units, decimal units refused) and one calendar retention period ("P5Y", "P1830D", unset keeps forever). An optional feature is on when its table is present: [terminology], [pixm], [auth.scim], [audit.syslog], [audit.fhir_feed], [privacy.identifier_protection] and [privacy.access_delay] carry no enabled; [authz.abac], [record.multimedia] and [auth.smart] keep a switch because some of their keys apply while off. record.subject_index has no default, so the server refuses to start until "pixm" or "local" is chosen. [api.cors] replaces cors_permissive with the Fetch Standard response headers (allow_origins, allow_methods, allow_headers, expose_headers, allow_credentials, max_age). The environment follows the tree (FERROEHR__DB__URL is FERROEHR__DATABASE__URL, FERROEHR__EHR_INDEX__PROVIDER is FERROEHR__RECORD__SUBJECT_INDEX). The Helm chart moves to its 25 line with config: on the same tree. Every key a release before v5.7.0 read is refused at load with its replacement named, all at once, so ferroehr config check --config <old file> prints the whole move list; the refusals ship in v5.7.0 only and the next release removes them. To upgrade, run config check on the old file and move each key it names: spec_profile, deployment_profile, deployment_accepts and licence.file are deployment.spec_profile, deployment.profile, deployment.accepts and deployment.licence_file; [rest] is [api] (swagger_ui is openapi_ui, body_limit_bytes is body_limit, cors_permissive = true is [api.cors] allow_origins = ["*"], [rest.identity] is [api.options] with solution and vendor; the version keys are facts of the build), admin.enabled is api.admin and terminology.api_enabled is api.terminology; [db] is [database] (acquire_timeout_secs is acquire_timeout, statement_timeout_ms is statement_timeout, migrate is migrations), [storage.<domain>] is [database.<domain>] and [storage.migrate.<domain>] url is [database.<domain>] migrate_url; [smart] is [auth.smart] (enabled is discovery, episode.enabled is episode_launch, launch_base64_json = true is "launch-base64-json" in auth.capabilities); [scim] is [auth.scim] (max_results is filter_max_results, deactivation_window_minutes is deactivation_window, contact_erasure_minutes is contact_erasure_delay); [authz.abac.cedar] reload_secs is reload_interval, [authz.abac.remote] server, connect_timeout_ms and request_timeout_ms are url, connect_timeout and timeout, authz.party_binding.provisional_report_days is provisional_report_after, and an agreement’s discard_window_hours and chained_log_retention_years are discard_window and chained_log_retention; [[origins.registered]] is [[authz.origin]]; management.port is management.listen (an address:port) and management.profiling.max_seconds is max_duration; audit.store.retention_days and retention_years are audit.store.retention, verify_interval_seconds and full_verify_interval_seconds are verify_interval and full_verify_interval; [audit.syslog] host and port are url = "tls://host:port", spool_max_messages is spool_max_events, spool_max_bytes is spool_max_size, retry_max_ms is retry.max_backoff, and [audit.fhir_feed] poll_interval_ms and retry_max_ms are poll_interval and retry.max_backoff; [demographic.identifier_protection] is [privacy.identifier_protection], [cohort] is [privacy.cohort], and [access_delay] is [privacy.access_delay] with hours as max_delay; ehr_index.provider is record.subject_index ("ihe" is "pixm") and the rest of [ehr_index] is [pixm] (pixm_url is url, invocation is method, request_timeout_ms is timeout); versioning.skip_identical_resend is record.versioning.idempotent_resend; [signing] is [record.signing] (key_path is key_file, retired_key_paths is retired_key_files), [query] is [record.query] (timeout_ms is timeout), [outbox] is [record.change_feed] (retention_days is retention, prune_interval_secs is prune_interval), and [multimedia] is [record.multimedia] (endpoint is url, threshold_bytes is threshold, and allow_http is retired: an http:// url is plain HTTP); [terminology.server] is [terminology] (fail_on_error is on_error = "accept" | "refuse", request_timeout_ms is timeout). The family sections take the same types, from ferrohealth-server 0.8, ferrohealth-telemetry 0.4, ferrohealth-smart 0.15 and ferrohealth-session 0.9: [server] request_timeout_ms, header_read_timeout_ms, idle_timeout_ms, drain_delay_ms, shutdown_timeout_ms and retry_after_s are request_timeout, header_read_timeout, idle_timeout, drain_delay, shutdown_timeout and retry_after, body_limit_bytes and compression_min_bytes are body_limit and compression_min_size, and [server.tls] reload_interval_s is reload_interval; [telemetry.otlp] trust_roots_file and client_identity_file move into [telemetry.otlp.tls], and an https:// collector is dialed over it; [auth] ca_file is [auth.tls] trust_roots_file, an [[auth.issuer]] table’s clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds and timeout_ms are clock_skew, jwks_max_age, jwks_refresh_interval and timeout, its introspection cache_max_age_seconds is cache_max_age, [auth.basic] cache_ttl_seconds is cache_ttl, [auth.dpop] max_age_seconds and clock_skew_seconds are max_age and clock_skew, [auth.idp.entra] timeout_ms and max_response_bytes are timeout and max_response_size, and an oauth grant’s ca_file, client_certificate_file and client_key_file are its tls table’s trust_roots_file and client_identity_file (one PEM), with refresh_margin_ms, timeout_ms and previous_key_overlap_ms as refresh_margin, timeout and previous_key_overlap. The [auth] keys v5.6 read for every issuer (audience, algorithms, token_profile, jwks_file, mode, introspection, role_claims and the rest) live on each [[auth.issuer]] table, and scopeless_subjects entries are { issuer, sub } tables. In the viewer, cdr.request_timeout_secs is cdr.timeout, [oidc] ca_file is [oidc.tls] trust_roots_file, [oidc.tls] client_certificate and client_key are client_identity_file, and [session] idle_timeout_s, absolute_timeout_s and sign_in_timeout_s are idle_timeout, absolute_timeout and sign_in_timeout. Each of these is refused in v5.7.0 naming its replacement, in the CDR and the viewer alike (#5259).

  • Breaking: a patients token that names an actor is refused. With ferrohealth-smart 0.15, a patients issuer reads a token whose top-level act (RFC 8693 §4.1) names someone other than the patient only when its [[auth.issuer]] table sets act_represents_patient = true; without it the token is refused 401, where v5.6 admitted it as the patient’s own access. With it, FerroEHR refuses the token 403 (refused-patient-representative), because it binds no party to someone acting for a patient, and the access log records the refusal (#5259).

  • SECURITY.md, security.txt (its Contact: mailto: field), the post-market record, the CRA technical documentation and the book’s compliance pages name security@cadasto.com as the email route for a vulnerability report, in place of info@cadasto.com, which stays the single point of contact for complaints and incidents, the manufacturer’s address on the running server and the address for licensing and other business questions (#5272).

  • ferroehr and ferroehr-viewer run through the FerroHEALTH family’s run path. A configuration either binary refuses (an unknown or invalid key, a broken semantic rule, an [auth] table with no mechanism, a SCIM mapping it cannot serve, or for the viewer a TLS file that does not read or an unusable [oidc] or [session] table) now exits 78 (EX_CONFIG of sysexits(3)) where it exited 1, so an orchestrator can tell a configuration to fix from a failure to retry. Every other failure still exits 1, and the failure is written to stderr as one line with its whole cause, prefixed with the binary’s name. Upgrade: a script or probe that expects exactly 1 from config check on a refused file expects 78 (#5321).

  • Breaking: the outbound integrations take the family’s endpoint table. [terminology], [pixm], [audit.fhir_feed] and [authz.abac.remote] read where they call and how from one [<x>.endpoint] table, the same keys for every outbound integration of every FerroHEALTH product: url, timeout, connect_timeout, retry (max_attempts, initial_backoff, max_backoff), tls (trust_roots_file, client_identity_file, min_version) and credentials (HTTP Basic user with password or password_file, or the client credentials grant [...credentials.oauth]). terminology.url, timeout and [terminology.oauth] are terminology.endpoint.url, timeout and [terminology.endpoint.credentials.oauth]; pixm.url, timeout, [pixm.tls] and [pixm.oauth] are pixm.endpoint.url, timeout, [pixm.endpoint.tls] and [pixm.endpoint.credentials.oauth]; audit.fhir_feed.url, [audit.fhir_feed.retry] and [audit.fhir_feed.oauth] are audit.fhir_feed.endpoint.url, [audit.fhir_feed.endpoint.retry] and [audit.fhir_feed.endpoint.credentials.oauth]; authz.abac.remote.url, connect_timeout, timeout and [authz.abac.remote.oauth] are authz.abac.remote.endpoint.url, connect_timeout, timeout and [authz.abac.remote.endpoint.credentials.oauth]. The v5.6 keys are refused naming their new home. Every client now honours retry: the terminology, PIXm and remote policy clients try a call that got no answer or a 5xx up to max_attempts times (three by default; each tried once before), so set max_attempts = 1 to keep one attempt, and the FHIR feed’s delivery rounds take max_attempts tries each, from initial_backoff, before the round waits max_backoff; the feed still never gives a record up. The environment follows the tree (FERROEHR__TERMINOLOGY__ENDPOINT__URL, FERROEHR__AUDIT__FHIR_FEED__ENDPOINT__URL), and so do the Helm chart’s values and its client secret routing, in the chart’s 25 line (#5331).

  • The FerroEHR PostgreSQL image is built on the 2026-11-09 respin of postgres:18.6 (sha256:e3d24f69…), a same-version respin with rebuilt base packages. A full scan of the rebuilt image on amd64 and arm64 finds the same findings as the image of the previous base, each judged in security/vex/postgres-os.openvex.json, and a gate scan finds no fixable HIGH or CRITICAL finding (#5332).

  • The development and Keycloak example compose files run Keycloak 26.8.2 (quay.io/keycloak/keycloak:26.8.2@sha256:723b1d6e…). Of the 26.8.2 upgrading notes, the shipped realm meets the ones that touch it: the Admin REST API’s new audience check passes for the master realm’s admin-cli token the browser-test harness uses, and the realm uses no token exchange, identity broker, SAML client, CIBA or dynamic client registration. The ferroehr-service client-credentials token carries the same audience and scopes as under 26.8.1 (#5333).

Removed

  • [authz.professional] (claim, client_tokens, acting_for_claim) and its 403 for a token that names no natural person. The issuer’s kind and profile classify the caller, and the access record’s acting mode and professional come from that classification: a person’s own token is the person, and a delegated token is the client acting for the person it names. Upgrade step: remove the table, which refuses the start, and list a client that acts for a person in its issuer’s delegates (#5003).

  • The named refusal of authz.rbac.user_role, which v5.6 introduced, is gone under the one-release rule for removal aids: the key is now refused as an unknown key when the configuration loads (#5259).

Fixed

  • A 553|incomplete| commit no longer refuses an empty mandatory archetype slot: RM common master06 §Incomplete Content sets every occurrences lower limit to zero, a slot’s included, while too many fillers are still refused. The OPT 1.4 writer now closes a mandatory slot once the fillers placed in it carry its lower bound, as ADL 2 §Slot Filling and Redefinition allows, and still refuses to close one they leave unmet (#5042).

  • The example of a party template whose identity the template marks optional is a party of its class carrying one identity at every detail level, where the required level answered 500: RM demographic PARTY.identities admits no party without one. A party template that constrains no identity at all answers 400 Bad Request naming its class (#5057).

  • An AQL VERSION standard predicate on commit_audit/time_committed applies its operator: VERSION v[commit_audit/time_committed/value > $t] returns every version committed after $t, where it returned the trunk version current at $t whatever the operator (QUERY 1.1.0 master03 §Standard predicate). A number, boolean or NULL literal as the bound, a string bound no ISO 8601 form reads, and a time_committed path other than the whole value or /value are refused with 400 (#5062).

  • AQL refuses a number, boolean or null parameter or literal compared with a date/time path, in WHERE and in a VERSION predicate, with the same 400 a non-ISO 8601 string gets; such a bound used to become NULL and match nothing (#5064).

  • The served OpenAPI document now lists, on both template example operations, the 400 answered for a party template that constrains no PARTY_IDENTITY (#5066).

  • The compose quickstart, the hosted sandbox and the Helm chart run FerroPIX 0.0.18 (ghcr.io/ferrohealth/ferropix:0.0.18@sha256:6fb622b5b9ade34c8a4053112c8f166ff495d8b32bf99f35529d01e512b7cc42), with the FerroPIX chart 0.3.10 built for it. The 0.0.16 they pinned had an outbox sender that never slept while its outbox was empty, so an idle install kept about 1.2 CPU cores busy between FerroPIX and its PostgreSQL at about 8,000 transactions a second (FerroHEALTH/FerroPIX#490). 0.0.18 also makes FerroPIX’s [auth] scopeless_subjects and role_claims hold per issuer, which changes nothing for a FerroPIX that trusts one issuer, as every configuration FerroEHR ships does. The chart moves to 21.0.5 (#5073).

  • An OPT2 built from a template that specialises a published archetype now carries that archetype, by its full id, as parent_archetype_id, as the OPERATIONAL_TEMPLATE invariant is_specialised requires (AOM2 §3.6.6), so a reader of the OPT finds the archetype its root specialises. The ADL form of an OPT still prints no specialize section (OPT2 §Artefact Structure), so the parent travels in the canonical JSON form (#5075).

  • A refused-key-claim-missing access record names the key claim the issuer’s binding reads (its person_key) in binding_detail.claim, so an operator sees which claim the token lacked. It records the name only, never a claim value (#5139).

  • A 403 for a caller kind is recorded with its own binding code: refused-system-personal-scope for a system caller holding a user/ or patient/ scope, refused-no-agreement for an external caller whose issuer has no agreement for its accessor kind, and refused-scope-beyond-agreement for an external caller holding a scope its agreement does not read. Audit migration 0024 widens the binding constraint; the problem text of each refusal is unchanged (#5140).

  • openehr-sdt builds no /context node, and no start_time or setting leaf, in the Web Template of an OPT 1.4 whose COMPOSITION prohibits context (existence 0..0, or an EVENT_CONTEXT of occurrences 0..0), and the template example of such a template carries no context. A template that leaves context open or optional keeps the node as before (#5147).

  • openehr-sdt gives the /context node it synthesizes in a Web Template the min of the template’s existence on COMPOSITION.context, and 0 where the template leaves context unconstrained, since the RM makes context optional (0..1). A client following the Web Template no longer treats the context as mandatory when the template does not require it, and the required template example of such a template carries no context (#5154).

  • An ADL2 template that deletes an attribute (existence matches {0}), for example COMPOSITION.context, keeps the prohibition after it compiles to its operational template: openehr-adl’s create_opt now returns a RawOpt that names each removed attribute beside the OPT, and build_web_template_v2_4 takes that list. A template that deletes context gets no /context node, and a commit that carries one is refused, as it is for an OPT 1.4 template that prohibits context. An OPT2 uploaded already compiled carries no such list and reads a deleted attribute at its reference-model existence. Breaking for crate users: create_opt returns RawOpt and build_web_template_v2_4 takes the prohibited attributes (#5154).

  • A composition commit whose template_id names a stored ADL2 archetype rather than a template is refused with 422 as an unknown operational template; it failed at the insert with a server error before (#5177).

  • openehr-rm types REFERENCE_RANGE.range as DV_INTERVAL<T>: ReferenceRange is now generic over its limit type (ReferenceRange<DvQuantity> on a DV_QUANTITY, ReferenceRange<DvOrdered> on the other ordered types), its is_in_range and Range_is_simple work at that type, and the static RM model gives REFERENCE_RANGE.range and IMPORTED_VERSION.item their argument. Each model::RmAttribute and model::RmTypeRef now keeps the formal parameter it is written as (formal, Some("T") for DV_INTERVAL.lower) beside the resolved bound. The RM class table declares REFERENCE_RANGE<T> and leaves T unused; the generator restores it from a cited table and refuses a schema that no longer matches (#5196) (#5188).

  • An access record names the client the token’s issuer profile reads from client_claims (azp under the entra and keycloak profiles), the one the caller was admitted as. A token carrying a different client_id claim beside it was recorded under that claim. The [origins] register and the SCIM client lookup read the same client (#5194).

  • An Interval or DV_INTERVAL read from canonical JSON or XML that omits a boundedness or inclusion flag now takes it from its bound. An absent bound reads as unbounded and not included, and a present bound as bounded and included. A flag the wire states is kept as written. Before, an omitted flag always read as bounded and included, so an interval with only an upper came out with a bounded, included lower side that had no value. ITS-JSON requires all four flags, and BASE ties none of them to its bound, so reading the omission is our own design (#5203).

  • The viewer’s New contribution commits as the signed-in login. Each audit carries the placeholder naming that login, and the server records the committer it derives from the login. The free-text committer field is gone. For a login bound to a demographic PERSON, a stated name that differs from the PERSON’s name was refused with 403 (#5222).

  • The management surface answers a caller the login binding refuses with 403 and its binding code, such as refused-unbound, as every other route does, in place of 401: the credential authenticated, so new credentials cannot help (RFC 9110 §15.5.4). An absent or invalid credential still answers 401 with its challenge (#5250).

  • The party binding posture in ferroehr report and the boot log masks the user name and password of an issuer URL, as every other rendering of the configuration does (#5259).

  • GET /definition/query/{namespace}:: lists every stored query in the namespace again, where it answered 400 since the typed stored-query name: ITS-REST 1.1.0 definition_query_list reads its name as a pattern in which an empty part is a wildcard, so ::{name} also lists the names that start with it in any namespace. openehr-its gains rest::stored_query::QueryNamePattern, the type the generated DefinitionQueryListParams carries. The store and the single-query reads still refuse an empty name or namespace (#5268).

  • The identifier scan ([privacy.identifier_scan]) reads the name of the ELEMENT that holds a DV_IDENTIFIER when the identifier’s type is absent or names neither a person nor an item, with the same word rules. An untyped laboratory order number in the Receiver order identifier or Requester order identifier element of laboratory_test_result.v1 that passes the NHS Number check by chance is now stored under the default rule set. A person word still wins, so an untyped value in an element named NHS number or Patient identifier stays scanned, and a type that names a kind takes precedence over the element name (#5291).

  • The Helm chart’s description and README name config.deployment.profile, the key the chart refuses to render without since 5.7.0, and the README’s install command sets it; the command it showed set the old config.deployment_profile and did not render (chart 25.0.1) (#5341).

Security

Shipped with functional changes because: the advisory reaches only a deployment that trusts several issuers, which no configuration FerroEHR ships does, and its fix changes what such a configuration must state at boot, so it lands on main beside the AQL version attestations already merged for this release; a security-only patch of the last release remains open to the release cut.

  • A Basic user never carries a system/ scope, through ferrohealth-smart 0.9.1 (GHSA-jhg2-jwjp-hq74, CWE-269, CWE-863). A system/ scope granted to a Basic user gave that person every resource of a type, a backend service’s reach. Such a user now refuses the start, naming the user and the scope. Upgrade step: give each Basic user the user/ scopes a person holds, and move each machine on Basic to a client of the authorization server under the client credentials grant (#5003).

  • [auth] scopeless_subjects and role_claims hold per issuer, through ferrohealth-smart 0.8.9 (GHSA-pfgx-xq28-2629, CWE-863, CWE-639). Before it, beside several [[auth.issuer]] tables both lists applied to every trusted issuer’s tokens: a token from any of them whose sub was listed met every SMART on openEHR requirement without a scope, and any of them could assert every role the RBAC checks, ADMIN included. An entry is now written <issuer>#<value> and holds for that issuer alone; beside several issuers a bare entry, or one naming an issuer the section does not trust, refuses the boot. Upgrade step for a deployment with several issuers: write each scopeless_subjects entry as <issuer>#<sub>, and write each role_claims path as <issuer>#<path> for the issuers whose roles count (or set role_claims = [] to read none), since the default role_claims is bare and refused. A deployment with one issuer changes nothing (#5047).

  • The ferroehr-postgres image carries zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 from trixie-security, which fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). The image of 5.6.3 and earlier releases carries 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action (security/vex/postgres-os.openvex.json). The ferroehr and ferroehr-viewer images keep the distroless base’s zlib, which neither program loads (security/vex/ferroehr-os.openvex.json) (#5137).

  • The ferroehr and ferroehr-viewer images move to the gcr.io/distroless/cc-debian13:nonroot build of 2026-11-08 (sha256:cc66d222…), whose zlib1g 1:1.3.dfsg+really1.3.1-1+deb13u1 fixes CVE-2026-76844 and CVE-2026-85091 (HIGH, heap overflows in zlib’s gzip writer) and CVE-2026-27171 (MEDIUM, an infinite loop in crc32_combine64). Neither program loads zlib. The images of 5.6.3 and earlier releases carry 1:1.3.dfsg+really1.3.1-1+b1, judged affected with the upgrade as its action, as the ferroehr-postgres image of those releases is (security/vex/ferroehr-os.openvex.json) (#5152, #5329).

  • CVE-2026-112321, CVE-2026-112322 and CVE-2026-112323 (memory and CPU exhaustion in kadmind through the AUTH_GSSAPI handler) and CVE-2026-119079 (CPU exhaustion in the KDC decoding an AS-REQ or TGS-REQ) in MIT Kerberos 5 have no Debian fix for trixie and are judged not_affected in ferroehr-postgres. The image carries neither kadmind, a KDC nor libgssrpc, and the PostgreSQL server never decodes either request; the configuration FerroEHR ships enables no GSS-API authentication (security/vex/postgres-os.openvex.json) (#5344).

5.6.3 - 2026-11-05

5.6.3 is a patch release of 5.6.0 and the first published release since it. The v5.6.1 and v5.6.2 tags exist, but neither release published: the v5.6.1 vulnerability record found CVE-2026-76844 in the zlib package of every image without a judgement, and the v5.6.2 performance guard took the unpublished v5.6.1 as its baseline, whose images do not exist. The guard now compares against the newest published release (#4998). 5.6.3 therefore carries the changes of the 5.6.1 and 5.6.2 sections below: the hosted sandbox’s viewer posture, the FerroHEALTH releases the deployment artifacts pin, and the not_affected judgement of CVE-2026-76844 in every image. It adds ferrohealth-server 0.5.0, FerroBRIDGE 0.0.21 in the Helm chart and the judgement of RUSTSEC-2026-0440, which reaches only the test harness. The REST surface, AQL, storage and configuration are those of 5.6.0.

Changed

  • The server and the viewer take ferrohealth-server 0.5.0, whose release removes five forwarding functions from its public API and changes no behaviour, so the HTTP stack, TLS, the same-origin check and the health check answer as before (#5000).

  • The Helm chart deploys FerroBRIDGE 0.0.21 (ghcr.io/ferrohealth/ferrobridge:0.0.21@sha256:83bd9d2cb5e4ef6b19617c9434242f88489b0fd4150b2238a22acf17435c9a35), its newest release, which renames or removes no configuration key the chart sets. The chart moves to 21.0.4 (#5017).

Security

Shipped with functional changes because: no shipped artefact is affected (the VEX statement below records why), so there is no security update to separate from the other changes this release carries.

  • RUSTSEC-2026-0440 (astral-tokio-tar is unmaintained, with no patched version) does not affect FerroEHR: testcontainers reaches the crate only in the test harness, and the server binary, the ferroehr image and the ferroehr-viewer image do not contain it. deny.toml accepts the advisory until a testcontainers release drops the crate (security/vex/rust-advisories.openvex.json) (#4996).

5.6.2 - 2026-11-05

5.6.2 is a security-only patch release over 5.6.1. The v5.6.1 tag exists, but its release never published: its vulnerability record found CVE-2026-76844 in the zlib package of every image without a judgement, and Debian has no fixed package for trixie. 5.6.2 adds that judgement and nothing else. It is the first published release since 5.6.0, so it also brings the 5.6.1 changes of this changelog: the hosted sandbox’s viewer posture and the FerroHEALTH releases the deployment artifacts pin. The REST surface, AQL, storage and configuration are those of 5.6.0.

Security

  • CVE-2026-76844, a heap overflow in zlib’s gzip file writer after a failed write, has no Debian fix for trixie and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never load the Debian zlib (security/vex/ferroehr-os.openvex.json), and in ferroehr-postgres the zlib 1.3.1 overflow needs a gzclearerr() call that nothing in the image makes (security/vex/postgres-os.openvex.json) (#4993).

5.6.1 - 2026-11-05

5.6.1 is a patch release of 5.6.0. The hosted sandbox’s viewer starts again, since its posture carries the session keys the 5.6.0 viewer reads, and the deployment artifacts pin the FerroHEALTH releases published with or after 5.6.0: FerroPIX 0.0.16, FerroBRIDGE 0.0.20, ferroehr-postgres 5.6.0 on the hosted database box and four FerroSYS crates whose releases change no behaviour. The REST surface, AQL, storage and configuration are unchanged from 5.6.0.

Changed

  • The deployment artifacts run the FerroHEALTH releases published with or after 5.6.0, each pinned by its index digest: FerroPIX 0.0.16 (ghcr.io/ferrohealth/ferropix:0.0.16@sha256:0452371913ce8ed5c5c379693b8bd78683ffaf43d4f51c5278c5444d01d30318) in the compose quickstart, the hosted sandbox and the Helm chart, with the FerroPIX chart 0.3.8 built for it; FerroBRIDGE 0.0.20 (ghcr.io/ferrohealth/ferrobridge:0.0.20@sha256:7dd5550d57b7f0bdfdf194453feff95b093152fd235b9378877bebb8e8d055e2) in the chart; and on the hosted database box ghcr.io/ferrohealth/ferroehr-postgres:5.6.0@sha256:7e4ee7a8a18a5bd3bf4c1b03afd9f584131f717d088cd6844a4a845ece6a9c1b. FerroTERM stays at 0.4.1, its newest release. The server and the viewer take ferrohealth-config 0.2.9, ferrohealth-server 0.4.4, ferrohealth-telemetry 0.2.5 and ferrohealth-usage-report 0.2.1, which change no public API or behaviour. None of these renames or removes a configuration key FerroEHR sets. The chart moves to 21.0.1 (#4987).

Fixed

  • The hosted sandbox’s viewer starts again: its posture (deploy/hosted/ferroehr-viewer.sandbox.toml) still spelled the 5.5 [session] keys cookie_secure and secret, which the 5.6.0 viewer refuses, so sandbox.ferroehr.eu answered 502 after the 5.6.0 deploy. The posture now sets store = "sealed", secure_cookie and key, and a viewer test reads it the way the box’s viewer does at boot (#4985).

5.6.0 - 2026-11-04

5.6.0 follows 5.5.0 as the stable release and carries every change of its ten release candidates: the EHDS Art. 3(3) delay of a person’s own access, the Art. 5 mark of what a person inserts and the Art. 10 opt-out from the access services, each off by default; ADL 2 validation, conversion and Web Templates checked against the AOM 2 rules; storage generations 3 to 6 and AQL plans that keep reads fast as a store grows; the outbox feed that replaces FerroEHR’s own broker publisher; RFC 9457 problem details on every error; and the viewer drawn on its server with FerroUI 0.6. It breaks deployments and clients, each break with its upgrade step under Changed or Removed. Before you upgrade from 5.5.0: export the audit records you must keep over ITI-81 with 5.5.0, then drop the audit schema, because audit migration 0017 starts 5.6.0 on an empty audit store and refuses one that holds records written by 5.5.0 or earlier; move every client of a route ITS-REST 1.1.0 does not define from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, since the old paths answer no redirect; read an error body as application/problem+json, detail where a client read message, and send Prefer: return=representation and read errors where it read validationErrors; send POST /ehr/{ehr_id}/contribution and POST /demographic/contribution bodies in the released NewContribution shape, since every other shape answers 400, and send no empty system_id, which every write path refuses with 422; write a space in a query string as %20; remove [events], the chart’s secrets.eventsUrl and plaintext_broker from deployment_accepts, which refuse the start, and deliver broker events through FerroBRIDGE 0.0.15 or later reading the outbox feed; set client_auth beside [server.tls] client_ca_file; and set config.deployment_profile for Helm chart 21.0.0, which renders no default profile, and bridge.config.auth for a chart FerroBRIDGE that serves a mapping set. Clinical migrations 0022 and 0023 and party migrations 0013 and 0014 read every version row of an existing store once, and every viewer user is signed out once. Helm chart 21.0.0 carries appVersion 5.6.0. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j, and the fix of CVE-2026-115905 in the zip crate that 5.5.0 shipped.

Added

  • Every signed-in viewer page carries FerroUI’s session watch: two minutes before the idle window ends a dialog offers Stay signed in or Sign out, typing or clicking on the page keeps the session going at most once a minute, and when the session has ended the page leaves for the sign-in page by itself, which returns to the page once you sign in again; every other viewer tab in the browser leaves with it, and signing out in one tab ends the others. The watch asks the new session route /api/session, whose reads never extend the session (#4223).

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The operations page gains a storage sizing section, measured on 5.6.0-rc5: about 21.6 KB of clinical storage per version of a vital signs composition (2.0 times its canonical JSON, 85 % of it node rows) and 155 KB for an International Patient Summary, about 1.2 KB per audit record of a read, the effect of [audit.store] retention on growth, pg_wal up to max_wal_size, the REINDEX INDEX CONCURRENTLY clinical.node_hot_pkey step (20 % of that index after heavy concurrent writing) and a query for the size per relation, with the method to reproduce the figures (#4304).

  • EHDS Art. 5: what a person inserts in their own EHR is marked as theirs, and they cannot alter what a professional committed. A write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as its AUDIT_DETAILS.committer, its external_ref naming the login, whatever committer the request names; a COMPOSITION it commits must name PARTY_SELF or PARTY_RELATED as composer; and its update or deletion of a version whose committer is not PARTY_SELF, directly or inside a CONTRIBUTION, is refused with 403. A professional’s writes are unchanged. The contribution list extension gains inserted_by_person per row, and the viewer names a PARTY_SELF committer “The person or their representative”. A service auth_failure status now answers 403 where it answered 500 (#4350).

  • The EHDS Art. 3(3) delay of the person’s own access to newly registered data, a Member State option and off by default. The new [access_delay] section (enabled, categories, hours) holds every new COMPOSITION of a listed [audit.categories] category from requests that may be the person’s own (a bearer token with a SMART patient/ grant or the ABAC patient claim, also beside user/ or system/ grants, as ferrohealth-smart 0.8.7 decides it) for at most hours: a read answers 404 as for an unknown composition, and a query, an export and a CONTRIBUTION leave it out, while a professional reads it as before. The treating professional releases a held composition from the clinical route POST {ext}/ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/access-delay/release, with the checks of a write to that EHR, and the controller with POST /admin/access-delay/release; a professional holds one by hand within the limit with POST /admin/access-delay and reads the register with GET /admin/access-delay; every hold ends at its limit, and placing, releasing and expiry are each an access record. The subject-scoped ITI-81 read of the person’s own trail (EHDS Art. 9) leaves out the records about a held composition until it is released or expires, then serves them with their original times (ehds-logging 0.3.3) (clinical migration 0027_access_delay.sql, party migration 0019_access_delay_column.sql) (#4351).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • [versioning] skip_identical_resend (default false): when on, an update whose content equals the current version’s commits nothing and answers like a successful update with the existing version’s ETag and Location. The access log keeps one resend record naming the version matched and the content digest; the book names the audit consequence: no commit audit and no contribution record the repeated send (#4415).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • The Conformance workflow runs the Veredictum smart documents (smart_discovery, smart_scopes) against the same FerroEHR build after the Veredictum Action, which forwards no SMART Platform base URL and no scoped tokens. The same pinned image runs veredictum test with --platform-url and docs/conformance/party/ferroehr/ixit-smart.json, minting each access token in memory from the committed test issuer the SMART overlay trusts, and the job uploads its results, JUnit and CTRF views beside the catalogue’s and adds its summary under a heading of its own. scripts/conformance.sh gains the smart stage, and run drives it after the catalogue, so a local run covers SMART too (#4450).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

  • The EHDS Art. 10 opt-out from access through the access services of Arts. 4 and 12, a Member State option and off by default. The new [authz.rbac] access_service_roles key declares the roles of the access-service callers; while an EHR carries the opt-out, a caller holding one reads it as absent: a route naming the EHR or an object in it answers 404 as for an unknown EHR, a population query leaves its data out, a scoped query, an export and the lookup by subject find no such EHR, and a write to it is refused as absent. Every other caller, the provider’s own care included, reads it as before. The controller records the opt-out on the person’s instruction with POST /admin/access-opt-out, reverses it with POST /admin/access-opt-out/reverse (the register keeps the reversal) and reads the register with GET /admin/access-opt-out; recording is refused with 400 while no access-service role is declared. An access-service caller that declares one of [audit] emergency_purpose_codes and does not act for the person reads the EHR under the Art. 10(2) vital-interest override, with the emergency mark on its access record and an access-opt-out-override record against the EHR. Recording, reversing and reading the register are access records (clinical migration 0029_access_opt_out.sql) (#4516).

  • Two read-only extension routes convert stored ADL 1.4 definitions to ADL 2: GET /ferroehr/rest/extensions/v1/definition/archetype/adl1.4/{archetype_id}/adl2 for a stored archetype and GET /ferroehr/rest/extensions/v1/definition/template/adl1.4/{template_id}/adl2 for a stored operational template, which converts to one ADL 2 archetype per embedded archetype root. Each answers 200 with application/json: (#4655)

    • the ADL 2 source of every root;
    • each root’s conversion log and its conversion_details, which name what the ADL 2 form drops or changes;
    • the slot fills that joined the roots. An unknown id answers 404. A stored source that no longer converts answers 422. No openEHR specification defines the conversion, so both routes are FerroEHR’s own design.
  • The commit-event outbox feed. [outbox] readers names the readers (FERROEHR__OUTBOX__READERS); boot registers each one and marks every reader the list no longer names inactive, so it holds the retention prune back no longer (#4766). GET /admin/outbox/{reader}?limit=&domain= answers the PHI-free envelopes after the reader’s cursor in seq order, and POST /admin/outbox/{reader}/ack with {"last_seq": n} moves the cursor (204; 409 below the cursor or past the highest row; 404 for an unknown or inactive reader). Both routes need the __ferroehr.outbox scope, every read is an access record, and the clinical and party outboxes each keep their own seq and cursor. A read waits for a commit that drew a lower seq to end, so a reader never acknowledges past a row that becomes visible later. The prune now counts from the commit instant and keeps every row an active reader has not acknowledged (migrations clinical/0030_outbox_feed.sql, party/0023_outbox_feed.sql). With no reader named, no commit writes an outbox row (#4786).

  • An administrator repairs an EHR held closed by EHR_ACCESS settings that do not parse with POST /admin/ehr-access/repair under the extensions root, naming the EHR and the new settings in the body. The repair commits a new EHR_ACCESS version with its commit audit and contribution, writes an access record, and is refused with 409 for an EHR that is not held (#4853).

  • openehr-am: ArchetypeHrid implements Display, writing the physical form [ns::]publisher-package-class.concept.vMAJOR.MINOR.PATCH[-status[.build]] that its FromStr reads back, so a consumer no longer rebuilds an archetype or template id from the fields. openehr-adl writes every identifier through it. version_id and physical_id no longer write a trailing . after a version modifier that carries no build count (#4938).

Changed

  • A viewer session now also ends session.absolute_minutes after sign-in, 12 hours by default, whatever the activity; the idle window (session.idle_minutes) can no longer carry it further. A lifetime of zero or one shorter than the idle window refuses the start. A session cookie sealed by an earlier release carries no sign-in time and no longer opens, so the upgrade signs every viewer user out once (#3057).

  • Breaking for clients that relied on the old leniencies: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution now read the body through the released NewContribution schema, and refuse every shape it does not define. Each refused shape answers 400: (#3595)

    • a lifecycle_state in the CODE_PHRASE shape {"terminology_id": …, "code_string": …}, as a plain string, or as a bare {"value": …}. Send a DV_CODED_TEXT: {"value": "complete", "defining_code": {"terminology_id": {"value": "openehr"}, "code_string": "532"}};
    • a description sent as a plain string, on the contribution audit, on a commit_audit or on an attestation. Send a DV_TEXT or DV_CODED_TEXT object: {"_type": "DV_TEXT", "value": "…"};
    • a commit_audit or contribution audit without change_type. The server no longer infers creation or modification from preceding_version_uid; send the code of the change, for example 249 (creation), 251 (modification), 523 (deleted) or 666 (attestation);
    • a commit_audit or contribution audit whose _type is ATTESTATION or UPDATE_ATTESTATION, or any value other than UPDATE_AUDIT or AUDIT_DETAILS. Send the audit as UPDATE_AUDIT (or with no _type) and put each attestation in the version’s attestations list as an UPDATE_ATTESTATION. A 666 (attestation) version carries its attestations there too, and one with no attestation is refused;
    • a missing audit, which answered 422 before. A missing versions list, an attestation without reason or is_pending, a system_id that is not a string, and a preceding_version_uid sent as a plain string instead of {"value": "…"} are refused the same way.
  • An attestation’s own system_id is kept, and one that states none takes the system_id of the version audit it belongs to instead of the server’s. An attestation committer is now checked against its PARTY_PROXY rules like every other committer, so an invalid one answers 422 (#3595).

  • The served OpenAPI document describes each extension route by what it does and by the ITS-REST text that governs it, and no longer names openEHR Service Model operations, interfaces or status codes. Routes, request and response fields and status codes are unchanged (#4213).

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The storage-bench lane measures the baseline and the candidate on one runner in eight interleaved rounds each and gives two verdicts. The plan verdict fails an operation whose median pages per call rose more than 10 % and one page above the baseline’s highest round, whatever its time. The time verdict divides each round by its machine factor, so a slow spell of the runner no longer reads as the code, and fails an operation whose median p50 or p99 rose more than 15 % with every candidate round slower than every baseline round, or a candidate build that ran more than 15 % slower across the board in every round. A latency is no longer compared with a record measured on another runner (#4318).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • The LANG specification text is re-vendored at f80763ab (#4377). SPECLANG-18 states how a P_BMM schema’s JSON form encodes a cardinality interval, which the JSON schema reader already followed; its documentation now cites the rule, and a test pins it. SPECLANG-16 types the P_BMM class lists as keyed collections; the generated P_BMM model takes that change with #4497. The nine openehr-* crates step to 0.0.122.

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • A version whose content equals its predecessor’s stores that content once: the version row, its commit audit, its contribution, its attestations and its uid-stamped body are written as before, and the node rows are not, the new version reading its predecessor’s through the new version.content_sys_version and the stored version.content_key every reader joins on (migrations clinical/0032, party/0025 and clinical/0033, which re-creates the retention_due view). Every reader, AQL included, answers each version with its own uid. The digest the content is compared on is version.content_digest, SHA-256 over the canonical body with the root uid removed; a row written before keeps its own rows and is never shared (#4415).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • The viewer runs on ferrohealth-server 0.4.2 and takes its Content-Security-Policy from the family’s server-rendered console headers (FerroHEALTH/FerroSYS#202) in place of a layer of its own. The policy a deployment sees is unchanged: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; object-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none', with no nonce. FerroUI’s assets under /ferrohealth-ui/ keep the Cache-Control FerroUI sets, and a refusal under that path is no-store whatever FerroUI’s router sets, so a cache in front of the viewer keeps no error answer for an asset. The REST server takes the same ferrohealth-server release with no change in what it serves (#4439).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The published conformance record is a Veredictum 0.3.3 released-profile run, committed as docs/conformance/ferroehr/results.json with its summary.md. The landing page and the book’s conformance page show the Veredictum badges of the released and development runs on main and the record’s counts per part (BASE and OPTIONS), in place of the Veredictum 0.1.6 heat grid, chapter bars and tier pills. The docs and landing pages admit images from raw.githubusercontent.com/FerroHEALTH/FerroEHR/ for the badges (#4451).

  • The measured performance lane runs Veredictum 0.3.3’s perf. scripts/conformance.sh perf composes the conformance deployment with the measurement overlay, runs the pinned clinical-mix workload, checks the record with perf-judge and redraws the book’s charts from docs/conformance/ferroehr/perf-result.json. scripts/conformance.sh perf-aa measures the deployment against itself, so the record carries its between-run noise floor. The measurement overlay also switches response compression off, and the IXIT names the deployment’s containers for the run’s server-side readings. The Performance and Benchmarks pages describe the new instrument. They keep the earlier figures until a record measured on the reference host replaces them (#4452).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

  • The P_BMM model follows SPECLANG-16. The LANG 1.1.0 BMM that openehr-codegen reads is now the specifications-LANG file at f80763ab, so P_BMM_SCHEMA.primitive_types and class_definitions are maps keyed by class name, P_BMM_CLASS.ancestor_defs is a map of P_BMM_BASE_TYPE keyed by type signature (a simple ancestor included), and P_BMM_GENERIC_TYPE.generic_parameter_defs is an optional map keyed by formal parameter name. The formal and actual generic parameters (P_BMM_CLASS.generic_parameter_defs, P_BMM_GENERIC_TYPE.generic_parameter_defs, BMM_GENERIC_CLASS.generic_parameters) are insertion-ordered IndexMaps, because the specification fixes their declaration order, and the model builders bind each actual parameter to its formal by name. In the released v1_0 generation, P_BMM_GENERIC_TYPE.generic_parameter_defs now carries the 1..* bound its BMM states. The nine openehr-* crates step to 0.0.123 (#4497).

  • The conformance pipeline runs Veredictum 0.3.1. The Conformance workflow drives the catalogue through the Veredictum Action twice, each run on fresh volumes: under the released profile, the frozen openEHR release, which is the record and decides the job, and under the development profile, the pinned master of each openEHR repository, which is published beside it under its own job summary heading and artifact and never counted toward it. Both runs include the SMART documents: the Action now takes the SMART Platform base URL and the committed test issuer’s key. The workflow reads the result paths and the exit status from the Action’s outputs. scripts/conformance.sh takes the profile as CONF_PROFILE (released by default; a development run writes to docs/conformance/ferroehr-development/) and drives the SMART documents in the same run, and the IXIT carries the test issuer’s scoped_tokens declaration (#4503).

  • The viewer groups a template’s releases. The template list draws one group per template, the newest release first with the number of releases and the earlier releases under it, in both the ADL 1.4 and the ADL 2 tab. A template’s page names its release beside its title, where every other release is a link with the newest marked, also with FerroUI’s scripts refused; the release select and its Open button above the tabs are gone. The Overview lists the releases with the compositions committed against each, and a new Compare tab lists the paths one release adds, removes and changes against another (#4506).

  • Every write checks the committer against the login that authenticated the request where the server can compare them: a committer naming another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, is refused with a 403. A name only, or an identifier of another authority, is stored as sent. The direct writes (composition, EHR status, directory, EHR, demographic) now also refuse a committer that breaks its own rules with a 422, as the contribution path already did (#4520).

  • An audit record no longer adds a row to audit.audit_search_token for its request id or for the version uid a create wrote. ITI-81 entity.identifier finds both from the record’s own request_id and resource_id columns, through a new hash index on request_id (audit migration 0020), so the token table stays flat as requests are audited. Retention now deletes the search tokens that no remaining record names, among them the request id tokens of records written by 5.6.0-rc5 and earlier, and the database still refuses to delete a token a kept record names (#4524).

  • A clinical or party node row below the root no longer repeats its archetype header, language and encoding: an archetype root’s archetype_details and archetype_node_id are rebuilt from the archetype column and a new rm_version column, and a language in ISO_639-1 and an encoding in IANA_character-sets keep only their code, in the new language and encoding columns. Every reader puts the keys back where they stood, so the canonical JSON, every AQL result and the served key order are unchanged; a header of any other shape stays in the fragment. A vital signs version takes 17.1 KB instead of 21.6 KB and an International Patient Summary 131 KB instead of 159 KB. Migrations ext/0007, clinical/0028 and party/0020 add the columns and the readers; no data is migrated, and a row written before keeps its whole fragment and reads the same (#4525).

  • ext.storage_generation() reports generation-4, the storage generation with the node header columns, and the storage benchmark records for it are committed under docs/benchmarks/storage/generation-4/. Migration ext/0008 changes only the label; nothing in the serving path reads it (#4525).

  • Five response messages no longer name the openEHR Service Model. The admin dump route’s 400 for a missing file_sys_loc now reads “request body must carry a non-empty file_sys_loc string, a location on the server’s file system (FerroEHR’s own dump/load extension)”, and its 400 for an encoding member reads “encoding is not accepted: FerroEHR’s own dump/load extension defines no encoding, so no value names one”. The admin activity-report 400s for a_service now list the accepted service names (“query parameter a_service is required: one of Admin, Definitions, …” and “query parameter a_service names no platform service: … (expected one of …)”). The not-implemented answer of store_query_set reads “store_query_set is not implemented: no openEHR specification defines a query set”. The openehr-rm validation docs for the three PARTY and PARTY_RELATIONSHIP relationship invariants now name the separate relationship container as FerroEHR’s own design against RM demographic §Party Relationships (#4531).

  • Breaking for clients that relied on the remaining commit-audit leniencies: the audit rules of the CONTRIBUTION commit now hold on every write path (#4543).

    • A description object without _type, such as {"value": "…"}, is refused with 400 on a contribution audit, a commit_audit and an attestation, and so is an attestation reason without _type. Send {"_type": "DV_TEXT", "value": "…"} or a DV_CODED_TEXT with its _type.
    • A stated empty system_id is refused with 422 on the direct writes too (openehr-audit-details: system_id="" on a create, update or delete), where the server used to replace it with its own. Leave system_id out to get the server’s, or send a non-empty one. A system_id that is not a string stays a 400 on both CONTRIBUTION routes.
    • A commit audit of type UPDATE_ATTESTATION handed to the library’s direct write calls (create_composition, update_composition, the EHR_STATUS, DIRECTORY and party writes) is refused with 400, as on the CONTRIBUTION routes. Send the audit as UPDATE_AUDIT and put each attestation in the version’s attestations list.
  • Members that NewContribution or UPDATE_AUDIT do not declare, such as a reason or is_pending on a plain audit or an extra top-level member of a contribution, are accepted and ignored: the released schemas leave both objects open (#4543).

  • The openehr-* crates take their version when they are published. A pull request that changes the crates no longer bumps their lockstep 0.0.x version: the crates publish reads crates.io, takes the next version when packaged content changed since the newest published one, and refuses to publish changed content under a version crates.io already holds. The version in the manifests on main can therefore trail the newest version on crates.io (#4555).

  • The served Web Template of an ADL 2 operational template writes the aqlPath of an ELEMENT’s value and of a data value’s parts without an [idN] predicate, since a data value carries no archetype_node_id (ADL2 master05-paths §“Relationship with W3C Xpath”). A DV_INTERVAL and its parts read …/value, …/value/lower and …/value/upper, the form the ADL 1.4 Web Template writes, where they read …/value[id6]/lower[id7] before, and the parts’ ids are lower and upper (#4559).

  • The served Web Template gives every DV_INTERVAL<T> node its lower and upper children, each of type T, also when the operational template leaves them unconstrained (RM data_types §DV_INTERVAL). Such a node had no children before. An unconstrained part carries the open inputs of its type, and its aqlPath is …/value/lower or …/value/upper. This holds for ADL 1.4 and ADL 2 templates (#4559).

  • The template example fills such an interval: an unconstrained part takes the example value of its constrained sibling, so the limits stay consistent (#4559).

  • The served Web Template of an ADL 2 operational template writes a coded input whose at-codes the template binds to a terminology as the bound terms: its code list carries each binding’s code and label and its terminology names the bound terminology (ADL2 master08-terminology_integration §“From Constraints to Concrete Codes in Data”). COMPOSITION.category bound at1 to openehr::433 reads {"value": "433", "label": "event"} with terminology: "openehr", where it read the local at1 before, and an [acN@ttttt] constraint takes the bound terms of the terminology it names. A FLAT or STRUCTURED composition without a category, and the example composition, now carry openehr::433. The aqlPath of every data value drops its [idN] predicate, so a constrained category and context/setting read /category and /context/setting (#4574).

  • A CONTRIBUTION’s versions[] entry records its merge inputs: the other_input_version_uids it states are stored on the committed ORIGINAL_VERSION and served on every read, where the server used to refuse the member with 400. The uids may name versions the server does not hold; an empty list or an entry that is not an OBJECT_VERSION_ID is a 422 (#4577).

  • Other members UPDATE_VERSION does not declare, such as item or an extension member, are accepted and ignored, where the server used to refuse them with 400: the released schema leaves the object open, as it does NewContribution and UPDATE_AUDIT. A uid or contribution on a version entry is still refused with 400, because the server assigns both, and so is a _type other than UPDATE_VERSION or ORIGINAL_VERSION (#4577).

  • The admin dump route’s refusals name its parameters by their wire names and list the accepted values, not the Service Model class names. A bad logical_format or compression_format now reads “<field> must be a string, one of …” or “<field> names no accepted value: … (expected one of …)”, and a non-integer segment_split_size reads “segment_split_size must be an integer size in kb”. The served OpenAPI text for the route no longer names EXPORT_SPEC or EXPORT_FORMAT (#4634).

  • Five archive dump and load failures still answer 500, and their bodies are now fixed text that echoes no value from the archive: (#4657)

    • an unreadable data segment no longer names the segment, though manifest.json and demographic-commons.json are still named;
    • an unknown logical format in the manifest no longer echoes the declared value;
    • a version whose id carries a path separator, and a version whose commit audit the export lacks, no longer name the version.
  • A failed expansion of a party’s protected national identifiers still answers 500, and its body now reads “expanding the protected identifiers of a party version failed” without echoing the requested version uid: a 500 body carries curated text only (#4657).

  • Two query failures still answer 500, and their bodies now carry curated text only. A stored query whose record holds no query text reads “the stored query has no query text” without echoing its name, and a terminology server failure while a TERMINOLOGY('expand', …) operand is resolved reads “the terminology service failed while the query was being planned” instead of the upstream message (#4657).

  • The db_transactions counter (outcome="commit") now counts every committed single-object write. It used to leave out EHR_STATUS and DIRECTORY writes (#4659).

  • The access record of a VERSION read (GET /ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/version, its /version/{version_uid} form, and the same two versioned_ehr_status routes) now names the content it served, and for a COMPOSITION the origins of its data, as the bare reads already did. Its ehr_id is the parsed EHR id, as on every other read, rather than the path text as sent. The response is unchanged (#4660).

  • A query request whose ehr_id, as the query parameter or the openehr-ehr-id header, is not a UUID is refused with the same 400 and text as before, now before the AQL is parsed and planned. The refusal therefore comes ahead of the refusals planning raises, such as an EHR scope on a query rooted at a demographic class, and aql_queries_total{outcome="analysis_error"} no longer counts it (#4661).

  • OPTIONS on the API base path sends conformance_profile as BASE+OPTIONS where it sent STANDARD. The build derives the value from the committed Veredictum record, docs/conformance/ferroehr/summary.md: BASE+OPTIONS when the BASE rows and the run OPTIONS rows have none failed and none errored, BASE when only the BASE rows do, and the member left out otherwise. ITS-REST 1.1.0 types the member as a free string and CNF names no profiles, so the values are FerroEHR’s own. The [rest.identity] conformance_profile configuration key is removed: a configuration that still sets it is refused at boot as an unknown key (#4672).

  • The served OpenAPI description of POST /query/cohort no longer names the linkage domain, which the server no longer has. It says the predicate selects parties on the demographic database credential, and that those parties are resolved to the EHRs whose EHR_STATUS.subject names one of them on the clinical credential. The route, its fields and its status codes are unchanged (#4695).

  • Every route now reads Prefer as RFC 7240 §2 defines it, as error responses already did: a preference in a second Prefer field, one that carries parameters after ; (return=representation; x=1), and one with whitespace around = (return = representation) are honoured. Before, only the first Prefer field was read, and such a preference was ignored (#4722).

  • The viewer answers every request other than GET and HEAD that arrives without a session, to any path but the sign-in page, the OIDC routes, sign-out and the search answer, with the sign-in page that says the session ended, before any handler runs. A post to a page returns there after signing in; a post to the export, preference, sidebar or paging route returns to the dashboard. The export post used to answer 303 to /login, and the preference, sidebar and paging posts ran without a session (#4724).

  • Breaking: POST /admin/load answers a malformed archive with 400 (#4738). A container, manifest.json, segment or blob entry that is absent, truncated or does not parse as an export archive, and a manifest that declares an unknown logical format, are the caller’s input, so the load refuses them with the new archive-invalid problem type instead of a 500. The detail names the entry, and with Prefer: return=representation the archive member carries it as {"entry": …} or {"format": …}. A location with no archive, or one the server’s filesystem refuses to read, still answers 500.

  • The conformance instrument is Veredictum 0.4.0, and the published conformance record (docs/conformance/ferroehr/results.json with its summary.md) is a 0.4.0 released-profile run. The party statement answers the 0.4.0 catalogue: it keeps the xml-namespace-v2 arm, which now asks only whether the TRIAL http://schemas.openehr.org/v2 namespace is also accepted, and selects the new development-profile option arm adl2-id-coded-supported (#4771).

  • The RM 1.2.0 TRANSLATION_DETAILS of the retained ADL 1.4 resource package, the development profile’s generation, spells its translator accreditation accreditation now, following the RM 1.2.0 BMM at 09e481e9: its canonical JSON and canonical XML carry the key accreditation in place of accreditaton, and the XML reader binds an <accreditation> element it used to skip. The openehr-rm field v1_2::common::resource::translation_details::TranslationDetails::accreditaton is renamed accreditation; the released v1_1 generation keeps the RM 1.1.0 spelling (#4813).

  • The regulatory position, the technical documentation, the audit trail page and the compliance overview name FerroBRIDGE’s recording of its FHIR facade’s accesses as part of the logging component, from FerroBRIDGE 0.0.16: FerroEHR’s record of a facade access names the bridge’s credential, and the bridge’s own Annex II 3.2 record, sent as an IHE BALP AuditEvent, names the natural person. The pages no longer say FerroBRIDGE holds no part of the log, and say how the two harmonised components stay independent with it (#4822).

  • The FerroEHR PostgreSQL image is built on the 2026-11-02 respin of postgres:18.6 (#4836, sha256:3a426452…), a same-version respin with rebuilt base packages. Every finding of a full scan of the rebuilt image keeps its judgement in security/vex/postgres-os.openvex.json, and none goes stale.

  • A CONTRIBUTION that writes ferroehr.access_control.v1 settings whose privacy override uid is neither a VERSIONED_COMPOSITION uid nor a version uid whose first part is one is refused with 400, naming the override. Such an override matched no composition, since every composition id is a UUID; stored settings that still carry one are read without it (#4837).

  • An archetype root converted from an ADL 1.4 operational template records each language its terminology carries beside the template’s own as a conversion_details entry, translation.<language>, naming the source template. An OPT 1.4 has no translation record, so the conversion writes no translations entry for these languages (#4861).

  • The viewer signs in and keeps its session through ferrohealth-session 0.2.1, FerroAUTH’s console sign-in (#4873). Its own OpenID Connect client and sealed-session code are gone. The sign-in page moves to /sign-in, and the sign-in routes are /login (which begins the OpenID Connect sign-in), /auth/callback, POST /logout (which now also revokes the token and, for an OpenID Connect session, uses the provider’s end-session endpoint when it has one) and /session, the session watch. Sessions are kept in the viewer process by default (session.store = "memory"), and in one sealed cookie with session.store = "sealed" and a key. A sealed store without a key refuses the start, so there is no ephemeral key any more. FerroEHR pins ferrohealth-smart 0.8.8. Upgrade note:

    • The provider must allow the redirect URI <public_base_url>/auth/callback in place of /auth/oidc/callback, and the post-logout redirect URI <public_base_url>/sign-in?signed-out=1, where an OpenID Connect sign-out returns. A Basic sign-out lands there directly.
    • [auth.oidc] becomes [oidc]. enabled moves to auth.oidc_enabled and public_base_url to auth.public_base_url. tls.trust_roots_file becomes oidc.ca_file, and resolve = "host=ip:port" becomes oidc.resolve = ["host=ip:port"]. tls.client_identity (one PEM) becomes oidc.tls.client_certificate with oidc.tls.client_key (or client_key_file), and tls.min_version becomes oidc.tls.min_version, still "1.3" by default.
    • In [session], idle_minutes becomes idle_timeout_s and absolute_minutes becomes absolute_timeout_s, both in seconds; cookie_secure becomes secure_cookie. secret and secret_file become key and key_file, which only a sealed store takes, so a scaled deployment sets session.store = "sealed" beside its key.
    • The variables follow: for example FERROEHR_VIEWER__AUTH__OIDC__ISSUER is now FERROEHR_VIEWER__OIDC__ISSUER. Bookmarks to /login reach the OpenID Connect sign-in, not the sign-in page.
    • The Helm chart (21.0.0) renders the new variables from the same viewer.auth.oidc.* values.
  • A clinical or party node row stores a HISTORY origin and an EVENT or ACTION time of the plain DV_DATE_TIME shape as their value string, leaves out an ENTRY subject that is a bare PARTY_SELF, and the root row leaves out its uid, which the version row already holds, and keeps its archetype header, language and encoding in the columns a row below the root used since generation 4, with the version’s template_id put back into the header. Four new boolean markers on the row (origin_compact, time_compact, party_self, uid_derived) say which keys it left out, so every reader puts them back exactly where they stood and a fragment of any other shape stays as written; the canonical JSON, every AQL result and the served key order are unchanged. Migrations ext/0010, clinical/0031 and party/0024 add the markers and the readers; no data is migrated, and a row written before keeps its keys and reads the same (#4904).

  • ext.storage_generation() reports generation-5, the storage generation with the node compact markers, and the storage benchmark records for it are committed under docs/benchmarks/storage/generation-5/. Migration ext/0011 changes only the label; nothing in the serving path reads it (#4904).

  • The terminology overlay (docker-compose.terminology.yml), the Helm chart and the hosted sandbox run FerroTERM 0.3.3 (ghcr.io/ferrohealth/ferroterm:0.3.3@sha256:f25ddad886fd218ae1eb82af6b81b8ed16651333f5bd802e621c5ae83b0913f1), where the overlay and the chart ran 0.1.6 and the sandbox 0.3.1. FerroTERM 0.3 reads only its FERROTERM__… configuration tree, so the overlay and the chart now set FERROTERM__TERMINOLOGY__CODE_SYSTEMS, FERROTERM__TERMINOLOGY__INDEX, FERROTERM__TERMINOLOGY__DEFAULT_LANGUAGE and FERROTERM__TELEMETRY__FORMAT. The overlay’s own FERROTERM_INDEX, FERROTERM_LOG_FORMAT and FERROTERM_DEFAULT_LANGUAGE shell variables keep their meaning. The index layout is unchanged, so a built index serves as it is. Upgrade: move any terminology.extraEnv entry that sets a pre-0.3 FerroTERM variable to its FERROTERM__… key (#4931).

  • The deployment artifacts run the newest release of each FerroHEALTH product, each pinned by its index digest. The terminology overlay, the SUT stack, the hosted sandbox and the Helm chart run FerroTERM 0.4.0 (ghcr.io/ferrohealth/ferroterm:0.4.0@sha256:efaa80152721c3738e0d0c0695cadbfefe89c1cf6df3567b9d3bc1788767e047), whose index layout is unchanged, so a built index serves as it is. The compose quickstart, the hosted sandbox and the chart run FerroPIX 0.0.14 (ghcr.io/ferrohealth/ferropix:0.0.14@sha256:389f40d144fe2ef3d79c1c1628e9a662a058961b1650c5541e601addc27112d8), with the FerroPIX chart 0.3.6 built for it. The chart runs FerroBRIDGE 0.0.17 (ghcr.io/ferrohealth/ferrobridge:0.0.17@sha256:4cdf8d4b4c4d626b99a3929a363198067f87ac70875f4ce6db1d0614884f12af) where it ran 0.0.5, and now writes FerroBRIDGE’s [usage_report] from bridge.usageReport.enabled. FerroBRIDGE 0.0.17 checks the callers of its FHIRconnect operations, FHIR facade and operator API through [auth], and refuses to start them until [auth] names issuer and audience or sets enabled = false. Upgrade: a values file that sets bridge.mappings adds bridge.config.auth.issuer and bridge.config.auth.audience, or bridge.config.auth.enabled: false to serve the operations to every peer bridge.networkPolicy.ingressFrom admits (#4933).

  • The viewer draws with FerroUI 0.6.6 (ferrohealth-ui), and serves its stylesheet and scripts under /ferrohealth-ui/0.6.6/, with FerroUI’s fixes since 0.6.0: a mapping row’s relationship words and the top bar’s search submit fit a 320 px screen, and a combobox is not spell-checked (#4933).

  • An ADL 2 upload that constrains property on DV_QUANTITY is still refused with VCARM, and the error now says that property is the ADL 1.4 archetype-profile attribute C_DV_QUANTITY.property, which no ADL 2 or RM release defines, and that the archetype’s ADL 1.4 source uploads and converts to ADL 2 without it (#4964).

  • The server, the viewer and the EHR index take the newest FerroHEALTH family crates: ferrohealth-config 0.2.8, ferrohealth-server 0.4.3, ferrohealth-session 0.2.2, ferrohealth-telemetry 0.2.4, ihe-iti 0.0.74 and ferrobridge-term 0.1.141. What changes for you: [telemetry.otlp] now refuses trust_roots_file or client_identity_file beside an endpoint that is not https://, and a file that is no PEM of its kind, at boot and naming the key, where it exported in plain text with the files ignored. Upgrade: point the endpoint at https:// when the collector speaks TLS, or drop the two keys when it does not. Each request’s exported span is a server span named {method} {route} whose status is an error on a 5xx answer, where it was named request. /management/prometheus gains http_server_active_requests, the requests in flight by method and scheme. A panic line names the server’s id of the request it happened in. A configuration refusal no longer echoes a secret written unquoted in the file, and a refusal inside an array of tables names its key with the element’s index (#4968).

  • The conformance instrument is Veredictum 0.4.1, and the published conformance record (docs/conformance/ferroehr/results.json with its summary.md) is a 0.4.1 released-profile run. The record gains the three aqlc_select cases of the aql-conformance corpus at 3ecd250 (a quoted 'True' and a quoted '3e102' stay String literals, and a quoted date/time literal projects as the value it writes). Veredictum 0.4.1 skips a refusal row whose every control was skipped, where 0.4.0 errored it, so on the development profile template_adl2/get_template_adl2_example-invalid_type and -invalid_detail_level are no longer errored. The development snapshot moves RM to 09e481e9 and ITS-XML to 27c7a78b (#4971).

  • The deployment artifacts run the newest FerroHEALTH releases, each pinned by its index digest: FerroTERM 0.4.1 (ghcr.io/ferrohealth/ferroterm:0.4.1@sha256:b7f10609992377a7b8bca666254bceca79fe4366e2e54fc5fbce079eb9abf619) in the terminology overlay, the SUT stack, the hosted sandbox and the Helm chart, with the index layout unchanged; FerroPIX 0.0.15 (ghcr.io/ferrohealth/ferropix:0.0.15@sha256:5572f80a3a17fe83228a2a6fcf40f1dcaf8cacc48275136927286ba6ded5c74b) in the compose quickstart, the hosted sandbox and the chart, with the FerroPIX chart 0.3.7 built for it; and FerroBRIDGE 0.0.19 (ghcr.io/ferrohealth/ferrobridge:0.0.19@sha256:bdf4bd416f772ac36e4b3a940eb67841989552c0a0950d680049d959bd54dcba) in the chart. None of the three renames or removes a configuration key FerroEHR sets. Helm chart 21.0.0 carries the three pins (#4981).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

  • The Veredictum 0.1.6 conformance documents: CONFORMANCE_REPORT.md, CONFORMANCE_STATEMENT.md, CONFORMANCE_CERTIFICATE.md, verdicts.json, run-exceptions.json, the CORE, STANDARD, OPTIONS, SEC-BASIC and performance badge files, and the heat grid and chapter bars the book and the landing page drew from them (#4451).

  • The CONF_PERF_CLASS, CONF_PERF_HOURS and CONF_STRESS settings of the measured lanes are gone, along with the class runs, the step-load stress ladder and the AQL probe, which Veredictum 0.3.1 removed (#4452).

  • The separate SMART step of the Conformance workflow, the smart stage of scripts/conformance.sh with its CONF_SMART_OUT, and docs/conformance/party/ferroehr/ixit-smart.json are gone: the one catalogue run covers the SMART documents. The party statement no longer carries the performance block, which the Veredictum 0.3.1 statement schema refuses (#4503).

  • The empty linkage storage domain is gone: its schema, its ferroehr_linkage role, the [storage.linkage] and [storage.migrate.linkage] configuration tables, its pool, the database.linkage and backup.linkage chart values, the ferroehr-backup-linkage compose job and the btree_gist extension only its table needed. A configuration that still sets [storage.linkage] or [storage.migrate.linkage] is refused at boot as an unknown key. Cohort and AQL demographic join crossings still write their linkage-domain access records. Upgrade note: remove those keys and values before you upgrade. On a co-located database, migration audit/0021_drop_linkage_domain.sql drops the schema, the role and the extension at the first boot, and raises a NOTICE for any of them it cannot drop. When the linkage domain lived in a database of its own, [storage.audit] names another database, or the cluster holds more than one FerroEHR database, run in each such database, as a superuser or a member of ferroehr_linkage: DROP SCHEMA IF EXISTS linkage CASCADE; REVOKE ALL ON SCHEMA ext FROM ferroehr_linkage; REVOKE ALL ON ALL FUNCTIONS IN SCHEMA ext FROM ferroehr_linkage; ALTER DEFAULT PRIVILEGES IN SCHEMA ext REVOKE EXECUTE ON FUNCTIONS FROM ferroehr_linkage; DROP OWNED BY ferroehr_linkage; DROP EXTENSION IF EXISTS btree_gist; then DROP ROLE ferroehr_linkage; once per cluster. The released grant files still name ferroehr_linkage, so a first install whose migrator holds no CREATEROLE still provisions that role and drops it after the first boot (#4684).

  • A party migration drops the database function party.resolve_national_identifier. No server path has called it since the linkage domain was removed (#4697).

  • The viewer no longer reads LEPTOS_SITE_ADDR: a value naming another address than server.listen no longer refuses the start. server.listen (FERROEHR_VIEWER__SERVER__LISTEN) sets the listen address (#4725).

  • Migration audit/0022 drops the SQL functions audit.reap_audit_events(integer), audit.reap_audit_events_years(integer) and audit.reap_audit_events_before(timestamptz), along with their EXECUTE grants. The server reaps the audit trail through audit.reap_audit_events_due, which honours each record’s own retention end, and nothing called the three any longer. The spec_profile documentation no longer names a LANG generation, since the profile selects the RM and BASE generations only (#4770).

  • Breaking: FerroEHR no longer publishes change events to a message broker. The AMQP publisher, the lapin dependency, the events cargo feature, the [events] section, the /admin/event_subscription routes with their subscription store (clinical migration 0030_outbox_feed.sql drops the table), the events health indicator and events_published metric, the plaintext_broker deployment gap, the chart’s secrets.eventsUrl and config.events values, and the viewer’s Subscriptions page are gone. A configuration that still carries [events], or plaintext_broker in deployment_accepts, is refused at boot. Broker delivery moves to FerroBRIDGE, where it is in progress (FerroHEALTH/FerroBRIDGE#1367): it is to read the new outbox feed and publish to the broker, so a deployment that needs broker events stays on a release with FerroEHR’s own publisher until a FerroBRIDGE release carries it. Then name its reader in [outbox] readers and configure FerroBRIDGE’s events lane with the broker URL, exchange and queues (Change events (outbox feed)) (#4786).

  • The chart’s terminology.ui value and the terminology overlay’s FERROTERM_UI variable: FerroTERM 0.3 serves no UI of its own, so neither had an effect. A values file that still sets terminology.ui is refused by the values schema; delete the key (#4931).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • An AQL query aimed at named EHRs (the ITS-REST ehr_id parameter) reads those EHRs’ objects first and each object’s own content after them, whatever the planner statistics say. When an archetype’s rows fell out of the most-common-value list of the archetype statistics, the planner estimated them at one row and could start from that archetype over the whole population, evaluating every projected path for every EHR’s rows before the EHR filter: the storage bench’s wide projection over one EHR read about 10 000 pages per call in some rounds and 1 200 in others. A node contained in a versioned object is now read inside its parent on such a query; results do not change (#4487).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

  • AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri argument whole when its value carries a ?, as the QUERY master03 example url=http://snomed.info/sct?fhir_vs does. Before, the arguments were read from after the first ? anywhere in params_uri, so the value-set URL reached the terminology server truncated and the arguments before it were lost. A params_uri that starts with a path still has its arguments read from after the path’s ? (#4490).

  • An ADL 2 template whose constituent archetype the CDR does not hold is refused at upload, with a 422 whose errors carry one VARXR entry per missing archetype (AOM2 C_ARCHETYPE_ROOT validity rule VARXR); the 2013 CKM t_encounter_opt_test, built on the refused apgar, no longer loads, so the sandbox library pin is 231 accepted / 91 refused. A stored template whose constituent was deleted later answers its JSON form and example with the same VARXR list, and the viewer says so on every tab of the template, names each missing archetype and links to uploading it, instead of a raw 422 (#4505).

  • ADL 2 validation admits a constraint on an RM computed attribute. VCARM accepts an attribute “stored or computed” (AOM2 master04.5 §Validity Rules: C_ATTRIBUTE), and a function with no arguments is a computed attribute (master04.2 §Attribute Nodes), so an archetype that constrains DV_PROPORTION.is_integral, DV_QUANTITY.is_integral or EVENT.offset is no longer refused with 422. The generated RM model (openehr_rm::v1_2::model) now carries each class’s computed attributes from the BMM functions (RmClass::computed_attributes, computed_attribute). openEHR-EHR-CLUSTER.microscopy_prostate_carcinoma.v1.0.0 from the 2013 CKM ADL 2 export now uploads, and the sandbox archetype-library pin moves to 232 accepted and 90 refused. DV_QUANTITY.property, which no RM release defines, stays refused. The nine openehr-* crates step to 0.0.124 (#4513).

  • DELETE /ferroehr/rest/extensions/v1/definition/artefact/adl2/{artefact_id} refuses with 409 Conflict while another stored ADL 2 artefact does not compile without the one deleted (a slot filler, a use_archetype reference or a specialisation parent, directly or through another artefact), and the message names every one of them. Before, the delete succeeded and left each template built on the archetype unable to compile (#4522).

  • The viewer’s ADL 2 template upload names each violation the CDR found when it refuses the source: every AOM2 rule code with the CDR’s words for it, so a template whose constituent archetype the CDR does not hold names that archetype (VARXR). The upload now asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details, and reads the stored template’s id from the ETag of the 201 (#4523).

  • An ADL 2 operational template whose included archetype reuses the template’s node ids uploads again, where it was refused with 422 under VCOSU. Node-id uniqueness now holds per archetype and restarts at each C_ARCHETYPE_ROOT (AOM 2 VCOSU, “unique within the archetype”), and on an operational template the codes and interior node ids below a root are judged against that constituent’s entry in component_terminologies (OPT 2 master03-opt_raw). A duplicate node id inside one archetype is still refused. The nine openehr-* crates step to 0.0.125 (#4529).

  • The viewer’s dashboard draws the EHR or composition count as unavailable, with the reason in plain words, when the CDR answers its count query with a result set of another shape (no row, several rows, an empty row, several values, or a value that is no whole number), where it drew 0 before. The template Releases table draws such a count the same way. A real zero still draws 0 (#4533).

  • An ADL 2 operational template upload runs the terminology integrity rules (VTLC, VATCV, VTVSID, VTVSUQ, VTVSMD, VTTBK, VTCBK, and VTSD where the level is carried) on each component_terminologies entry, and each violation names the component’s archetype id (#4536).

  • A COMPOSITION committed against an ADL 2 operational template is validated against the template’s leaf constraints, where it was stored with 201 whatever it held. An inlined archetype recorded by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability) now matches its template node, so the C_INTEGER, C_REAL, C_STRING, C_BOOLEAN, tuple, value-set and cardinality constraints below it answer 422 (AOM 2 master04.5). An archetype root the template does not inline is refused as a node the template does not define, and a C_INTEGER or C_REAL list ({10, 15, 20}) admits each of its values where only the first was read. Canonical, FLAT and STRUCTURED commits take the same check. The nine openehr-* crates step to 0.0.127 (#4540).

  • A versioned read whose version path segment is no SEMVER version and no {major} or {major}.{minor} prefix, such as v1.x, 1.x or 01, now answers 400 before any lookup, where it answered 404 (parameters/path/version.yaml, responses/400.yaml). This covers the ADL 2 template read GET /definition/template/adl2/{template_id}/{version}, the stored-query read GET /definition/query/{qualified_query_name}/{version} and both stored-query executions at a version, GET and POST /query/{qualified_query_name}/{version}. A well-formed version the server does not hold still answers 404 (#4541).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the content below an inlined archetype that the composition records by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability). The read left out everything below such a root with no error; it now matches the root to its template node the way the commit does. The nine openehr-* crates step to 0.0.128 (#4551).

  • A COMPOSITION committed against an ADL 2 operational template whose C_INTEGER or C_REAL constraint is a union of ranges (for example magnitude matches {|0..5|, |10..15|}) is refused with 422 when the value lies outside every range. Such a value was stored. The served Web Template does not change: its input validation.range holds one range, so a union leaves it out, as before, and the check runs on the commit (#4552).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the value of an ELEMENT that holds a DV_INTERVAL whose lower and upper the template constrains. The read left the value out with no error (#4559).

  • A FLAT or STRUCTURED read of a COMPOSITION returns the value of an ELEMENT that holds a DV_INTERVAL whose template leaves the interval’s lower and upper unconstrained, for ADL 1.4 and ADL 2 templates. The read left the value out with no error, and a FLAT or STRUCTURED commit could not name the parts (#4559).

  • A COMPOSITION written from FLAT or STRUCTURED against an ADL 2 operational template carries what a canonical body built to that template carries. A node the template gives no term text, such as the HISTORY and ITEM_TREE of an OBSERVATION, is named for its RM type (History, Tree) instead of its bare id-code (id2). Every inlined archetype root records the complete identifier the template names (…cnf_minimal.v1.0.0, BASE master07-referencing §Reconstitutability) in its archetype_node_id and archetype_details.archetype_id, and archetype_details.rm_version is the template’s rm_release. Before this change the root recorded the interface form (…v1) and the server’s RM 1.2.0. Compositions built against an ADL 1.4 template keep their major-version archetype ids and RM 1.2.0 (#4560).

  • ferroehr report in a release build names the openehr-* crate versions that release published to crates.io: the release lanes stamp the published version into the checkout before they build the server, where the report used to name the floor the source tree holds. A new openehr_crates_origin field says published for a release build and source_tree for any other, a pre-release included (#4565).

  • An ADL 2 archetype reference resolves to one defined stored release: an interface reference (….v1) takes the newest release of its major version, a reference with a minor or patch version narrows the match to it, and an alpha or beta version matches only its full identifier (BASE Artefact Identification, Referencing). With several releases of one archetype stored, a template upload, an operational template and the delete’s dependents check used whichever release PostgreSQL returned last, and a reference to a major version no stored release carries resolved to another major version (#4567).

  • An ADL 2 upload and an ADL 2 delete no longer race. An upload validated before a concurrent delete of an archetype it uses committed could store a template whose constituent was gone. Both now hold one transaction-scoped advisory lock from the repository read to the commit, so the second write decides on what the first committed: an upload after the delete is refused with VARXR, and a delete after the upload is refused with a 409 naming the template (#4568).

  • The syslog form (DICOM PS3.15, IHE ITI-20) of an EHR-level audit record names its object beside the patient, as a URI participant object with the id its FHIR AuditEvent names. The record of placing, releasing or expiring an EHDS Art. 3(3) access delay names the held composition as access-delay:<vo>, so an Audit Record Repository fed by syslog can tell which composition a hold covers (#4569).

  • An ADL 2 template or archetype that references two major versions of one archetype (….v1 and ….v2), or one archetype under two namespaces, now compiles against both: the openehr-adl repository holds an artefact under its namespace, archetype and major version, and resolves each reference within that key (BASE Artefact Identification, Referencing). A reference without a namespace now resolves in the namespace of the artefact that makes it, and never in another namespace (#4579).

  • A read of a stored ADL 2 template by a partial id or a version (GET /definition/template/adl2/{template_id}) now resolves the way template compilation does: with v1.0.0 stored beside v1.0.0-rc.1, ….v1, ….v1.0 and version=1.0.0 answer the release in either upload order, and the release candidate is read only by its full version (BASE Artefact Identification, Referencing) (#4580).

  • A write that sends no committer, on a route that takes a bare RM body (the EHR_STATUS, COMPOSITION and DIRECTORY writes, and the demographic party and relationship writes), now records a Bearer caller’s login as issued by the token issuer, as EHR creation and every other default committer already did. It recorded the issuer ferroehr for every caller before; a Basic caller’s login is still issued by ferroehr (#4585).

  • The viewer’s ADL 1.4 template upload opens the stored template’s page, read from the ETag (else the Location) of the CDR’s 201, where it went to an address with no template id. A refused upload names each violation the CDR found, its rule code or path and the CDR’s words for it, as the ADL 2 upload does: the upload asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details (#4596).

  • A 422 validation refusal tells the client to send Prefer: return=representation only when the request did not send it; a request that sent it receives the violations in errors and a detail without the advice (#4597).

  • An EHR Extract import refusal for a received version now names that version’s position in the extract (chapters[i]/items[j]/item/versions[k]), so an importer can find the defective ORIGINAL_VERSION in a large extract (#4602).

  • The import reads a version’s other_input_version_uids with the parser the CONTRIBUTION path uses and refuses an entry that is not an OBJECT_VERSION_ID, where it used to drop the entry. On the wire the extract decode already refuses such an entry with 422, so the import stays all or nothing (#4602).

  • An AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts within the major: org.a::… no longer finds data recorded in org.b, an id without a namespace finds only data recorded without one, and .v1.4.2 no longer finds data recorded under v1.0.0 (BASE Artefact Identification §Referencing and §Supporting Archetype-based Querying; #4629, #4630).

  • A query naming a parent archetype finds the data of a specialised child release only when that release declares a parent release at or after the queried one. Before, once one stored release of the child qualified, the data of every release of it was returned (BASE Artefact Identification, Referencing §Supporting Archetype-based Querying; #4637).

  • POST /definition/template/adl1.4 and POST /definition/template/adl2 answer 400 for every invalid template, the only refusal ITS-REST 1.1.0 declares for both uploads besides the 409 duplicate. An OPT that breaks the OPT 1.4 schema, an ADL 2 source that breaks an AOM2 rule such as VCORM or VCARM, and an ADL 2 template whose constituent archetype is not held were refused with 422. The problem keeps its detail and, with Prefer: return=representation, its errors. Upgrade note: a client that read a 422 from a template upload reads a 400, with the same detail and errors, under the about:blank problem type (#4643).

  • A directory create or update whose FOLDER carries an empty folders array, at the root or in a sub-folder, is refused with 400 and commits nothing. The ITS-JSON FOLDER schema gives folders minItems: 1, and the canonical JSON reader now refuses the empty array the same way in every document that carries a FOLDER, a CONTRIBUTION included. The viewer leaves folders out of a folder with no sub-folders (#4644).

  • A canonical XML FOLDER, EHR_STATUS, PERSON, AGENT, GROUP, ORGANISATION, ROLE or PARTY_RELATIONSHIP document now has the root element items with xsi:type naming its class, the element the ITS-XML schemas publish for a LOCATABLE (Structure.xsd). The directory reads (latest, by version and at a time), the EHR_STATUS reads, the demographic reads and the demographic template examples served <folder>, <ehr_status>, <person> and the like, which no ITS-XML schema declares. XML input accepts the same form (#4645).

  • POST /definition/template/adl1.4 builds the template’s Web Template before it stores the template, so a template it cannot build is refused with 400 and nothing is stored. Once the template is stored, the upload answers 201 with its Location; with Prefer: return=representation and a Web Template Accept it serves the Web Template built before the store, and the OPT if that cannot be served. It no longer answers 422 for a template it has already stored (#4647).

  • A composition whose HISTORY carries events: [], and an AGENT, GROUP, ORGANISATION or PERSON carrying languages: [], are refused with 400 and commit nothing. The ITS-JSON schemas give both arrays minItems: 1, and the canonical JSON reader now refuses the empty array as it does an empty FOLDER folders. The FLAT builder no longer writes an empty events for a HISTORY that has no event (#4648).

  • A write to a restricted COMPOSITION (POST, PUT or DELETE on /ehr/{ehr_id}/composition) refused under a GDPR Art. 18 restriction now answers with the processing-restricted problem type, as the reads of a restricted record already did, in place of the generic forbidden type. The status stays 403 (#4657).

  • When an EXTRACT_SPEC criterion’s AQL runs past the query time budget, the 400 body now states the timeout as plain text. It used to start with a control-character marker that was only ever meant for the server itself (#4657).

  • Under a separated clinical credential, POST /ehr, PUT /ehr/{ehr_id} and an EHR_STATUS update that name a subject no longer fail with a database error: the subject pseudonym guard reads the deployment posture, and the clinical role now holds SELECT on ext.posture through the ext migration 0009 (#4675).

  • A cohort that names its subjects now finds, under [ehr_index] provider = "local", the EHR created over POST /ehr with that subject in its EHR_STATUS; it read a cross-reference table no route wrote, so it matched nothing. A cohort predicate and the AQL demographic join resolve a party to the EHRs whose EHR_STATUS.subject names it. The EHR_STATUS subject is now the one source of which EHR belongs to which subject. Upgrade note: linkage.subject_ehr and the linkage.erase_ehr function go with the linkage schema, which audit migration 0021 drops with the linkage domain. Rows written there through the service API are dropped with the table; the EHRs they named keep their subject in their own EHR_STATUS (#4675).

  • The identifier scan ([privacy.identifier_scan]) no longer refuses a DV_IDENTIFIER whose type declares an item rather than a person, such as a laboratory filler_order_number that passes the NHS Number check by chance. A DV_IDENTIFIER in FEEDER_AUDIT.originating_system_item_ids or feeder_system_item_ids is scanned only when its type names a person identifier. A type naming a person (patient, NHS, BSN, social security and similar) keeps the value scanned in every slot, and free text and untyped identifiers in clinical content are scanned as before. The default rule set is unchanged; a deployment that knows its jurisdictions still narrows rules, so a Dutch one with rules = ["nl-bsn"] never claims an NHS Number (#4702).

  • An inline primitive constraint in a parsed archetype or template now carries the node id AOM2 ADL_CODE_DEFINITIONS.Primitive_node_id defines, at9999 in an at-coded archetype and id9999 in an id-coded one, where it carried the constant’s name Primitive_node_id. An ADL 1.4 archetype converted to ADL2 gets the id-coded form (#4739).

  • The ABAC policy-enforcement point checks every item-tag operation as it checks its target: a COMPOSITION tag read after the call like the COMPOSITION read, a COMPOSITION tag write before the call with the tagged version’s template, category and origin, and the EHR tag read before the call like the EHR read. These operations were unchecked before, so a patient-scoped or template-scoped rule now refuses a tag read or write wherever it refuses the target’s read or write. A COMPOSITION tag read also falls under the EHR_ACCESS privacy ceiling and its audit record names the tagged COMPOSITION (#4768).

  • A primitive constraint inside a rule or a slot assertion of an id-coded ADL 2 archetype carries id9999 as its node id, as org.openehr.am.aom2.adl_code_definitions §Constants defines Primitive_node_id for id-coded archetypes, where it carried at9999 before. An ADL 1.4 archetype converted to ADL 2 recodes its slot assertions the same way. openehr_adl::rules::parse_rules_body and parse_slot_assertions take the archetype’s NodeCoding, which openehr_adl::codes now exports (#4770).

  • Under spec_profile = "stable", a canonical-JSON party read lists in reverse_relationships every PARTY_RELATIONSHIP another party’s current version carries with this party as its target, one LOCATABLE_REF each naming the source party’s version container and the relationship’s path, as RM 1.1.0 §Party Relationships requires. The development profile, whose RM 1.2.0 makes the attribute a function, is unchanged. The released-profile conformance run now judges the server on the stable profile (#4775).

  • Under the development profile, POST /definition/template/adl2 refuses with 400 a template whose description has no original_author or an empty lifecycle_state, naming RESOURCE_DESCRIPTION.Original_author_valid or RESOURCE_DESCRIPTION.Lifecycle_state_valid. BASE 1.3.0 states both invariants (org.openehr.base.resource.resource_description.adoc §Invariants). The stable profile’s BASE 1.2.0 states neither, so it still accepts such a template (#4776).

  • A stable-profile party read in canonical XML carries reverse_relationships (#4778), as the canonical-JSON read does: the XML response is written through the RM 1.1.0 generation, whose PARTY declares the element the ITS-XML RM Release-1.1.0 Demographic.xsd carries. A development-profile read stays without it.

  • A database that accepts connections slowly is no longer recorded as down. A probe connect still in progress after its two-second bound now leaves the acquire queueing on the pool within acquire_timeout_secs, so the request is served once the connect completes, or answers the pool-timeout 503 with Retry-After. Only a refused or reset connect, or a server shutting down (57P01, 57P02, 57P03), marks the database down and answers the outage 503 at once (#4782).

  • A database that resets connections is now answered as an outage at once on every request, the 503 with Retry-After a refused connect gets, instead of each request opening a connection of its own that the database resets. A connect of the pool’s own that is refused, reset or turned away as shutting down (57P01, 57P02, 57P03) records the outage as the probe does, and a probe that connects again clears it. A connect that is only slow still queues (#4807).

  • An EHR whose ferroehr.access_control.v1 settings carry the scheme’s _type but do not parse is held closed. Before, such an EHR was served at the default level, as if it carried no settings: open to every caller unless the server default was restricted, with no privacy ceiling on its compositions. Now every request that names the EHR, its reads and its export included, is refused with 500, no query reaches it, and the server logs an error naming the EHR (#4846).

  • A CONTRIBUTION or an EHR Extract import that carries ferroehr.access_control.v1 settings which do not parse is refused with 400, naming the first setting that does not parse. Before, such settings were stored, and since #4846 they hold the EHR closed, with no way to repair it over REST (#4854).

  • POST /message/export checks the EHR_ACCESS settings of every EHR its manifest names, by ehr_id or subject_id, before it reads any content, and refuses the whole export with the 403 or 500 a read of that EHR gets. Before, the manifest form passed no EHR_ACCESS gate: any caller the route admitted could export an EHR whose access list refused them, or one held closed because its settings do not parse (#4864).

  • An admin archive load refuses an archive whose ferroehr.access_control.v1 settings do not parse with 400, naming the setting and the EHR, before anything of it is written. Before, the load stored such settings, and the EHR they belong to was then held closed (#4865).

  • With attribute-based access control on, both EHR Extract export routes pass the ABAC pre-check: GET /message/export/{ehr_id} the one a read of the EHR passes, and POST /message/export the patient scope on every EHR its manifest names, before any content is read. Before, neither route reached the ABAC policy-enforcement point, so a caller bound to one patient could export any EHR the coarse layers admitted (#4869).

  • An access service that exports an opted-out EHR through a POST /message/export manifest under the EHDS Art. 10(2) vital-interest override now records the override against that EHR, once, before any content is read, as a read of the EHR by its path does. Before, the manifest form exported the EHR under the override and wrote no override record (#4870).

  • With attribute-based access control on, both EHR Extract import routes (POST /message/import and POST /message/import/{ehr_id}) pass the ABAC pre-check a direct write passes, the patient scope on the target EHR included, and are audited as EHR writes. Before, neither route reached the ABAC policy-enforcement point, so a caller bound to one patient could import into any EHR the coarse layers admitted, and the request was audited under the generic application-activity class (#4872).

  • A REST request whose body cannot be read is refused with 400, and one whose path parameter is not valid UTF-8 with 400, instead of running as a request with an empty body or no path parameters. The body-limit check answers 413 only for a body over its limit; any other read failure is 400 (#4895).

  • An AQL request abandoned while the EHR_ACCESS settings index loads, for example behind a table lock, no longer keeps its pool connection until the statement ends: each waiting request started the load again, so ten abandoned queries held nine of ten connections (#4903).

  • An archetype root converted from an ADL 1.4 operational template defines each slot that replaced an embedded root in every language of its terminology, with the embedded root’s concept term, and drops a term keyed by no at or ac code (such as a tool’s °C unit label) with a conversion_details entry. Before, such roots failed ADL 2 validation with VATID and VATCV (#4917).

  • The ADL 1.4 to 2 converter no longer writes a property attribute on DV_QUANTITY, which no RM release defines, from an ADL 1.4 archetype or an operational template; it drops the C_DV_QUANTITY property constraint and records it in conversion_details. Before, a converted archetype carrying one was refused with VCARM (#4920).

  • The ADL 2 WebTemplate builder gives each use_archetype sibling of one archetype that the template names with a fixed name/value its own name-qualified path ([openEHR-EHR-SECTION.adhoc.v1,'Medication summary']), as the ADL 1.4 builder does, so a composition holding several such sections, such as an International Patient Summary, validates instead of being refused with “too many occurrences” (#4927).

  • openehr-adl: print::print refuses a terminology constraint ADL2 has no syntax for, such as the ADL 1.4 [openehr::433] an unconverted 1.4 archetype keeps, with PrintError::TerminologyConstraint, where it used to write text that parse_artefact then refused (ADL2 master04.5 §Terminology Constraints) (#4937).

  • A LATEST_VERSION AQL read joins an object’s node rows on its head’s content key. Since a version identical to its predecessor reads its predecessor’s node rows, such a read could reach a shared row through every version of a long identical chain before keeping the head. vo_head.trunk_head_content_key now names the head’s node rows and every LATEST_VERSION read joins on it, and the version partitions declare a vo_id distinct count (n_distinct = -0.5) so the planner never prices a probe of one object’s versions at one row (migrations clinical/0034, party/0026). The storage generation moves to generation-6 (ext/0012). The flat population read the storage bench measures stops at its limit only with the planner fix of the next entry (#4944).

  • A population AQL read ordered by the composition uid with a LIMIT (SELECT … FROM EHR e CONTAINS COMPOSITION c CONTAINS OBSERVATION o[…] ORDER BY c/uid/value LIMIT 50) stops at its limit again. Since a version identical to its predecessor reads its predecessor’s node rows, the store holds fewer node rows, and the planner’s estimate of the answer fell to a tenth of its size: it then joined the whole population into a sort, and the storage bench’s flat two-leaf read took thirteen times the pages. The ADL 1.4 specialisation test of an archetype predicate is now a byte range the archetype extended statistics read (arch_concept ~>=~ 'concept-' AND arch_concept ~<~ 'concept.', the same strings as LIKE 'concept-%'), and a node contained in a COMPOSITION, EHR_STATUS or EHR_ACCESS root joins it on num > 0 in place of the root’s nested-set interval (#4944).

  • The viewer refuses, before sending it, a CDR request whose URL would leave the configured CDR: another scheme, host, port or user, a path outside the product root of [cdr] base_path, the management base or the health family, or a path that holds a . or .. segment in any spelling, a \ or a control character. A browser-supplied id, stored-query name or version of .. was percent-encoded unchanged and resolved away from the ITS-REST base; it now draws the refusal instead. A [cdr] URL that makes no absolute URL stops the viewer at start (#4954).

  • openehr-adl: opt::create_opt takes unconverted ADL 1.4 constituents again. It keeps their 1.4 node codes, so 1.4 template paths such as /activities[at0001]/description[at0002] still resolve, and rewrites each 1.4 term constraint to the form ADL2 spells (master04.5 §Formal Terminology Constraint), the way adl14::convert does. An external code such as [openehr::433] becomes a new at-code bound to http://openehr.org/id/433, and a code list such as [local::at0168, at0169] becomes an ac-code value set. The OPT2 then prints and parses back (#4956).

  • An AQL OR containment branch that binds a versioned object under LATEST_VERSION answers the object once, with the head’s uid, when the head shares its predecessor’s node rows; it answered once for every version reading the shared tree (#4957).

  • Under spec_profile = "stable", an AQL whole-object result whose stored body the profile gate cannot read is refused with an error naming the version; it was served without the assessment (#4960).

  • openehr-lang lexes a long string literal, -- comment, <# … #> plug-in block, embedded URI or contained regexp in constant stack in a debug build. Before, the logos lexer took one stack frame per multi-byte character or escape of such a token, so building a template from ADL archetypes could overflow a thread’s default 2 MiB stack and abort. A 1 MiB token now lexes on a 256 KiB thread. The token stream and its spans are unchanged (#4961).

  • The ADL 1.4 to ADL 2 conversion gives a terminology constraint that names no codes ([openEHR::], [local::]) an ac-code per terminology, with no value set and no binding, and records the terminology in the conversion log. Before, every such constraint in an archetype shared one ac-code labelled as a synthesised value set, whatever terminology it named (#4965).

  • Converting a stored specialised ADL 1.4 archetype to ADL 2 answers 422 when the archetype introduces a term code its stored parent lacks without the specialisation dots its depth requires (at0006 where depth 1 needs at0.1), naming each code. AOM 1.4 requires the dots, and the conversion used to carry such codes into ADL 2 (#4966).

  • The ADL 1.4 to ADL 2 conversion gives a slot that redefines a parent slot (at0004.1) the parent slot’s node id, as AOM 2 VDSSID requires, and the differential form it builds against a converted parent drops archetype slots and proxies the parent holds unchanged and keeps only terms of the archetype’s own specialisation level (VTSD) (#4967).

  • openehr-adl’s opt::create_opt no longer flattens an ADL 1.4 specialised archetype over its parent. An ADL 1.4 archetype is authored in flat form (ADL2 master07.02 §File-naming Convention, AOM2 master03 §Structural Variants), so overlaying a child such as CKM openEHR-EHR-COMPOSITION.report-result.v1 on report.v1 appended every object without a node id a second time: the OPT held two EVENT_CONTEXT objects and its Web Template two /context nodes. An artefact whose adl_version is 1.4 or earlier is now its own flat form, keeps its parent_archetype_id, and needs no parent in the repository; its 1.4 term constraints are still rewritten into the ADL2 form. ADL 2 differential flattening is unchanged (#4979).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names. The zip move is a lock-only update that a security-only 5.5.1 could have carried; the published-image scan found CVE-2026-115905 in 5.5.0 on 2026-11-03, the day before this release was cut, so 5.6.0 ships the fix in place of a separate patch.

  • FerroEHR takes zip 9.1.3 in place of 9.0.4, and the workspace requires 9.0.5 or later. Through 9.0.4 the crate reads a file’s name from the central directory in ZipArchive and from the local file header in its streaming reader, so one crafted archive can name two different files (CVE-2026-115905, CRITICAL). The server links the crate for the administrator’s dump and load archive, so the server binary and the ferroehr image of 5.5.0 and earlier carry the affected version; the viewer image does not. Upgrade from 5.5.0, and until then load through POST /admin/load only archives this FerroEHR wrote (security/vex/rust-releases.openvex.json) (#4892).

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

  • CVE-2026-84243, the glibc gettext LANGUAGE directory traversal (GLIBC-SA-2026-0025), has no Debian fix for trixie yet and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never leave the C locale, so glibc ignores LANGUAGE and opens no message catalog (security/vex/ferroehr-os.openvex.json). In ferroehr-postgres only the container environment the operator writes can set LANGUAGE; no database connection can (security/vex/postgres-os.openvex.json) (#4500).

  • CVE-2026-114690, CVE-2026-114691 and CVE-2026-115455, three libstdc++ defects in the gcc-14 packages of every image, have no Debian fix for trixie and are judged not_affected. Each fix changes only a header template (the parallel-mode search, the std::vector<bool> constructors, wide-stream extraction into std::bitset), so libstdc++.so.6 holds no copy; the ferroehr and ferroehr-viewer programs contain no C++ code (security/vex/ferroehr-os.openvex.json), and in ferroehr-postgres nothing is built in parallel mode and PostgreSQL does not link libstdc++ (security/vex/postgres-os.openvex.json) (#4789).

  • CVE-2026-119371 to CVE-2026-119379 and CVE-2026-119591, ten SQLite defects without a Debian fix, are judged not_affected in ferroehr-postgres: only lslogins reaches SQLite there, through liblastlog2, and nothing in the image runs it (security/vex/postgres-os.openvex.json) (#4789).

5.6.0-rc10 - 2026-11-02

5.6.0-rc10 is the tenth release candidate of 5.6.0, published for testing, and the first candidate published since 5.6.0-rc7. 5.6.0-rc8 and 5.6.0-rc9 were tagged and never published: the vulnerability record of 5.6.0-rc8 refused 86 findings that had no OpenVEX judgement, and the release plan of 5.6.0-rc9 refused the tag because two fragments merged ahead of it were missing from its section. A tag is never moved, so 5.6.0-rc10 carries everything both carried. It carries the CONTRIBUTION break of 5.6.0-rc6: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution read the body through the released NewContribution schema and answer 400 for a lifecycle_state that is no DV_CODED_TEXT, a description that is a plain string or an object without _type, an audit without change_type, an audit typed ATTESTATION or UPDATE_ATTESTATION, a missing audit or versions, and a preceding_version_uid sent as a plain string, and a stated empty system_id is refused with 422 on every write path (#3595, #4543). These notes carry every change since 5.5.0, those of the 5.6.0-rc9, 5.6.0-rc8, 5.6.0-rc7, 5.6.0-rc6, 5.6.0-rc5, 5.6.0-rc4 and 5.6.0-rc3 sections below included. New since 5.6.0-rc7, and so in neither published candidate before this one: two read-only extension routes that convert a stored ADL 1.4 archetype or operational template to ADL 2 (#4655); template uploads that answer 400 for every invalid template and never refuse a template already stored (#4643, #4647); the ITS-JSON minItems of FOLDER.folders, HISTORY.events and ACTOR.languages enforced (#4644, #4648), and a canonical XML LOCATABLE rooted at the items element the ITS-XML schemas publish (#4645); an AQL query naming a parent archetype that finds a specialised child’s data only from a child release whose own parent release qualifies (#4637); conformance_profile derived from the committed conformance record (#4672); a subject resolved through the EHR_STATUS of its EHR, and the empty linkage domain removed with its configuration (#4675, #4684, #4697); a write refused under a GDPR Art. 18 restriction answering with the processing-restricted problem type, and 500 bodies that carry curated text only (#4657); an identifier scan that reads a DV_IDENTIFIER’s declared type (#4702); a validation refusal that advises Prefer: return=representation only to a request that did not send it (#4597); the AOM2 node id of an inline primitive constraint (#4739); a viewer that answers a request other than GET or HEAD without a session with the sign-in page and no longer reads LEPTOS_SITE_ADDR (#4724, #4725); every item-tag operation checked by ABAC in its target’s mode (#4768); and, under spec_profile = "stable", a canonical-JSON party read that lists the relationships naming it as their target in reverse_relationships (#4775); since 5.6.0-rc8, the OpenVEX judgements of the gcc-14 and SQLite findings that stopped it (#4789), the AOM2 node id of a primitive constraint inside a rule or slot assertion of an id-coded ADL 2 archetype, and the three unused horizon reap functions of the audit store removed (#4770); since 5.6.0-rc9, the outbox served to its named readers over a cursor feed in place of FerroEHR’s own broker publisher (#4786, #4766), the access record of a VERSION read naming the content it served (#4660), the BASE 1.3.0 description invariants of an ADL 2 template under the development profile (#4776), a slow database connect no longer recorded as an outage (#4782) while a database that resets connections is recorded as an outage on the pool’s own connect too (#4807), and Prefer read on every route as RFC 7240 §2 defines it (#4722). The published conformance record is a Veredictum 0.4.0 released-profile run against FerroEHR on spec_profile = "stable": 2872 of 3071 rows passed, none failed and none errored, and 199 were skipped, 116 for an option the statement does not select and 83 that rest on development text alone; it replaces the Veredictum 0.1.6 documents (#4451, #4771). The released-profile verdict is now required: a failed or errored row fails the merge queue, the image run on main and the release (#4455). The development-profile run, which is never a verdict, has 2946 rows passed, 3 failed (#4776, #4777) and 2 errored by Veredictum itself. 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: FerroEHR no longer publishes to a message broker, so [events], the /admin/event_subscription routes and the viewer’s Subscriptions page are gone, a configuration that still carries [events] or plaintext_broker in deployment_accepts is refused at boot, and Helm chart 21.0.0-rc9 drops secrets.eventsUrl and config.events (clinical migration 0030 and party migration 0023); the outbox is served over GET /admin/outbox/{reader} and its ack, POST /admin/outbox/{reader}/ack, to the readers [outbox] readers names, under the __ferroehr.outbox scope; broker publishing moves to FerroBRIDGE, where it is in progress (FerroHEALTH/FerroBRIDGE#1367), so a deployment that needs RabbitMQ events stays on 5.6.0-rc7 until a FerroBRIDGE release carries it; under the development profile an ADL 2 template upload whose description has no original_author or an empty lifecycle_state answers 400; an invalid template on POST /definition/template/adl1.4 or POST /definition/template/adl2, an ADL 2 template whose constituent archetype is not held among them, answers 400 where it answered 422, with the same detail and errors under the about:blank problem type; a FOLDER with an empty folders, a HISTORY with an empty events and an AGENT, GROUP, ORGANISATION or PERSON with an empty languages answer 400; a canonical XML FOLDER, EHR_STATUS, PERSON, AGENT, GROUP, ORGANISATION, ROLE or PARTY_RELATIONSHIP document has the root element items with xsi:type naming its class, where it had <folder>, <ehr_status>, <person> and the like; OPTIONS on the API base path sends conformance_profile as BASE+OPTIONS where it sent STANDARD, and a configuration that sets [rest.identity] conformance_profile is refused at boot as an unknown key; an AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts, and a query naming a parent archetype finds a specialised child’s data only from a child release whose own parent release qualifies, so a query may find less than before; a cohort that names its subjects and the AQL demographic join resolve a party to the EHRs whose EHR_STATUS.subject names it, and a configuration that sets [storage.linkage] or [storage.migrate.linkage] is refused at boot and Helm chart 21.0.0-rc9 refuses a values file that sets database.linkage or backup.linkage, so remove those keys and values before you upgrade and, where the linkage domain lived in a database of its own, run the SQL the Removed entry gives; a POST, PUT or DELETE on a COMPOSITION under a GDPR Art. 18 restriction still answers 403, now with the processing-restricted problem type in place of forbidden; the 500 bodies of the archive dump and load, a failed expansion of a party’s protected identifiers and two query failures carry fixed text that echoes no value from the request, the archive or the terminology server; the identifier scan no longer refuses a DV_IDENTIFIER whose type declares an item rather than a person, and scans one in a FEEDER_AUDIT only when its type names a person identifier; the Veredictum 0.1.6 conformance documents (CONFORMANCE_STATEMENT.md and the rest) and the CONF_PERF_CLASS, CONF_PERF_HOURS and CONF_STRESS settings of the measured lanes are gone; with ABAC on, a patient-scoped or template-scoped rule refuses an item-tag read or write wherever it refuses the read or write of the tag’s target, and a COMPOSITION tag read falls under the EHR_ACCESS privacy ceiling; under spec_profile = "stable" a canonical-JSON party read carries reverse_relationships, while the canonical-XML read does not yet (#4778) and the development profile is unchanged; a write whose committer names another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, answers 403, and a direct write whose committer breaks its own rules answers 422; a write that sends no committer on a bare RM body route records a Bearer caller’s login as issued by the token issuer, where it recorded ferroehr; a FLAT or STRUCTURED composition under an ADL 2 template records the complete archetype identifier and the template’s rm_release, and the served Web Template writes a bound coded input, COMPOSITION.category among them, as the bound terms (openehr::433) where it wrote the local at-code; a CONTRIBUTION commit must take the shapes above; a write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as committer and may change only what the person committed, and a service auth_failure answers 403 where it answered 500; a versioned read whose version is no SEMVER version or prefix answers 400 where it answered 404; the served Web Template of an ADL 2 template writes a data value’s aqlPath without an [idN] predicate; a composition that breaks an ADL 2 template’s leaf constraint answers 422; every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes, so keep it longer than the idle timeout of your clients and load balancers; a client that wrote a space as + in a query string sends %20; a party delete names the latest OBJECT_VERSION_ID in its path and answers 400 for a versioned object id; a body the RM refuses after the ITS-JSON schema admits it answers 422, and a misplaced CONTRIBUTION member, a Simplified Format commit without openehr-template-id and a flat |raw value without _type answer 400; a tag update under Prefer: return=identifier answers 200 with the tag list; with [authz.rbac] enabled a token holding only system/ SMART on openEHR scopes reaches clinical operations without a role; the viewer image serves no /pkg/ and reads no LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, and its Content-Security-Policy is script-src 'self'; style-src 'self'; Helm chart 21.0.0-rc9 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data, and ext migration 0008 names generation 4, the generation with the node header columns, and changes only the label; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; clinical 0024 and party 0015 index the version uid, ext 0006, clinical 0025 and party 0016 add the ELEMENT value search index, built over every hot node row of an existing store, party 0017 adds the archetype statistics, and clinical 0026 and party 0018 grant MAINTAIN on the partitioned tables to the serving roles; clinical 0027 and party 0019 add the access delay, and ext 0007, clinical 0028 and party 0020 add the node header columns with no data migrated; audit migration 0019 gives the search token index a pending list of 512 kB, and audit 0020 adds the hash index on request_id; clinical 0029 adds the access opt-out; ext migration 0009 grants the clinical role SELECT on ext.posture, party 0021 drops the database function party.resolve_national_identifier, audit 0021 drops the linkage schema, the ferroehr_linkage role and the btree_gist extension, raising a NOTICE for any of them it cannot drop, party 0022 restates the comment on party_relationship_target.source_vo_id with no change to data, and audit 0022 drops the functions audit.reap_audit_events, audit.reap_audit_events_years and audit.reap_audit_events_before, which nothing called; clinical migration 0030 drops the event subscription table, and it and party migration 0023 replace the outbox’s published_at and its two indexes with an index on committed_at; the linkage migration set is gone. Helm chart 21.0.0-rc9 carries appVersion 5.6.0-rc10. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j, and the OpenVEX judgements of CVE-2026-84243, of three gcc-14 libstdc++ findings and of ten SQLite findings.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The operations page gains a storage sizing section, measured on 5.6.0-rc5: about 21.6 KB of clinical storage per version of a vital signs composition (2.0 times its canonical JSON, 85 % of it node rows) and 155 KB for an International Patient Summary, about 1.2 KB per audit record of a read, the effect of [audit.store] retention on growth, pg_wal up to max_wal_size, the REINDEX INDEX CONCURRENTLY clinical.node_hot_pkey step (20 % of that index after heavy concurrent writing) and a query for the size per relation, with the method to reproduce the figures (#4304).

  • EHDS Art. 5: what a person inserts in their own EHR is marked as theirs, and they cannot alter what a professional committed. A write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as its AUDIT_DETAILS.committer, its external_ref naming the login, whatever committer the request names; a COMPOSITION it commits must name PARTY_SELF or PARTY_RELATED as composer; and its update or deletion of a version whose committer is not PARTY_SELF, directly or inside a CONTRIBUTION, is refused with 403. A professional’s writes are unchanged. The contribution list extension gains inserted_by_person per row, and the viewer names a PARTY_SELF committer “The person or their representative”. A service auth_failure status now answers 403 where it answered 500 (#4350).

  • The EHDS Art. 3(3) delay of the person’s own access to newly registered data, a Member State option and off by default. The new [access_delay] section (enabled, categories, hours) holds every new COMPOSITION of a listed [audit.categories] category from requests that may be the person’s own (a bearer token with a SMART patient/ grant or the ABAC patient claim, also beside user/ or system/ grants, as ferrohealth-smart 0.8.7 decides it) for at most hours: a read answers 404 as for an unknown composition, and a query, an export and a CONTRIBUTION leave it out, while a professional reads it as before. The treating professional releases a held composition from the clinical route POST {ext}/ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/access-delay/release, with the checks of a write to that EHR, and the controller with POST /admin/access-delay/release; a professional holds one by hand within the limit with POST /admin/access-delay and reads the register with GET /admin/access-delay; every hold ends at its limit, and placing, releasing and expiry are each an access record. The subject-scoped ITI-81 read of the person’s own trail (EHDS Art. 9) leaves out the records about a held composition until it is released or expires, then serves them with their original times (ehds-logging 0.3.3) (clinical migration 0027_access_delay.sql, party migration 0019_access_delay_column.sql) (#4351).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • The Conformance workflow runs the Veredictum smart documents (smart_discovery, smart_scopes) against the same FerroEHR build after the Veredictum Action, which forwards no SMART Platform base URL and no scoped tokens. The same pinned image runs veredictum test with --platform-url and docs/conformance/party/ferroehr/ixit-smart.json, minting each access token in memory from the committed test issuer the SMART overlay trusts, and the job uploads its results, JUnit and CTRF views beside the catalogue’s and adds its summary under a heading of its own. scripts/conformance.sh gains the smart stage, and run drives it after the catalogue, so a local run covers SMART too (#4450).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

  • The EHDS Art. 10 opt-out from access through the access services of Arts. 4 and 12, a Member State option and off by default. The new [authz.rbac] access_service_roles key declares the roles of the access-service callers; while an EHR carries the opt-out, a caller holding one reads it as absent: a route naming the EHR or an object in it answers 404 as for an unknown EHR, a population query leaves its data out, a scoped query, an export and the lookup by subject find no such EHR, and a write to it is refused as absent. Every other caller, the provider’s own care included, reads it as before. The controller records the opt-out on the person’s instruction with POST /admin/access-opt-out, reverses it with POST /admin/access-opt-out/reverse (the register keeps the reversal) and reads the register with GET /admin/access-opt-out; recording is refused with 400 while no access-service role is declared. An access-service caller that declares one of [audit] emergency_purpose_codes and does not act for the person reads the EHR under the Art. 10(2) vital-interest override, with the emergency mark on its access record and an access-opt-out-override record against the EHR. Recording, reversing and reading the register are access records (clinical migration 0029_access_opt_out.sql) (#4516).

  • Two read-only extension routes convert stored ADL 1.4 definitions to ADL 2: GET /ferroehr/rest/extensions/v1/definition/archetype/adl1.4/{archetype_id}/adl2 for a stored archetype and GET /ferroehr/rest/extensions/v1/definition/template/adl1.4/{template_id}/adl2 for a stored operational template, which converts to one ADL 2 archetype per embedded archetype root. Each answers 200 with application/json: (#4655)

    • the ADL 2 source of every root;
    • each root’s conversion log and its conversion_details, which name what the ADL 2 form drops or changes;
    • the slot fills that joined the roots. An unknown id answers 404. A stored source that no longer converts answers 422. No openEHR specification defines the conversion, so both routes are FerroEHR’s own design.
  • The commit-event outbox feed. [outbox] readers names the readers (FERROEHR__OUTBOX__READERS); boot registers each one and marks every reader the list no longer names inactive, so it holds the retention prune back no longer (#4766). GET /admin/outbox/{reader}?limit=&domain= answers the PHI-free envelopes after the reader’s cursor in seq order, and POST /admin/outbox/{reader}/ack with {"last_seq": n} moves the cursor (204; 409 below the cursor or past the highest row; 404 for an unknown or inactive reader). Both routes need the __ferroehr.outbox scope, every read is an access record, and the clinical and party outboxes each keep their own seq and cursor. A read waits for a commit that drew a lower seq to end, so a reader never acknowledges past a row that becomes visible later. The prune now counts from the commit instant and keeps every row an active reader has not acknowledged (migrations clinical/0030_outbox_feed.sql, party/0023_outbox_feed.sql). With no reader named, no commit writes an outbox row (#4786).

Changed

  • Breaking for clients that relied on the old leniencies: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution now read the body through the released NewContribution schema, and refuse every shape it does not define. Each refused shape answers 400: (#3595)

    • a lifecycle_state in the CODE_PHRASE shape {"terminology_id": …, "code_string": …}, as a plain string, or as a bare {"value": …}. Send a DV_CODED_TEXT: {"value": "complete", "defining_code": {"terminology_id": {"value": "openehr"}, "code_string": "532"}};
    • a description sent as a plain string, on the contribution audit, on a commit_audit or on an attestation. Send a DV_TEXT or DV_CODED_TEXT object: {"_type": "DV_TEXT", "value": "…"};
    • a commit_audit or contribution audit without change_type. The server no longer infers creation or modification from preceding_version_uid; send the code of the change, for example 249 (creation), 251 (modification), 523 (deleted) or 666 (attestation);
    • a commit_audit or contribution audit whose _type is ATTESTATION or UPDATE_ATTESTATION, or any value other than UPDATE_AUDIT or AUDIT_DETAILS. Send the audit as UPDATE_AUDIT (or with no _type) and put each attestation in the version’s attestations list as an UPDATE_ATTESTATION. A 666 (attestation) version carries its attestations there too, and one with no attestation is refused;
    • a missing audit, which answered 422 before. A missing versions list, an attestation without reason or is_pending, a system_id that is not a string, and a preceding_version_uid sent as a plain string instead of {"value": "…"} are refused the same way.
  • An attestation’s own system_id is kept, and one that states none takes the system_id of the version audit it belongs to instead of the server’s. An attestation committer is now checked against its PARTY_PROXY rules like every other committer, so an invalid one answers 422 (#3595).

  • The served OpenAPI document describes each extension route by what it does and by the ITS-REST text that governs it, and no longer names openEHR Service Model operations, interfaces or status codes. Routes, request and response fields and status codes are unchanged (#4213).

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The storage-bench lane measures the baseline and the candidate on one runner in eight interleaved rounds each and gives two verdicts. The plan verdict fails an operation whose median pages per call rose more than 10 % and one page above the baseline’s highest round, whatever its time. The time verdict divides each round by its machine factor, so a slow spell of the runner no longer reads as the code, and fails an operation whose median p50 or p99 rose more than 15 % with every candidate round slower than every baseline round, or a candidate build that ran more than 15 % slower across the board in every round. A latency is no longer compared with a record measured on another runner (#4318).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • The LANG specification text is re-vendored at f80763ab (#4377). SPECLANG-18 states how a P_BMM schema’s JSON form encodes a cardinality interval, which the JSON schema reader already followed; its documentation now cites the rule, and a test pins it. SPECLANG-16 types the P_BMM class lists as keyed collections; the generated P_BMM model takes that change with #4497. The nine openehr-* crates step to 0.0.122.

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • The viewer runs on ferrohealth-server 0.4.2 and takes its Content-Security-Policy from the family’s server-rendered console headers (FerroHEALTH/FerroSYS#202) in place of a layer of its own. The policy a deployment sees is unchanged: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; object-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none', with no nonce. FerroUI’s assets under /ferrohealth-ui/ keep the Cache-Control FerroUI sets, and a refusal under that path is no-store whatever FerroUI’s router sets, so a cache in front of the viewer keeps no error answer for an asset. The REST server takes the same ferrohealth-server release with no change in what it serves (#4439).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The published conformance record is a Veredictum 0.3.3 released-profile run, committed as docs/conformance/ferroehr/results.json with its summary.md. The landing page and the book’s conformance page show the Veredictum badges of the released and development runs on main and the record’s counts per part (BASE and OPTIONS), in place of the Veredictum 0.1.6 heat grid, chapter bars and tier pills. The docs and landing pages admit images from raw.githubusercontent.com/FerroHEALTH/FerroEHR/ for the badges (#4451).

  • The measured performance lane runs Veredictum 0.3.3’s perf. scripts/conformance.sh perf composes the conformance deployment with the measurement overlay, runs the pinned clinical-mix workload, checks the record with perf-judge and redraws the book’s charts from docs/conformance/ferroehr/perf-result.json. scripts/conformance.sh perf-aa measures the deployment against itself, so the record carries its between-run noise floor. The measurement overlay also switches response compression off, and the IXIT names the deployment’s containers for the run’s server-side readings. The Performance and Benchmarks pages describe the new instrument. They keep the earlier figures until a record measured on the reference host replaces them (#4452).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

  • The P_BMM model follows SPECLANG-16. The LANG 1.1.0 BMM that openehr-codegen reads is now the specifications-LANG file at f80763ab, so P_BMM_SCHEMA.primitive_types and class_definitions are maps keyed by class name, P_BMM_CLASS.ancestor_defs is a map of P_BMM_BASE_TYPE keyed by type signature (a simple ancestor included), and P_BMM_GENERIC_TYPE.generic_parameter_defs is an optional map keyed by formal parameter name. The formal and actual generic parameters (P_BMM_CLASS.generic_parameter_defs, P_BMM_GENERIC_TYPE.generic_parameter_defs, BMM_GENERIC_CLASS.generic_parameters) are insertion-ordered IndexMaps, because the specification fixes their declaration order, and the model builders bind each actual parameter to its formal by name. In the released v1_0 generation, P_BMM_GENERIC_TYPE.generic_parameter_defs now carries the 1..* bound its BMM states. The nine openehr-* crates step to 0.0.123 (#4497).

  • The conformance pipeline runs Veredictum 0.3.1. The Conformance workflow drives the catalogue through the Veredictum Action twice, each run on fresh volumes: under the released profile, the frozen openEHR release, which is the record and decides the job, and under the development profile, the pinned master of each openEHR repository, which is published beside it under its own job summary heading and artifact and never counted toward it. Both runs include the SMART documents: the Action now takes the SMART Platform base URL and the committed test issuer’s key. The workflow reads the result paths and the exit status from the Action’s outputs. scripts/conformance.sh takes the profile as CONF_PROFILE (released by default; a development run writes to docs/conformance/ferroehr-development/) and drives the SMART documents in the same run, and the IXIT carries the test issuer’s scoped_tokens declaration (#4503).

  • The viewer groups a template’s releases. The template list draws one group per template, the newest release first with the number of releases and the earlier releases under it, in both the ADL 1.4 and the ADL 2 tab. A template’s page names its release beside its title, where every other release is a link with the newest marked, also with FerroUI’s scripts refused; the release select and its Open button above the tabs are gone. The Overview lists the releases with the compositions committed against each, and a new Compare tab lists the paths one release adds, removes and changes against another (#4506).

  • Every write checks the committer against the login that authenticated the request where the server can compare them: a committer naming another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, is refused with a 403. A name only, or an identifier of another authority, is stored as sent. The direct writes (composition, EHR status, directory, EHR, demographic) now also refuse a committer that breaks its own rules with a 422, as the contribution path already did (#4520).

  • An audit record no longer adds a row to audit.audit_search_token for its request id or for the version uid a create wrote. ITI-81 entity.identifier finds both from the record’s own request_id and resource_id columns, through a new hash index on request_id (audit migration 0020), so the token table stays flat as requests are audited. Retention now deletes the search tokens that no remaining record names, among them the request id tokens of records written by 5.6.0-rc5 and earlier, and the database still refuses to delete a token a kept record names (#4524).

  • A clinical or party node row below the root no longer repeats its archetype header, language and encoding: an archetype root’s archetype_details and archetype_node_id are rebuilt from the archetype column and a new rm_version column, and a language in ISO_639-1 and an encoding in IANA_character-sets keep only their code, in the new language and encoding columns. Every reader puts the keys back where they stood, so the canonical JSON, every AQL result and the served key order are unchanged; a header of any other shape stays in the fragment. A vital signs version takes 17.1 KB instead of 21.6 KB and an International Patient Summary 131 KB instead of 159 KB. Migrations ext/0007, clinical/0028 and party/0020 add the columns and the readers; no data is migrated, and a row written before keeps its whole fragment and reads the same (#4525).

  • ext.storage_generation() reports generation-4, the storage generation with the node header columns, and the storage benchmark records for it are committed under docs/benchmarks/storage/generation-4/. Migration ext/0008 changes only the label; nothing in the serving path reads it (#4525).

  • Five response messages no longer name the openEHR Service Model. The admin dump route’s 400 for a missing file_sys_loc now reads “request body must carry a non-empty file_sys_loc string, a location on the server’s file system (FerroEHR’s own dump/load extension)”, and its 400 for an encoding member reads “encoding is not accepted: FerroEHR’s own dump/load extension defines no encoding, so no value names one”. The admin activity-report 400s for a_service now list the accepted service names (“query parameter a_service is required: one of Admin, Definitions, …” and “query parameter a_service names no platform service: … (expected one of …)”). The not-implemented answer of store_query_set reads “store_query_set is not implemented: no openEHR specification defines a query set”. The openehr-rm validation docs for the three PARTY and PARTY_RELATIONSHIP relationship invariants now name the separate relationship container as FerroEHR’s own design against RM demographic §Party Relationships (#4531).

  • Breaking for clients that relied on the remaining commit-audit leniencies: the audit rules of the CONTRIBUTION commit now hold on every write path (#4543).

    • A description object without _type, such as {"value": "…"}, is refused with 400 on a contribution audit, a commit_audit and an attestation, and so is an attestation reason without _type. Send {"_type": "DV_TEXT", "value": "…"} or a DV_CODED_TEXT with its _type.
    • A stated empty system_id is refused with 422 on the direct writes too (openehr-audit-details: system_id="" on a create, update or delete), where the server used to replace it with its own. Leave system_id out to get the server’s, or send a non-empty one. A system_id that is not a string stays a 400 on both CONTRIBUTION routes.
    • A commit audit of type UPDATE_ATTESTATION handed to the library’s direct write calls (create_composition, update_composition, the EHR_STATUS, DIRECTORY and party writes) is refused with 400, as on the CONTRIBUTION routes. Send the audit as UPDATE_AUDIT and put each attestation in the version’s attestations list.
  • Members that NewContribution or UPDATE_AUDIT do not declare, such as a reason or is_pending on a plain audit or an extra top-level member of a contribution, are accepted and ignored: the released schemas leave both objects open (#4543).

  • The openehr-* crates take their version when they are published. A pull request that changes the crates no longer bumps their lockstep 0.0.x version: the crates publish reads crates.io, takes the next version when packaged content changed since the newest published one, and refuses to publish changed content under a version crates.io already holds. The version in the manifests on main can therefore trail the newest version on crates.io (#4555).

  • The served Web Template of an ADL 2 operational template writes the aqlPath of an ELEMENT’s value and of a data value’s parts without an [idN] predicate, since a data value carries no archetype_node_id (ADL2 master05-paths §“Relationship with W3C Xpath”). A DV_INTERVAL and its parts read …/value, …/value/lower and …/value/upper, the form the ADL 1.4 Web Template writes, where they read …/value[id6]/lower[id7] before, and the parts’ ids are lower and upper (#4559).

  • The served Web Template gives every DV_INTERVAL<T> node its lower and upper children, each of type T, also when the operational template leaves them unconstrained (RM data_types §DV_INTERVAL). Such a node had no children before. An unconstrained part carries the open inputs of its type, and its aqlPath is …/value/lower or …/value/upper. This holds for ADL 1.4 and ADL 2 templates (#4559).

  • The template example fills such an interval: an unconstrained part takes the example value of its constrained sibling, so the limits stay consistent (#4559).

  • The served Web Template of an ADL 2 operational template writes a coded input whose at-codes the template binds to a terminology as the bound terms: its code list carries each binding’s code and label and its terminology names the bound terminology (ADL2 master08-terminology_integration §“From Constraints to Concrete Codes in Data”). COMPOSITION.category bound at1 to openehr::433 reads {"value": "433", "label": "event"} with terminology: "openehr", where it read the local at1 before, and an [acN@ttttt] constraint takes the bound terms of the terminology it names. A FLAT or STRUCTURED composition without a category, and the example composition, now carry openehr::433. The aqlPath of every data value drops its [idN] predicate, so a constrained category and context/setting read /category and /context/setting (#4574).

  • A CONTRIBUTION’s versions[] entry records its merge inputs: the other_input_version_uids it states are stored on the committed ORIGINAL_VERSION and served on every read, where the server used to refuse the member with 400. The uids may name versions the server does not hold; an empty list or an entry that is not an OBJECT_VERSION_ID is a 422 (#4577).

  • Other members UPDATE_VERSION does not declare, such as item or an extension member, are accepted and ignored, where the server used to refuse them with 400: the released schema leaves the object open, as it does NewContribution and UPDATE_AUDIT. A uid or contribution on a version entry is still refused with 400, because the server assigns both, and so is a _type other than UPDATE_VERSION or ORIGINAL_VERSION (#4577).

  • The admin dump route’s refusals name its parameters by their wire names and list the accepted values, not the Service Model class names. A bad logical_format or compression_format now reads “<field> must be a string, one of …” or “<field> names no accepted value: … (expected one of …)”, and a non-integer segment_split_size reads “segment_split_size must be an integer size in kb”. The served OpenAPI text for the route no longer names EXPORT_SPEC or EXPORT_FORMAT (#4634).

  • Five archive dump and load failures still answer 500, and their bodies are now fixed text that echoes no value from the archive: (#4657)

    • an unreadable data segment no longer names the segment, though manifest.json and demographic-commons.json are still named;
    • an unknown logical format in the manifest no longer echoes the declared value;
    • a version whose id carries a path separator, and a version whose commit audit the export lacks, no longer name the version.
  • A failed expansion of a party’s protected national identifiers still answers 500, and its body now reads “expanding the protected identifiers of a party version failed” without echoing the requested version uid: a 500 body carries curated text only (#4657).

  • Two query failures still answer 500, and their bodies now carry curated text only. A stored query whose record holds no query text reads “the stored query has no query text” without echoing its name, and a terminology server failure while a TERMINOLOGY('expand', …) operand is resolved reads “the terminology service failed while the query was being planned” instead of the upstream message (#4657).

  • The db_transactions counter (outcome="commit") now counts every committed single-object write. It used to leave out EHR_STATUS and DIRECTORY writes (#4659).

  • The access record of a VERSION read (GET /ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/version, its /version/{version_uid} form, and the same two versioned_ehr_status routes) now names the content it served, and for a COMPOSITION the origins of its data, as the bare reads already did. Its ehr_id is the parsed EHR id, as on every other read, rather than the path text as sent. The response is unchanged (#4660).

  • OPTIONS on the API base path sends conformance_profile as BASE+OPTIONS where it sent STANDARD. The build derives the value from the committed Veredictum record, docs/conformance/ferroehr/summary.md: BASE+OPTIONS when the BASE rows and the run OPTIONS rows have none failed and none errored, BASE when only the BASE rows do, and the member left out otherwise. ITS-REST 1.1.0 types the member as a free string and CNF names no profiles, so the values are FerroEHR’s own. The [rest.identity] conformance_profile configuration key is removed: a configuration that still sets it is refused at boot as an unknown key (#4672).

  • Every route now reads Prefer as RFC 7240 §2 defines it, as error responses already did: a preference in a second Prefer field, one that carries parameters after ; (return=representation; x=1), and one with whitespace around = (return = representation) are honoured. Before, only the first Prefer field was read, and such a preference was ignored (#4722).

  • The viewer answers every request other than GET and HEAD that arrives without a session, to any path but the sign-in page, the OIDC routes, sign-out and the search answer, with the sign-in page that says the session ended, before any handler runs. A post to a page returns there after signing in; a post to the export, preference, sidebar or paging route returns to the dashboard. The export post used to answer 303 to /login, and the preference, sidebar and paging posts ran without a session (#4724).

  • The conformance instrument is Veredictum 0.4.0, and the published conformance record (docs/conformance/ferroehr/results.json with its summary.md) is a 0.4.0 released-profile run. The party statement answers the 0.4.0 catalogue: it keeps the xml-namespace-v2 arm, which now asks only whether the TRIAL http://schemas.openehr.org/v2 namespace is also accepted, and selects the new development-profile option arm adl2-id-coded-supported (#4771).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

  • The Veredictum 0.1.6 conformance documents: CONFORMANCE_REPORT.md, CONFORMANCE_STATEMENT.md, CONFORMANCE_CERTIFICATE.md, verdicts.json, run-exceptions.json, the CORE, STANDARD, OPTIONS, SEC-BASIC and performance badge files, and the heat grid and chapter bars the book and the landing page drew from them (#4451).

  • The CONF_PERF_CLASS, CONF_PERF_HOURS and CONF_STRESS settings of the measured lanes are gone, along with the class runs, the step-load stress ladder and the AQL probe, which Veredictum 0.3.1 removed (#4452).

  • The separate SMART step of the Conformance workflow, the smart stage of scripts/conformance.sh with its CONF_SMART_OUT, and docs/conformance/party/ferroehr/ixit-smart.json are gone: the one catalogue run covers the SMART documents. The party statement no longer carries the performance block, which the Veredictum 0.3.1 statement schema refuses (#4503).

  • The empty linkage storage domain is gone: its schema, its ferroehr_linkage role, the [storage.linkage] and [storage.migrate.linkage] configuration tables, its pool, the database.linkage and backup.linkage chart values, the ferroehr-backup-linkage compose job and the btree_gist extension only its table needed. A configuration that still sets [storage.linkage] or [storage.migrate.linkage] is refused at boot as an unknown key. Cohort and AQL demographic join crossings still write their linkage-domain access records. Upgrade note: remove those keys and values before you upgrade. On a co-located database, migration audit/0021_drop_linkage_domain.sql drops the schema, the role and the extension at the first boot, and raises a NOTICE for any of them it cannot drop. When the linkage domain lived in a database of its own, [storage.audit] names another database, or the cluster holds more than one FerroEHR database, run in each such database, as a superuser or a member of ferroehr_linkage: DROP SCHEMA IF EXISTS linkage CASCADE; REVOKE ALL ON SCHEMA ext FROM ferroehr_linkage; REVOKE ALL ON ALL FUNCTIONS IN SCHEMA ext FROM ferroehr_linkage; ALTER DEFAULT PRIVILEGES IN SCHEMA ext REVOKE EXECUTE ON FUNCTIONS FROM ferroehr_linkage; DROP OWNED BY ferroehr_linkage; DROP EXTENSION IF EXISTS btree_gist; then DROP ROLE ferroehr_linkage; once per cluster. The released grant files still name ferroehr_linkage, so a first install whose migrator holds no CREATEROLE still provisions that role and drops it after the first boot (#4684).

  • A party migration drops the database function party.resolve_national_identifier. No server path has called it since the linkage domain was removed (#4697).

  • The viewer no longer reads LEPTOS_SITE_ADDR: a value naming another address than server.listen no longer refuses the start. server.listen (FERROEHR_VIEWER__SERVER__LISTEN) sets the listen address (#4725).

  • Migration audit/0022 drops the SQL functions audit.reap_audit_events(integer), audit.reap_audit_events_years(integer) and audit.reap_audit_events_before(timestamptz), along with their EXECUTE grants. The server reaps the audit trail through audit.reap_audit_events_due, which honours each record’s own retention end, and nothing called the three any longer. The spec_profile documentation no longer names a LANG generation, since the profile selects the RM and BASE generations only (#4770).

  • Breaking: FerroEHR no longer publishes change events to a message broker. The AMQP publisher, the lapin dependency, the events cargo feature, the [events] section, the /admin/event_subscription routes with their subscription store (clinical migration 0030_outbox_feed.sql drops the table), the events health indicator and events_published metric, the plaintext_broker deployment gap, the chart’s secrets.eventsUrl and config.events values, and the viewer’s Subscriptions page are gone. A configuration that still carries [events], or plaintext_broker in deployment_accepts, is refused at boot. Broker delivery moves to FerroBRIDGE, where it is in progress (FerroHEALTH/FerroBRIDGE#1367): it is to read the new outbox feed and publish to the broker, so a deployment that needs broker events stays on a release with FerroEHR’s own publisher until a FerroBRIDGE release carries it. Then name its reader in [outbox] readers and configure FerroBRIDGE’s events lane with the broker URL, exchange and queues (Change events (outbox feed)) (#4786).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • An AQL query aimed at named EHRs (the ITS-REST ehr_id parameter) reads those EHRs’ objects first and each object’s own content after them, whatever the planner statistics say. When an archetype’s rows fell out of the most-common-value list of the archetype statistics, the planner estimated them at one row and could start from that archetype over the whole population, evaluating every projected path for every EHR’s rows before the EHR filter: the storage bench’s wide projection over one EHR read about 10 000 pages per call in some rounds and 1 200 in others. A node contained in a versioned object is now read inside its parent on such a query; results do not change (#4487).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

  • AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri argument whole when its value carries a ?, as the QUERY master03 example url=http://snomed.info/sct?fhir_vs does. Before, the arguments were read from after the first ? anywhere in params_uri, so the value-set URL reached the terminology server truncated and the arguments before it were lost. A params_uri that starts with a path still has its arguments read from after the path’s ? (#4490).

  • An ADL 2 template whose constituent archetype the CDR does not hold is refused at upload, with a 422 whose errors carry one VARXR entry per missing archetype (AOM2 C_ARCHETYPE_ROOT validity rule VARXR); the 2013 CKM t_encounter_opt_test, built on the refused apgar, no longer loads, so the sandbox library pin is 231 accepted / 91 refused. A stored template whose constituent was deleted later answers its JSON form and example with the same VARXR list, and the viewer says so on every tab of the template, names each missing archetype and links to uploading it, instead of a raw 422 (#4505).

  • ADL 2 validation admits a constraint on an RM computed attribute. VCARM accepts an attribute “stored or computed” (AOM2 master04.5 §Validity Rules: C_ATTRIBUTE), and a function with no arguments is a computed attribute (master04.2 §Attribute Nodes), so an archetype that constrains DV_PROPORTION.is_integral, DV_QUANTITY.is_integral or EVENT.offset is no longer refused with 422. The generated RM model (openehr_rm::v1_2::model) now carries each class’s computed attributes from the BMM functions (RmClass::computed_attributes, computed_attribute). openEHR-EHR-CLUSTER.microscopy_prostate_carcinoma.v1.0.0 from the 2013 CKM ADL 2 export now uploads, and the sandbox archetype-library pin moves to 232 accepted and 90 refused. DV_QUANTITY.property, which no RM release defines, stays refused. The nine openehr-* crates step to 0.0.124 (#4513).

  • DELETE /ferroehr/rest/extensions/v1/definition/artefact/adl2/{artefact_id} refuses with 409 Conflict while another stored ADL 2 artefact does not compile without the one deleted (a slot filler, a use_archetype reference or a specialisation parent, directly or through another artefact), and the message names every one of them. Before, the delete succeeded and left each template built on the archetype unable to compile (#4522).

  • The viewer’s ADL 2 template upload names each violation the CDR found when it refuses the source: every AOM2 rule code with the CDR’s words for it, so a template whose constituent archetype the CDR does not hold names that archetype (VARXR). The upload now asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details, and reads the stored template’s id from the ETag of the 201 (#4523).

  • An ADL 2 operational template whose included archetype reuses the template’s node ids uploads again, where it was refused with 422 under VCOSU. Node-id uniqueness now holds per archetype and restarts at each C_ARCHETYPE_ROOT (AOM 2 VCOSU, “unique within the archetype”), and on an operational template the codes and interior node ids below a root are judged against that constituent’s entry in component_terminologies (OPT 2 master03-opt_raw). A duplicate node id inside one archetype is still refused. The nine openehr-* crates step to 0.0.125 (#4529).

  • The viewer’s dashboard draws the EHR or composition count as unavailable, with the reason in plain words, when the CDR answers its count query with a result set of another shape (no row, several rows, an empty row, several values, or a value that is no whole number), where it drew 0 before. The template Releases table draws such a count the same way. A real zero still draws 0 (#4533).

  • An ADL 2 operational template upload runs the terminology integrity rules (VTLC, VATCV, VTVSID, VTVSUQ, VTVSMD, VTTBK, VTCBK, and VTSD where the level is carried) on each component_terminologies entry, and each violation names the component’s archetype id (#4536).

  • A COMPOSITION committed against an ADL 2 operational template is validated against the template’s leaf constraints, where it was stored with 201 whatever it held. An inlined archetype recorded by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability) now matches its template node, so the C_INTEGER, C_REAL, C_STRING, C_BOOLEAN, tuple, value-set and cardinality constraints below it answer 422 (AOM 2 master04.5). An archetype root the template does not inline is refused as a node the template does not define, and a C_INTEGER or C_REAL list ({10, 15, 20}) admits each of its values where only the first was read. Canonical, FLAT and STRUCTURED commits take the same check. The nine openehr-* crates step to 0.0.127 (#4540).

  • A versioned read whose version path segment is no SEMVER version and no {major} or {major}.{minor} prefix, such as v1.x, 1.x or 01, now answers 400 before any lookup, where it answered 404 (parameters/path/version.yaml, responses/400.yaml). This covers the ADL 2 template read GET /definition/template/adl2/{template_id}/{version}, the stored-query read GET /definition/query/{qualified_query_name}/{version} and both stored-query executions at a version, GET and POST /query/{qualified_query_name}/{version}. A well-formed version the server does not hold still answers 404 (#4541).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the content below an inlined archetype that the composition records by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability). The read left out everything below such a root with no error; it now matches the root to its template node the way the commit does. The nine openehr-* crates step to 0.0.128 (#4551).

  • A COMPOSITION committed against an ADL 2 operational template whose C_INTEGER or C_REAL constraint is a union of ranges (for example magnitude matches {|0..5|, |10..15|}) is refused with 422 when the value lies outside every range. Such a value was stored. The served Web Template does not change: its input validation.range holds one range, so a union leaves it out, as before, and the check runs on the commit (#4552).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the value of an ELEMENT that holds a DV_INTERVAL whose lower and upper the template constrains. The read left the value out with no error (#4559).

  • A FLAT or STRUCTURED read of a COMPOSITION returns the value of an ELEMENT that holds a DV_INTERVAL whose template leaves the interval’s lower and upper unconstrained, for ADL 1.4 and ADL 2 templates. The read left the value out with no error, and a FLAT or STRUCTURED commit could not name the parts (#4559).

  • A COMPOSITION written from FLAT or STRUCTURED against an ADL 2 operational template carries what a canonical body built to that template carries. A node the template gives no term text, such as the HISTORY and ITEM_TREE of an OBSERVATION, is named for its RM type (History, Tree) instead of its bare id-code (id2). Every inlined archetype root records the complete identifier the template names (…cnf_minimal.v1.0.0, BASE master07-referencing §Reconstitutability) in its archetype_node_id and archetype_details.archetype_id, and archetype_details.rm_version is the template’s rm_release. Before this change the root recorded the interface form (…v1) and the server’s RM 1.2.0. Compositions built against an ADL 1.4 template keep their major-version archetype ids and RM 1.2.0 (#4560).

  • ferroehr report in a release build names the openehr-* crate versions that release published to crates.io: the release lanes stamp the published version into the checkout before they build the server, where the report used to name the floor the source tree holds. A new openehr_crates_origin field says published for a release build and source_tree for any other, a pre-release included (#4565).

  • An ADL 2 archetype reference resolves to one defined stored release: an interface reference (….v1) takes the newest release of its major version, a reference with a minor or patch version narrows the match to it, and an alpha or beta version matches only its full identifier (BASE Artefact Identification, Referencing). With several releases of one archetype stored, a template upload, an operational template and the delete’s dependents check used whichever release PostgreSQL returned last, and a reference to a major version no stored release carries resolved to another major version (#4567).

  • An ADL 2 upload and an ADL 2 delete no longer race. An upload validated before a concurrent delete of an archetype it uses committed could store a template whose constituent was gone. Both now hold one transaction-scoped advisory lock from the repository read to the commit, so the second write decides on what the first committed: an upload after the delete is refused with VARXR, and a delete after the upload is refused with a 409 naming the template (#4568).

  • The syslog form (DICOM PS3.15, IHE ITI-20) of an EHR-level audit record names its object beside the patient, as a URI participant object with the id its FHIR AuditEvent names. The record of placing, releasing or expiring an EHDS Art. 3(3) access delay names the held composition as access-delay:<vo>, so an Audit Record Repository fed by syslog can tell which composition a hold covers (#4569).

  • An ADL 2 template or archetype that references two major versions of one archetype (….v1 and ….v2), or one archetype under two namespaces, now compiles against both: the openehr-adl repository holds an artefact under its namespace, archetype and major version, and resolves each reference within that key (BASE Artefact Identification, Referencing). A reference without a namespace now resolves in the namespace of the artefact that makes it, and never in another namespace (#4579).

  • A read of a stored ADL 2 template by a partial id or a version (GET /definition/template/adl2/{template_id}) now resolves the way template compilation does: with v1.0.0 stored beside v1.0.0-rc.1, ….v1, ….v1.0 and version=1.0.0 answer the release in either upload order, and the release candidate is read only by its full version (BASE Artefact Identification, Referencing) (#4580).

  • A write that sends no committer, on a route that takes a bare RM body (the EHR_STATUS, COMPOSITION and DIRECTORY writes, and the demographic party and relationship writes), now records a Bearer caller’s login as issued by the token issuer, as EHR creation and every other default committer already did. It recorded the issuer ferroehr for every caller before; a Basic caller’s login is still issued by ferroehr (#4585).

  • The viewer’s ADL 1.4 template upload opens the stored template’s page, read from the ETag (else the Location) of the CDR’s 201, where it went to an address with no template id. A refused upload names each violation the CDR found, its rule code or path and the CDR’s words for it, as the ADL 2 upload does: the upload asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details (#4596).

  • A 422 validation refusal tells the client to send Prefer: return=representation only when the request did not send it; a request that sent it receives the violations in errors and a detail without the advice (#4597).

  • An EHR Extract import refusal for a received version now names that version’s position in the extract (chapters[i]/items[j]/item/versions[k]), so an importer can find the defective ORIGINAL_VERSION in a large extract (#4602).

  • The import reads a version’s other_input_version_uids with the parser the CONTRIBUTION path uses and refuses an entry that is not an OBJECT_VERSION_ID, where it used to drop the entry. On the wire the extract decode already refuses such an entry with 422, so the import stays all or nothing (#4602).

  • An AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts within the major: org.a::… no longer finds data recorded in org.b, an id without a namespace finds only data recorded without one, and .v1.4.2 no longer finds data recorded under v1.0.0 (BASE Artefact Identification §Referencing and §Supporting Archetype-based Querying; #4629, #4630).

  • A query naming a parent archetype finds the data of a specialised child release only when that release declares a parent release at or after the queried one. Before, once one stored release of the child qualified, the data of every release of it was returned (BASE Artefact Identification, Referencing §Supporting Archetype-based Querying; #4637).

  • POST /definition/template/adl1.4 and POST /definition/template/adl2 answer 400 for every invalid template, the only refusal ITS-REST 1.1.0 declares for both uploads besides the 409 duplicate. An OPT that breaks the OPT 1.4 schema, an ADL 2 source that breaks an AOM2 rule such as VCORM or VCARM, and an ADL 2 template whose constituent archetype is not held were refused with 422. The problem keeps its detail and, with Prefer: return=representation, its errors. Upgrade note: a client that read a 422 from a template upload reads a 400, with the same detail and errors, under the about:blank problem type (#4643).

  • A directory create or update whose FOLDER carries an empty folders array, at the root or in a sub-folder, is refused with 400 and commits nothing. The ITS-JSON FOLDER schema gives folders minItems: 1, and the canonical JSON reader now refuses the empty array the same way in every document that carries a FOLDER, a CONTRIBUTION included. The viewer leaves folders out of a folder with no sub-folders (#4644).

  • A canonical XML FOLDER, EHR_STATUS, PERSON, AGENT, GROUP, ORGANISATION, ROLE or PARTY_RELATIONSHIP document now has the root element items with xsi:type naming its class, the element the ITS-XML schemas publish for a LOCATABLE (Structure.xsd). The directory reads (latest, by version and at a time), the EHR_STATUS reads, the demographic reads and the demographic template examples served <folder>, <ehr_status>, <person> and the like, which no ITS-XML schema declares. XML input accepts the same form (#4645).

  • POST /definition/template/adl1.4 builds the template’s Web Template before it stores the template, so a template it cannot build is refused with 400 and nothing is stored. Once the template is stored, the upload answers 201 with its Location; with Prefer: return=representation and a Web Template Accept it serves the Web Template built before the store, and the OPT if that cannot be served. It no longer answers 422 for a template it has already stored (#4647).

  • A composition whose HISTORY carries events: [], and an AGENT, GROUP, ORGANISATION or PERSON carrying languages: [], are refused with 400 and commit nothing. The ITS-JSON schemas give both arrays minItems: 1, and the canonical JSON reader now refuses the empty array as it does an empty FOLDER folders. The FLAT builder no longer writes an empty events for a HISTORY that has no event (#4648).

  • A write to a restricted COMPOSITION (POST, PUT or DELETE on /ehr/{ehr_id}/composition) refused under a GDPR Art. 18 restriction now answers with the processing-restricted problem type, as the reads of a restricted record already did, in place of the generic forbidden type. The status stays 403 (#4657).

  • When an EXTRACT_SPEC criterion’s AQL runs past the query time budget, the 400 body now states the timeout as plain text. It used to start with a control-character marker that was only ever meant for the server itself (#4657).

  • Under a separated clinical credential, POST /ehr, PUT /ehr/{ehr_id} and an EHR_STATUS update that name a subject no longer fail with a database error: the subject pseudonym guard reads the deployment posture, and the clinical role now holds SELECT on ext.posture through the ext migration 0009 (#4675).

  • A cohort that names its subjects now finds, under [ehr_index] provider = "local", the EHR created over POST /ehr with that subject in its EHR_STATUS; it read a cross-reference table no route wrote, so it matched nothing. A cohort predicate and the AQL demographic join resolve a party to the EHRs whose EHR_STATUS.subject names it. The EHR_STATUS subject is now the one source of which EHR belongs to which subject. Upgrade note: linkage.subject_ehr and the linkage.erase_ehr function go with the linkage schema, which audit migration 0021 drops with the linkage domain. Rows written there through the service API are dropped with the table; the EHRs they named keep their subject in their own EHR_STATUS (#4675).

  • The identifier scan ([privacy.identifier_scan]) no longer refuses a DV_IDENTIFIER whose type declares an item rather than a person, such as a laboratory filler_order_number that passes the NHS Number check by chance. A DV_IDENTIFIER in FEEDER_AUDIT.originating_system_item_ids or feeder_system_item_ids is scanned only when its type names a person identifier. A type naming a person (patient, NHS, BSN, social security and similar) keeps the value scanned in every slot, and free text and untyped identifiers in clinical content are scanned as before. The default rule set is unchanged; a deployment that knows its jurisdictions still narrows rules, so a Dutch one with rules = ["nl-bsn"] never claims an NHS Number (#4702).

  • An inline primitive constraint in a parsed archetype or template now carries the node id AOM2 ADL_CODE_DEFINITIONS.Primitive_node_id defines, at9999 in an at-coded archetype and id9999 in an id-coded one, where it carried the constant’s name Primitive_node_id. An ADL 1.4 archetype converted to ADL2 gets the id-coded form (#4739).

  • The ABAC policy-enforcement point checks every item-tag operation as it checks its target: a COMPOSITION tag read after the call like the COMPOSITION read, a COMPOSITION tag write before the call with the tagged version’s template, category and origin, and the EHR tag read before the call like the EHR read. These operations were unchecked before, so a patient-scoped or template-scoped rule now refuses a tag read or write wherever it refuses the target’s read or write. A COMPOSITION tag read also falls under the EHR_ACCESS privacy ceiling and its audit record names the tagged COMPOSITION (#4768).

  • A primitive constraint inside a rule or a slot assertion of an id-coded ADL 2 archetype carries id9999 as its node id, as org.openehr.am.aom2.adl_code_definitions §Constants defines Primitive_node_id for id-coded archetypes, where it carried at9999 before. An ADL 1.4 archetype converted to ADL 2 recodes its slot assertions the same way. openehr_adl::rules::parse_rules_body and parse_slot_assertions take the archetype’s NodeCoding, which openehr_adl::codes now exports (#4770).

  • Under spec_profile = "stable", a canonical-JSON party read lists in reverse_relationships every PARTY_RELATIONSHIP another party’s current version carries with this party as its target, one LOCATABLE_REF each naming the source party’s version container and the relationship’s path, as RM 1.1.0 §Party Relationships requires. The development profile, whose RM 1.2.0 makes the attribute a function, is unchanged. The released-profile conformance run now judges the server on the stable profile (#4775).

  • Under the development profile, POST /definition/template/adl2 refuses with 400 a template whose description has no original_author or an empty lifecycle_state, naming RESOURCE_DESCRIPTION.Original_author_valid or RESOURCE_DESCRIPTION.Lifecycle_state_valid. BASE 1.3.0 states both invariants (org.openehr.base.resource.resource_description.adoc §Invariants). The stable profile’s BASE 1.2.0 states neither, so it still accepts such a template (#4776).

  • A database that accepts connections slowly is no longer recorded as down. A probe connect still in progress after its two-second bound now leaves the acquire queueing on the pool within acquire_timeout_secs, so the request is served once the connect completes, or answers the pool-timeout 503 with Retry-After. Only a refused or reset connect, or a server shutting down (57P01, 57P02, 57P03), marks the database down and answers the outage 503 at once (#4782).

  • A database that resets connections is now answered as an outage at once on every request, the 503 with Retry-After a refused connect gets, instead of each request opening a connection of its own that the database resets. A connect of the pool’s own that is refused, reset or turned away as shutting down (57P01, 57P02, 57P03) records the outage as the probe does, and a probe that connects again clears it. A connect that is only slow still queues (#4807).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

  • CVE-2026-84243, the glibc gettext LANGUAGE directory traversal (GLIBC-SA-2026-0025), has no Debian fix for trixie yet and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never leave the C locale, so glibc ignores LANGUAGE and opens no message catalog (security/vex/ferroehr-os.openvex.json). In ferroehr-postgres only the container environment the operator writes can set LANGUAGE; no database connection can (security/vex/postgres-os.openvex.json) (#4500).

  • CVE-2026-114690, CVE-2026-114691 and CVE-2026-115455, three libstdc++ defects in the gcc-14 packages of every image, have no Debian fix for trixie and are judged not_affected. Each fix changes only a header template (the parallel-mode search, the std::vector<bool> constructors, wide-stream extraction into std::bitset), so libstdc++.so.6 holds no copy; the ferroehr and ferroehr-viewer programs contain no C++ code (security/vex/ferroehr-os.openvex.json), and in ferroehr-postgres nothing is built in parallel mode and PostgreSQL does not link libstdc++ (security/vex/postgres-os.openvex.json) (#4789).

  • CVE-2026-119371 to CVE-2026-119379 and CVE-2026-119591, ten SQLite defects without a Debian fix, are judged not_affected in ferroehr-postgres: only lslogins reaches SQLite there, through liblastlog2, and nothing in the image runs it (security/vex/postgres-os.openvex.json) (#4789).

5.6.0-rc9 - 2026-11-02

5.6.0-rc9 is the ninth release candidate of 5.6.0, published for testing, and the first candidate published since 5.6.0-rc7: 5.6.0-rc8 was tagged and never published, because its vulnerability record refused 86 findings that had no OpenVEX judgement, and a tag is never moved, so 5.6.0-rc9 carries everything 5.6.0-rc8 carried. It carries the CONTRIBUTION break of 5.6.0-rc6: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution read the body through the released NewContribution schema and answer 400 for a lifecycle_state that is no DV_CODED_TEXT, a description that is a plain string or an object without _type, an audit without change_type, an audit typed ATTESTATION or UPDATE_ATTESTATION, a missing audit or versions, and a preceding_version_uid sent as a plain string, and a stated empty system_id is refused with 422 on every write path (#3595, #4543). These notes carry every change since 5.5.0, those of the 5.6.0-rc8, 5.6.0-rc7, 5.6.0-rc6, 5.6.0-rc5, 5.6.0-rc4 and 5.6.0-rc3 sections below included. New since 5.6.0-rc7, and so in neither published candidate before this one: two read-only extension routes that convert a stored ADL 1.4 archetype or operational template to ADL 2 (#4655); template uploads that answer 400 for every invalid template and never refuse a template already stored (#4643, #4647); the ITS-JSON minItems of FOLDER.folders, HISTORY.events and ACTOR.languages enforced (#4644, #4648), and a canonical XML LOCATABLE rooted at the items element the ITS-XML schemas publish (#4645); an AQL query naming a parent archetype that finds a specialised child’s data only from a child release whose own parent release qualifies (#4637); conformance_profile derived from the committed conformance record (#4672); a subject resolved through the EHR_STATUS of its EHR, and the empty linkage domain removed with its configuration (#4675, #4684, #4697); a write refused under a GDPR Art. 18 restriction answering with the processing-restricted problem type, and 500 bodies that carry curated text only (#4657); an identifier scan that reads a DV_IDENTIFIER’s declared type (#4702); a validation refusal that advises Prefer: return=representation only to a request that did not send it (#4597); the AOM2 node id of an inline primitive constraint (#4739); a viewer that answers a request other than GET or HEAD without a session with the sign-in page and no longer reads LEPTOS_SITE_ADDR (#4724, #4725); every item-tag operation checked by ABAC in its target’s mode (#4768); and, under spec_profile = "stable", a canonical-JSON party read that lists the relationships naming it as their target in reverse_relationships (#4775); since 5.6.0-rc8, the OpenVEX judgements of the gcc-14 and SQLite findings that stopped it (#4789), the AOM2 node id of a primitive constraint inside a rule or slot assertion of an id-coded ADL 2 archetype, and the three unused horizon reap functions of the audit store removed (#4770). The published conformance record is a Veredictum 0.4.0 released-profile run against FerroEHR on spec_profile = "stable": 2872 of 3071 rows passed, none failed and none errored, and 199 were skipped, 116 for an option the statement does not select and 83 that rest on development text alone; it replaces the Veredictum 0.1.6 documents (#4451, #4771). The released-profile verdict is now required: a failed or errored row fails the merge queue, the image run on main and the release (#4455). The development-profile run, which is never a verdict, has 2946 rows passed, 3 failed (#4776, #4777) and 2 errored by Veredictum itself. 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: an invalid template on POST /definition/template/adl1.4 or POST /definition/template/adl2, an ADL 2 template whose constituent archetype is not held among them, answers 400 where it answered 422, with the same detail and errors under the about:blank problem type; a FOLDER with an empty folders, a HISTORY with an empty events and an AGENT, GROUP, ORGANISATION or PERSON with an empty languages answer 400; a canonical XML FOLDER, EHR_STATUS, PERSON, AGENT, GROUP, ORGANISATION, ROLE or PARTY_RELATIONSHIP document has the root element items with xsi:type naming its class, where it had <folder>, <ehr_status>, <person> and the like; OPTIONS on the API base path sends conformance_profile as BASE+OPTIONS where it sent STANDARD, and a configuration that sets [rest.identity] conformance_profile is refused at boot as an unknown key; an AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts, and a query naming a parent archetype finds a specialised child’s data only from a child release whose own parent release qualifies, so a query may find less than before; a cohort that names its subjects and the AQL demographic join resolve a party to the EHRs whose EHR_STATUS.subject names it, and a configuration that sets [storage.linkage] or [storage.migrate.linkage] is refused at boot and Helm chart 21.0.0-rc8 refuses a values file that sets database.linkage or backup.linkage, so remove those keys and values before you upgrade and, where the linkage domain lived in a database of its own, run the SQL the Removed entry gives; a POST, PUT or DELETE on a COMPOSITION under a GDPR Art. 18 restriction still answers 403, now with the processing-restricted problem type in place of forbidden; the 500 bodies of the archive dump and load, a failed expansion of a party’s protected identifiers and two query failures carry fixed text that echoes no value from the request, the archive or the terminology server; the identifier scan no longer refuses a DV_IDENTIFIER whose type declares an item rather than a person, and scans one in a FEEDER_AUDIT only when its type names a person identifier; the Veredictum 0.1.6 conformance documents (CONFORMANCE_STATEMENT.md and the rest) and the CONF_PERF_CLASS, CONF_PERF_HOURS and CONF_STRESS settings of the measured lanes are gone; with ABAC on, a patient-scoped or template-scoped rule refuses an item-tag read or write wherever it refuses the read or write of the tag’s target, and a COMPOSITION tag read falls under the EHR_ACCESS privacy ceiling; under spec_profile = "stable" a canonical-JSON party read carries reverse_relationships, while the canonical-XML read does not yet (#4778) and the development profile is unchanged; a write whose committer names another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, answers 403, and a direct write whose committer breaks its own rules answers 422; a write that sends no committer on a bare RM body route records a Bearer caller’s login as issued by the token issuer, where it recorded ferroehr; a FLAT or STRUCTURED composition under an ADL 2 template records the complete archetype identifier and the template’s rm_release, and the served Web Template writes a bound coded input, COMPOSITION.category among them, as the bound terms (openehr::433) where it wrote the local at-code; a CONTRIBUTION commit must take the shapes above; a write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as committer and may change only what the person committed, and a service auth_failure answers 403 where it answered 500; a versioned read whose version is no SEMVER version or prefix answers 400 where it answered 404; the served Web Template of an ADL 2 template writes a data value’s aqlPath without an [idN] predicate; a composition that breaks an ADL 2 template’s leaf constraint answers 422; every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes, so keep it longer than the idle timeout of your clients and load balancers; a client that wrote a space as + in a query string sends %20; a party delete names the latest OBJECT_VERSION_ID in its path and answers 400 for a versioned object id; a body the RM refuses after the ITS-JSON schema admits it answers 422, and a misplaced CONTRIBUTION member, a Simplified Format commit without openehr-template-id and a flat |raw value without _type answer 400; a tag update under Prefer: return=identifier answers 200 with the tag list; with [authz.rbac] enabled a token holding only system/ SMART on openEHR scopes reaches clinical operations without a role; the viewer image serves no /pkg/ and reads no LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, and its Content-Security-Policy is script-src 'self'; style-src 'self'; Helm chart 21.0.0-rc8 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data, and ext migration 0008 names generation 4, the generation with the node header columns, and changes only the label; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; clinical 0024 and party 0015 index the version uid, ext 0006, clinical 0025 and party 0016 add the ELEMENT value search index, built over every hot node row of an existing store, party 0017 adds the archetype statistics, and clinical 0026 and party 0018 grant MAINTAIN on the partitioned tables to the serving roles; clinical 0027 and party 0019 add the access delay, and ext 0007, clinical 0028 and party 0020 add the node header columns with no data migrated; audit migration 0019 gives the search token index a pending list of 512 kB, and audit 0020 adds the hash index on request_id; clinical 0029 adds the access opt-out; ext migration 0009 grants the clinical role SELECT on ext.posture, party 0021 drops the database function party.resolve_national_identifier, audit 0021 drops the linkage schema, the ferroehr_linkage role and the btree_gist extension, raising a NOTICE for any of them it cannot drop, party 0022 restates the comment on party_relationship_target.source_vo_id with no change to data, and audit 0022 drops the functions audit.reap_audit_events, audit.reap_audit_events_years and audit.reap_audit_events_before, which nothing called; the linkage migration set is gone. Helm chart 21.0.0-rc8 carries appVersion 5.6.0-rc9. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j, and the OpenVEX judgements of CVE-2026-84243, of three gcc-14 libstdc++ findings and of ten SQLite findings.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The operations page gains a storage sizing section, measured on 5.6.0-rc5: about 21.6 KB of clinical storage per version of a vital signs composition (2.0 times its canonical JSON, 85 % of it node rows) and 155 KB for an International Patient Summary, about 1.2 KB per audit record of a read, the effect of [audit.store] retention on growth, pg_wal up to max_wal_size, the REINDEX INDEX CONCURRENTLY clinical.node_hot_pkey step (20 % of that index after heavy concurrent writing) and a query for the size per relation, with the method to reproduce the figures (#4304).

  • EHDS Art. 5: what a person inserts in their own EHR is marked as theirs, and they cannot alter what a professional committed. A write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as its AUDIT_DETAILS.committer, its external_ref naming the login, whatever committer the request names; a COMPOSITION it commits must name PARTY_SELF or PARTY_RELATED as composer; and its update or deletion of a version whose committer is not PARTY_SELF, directly or inside a CONTRIBUTION, is refused with 403. A professional’s writes are unchanged. The contribution list extension gains inserted_by_person per row, and the viewer names a PARTY_SELF committer “The person or their representative”. A service auth_failure status now answers 403 where it answered 500 (#4350).

  • The EHDS Art. 3(3) delay of the person’s own access to newly registered data, a Member State option and off by default. The new [access_delay] section (enabled, categories, hours) holds every new COMPOSITION of a listed [audit.categories] category from requests that may be the person’s own (a bearer token with a SMART patient/ grant or the ABAC patient claim, also beside user/ or system/ grants, as ferrohealth-smart 0.8.7 decides it) for at most hours: a read answers 404 as for an unknown composition, and a query, an export and a CONTRIBUTION leave it out, while a professional reads it as before. The treating professional releases a held composition from the clinical route POST {ext}/ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/access-delay/release, with the checks of a write to that EHR, and the controller with POST /admin/access-delay/release; a professional holds one by hand within the limit with POST /admin/access-delay and reads the register with GET /admin/access-delay; every hold ends at its limit, and placing, releasing and expiry are each an access record. The subject-scoped ITI-81 read of the person’s own trail (EHDS Art. 9) leaves out the records about a held composition until it is released or expires, then serves them with their original times (ehds-logging 0.3.3) (clinical migration 0027_access_delay.sql, party migration 0019_access_delay_column.sql) (#4351).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • The Conformance workflow runs the Veredictum smart documents (smart_discovery, smart_scopes) against the same FerroEHR build after the Veredictum Action, which forwards no SMART Platform base URL and no scoped tokens. The same pinned image runs veredictum test with --platform-url and docs/conformance/party/ferroehr/ixit-smart.json, minting each access token in memory from the committed test issuer the SMART overlay trusts, and the job uploads its results, JUnit and CTRF views beside the catalogue’s and adds its summary under a heading of its own. scripts/conformance.sh gains the smart stage, and run drives it after the catalogue, so a local run covers SMART too (#4450).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

  • The EHDS Art. 10 opt-out from access through the access services of Arts. 4 and 12, a Member State option and off by default. The new [authz.rbac] access_service_roles key declares the roles of the access-service callers; while an EHR carries the opt-out, a caller holding one reads it as absent: a route naming the EHR or an object in it answers 404 as for an unknown EHR, a population query leaves its data out, a scoped query, an export and the lookup by subject find no such EHR, and a write to it is refused as absent. Every other caller, the provider’s own care included, reads it as before. The controller records the opt-out on the person’s instruction with POST /admin/access-opt-out, reverses it with POST /admin/access-opt-out/reverse (the register keeps the reversal) and reads the register with GET /admin/access-opt-out; recording is refused with 400 while no access-service role is declared. An access-service caller that declares one of [audit] emergency_purpose_codes and does not act for the person reads the EHR under the Art. 10(2) vital-interest override, with the emergency mark on its access record and an access-opt-out-override record against the EHR. Recording, reversing and reading the register are access records (clinical migration 0029_access_opt_out.sql) (#4516).

  • Two read-only extension routes convert stored ADL 1.4 definitions to ADL 2: GET /ferroehr/rest/extensions/v1/definition/archetype/adl1.4/{archetype_id}/adl2 for a stored archetype and GET /ferroehr/rest/extensions/v1/definition/template/adl1.4/{template_id}/adl2 for a stored operational template, which converts to one ADL 2 archetype per embedded archetype root. Each answers 200 with application/json: (#4655)

    • the ADL 2 source of every root;
    • each root’s conversion log and its conversion_details, which name what the ADL 2 form drops or changes;
    • the slot fills that joined the roots. An unknown id answers 404. A stored source that no longer converts answers 422. No openEHR specification defines the conversion, so both routes are FerroEHR’s own design.

Changed

  • Breaking for clients that relied on the old leniencies: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution now read the body through the released NewContribution schema, and refuse every shape it does not define. Each refused shape answers 400: (#3595)

    • a lifecycle_state in the CODE_PHRASE shape {"terminology_id": …, "code_string": …}, as a plain string, or as a bare {"value": …}. Send a DV_CODED_TEXT: {"value": "complete", "defining_code": {"terminology_id": {"value": "openehr"}, "code_string": "532"}};
    • a description sent as a plain string, on the contribution audit, on a commit_audit or on an attestation. Send a DV_TEXT or DV_CODED_TEXT object: {"_type": "DV_TEXT", "value": "…"};
    • a commit_audit or contribution audit without change_type. The server no longer infers creation or modification from preceding_version_uid; send the code of the change, for example 249 (creation), 251 (modification), 523 (deleted) or 666 (attestation);
    • a commit_audit or contribution audit whose _type is ATTESTATION or UPDATE_ATTESTATION, or any value other than UPDATE_AUDIT or AUDIT_DETAILS. Send the audit as UPDATE_AUDIT (or with no _type) and put each attestation in the version’s attestations list as an UPDATE_ATTESTATION. A 666 (attestation) version carries its attestations there too, and one with no attestation is refused;
    • a missing audit, which answered 422 before. A missing versions list, an attestation without reason or is_pending, a system_id that is not a string, and a preceding_version_uid sent as a plain string instead of {"value": "…"} are refused the same way.
  • An attestation’s own system_id is kept, and one that states none takes the system_id of the version audit it belongs to instead of the server’s. An attestation committer is now checked against its PARTY_PROXY rules like every other committer, so an invalid one answers 422 (#3595).

  • The served OpenAPI document describes each extension route by what it does and by the ITS-REST text that governs it, and no longer names openEHR Service Model operations, interfaces or status codes. Routes, request and response fields and status codes are unchanged (#4213).

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The storage-bench lane measures the baseline and the candidate on one runner in eight interleaved rounds each and gives two verdicts. The plan verdict fails an operation whose median pages per call rose more than 10 % and one page above the baseline’s highest round, whatever its time. The time verdict divides each round by its machine factor, so a slow spell of the runner no longer reads as the code, and fails an operation whose median p50 or p99 rose more than 15 % with every candidate round slower than every baseline round, or a candidate build that ran more than 15 % slower across the board in every round. A latency is no longer compared with a record measured on another runner (#4318).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • The LANG specification text is re-vendored at f80763ab (#4377). SPECLANG-18 states how a P_BMM schema’s JSON form encodes a cardinality interval, which the JSON schema reader already followed; its documentation now cites the rule, and a test pins it. SPECLANG-16 types the P_BMM class lists as keyed collections; the generated P_BMM model takes that change with #4497. The nine openehr-* crates step to 0.0.122.

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • The viewer runs on ferrohealth-server 0.4.2 and takes its Content-Security-Policy from the family’s server-rendered console headers (FerroHEALTH/FerroSYS#202) in place of a layer of its own. The policy a deployment sees is unchanged: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; object-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none', with no nonce. FerroUI’s assets under /ferrohealth-ui/ keep the Cache-Control FerroUI sets, and a refusal under that path is no-store whatever FerroUI’s router sets, so a cache in front of the viewer keeps no error answer for an asset. The REST server takes the same ferrohealth-server release with no change in what it serves (#4439).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The published conformance record is a Veredictum 0.3.3 released-profile run, committed as docs/conformance/ferroehr/results.json with its summary.md. The landing page and the book’s conformance page show the Veredictum badges of the released and development runs on main and the record’s counts per part (BASE and OPTIONS), in place of the Veredictum 0.1.6 heat grid, chapter bars and tier pills. The docs and landing pages admit images from raw.githubusercontent.com/FerroHEALTH/FerroEHR/ for the badges (#4451).

  • The measured performance lane runs Veredictum 0.3.3’s perf. scripts/conformance.sh perf composes the conformance deployment with the measurement overlay, runs the pinned clinical-mix workload, checks the record with perf-judge and redraws the book’s charts from docs/conformance/ferroehr/perf-result.json. scripts/conformance.sh perf-aa measures the deployment against itself, so the record carries its between-run noise floor. The measurement overlay also switches response compression off, and the IXIT names the deployment’s containers for the run’s server-side readings. The Performance and Benchmarks pages describe the new instrument. They keep the earlier figures until a record measured on the reference host replaces them (#4452).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

  • The P_BMM model follows SPECLANG-16. The LANG 1.1.0 BMM that openehr-codegen reads is now the specifications-LANG file at f80763ab, so P_BMM_SCHEMA.primitive_types and class_definitions are maps keyed by class name, P_BMM_CLASS.ancestor_defs is a map of P_BMM_BASE_TYPE keyed by type signature (a simple ancestor included), and P_BMM_GENERIC_TYPE.generic_parameter_defs is an optional map keyed by formal parameter name. The formal and actual generic parameters (P_BMM_CLASS.generic_parameter_defs, P_BMM_GENERIC_TYPE.generic_parameter_defs, BMM_GENERIC_CLASS.generic_parameters) are insertion-ordered IndexMaps, because the specification fixes their declaration order, and the model builders bind each actual parameter to its formal by name. In the released v1_0 generation, P_BMM_GENERIC_TYPE.generic_parameter_defs now carries the 1..* bound its BMM states. The nine openehr-* crates step to 0.0.123 (#4497).

  • The conformance pipeline runs Veredictum 0.3.1. The Conformance workflow drives the catalogue through the Veredictum Action twice, each run on fresh volumes: under the released profile, the frozen openEHR release, which is the record and decides the job, and under the development profile, the pinned master of each openEHR repository, which is published beside it under its own job summary heading and artifact and never counted toward it. Both runs include the SMART documents: the Action now takes the SMART Platform base URL and the committed test issuer’s key. The workflow reads the result paths and the exit status from the Action’s outputs. scripts/conformance.sh takes the profile as CONF_PROFILE (released by default; a development run writes to docs/conformance/ferroehr-development/) and drives the SMART documents in the same run, and the IXIT carries the test issuer’s scoped_tokens declaration (#4503).

  • The viewer groups a template’s releases. The template list draws one group per template, the newest release first with the number of releases and the earlier releases under it, in both the ADL 1.4 and the ADL 2 tab. A template’s page names its release beside its title, where every other release is a link with the newest marked, also with FerroUI’s scripts refused; the release select and its Open button above the tabs are gone. The Overview lists the releases with the compositions committed against each, and a new Compare tab lists the paths one release adds, removes and changes against another (#4506).

  • Every write checks the committer against the login that authenticated the request where the server can compare them: a committer naming another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, is refused with a 403. A name only, or an identifier of another authority, is stored as sent. The direct writes (composition, EHR status, directory, EHR, demographic) now also refuse a committer that breaks its own rules with a 422, as the contribution path already did (#4520).

  • An audit record no longer adds a row to audit.audit_search_token for its request id or for the version uid a create wrote. ITI-81 entity.identifier finds both from the record’s own request_id and resource_id columns, through a new hash index on request_id (audit migration 0020), so the token table stays flat as requests are audited. Retention now deletes the search tokens that no remaining record names, among them the request id tokens of records written by 5.6.0-rc5 and earlier, and the database still refuses to delete a token a kept record names (#4524).

  • A clinical or party node row below the root no longer repeats its archetype header, language and encoding: an archetype root’s archetype_details and archetype_node_id are rebuilt from the archetype column and a new rm_version column, and a language in ISO_639-1 and an encoding in IANA_character-sets keep only their code, in the new language and encoding columns. Every reader puts the keys back where they stood, so the canonical JSON, every AQL result and the served key order are unchanged; a header of any other shape stays in the fragment. A vital signs version takes 17.1 KB instead of 21.6 KB and an International Patient Summary 131 KB instead of 159 KB. Migrations ext/0007, clinical/0028 and party/0020 add the columns and the readers; no data is migrated, and a row written before keeps its whole fragment and reads the same (#4525).

  • ext.storage_generation() reports generation-4, the storage generation with the node header columns, and the storage benchmark records for it are committed under docs/benchmarks/storage/generation-4/. Migration ext/0008 changes only the label; nothing in the serving path reads it (#4525).

  • Five response messages no longer name the openEHR Service Model. The admin dump route’s 400 for a missing file_sys_loc now reads “request body must carry a non-empty file_sys_loc string, a location on the server’s file system (FerroEHR’s own dump/load extension)”, and its 400 for an encoding member reads “encoding is not accepted: FerroEHR’s own dump/load extension defines no encoding, so no value names one”. The admin activity-report 400s for a_service now list the accepted service names (“query parameter a_service is required: one of Admin, Definitions, …” and “query parameter a_service names no platform service: … (expected one of …)”). The not-implemented answer of store_query_set reads “store_query_set is not implemented: no openEHR specification defines a query set”. The openehr-rm validation docs for the three PARTY and PARTY_RELATIONSHIP relationship invariants now name the separate relationship container as FerroEHR’s own design against RM demographic §Party Relationships (#4531).

  • Breaking for clients that relied on the remaining commit-audit leniencies: the audit rules of the CONTRIBUTION commit now hold on every write path (#4543).

    • A description object without _type, such as {"value": "…"}, is refused with 400 on a contribution audit, a commit_audit and an attestation, and so is an attestation reason without _type. Send {"_type": "DV_TEXT", "value": "…"} or a DV_CODED_TEXT with its _type.
    • A stated empty system_id is refused with 422 on the direct writes too (openehr-audit-details: system_id="" on a create, update or delete), where the server used to replace it with its own. Leave system_id out to get the server’s, or send a non-empty one. A system_id that is not a string stays a 400 on both CONTRIBUTION routes.
    • A commit audit of type UPDATE_ATTESTATION handed to the library’s direct write calls (create_composition, update_composition, the EHR_STATUS, DIRECTORY and party writes) is refused with 400, as on the CONTRIBUTION routes. Send the audit as UPDATE_AUDIT and put each attestation in the version’s attestations list.
  • Members that NewContribution or UPDATE_AUDIT do not declare, such as a reason or is_pending on a plain audit or an extra top-level member of a contribution, are accepted and ignored: the released schemas leave both objects open (#4543).

  • The openehr-* crates take their version when they are published. A pull request that changes the crates no longer bumps their lockstep 0.0.x version: the crates publish reads crates.io, takes the next version when packaged content changed since the newest published one, and refuses to publish changed content under a version crates.io already holds. The version in the manifests on main can therefore trail the newest version on crates.io (#4555).

  • The served Web Template of an ADL 2 operational template writes the aqlPath of an ELEMENT’s value and of a data value’s parts without an [idN] predicate, since a data value carries no archetype_node_id (ADL2 master05-paths §“Relationship with W3C Xpath”). A DV_INTERVAL and its parts read …/value, …/value/lower and …/value/upper, the form the ADL 1.4 Web Template writes, where they read …/value[id6]/lower[id7] before, and the parts’ ids are lower and upper (#4559).

  • The served Web Template gives every DV_INTERVAL<T> node its lower and upper children, each of type T, also when the operational template leaves them unconstrained (RM data_types §DV_INTERVAL). Such a node had no children before. An unconstrained part carries the open inputs of its type, and its aqlPath is …/value/lower or …/value/upper. This holds for ADL 1.4 and ADL 2 templates (#4559).

  • The template example fills such an interval: an unconstrained part takes the example value of its constrained sibling, so the limits stay consistent (#4559).

  • The served Web Template of an ADL 2 operational template writes a coded input whose at-codes the template binds to a terminology as the bound terms: its code list carries each binding’s code and label and its terminology names the bound terminology (ADL2 master08-terminology_integration §“From Constraints to Concrete Codes in Data”). COMPOSITION.category bound at1 to openehr::433 reads {"value": "433", "label": "event"} with terminology: "openehr", where it read the local at1 before, and an [acN@ttttt] constraint takes the bound terms of the terminology it names. A FLAT or STRUCTURED composition without a category, and the example composition, now carry openehr::433. The aqlPath of every data value drops its [idN] predicate, so a constrained category and context/setting read /category and /context/setting (#4574).

  • A CONTRIBUTION’s versions[] entry records its merge inputs: the other_input_version_uids it states are stored on the committed ORIGINAL_VERSION and served on every read, where the server used to refuse the member with 400. The uids may name versions the server does not hold; an empty list or an entry that is not an OBJECT_VERSION_ID is a 422 (#4577).

  • Other members UPDATE_VERSION does not declare, such as item or an extension member, are accepted and ignored, where the server used to refuse them with 400: the released schema leaves the object open, as it does NewContribution and UPDATE_AUDIT. A uid or contribution on a version entry is still refused with 400, because the server assigns both, and so is a _type other than UPDATE_VERSION or ORIGINAL_VERSION (#4577).

  • The admin dump route’s refusals name its parameters by their wire names and list the accepted values, not the Service Model class names. A bad logical_format or compression_format now reads “<field> must be a string, one of …” or “<field> names no accepted value: … (expected one of …)”, and a non-integer segment_split_size reads “segment_split_size must be an integer size in kb”. The served OpenAPI text for the route no longer names EXPORT_SPEC or EXPORT_FORMAT (#4634).

  • Five archive dump and load failures still answer 500, and their bodies are now fixed text that echoes no value from the archive: (#4657)

    • an unreadable data segment no longer names the segment, though manifest.json and demographic-commons.json are still named;
    • an unknown logical format in the manifest no longer echoes the declared value;
    • a version whose id carries a path separator, and a version whose commit audit the export lacks, no longer name the version.
  • A failed expansion of a party’s protected national identifiers still answers 500, and its body now reads “expanding the protected identifiers of a party version failed” without echoing the requested version uid: a 500 body carries curated text only (#4657).

  • Two query failures still answer 500, and their bodies now carry curated text only. A stored query whose record holds no query text reads “the stored query has no query text” without echoing its name, and a terminology server failure while a TERMINOLOGY('expand', …) operand is resolved reads “the terminology service failed while the query was being planned” instead of the upstream message (#4657).

  • The db_transactions counter (outcome="commit") now counts every committed single-object write. It used to leave out EHR_STATUS and DIRECTORY writes (#4659).

  • OPTIONS on the API base path sends conformance_profile as BASE+OPTIONS where it sent STANDARD. The build derives the value from the committed Veredictum record, docs/conformance/ferroehr/summary.md: BASE+OPTIONS when the BASE rows and the run OPTIONS rows have none failed and none errored, BASE when only the BASE rows do, and the member left out otherwise. ITS-REST 1.1.0 types the member as a free string and CNF names no profiles, so the values are FerroEHR’s own. The [rest.identity] conformance_profile configuration key is removed: a configuration that still sets it is refused at boot as an unknown key (#4672).

  • The viewer answers every request other than GET and HEAD that arrives without a session, to any path but the sign-in page, the OIDC routes, sign-out and the search answer, with the sign-in page that says the session ended, before any handler runs. A post to a page returns there after signing in; a post to the export, preference, sidebar or paging route returns to the dashboard. The export post used to answer 303 to /login, and the preference, sidebar and paging posts ran without a session (#4724).

  • The conformance instrument is Veredictum 0.4.0, and the published conformance record (docs/conformance/ferroehr/results.json with its summary.md) is a 0.4.0 released-profile run. The party statement answers the 0.4.0 catalogue: it keeps the xml-namespace-v2 arm, which now asks only whether the TRIAL http://schemas.openehr.org/v2 namespace is also accepted, and selects the new development-profile option arm adl2-id-coded-supported (#4771).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

  • The Veredictum 0.1.6 conformance documents: CONFORMANCE_REPORT.md, CONFORMANCE_STATEMENT.md, CONFORMANCE_CERTIFICATE.md, verdicts.json, run-exceptions.json, the CORE, STANDARD, OPTIONS, SEC-BASIC and performance badge files, and the heat grid and chapter bars the book and the landing page drew from them (#4451).

  • The CONF_PERF_CLASS, CONF_PERF_HOURS and CONF_STRESS settings of the measured lanes are gone, along with the class runs, the step-load stress ladder and the AQL probe, which Veredictum 0.3.1 removed (#4452).

  • The separate SMART step of the Conformance workflow, the smart stage of scripts/conformance.sh with its CONF_SMART_OUT, and docs/conformance/party/ferroehr/ixit-smart.json are gone: the one catalogue run covers the SMART documents. The party statement no longer carries the performance block, which the Veredictum 0.3.1 statement schema refuses (#4503).

  • The empty linkage storage domain is gone: its schema, its ferroehr_linkage role, the [storage.linkage] and [storage.migrate.linkage] configuration tables, its pool, the database.linkage and backup.linkage chart values, the ferroehr-backup-linkage compose job and the btree_gist extension only its table needed. A configuration that still sets [storage.linkage] or [storage.migrate.linkage] is refused at boot as an unknown key. Cohort and AQL demographic join crossings still write their linkage-domain access records. Upgrade note: remove those keys and values before you upgrade. On a co-located database, migration audit/0021_drop_linkage_domain.sql drops the schema, the role and the extension at the first boot, and raises a NOTICE for any of them it cannot drop. When the linkage domain lived in a database of its own, [storage.audit] names another database, or the cluster holds more than one FerroEHR database, run in each such database, as a superuser or a member of ferroehr_linkage: DROP SCHEMA IF EXISTS linkage CASCADE; REVOKE ALL ON SCHEMA ext FROM ferroehr_linkage; REVOKE ALL ON ALL FUNCTIONS IN SCHEMA ext FROM ferroehr_linkage; ALTER DEFAULT PRIVILEGES IN SCHEMA ext REVOKE EXECUTE ON FUNCTIONS FROM ferroehr_linkage; DROP OWNED BY ferroehr_linkage; DROP EXTENSION IF EXISTS btree_gist; then DROP ROLE ferroehr_linkage; once per cluster. The released grant files still name ferroehr_linkage, so a first install whose migrator holds no CREATEROLE still provisions that role and drops it after the first boot (#4684).

  • A party migration drops the database function party.resolve_national_identifier. No server path has called it since the linkage domain was removed (#4697).

  • The viewer no longer reads LEPTOS_SITE_ADDR: a value naming another address than server.listen no longer refuses the start. server.listen (FERROEHR_VIEWER__SERVER__LISTEN) sets the listen address (#4725).

  • Migration audit/0022 drops the SQL functions audit.reap_audit_events(integer), audit.reap_audit_events_years(integer) and audit.reap_audit_events_before(timestamptz), along with their EXECUTE grants. The server reaps the audit trail through audit.reap_audit_events_due, which honours each record’s own retention end, and nothing called the three any longer. The spec_profile documentation no longer names a LANG generation, since the profile selects the RM and BASE generations only (#4770).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • An AQL query aimed at named EHRs (the ITS-REST ehr_id parameter) reads those EHRs’ objects first and each object’s own content after them, whatever the planner statistics say. When an archetype’s rows fell out of the most-common-value list of the archetype statistics, the planner estimated them at one row and could start from that archetype over the whole population, evaluating every projected path for every EHR’s rows before the EHR filter: the storage bench’s wide projection over one EHR read about 10 000 pages per call in some rounds and 1 200 in others. A node contained in a versioned object is now read inside its parent on such a query; results do not change (#4487).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

  • AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri argument whole when its value carries a ?, as the QUERY master03 example url=http://snomed.info/sct?fhir_vs does. Before, the arguments were read from after the first ? anywhere in params_uri, so the value-set URL reached the terminology server truncated and the arguments before it were lost. A params_uri that starts with a path still has its arguments read from after the path’s ? (#4490).

  • An ADL 2 template whose constituent archetype the CDR does not hold is refused at upload, with a 422 whose errors carry one VARXR entry per missing archetype (AOM2 C_ARCHETYPE_ROOT validity rule VARXR); the 2013 CKM t_encounter_opt_test, built on the refused apgar, no longer loads, so the sandbox library pin is 231 accepted / 91 refused. A stored template whose constituent was deleted later answers its JSON form and example with the same VARXR list, and the viewer says so on every tab of the template, names each missing archetype and links to uploading it, instead of a raw 422 (#4505).

  • ADL 2 validation admits a constraint on an RM computed attribute. VCARM accepts an attribute “stored or computed” (AOM2 master04.5 §Validity Rules: C_ATTRIBUTE), and a function with no arguments is a computed attribute (master04.2 §Attribute Nodes), so an archetype that constrains DV_PROPORTION.is_integral, DV_QUANTITY.is_integral or EVENT.offset is no longer refused with 422. The generated RM model (openehr_rm::v1_2::model) now carries each class’s computed attributes from the BMM functions (RmClass::computed_attributes, computed_attribute). openEHR-EHR-CLUSTER.microscopy_prostate_carcinoma.v1.0.0 from the 2013 CKM ADL 2 export now uploads, and the sandbox archetype-library pin moves to 232 accepted and 90 refused. DV_QUANTITY.property, which no RM release defines, stays refused. The nine openehr-* crates step to 0.0.124 (#4513).

  • DELETE /ferroehr/rest/extensions/v1/definition/artefact/adl2/{artefact_id} refuses with 409 Conflict while another stored ADL 2 artefact does not compile without the one deleted (a slot filler, a use_archetype reference or a specialisation parent, directly or through another artefact), and the message names every one of them. Before, the delete succeeded and left each template built on the archetype unable to compile (#4522).

  • The viewer’s ADL 2 template upload names each violation the CDR found when it refuses the source: every AOM2 rule code with the CDR’s words for it, so a template whose constituent archetype the CDR does not hold names that archetype (VARXR). The upload now asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details, and reads the stored template’s id from the ETag of the 201 (#4523).

  • An ADL 2 operational template whose included archetype reuses the template’s node ids uploads again, where it was refused with 422 under VCOSU. Node-id uniqueness now holds per archetype and restarts at each C_ARCHETYPE_ROOT (AOM 2 VCOSU, “unique within the archetype”), and on an operational template the codes and interior node ids below a root are judged against that constituent’s entry in component_terminologies (OPT 2 master03-opt_raw). A duplicate node id inside one archetype is still refused. The nine openehr-* crates step to 0.0.125 (#4529).

  • The viewer’s dashboard draws the EHR or composition count as unavailable, with the reason in plain words, when the CDR answers its count query with a result set of another shape (no row, several rows, an empty row, several values, or a value that is no whole number), where it drew 0 before. The template Releases table draws such a count the same way. A real zero still draws 0 (#4533).

  • An ADL 2 operational template upload runs the terminology integrity rules (VTLC, VATCV, VTVSID, VTVSUQ, VTVSMD, VTTBK, VTCBK, and VTSD where the level is carried) on each component_terminologies entry, and each violation names the component’s archetype id (#4536).

  • A COMPOSITION committed against an ADL 2 operational template is validated against the template’s leaf constraints, where it was stored with 201 whatever it held. An inlined archetype recorded by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability) now matches its template node, so the C_INTEGER, C_REAL, C_STRING, C_BOOLEAN, tuple, value-set and cardinality constraints below it answer 422 (AOM 2 master04.5). An archetype root the template does not inline is refused as a node the template does not define, and a C_INTEGER or C_REAL list ({10, 15, 20}) admits each of its values where only the first was read. Canonical, FLAT and STRUCTURED commits take the same check. The nine openehr-* crates step to 0.0.127 (#4540).

  • A versioned read whose version path segment is no SEMVER version and no {major} or {major}.{minor} prefix, such as v1.x, 1.x or 01, now answers 400 before any lookup, where it answered 404 (parameters/path/version.yaml, responses/400.yaml). This covers the ADL 2 template read GET /definition/template/adl2/{template_id}/{version}, the stored-query read GET /definition/query/{qualified_query_name}/{version} and both stored-query executions at a version, GET and POST /query/{qualified_query_name}/{version}. A well-formed version the server does not hold still answers 404 (#4541).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the content below an inlined archetype that the composition records by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability). The read left out everything below such a root with no error; it now matches the root to its template node the way the commit does. The nine openehr-* crates step to 0.0.128 (#4551).

  • A COMPOSITION committed against an ADL 2 operational template whose C_INTEGER or C_REAL constraint is a union of ranges (for example magnitude matches {|0..5|, |10..15|}) is refused with 422 when the value lies outside every range. Such a value was stored. The served Web Template does not change: its input validation.range holds one range, so a union leaves it out, as before, and the check runs on the commit (#4552).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the value of an ELEMENT that holds a DV_INTERVAL whose lower and upper the template constrains. The read left the value out with no error (#4559).

  • A FLAT or STRUCTURED read of a COMPOSITION returns the value of an ELEMENT that holds a DV_INTERVAL whose template leaves the interval’s lower and upper unconstrained, for ADL 1.4 and ADL 2 templates. The read left the value out with no error, and a FLAT or STRUCTURED commit could not name the parts (#4559).

  • A COMPOSITION written from FLAT or STRUCTURED against an ADL 2 operational template carries what a canonical body built to that template carries. A node the template gives no term text, such as the HISTORY and ITEM_TREE of an OBSERVATION, is named for its RM type (History, Tree) instead of its bare id-code (id2). Every inlined archetype root records the complete identifier the template names (…cnf_minimal.v1.0.0, BASE master07-referencing §Reconstitutability) in its archetype_node_id and archetype_details.archetype_id, and archetype_details.rm_version is the template’s rm_release. Before this change the root recorded the interface form (…v1) and the server’s RM 1.2.0. Compositions built against an ADL 1.4 template keep their major-version archetype ids and RM 1.2.0 (#4560).

  • ferroehr report in a release build names the openehr-* crate versions that release published to crates.io: the release lanes stamp the published version into the checkout before they build the server, where the report used to name the floor the source tree holds. A new openehr_crates_origin field says published for a release build and source_tree for any other, a pre-release included (#4565).

  • An ADL 2 archetype reference resolves to one defined stored release: an interface reference (….v1) takes the newest release of its major version, a reference with a minor or patch version narrows the match to it, and an alpha or beta version matches only its full identifier (BASE Artefact Identification, Referencing). With several releases of one archetype stored, a template upload, an operational template and the delete’s dependents check used whichever release PostgreSQL returned last, and a reference to a major version no stored release carries resolved to another major version (#4567).

  • An ADL 2 upload and an ADL 2 delete no longer race. An upload validated before a concurrent delete of an archetype it uses committed could store a template whose constituent was gone. Both now hold one transaction-scoped advisory lock from the repository read to the commit, so the second write decides on what the first committed: an upload after the delete is refused with VARXR, and a delete after the upload is refused with a 409 naming the template (#4568).

  • The syslog form (DICOM PS3.15, IHE ITI-20) of an EHR-level audit record names its object beside the patient, as a URI participant object with the id its FHIR AuditEvent names. The record of placing, releasing or expiring an EHDS Art. 3(3) access delay names the held composition as access-delay:<vo>, so an Audit Record Repository fed by syslog can tell which composition a hold covers (#4569).

  • An ADL 2 template or archetype that references two major versions of one archetype (….v1 and ….v2), or one archetype under two namespaces, now compiles against both: the openehr-adl repository holds an artefact under its namespace, archetype and major version, and resolves each reference within that key (BASE Artefact Identification, Referencing). A reference without a namespace now resolves in the namespace of the artefact that makes it, and never in another namespace (#4579).

  • A read of a stored ADL 2 template by a partial id or a version (GET /definition/template/adl2/{template_id}) now resolves the way template compilation does: with v1.0.0 stored beside v1.0.0-rc.1, ….v1, ….v1.0 and version=1.0.0 answer the release in either upload order, and the release candidate is read only by its full version (BASE Artefact Identification, Referencing) (#4580).

  • A write that sends no committer, on a route that takes a bare RM body (the EHR_STATUS, COMPOSITION and DIRECTORY writes, and the demographic party and relationship writes), now records a Bearer caller’s login as issued by the token issuer, as EHR creation and every other default committer already did. It recorded the issuer ferroehr for every caller before; a Basic caller’s login is still issued by ferroehr (#4585).

  • The viewer’s ADL 1.4 template upload opens the stored template’s page, read from the ETag (else the Location) of the CDR’s 201, where it went to an address with no template id. A refused upload names each violation the CDR found, its rule code or path and the CDR’s words for it, as the ADL 2 upload does: the upload asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details (#4596).

  • A 422 validation refusal tells the client to send Prefer: return=representation only when the request did not send it; a request that sent it receives the violations in errors and a detail without the advice (#4597).

  • An EHR Extract import refusal for a received version now names that version’s position in the extract (chapters[i]/items[j]/item/versions[k]), so an importer can find the defective ORIGINAL_VERSION in a large extract (#4602).

  • The import reads a version’s other_input_version_uids with the parser the CONTRIBUTION path uses and refuses an entry that is not an OBJECT_VERSION_ID, where it used to drop the entry. On the wire the extract decode already refuses such an entry with 422, so the import stays all or nothing (#4602).

  • An AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts within the major: org.a::… no longer finds data recorded in org.b, an id without a namespace finds only data recorded without one, and .v1.4.2 no longer finds data recorded under v1.0.0 (BASE Artefact Identification §Referencing and §Supporting Archetype-based Querying; #4629, #4630).

  • A query naming a parent archetype finds the data of a specialised child release only when that release declares a parent release at or after the queried one. Before, once one stored release of the child qualified, the data of every release of it was returned (BASE Artefact Identification, Referencing §Supporting Archetype-based Querying; #4637).

  • POST /definition/template/adl1.4 and POST /definition/template/adl2 answer 400 for every invalid template, the only refusal ITS-REST 1.1.0 declares for both uploads besides the 409 duplicate. An OPT that breaks the OPT 1.4 schema, an ADL 2 source that breaks an AOM2 rule such as VCORM or VCARM, and an ADL 2 template whose constituent archetype is not held were refused with 422. The problem keeps its detail and, with Prefer: return=representation, its errors. Upgrade note: a client that read a 422 from a template upload reads a 400, with the same detail and errors, under the about:blank problem type (#4643).

  • A directory create or update whose FOLDER carries an empty folders array, at the root or in a sub-folder, is refused with 400 and commits nothing. The ITS-JSON FOLDER schema gives folders minItems: 1, and the canonical JSON reader now refuses the empty array the same way in every document that carries a FOLDER, a CONTRIBUTION included. The viewer leaves folders out of a folder with no sub-folders (#4644).

  • A canonical XML FOLDER, EHR_STATUS, PERSON, AGENT, GROUP, ORGANISATION, ROLE or PARTY_RELATIONSHIP document now has the root element items with xsi:type naming its class, the element the ITS-XML schemas publish for a LOCATABLE (Structure.xsd). The directory reads (latest, by version and at a time), the EHR_STATUS reads, the demographic reads and the demographic template examples served <folder>, <ehr_status>, <person> and the like, which no ITS-XML schema declares. XML input accepts the same form (#4645).

  • POST /definition/template/adl1.4 builds the template’s Web Template before it stores the template, so a template it cannot build is refused with 400 and nothing is stored. Once the template is stored, the upload answers 201 with its Location; with Prefer: return=representation and a Web Template Accept it serves the Web Template built before the store, and the OPT if that cannot be served. It no longer answers 422 for a template it has already stored (#4647).

  • A composition whose HISTORY carries events: [], and an AGENT, GROUP, ORGANISATION or PERSON carrying languages: [], are refused with 400 and commit nothing. The ITS-JSON schemas give both arrays minItems: 1, and the canonical JSON reader now refuses the empty array as it does an empty FOLDER folders. The FLAT builder no longer writes an empty events for a HISTORY that has no event (#4648).

  • A write to a restricted COMPOSITION (POST, PUT or DELETE on /ehr/{ehr_id}/composition) refused under a GDPR Art. 18 restriction now answers with the processing-restricted problem type, as the reads of a restricted record already did, in place of the generic forbidden type. The status stays 403 (#4657).

  • When an EXTRACT_SPEC criterion’s AQL runs past the query time budget, the 400 body now states the timeout as plain text. It used to start with a control-character marker that was only ever meant for the server itself (#4657).

  • Under a separated clinical credential, POST /ehr, PUT /ehr/{ehr_id} and an EHR_STATUS update that name a subject no longer fail with a database error: the subject pseudonym guard reads the deployment posture, and the clinical role now holds SELECT on ext.posture through the ext migration 0009 (#4675).

  • A cohort that names its subjects now finds, under [ehr_index] provider = "local", the EHR created over POST /ehr with that subject in its EHR_STATUS; it read a cross-reference table no route wrote, so it matched nothing. A cohort predicate and the AQL demographic join resolve a party to the EHRs whose EHR_STATUS.subject names it. The EHR_STATUS subject is now the one source of which EHR belongs to which subject. Upgrade note: linkage.subject_ehr and the linkage.erase_ehr function go with the linkage schema, which audit migration 0021 drops with the linkage domain. Rows written there through the service API are dropped with the table; the EHRs they named keep their subject in their own EHR_STATUS (#4675).

  • The identifier scan ([privacy.identifier_scan]) no longer refuses a DV_IDENTIFIER whose type declares an item rather than a person, such as a laboratory filler_order_number that passes the NHS Number check by chance. A DV_IDENTIFIER in FEEDER_AUDIT.originating_system_item_ids or feeder_system_item_ids is scanned only when its type names a person identifier. A type naming a person (patient, NHS, BSN, social security and similar) keeps the value scanned in every slot, and free text and untyped identifiers in clinical content are scanned as before. The default rule set is unchanged; a deployment that knows its jurisdictions still narrows rules, so a Dutch one with rules = ["nl-bsn"] never claims an NHS Number (#4702).

  • An inline primitive constraint in a parsed archetype or template now carries the node id AOM2 ADL_CODE_DEFINITIONS.Primitive_node_id defines, at9999 in an at-coded archetype and id9999 in an id-coded one, where it carried the constant’s name Primitive_node_id. An ADL 1.4 archetype converted to ADL2 gets the id-coded form (#4739).

  • The ABAC policy-enforcement point checks every item-tag operation as it checks its target: a COMPOSITION tag read after the call like the COMPOSITION read, a COMPOSITION tag write before the call with the tagged version’s template, category and origin, and the EHR tag read before the call like the EHR read. These operations were unchecked before, so a patient-scoped or template-scoped rule now refuses a tag read or write wherever it refuses the target’s read or write. A COMPOSITION tag read also falls under the EHR_ACCESS privacy ceiling and its audit record names the tagged COMPOSITION (#4768).

  • A primitive constraint inside a rule or a slot assertion of an id-coded ADL 2 archetype carries id9999 as its node id, as org.openehr.am.aom2.adl_code_definitions §Constants defines Primitive_node_id for id-coded archetypes, where it carried at9999 before. An ADL 1.4 archetype converted to ADL 2 recodes its slot assertions the same way. openehr_adl::rules::parse_rules_body and parse_slot_assertions take the archetype’s NodeCoding, which openehr_adl::codes now exports (#4770).

  • Under spec_profile = "stable", a canonical-JSON party read lists in reverse_relationships every PARTY_RELATIONSHIP another party’s current version carries with this party as its target, one LOCATABLE_REF each naming the source party’s version container and the relationship’s path, as RM 1.1.0 §Party Relationships requires. The development profile, whose RM 1.2.0 makes the attribute a function, is unchanged. The released-profile conformance run now judges the server on the stable profile (#4775).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

  • CVE-2026-84243, the glibc gettext LANGUAGE directory traversal (GLIBC-SA-2026-0025), has no Debian fix for trixie yet and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never leave the C locale, so glibc ignores LANGUAGE and opens no message catalog (security/vex/ferroehr-os.openvex.json). In ferroehr-postgres only the container environment the operator writes can set LANGUAGE; no database connection can (security/vex/postgres-os.openvex.json) (#4500).

  • CVE-2026-114690, CVE-2026-114691 and CVE-2026-115455, three libstdc++ defects in the gcc-14 packages of every image, have no Debian fix for trixie and are judged not_affected. Each fix changes only a header template (the parallel-mode search, the std::vector<bool> constructors, wide-stream extraction into std::bitset), so libstdc++.so.6 holds no copy; the ferroehr and ferroehr-viewer programs contain no C++ code (security/vex/ferroehr-os.openvex.json), and in ferroehr-postgres nothing is built in parallel mode and PostgreSQL does not link libstdc++ (security/vex/postgres-os.openvex.json) (#4789).

  • CVE-2026-119371 to CVE-2026-119379 and CVE-2026-119591, ten SQLite defects without a Debian fix, are judged not_affected in ferroehr-postgres: only lslogins reaches SQLite there, through liblastlog2, and nothing in the image runs it (security/vex/postgres-os.openvex.json) (#4789).

5.6.0-rc8 - 2026-11-02

5.6.0-rc8 is the eighth release candidate of 5.6.0, published for testing. It carries the CONTRIBUTION break of 5.6.0-rc6: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution read the body through the released NewContribution schema and answer 400 for a lifecycle_state that is no DV_CODED_TEXT, a description that is a plain string or an object without _type, an audit without change_type, an audit typed ATTESTATION or UPDATE_ATTESTATION, a missing audit or versions, and a preceding_version_uid sent as a plain string, and a stated empty system_id is refused with 422 on every write path (#3595, #4543). These notes carry every change since 5.5.0, those of the 5.6.0-rc7, 5.6.0-rc6, 5.6.0-rc5, 5.6.0-rc4 and 5.6.0-rc3 sections below included. New since 5.6.0-rc7: two read-only extension routes that convert a stored ADL 1.4 archetype or operational template to ADL 2 (#4655); template uploads that answer 400 for every invalid template and never refuse a template already stored (#4643, #4647); the ITS-JSON minItems of FOLDER.folders, HISTORY.events and ACTOR.languages enforced (#4644, #4648), and a canonical XML LOCATABLE rooted at the items element the ITS-XML schemas publish (#4645); an AQL query naming a parent archetype that finds a specialised child’s data only from a child release whose own parent release qualifies (#4637); conformance_profile derived from the committed conformance record (#4672); a subject resolved through the EHR_STATUS of its EHR, and the empty linkage domain removed with its configuration (#4675, #4684, #4697); a write refused under a GDPR Art. 18 restriction answering with the processing-restricted problem type, and 500 bodies that carry curated text only (#4657); an identifier scan that reads a DV_IDENTIFIER’s declared type (#4702); a validation refusal that advises Prefer: return=representation only to a request that did not send it (#4597); the AOM2 node id of an inline primitive constraint (#4739); a viewer that answers a request other than GET or HEAD without a session with the sign-in page and no longer reads LEPTOS_SITE_ADDR (#4724, #4725); every item-tag operation checked by ABAC in its target’s mode (#4768); and, under spec_profile = "stable", a canonical-JSON party read that lists the relationships naming it as their target in reverse_relationships (#4775). The published conformance record is a Veredictum 0.4.0 released-profile run against FerroEHR on spec_profile = "stable": 2872 of 3071 rows passed, none failed and none errored, and 199 were skipped, 116 for an option the statement does not select and 83 that rest on development text alone; it replaces the Veredictum 0.1.6 documents (#4451, #4771). The released-profile verdict is now required: a failed or errored row fails the merge queue, the image run on main and the release (#4455). The development-profile run, which is never a verdict, has 2946 rows passed, 3 failed (#4776, #4777) and 2 errored by Veredictum itself. 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: an invalid template on POST /definition/template/adl1.4 or POST /definition/template/adl2, an ADL 2 template whose constituent archetype is not held among them, answers 400 where it answered 422, with the same detail and errors under the about:blank problem type; a FOLDER with an empty folders, a HISTORY with an empty events and an AGENT, GROUP, ORGANISATION or PERSON with an empty languages answer 400; a canonical XML FOLDER, EHR_STATUS, PERSON, AGENT, GROUP, ORGANISATION, ROLE or PARTY_RELATIONSHIP document has the root element items with xsi:type naming its class, where it had <folder>, <ehr_status>, <person> and the like; OPTIONS on the API base path sends conformance_profile as BASE+OPTIONS where it sent STANDARD, and a configuration that sets [rest.identity] conformance_profile is refused at boot as an unknown key; an AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts, and a query naming a parent archetype finds a specialised child’s data only from a child release whose own parent release qualifies, so a query may find less than before; a cohort that names its subjects and the AQL demographic join resolve a party to the EHRs whose EHR_STATUS.subject names it, and a configuration that sets [storage.linkage] or [storage.migrate.linkage] is refused at boot and Helm chart 21.0.0-rc8 refuses a values file that sets database.linkage or backup.linkage, so remove those keys and values before you upgrade and, where the linkage domain lived in a database of its own, run the SQL the Removed entry gives; a POST, PUT or DELETE on a COMPOSITION under a GDPR Art. 18 restriction still answers 403, now with the processing-restricted problem type in place of forbidden; the 500 bodies of the archive dump and load, a failed expansion of a party’s protected identifiers and two query failures carry fixed text that echoes no value from the request, the archive or the terminology server; the identifier scan no longer refuses a DV_IDENTIFIER whose type declares an item rather than a person, and scans one in a FEEDER_AUDIT only when its type names a person identifier; the Veredictum 0.1.6 conformance documents (CONFORMANCE_STATEMENT.md and the rest) and the CONF_PERF_CLASS, CONF_PERF_HOURS and CONF_STRESS settings of the measured lanes are gone; with ABAC on, a patient-scoped or template-scoped rule refuses an item-tag read or write wherever it refuses the read or write of the tag’s target, and a COMPOSITION tag read falls under the EHR_ACCESS privacy ceiling; under spec_profile = "stable" a canonical-JSON party read carries reverse_relationships, while the canonical-XML read does not yet (#4778) and the development profile is unchanged; a write whose committer names another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, answers 403, and a direct write whose committer breaks its own rules answers 422; a write that sends no committer on a bare RM body route records a Bearer caller’s login as issued by the token issuer, where it recorded ferroehr; a FLAT or STRUCTURED composition under an ADL 2 template records the complete archetype identifier and the template’s rm_release, and the served Web Template writes a bound coded input, COMPOSITION.category among them, as the bound terms (openehr::433) where it wrote the local at-code; a CONTRIBUTION commit must take the shapes above; a write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as committer and may change only what the person committed, and a service auth_failure answers 403 where it answered 500; a versioned read whose version is no SEMVER version or prefix answers 400 where it answered 404; the served Web Template of an ADL 2 template writes a data value’s aqlPath without an [idN] predicate; a composition that breaks an ADL 2 template’s leaf constraint answers 422; every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes, so keep it longer than the idle timeout of your clients and load balancers; a client that wrote a space as + in a query string sends %20; a party delete names the latest OBJECT_VERSION_ID in its path and answers 400 for a versioned object id; a body the RM refuses after the ITS-JSON schema admits it answers 422, and a misplaced CONTRIBUTION member, a Simplified Format commit without openehr-template-id and a flat |raw value without _type answer 400; a tag update under Prefer: return=identifier answers 200 with the tag list; with [authz.rbac] enabled a token holding only system/ SMART on openEHR scopes reaches clinical operations without a role; the viewer image serves no /pkg/ and reads no LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, and its Content-Security-Policy is script-src 'self'; style-src 'self'; Helm chart 21.0.0-rc8 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data, and ext migration 0008 names generation 4, the generation with the node header columns, and changes only the label; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; clinical 0024 and party 0015 index the version uid, ext 0006, clinical 0025 and party 0016 add the ELEMENT value search index, built over every hot node row of an existing store, party 0017 adds the archetype statistics, and clinical 0026 and party 0018 grant MAINTAIN on the partitioned tables to the serving roles; clinical 0027 and party 0019 add the access delay, and ext 0007, clinical 0028 and party 0020 add the node header columns with no data migrated; audit migration 0019 gives the search token index a pending list of 512 kB, and audit 0020 adds the hash index on request_id; clinical 0029 adds the access opt-out; ext migration 0009 grants the clinical role SELECT on ext.posture, party 0021 drops the database function party.resolve_national_identifier, audit 0021 drops the linkage schema, the ferroehr_linkage role and the btree_gist extension, raising a NOTICE for any of them it cannot drop, and party 0022 restates the comment on party_relationship_target.source_vo_id with no change to data; the linkage migration set is gone. Helm chart 21.0.0-rc8 carries appVersion 5.6.0-rc8. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j, and the OpenVEX judgement of CVE-2026-84243.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The operations page gains a storage sizing section, measured on 5.6.0-rc5: about 21.6 KB of clinical storage per version of a vital signs composition (2.0 times its canonical JSON, 85 % of it node rows) and 155 KB for an International Patient Summary, about 1.2 KB per audit record of a read, the effect of [audit.store] retention on growth, pg_wal up to max_wal_size, the REINDEX INDEX CONCURRENTLY clinical.node_hot_pkey step (20 % of that index after heavy concurrent writing) and a query for the size per relation, with the method to reproduce the figures (#4304).

  • EHDS Art. 5: what a person inserts in their own EHR is marked as theirs, and they cannot alter what a professional committed. A write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as its AUDIT_DETAILS.committer, its external_ref naming the login, whatever committer the request names; a COMPOSITION it commits must name PARTY_SELF or PARTY_RELATED as composer; and its update or deletion of a version whose committer is not PARTY_SELF, directly or inside a CONTRIBUTION, is refused with 403. A professional’s writes are unchanged. The contribution list extension gains inserted_by_person per row, and the viewer names a PARTY_SELF committer “The person or their representative”. A service auth_failure status now answers 403 where it answered 500 (#4350).

  • The EHDS Art. 3(3) delay of the person’s own access to newly registered data, a Member State option and off by default. The new [access_delay] section (enabled, categories, hours) holds every new COMPOSITION of a listed [audit.categories] category from requests that may be the person’s own (a bearer token with a SMART patient/ grant or the ABAC patient claim, also beside user/ or system/ grants, as ferrohealth-smart 0.8.7 decides it) for at most hours: a read answers 404 as for an unknown composition, and a query, an export and a CONTRIBUTION leave it out, while a professional reads it as before. The treating professional releases a held composition from the clinical route POST {ext}/ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/access-delay/release, with the checks of a write to that EHR, and the controller with POST /admin/access-delay/release; a professional holds one by hand within the limit with POST /admin/access-delay and reads the register with GET /admin/access-delay; every hold ends at its limit, and placing, releasing and expiry are each an access record. The subject-scoped ITI-81 read of the person’s own trail (EHDS Art. 9) leaves out the records about a held composition until it is released or expires, then serves them with their original times (ehds-logging 0.3.3) (clinical migration 0027_access_delay.sql, party migration 0019_access_delay_column.sql) (#4351).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • The Conformance workflow runs the Veredictum smart documents (smart_discovery, smart_scopes) against the same FerroEHR build after the Veredictum Action, which forwards no SMART Platform base URL and no scoped tokens. The same pinned image runs veredictum test with --platform-url and docs/conformance/party/ferroehr/ixit-smart.json, minting each access token in memory from the committed test issuer the SMART overlay trusts, and the job uploads its results, JUnit and CTRF views beside the catalogue’s and adds its summary under a heading of its own. scripts/conformance.sh gains the smart stage, and run drives it after the catalogue, so a local run covers SMART too (#4450).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

  • The EHDS Art. 10 opt-out from access through the access services of Arts. 4 and 12, a Member State option and off by default. The new [authz.rbac] access_service_roles key declares the roles of the access-service callers; while an EHR carries the opt-out, a caller holding one reads it as absent: a route naming the EHR or an object in it answers 404 as for an unknown EHR, a population query leaves its data out, a scoped query, an export and the lookup by subject find no such EHR, and a write to it is refused as absent. Every other caller, the provider’s own care included, reads it as before. The controller records the opt-out on the person’s instruction with POST /admin/access-opt-out, reverses it with POST /admin/access-opt-out/reverse (the register keeps the reversal) and reads the register with GET /admin/access-opt-out; recording is refused with 400 while no access-service role is declared. An access-service caller that declares one of [audit] emergency_purpose_codes and does not act for the person reads the EHR under the Art. 10(2) vital-interest override, with the emergency mark on its access record and an access-opt-out-override record against the EHR. Recording, reversing and reading the register are access records (clinical migration 0029_access_opt_out.sql) (#4516).

  • Two read-only extension routes convert stored ADL 1.4 definitions to ADL 2: GET /ferroehr/rest/extensions/v1/definition/archetype/adl1.4/{archetype_id}/adl2 for a stored archetype and GET /ferroehr/rest/extensions/v1/definition/template/adl1.4/{template_id}/adl2 for a stored operational template, which converts to one ADL 2 archetype per embedded archetype root. Each answers 200 with application/json: (#4655)

    • the ADL 2 source of every root;
    • each root’s conversion log and its conversion_details, which name what the ADL 2 form drops or changes;
    • the slot fills that joined the roots. An unknown id answers 404. A stored source that no longer converts answers 422. No openEHR specification defines the conversion, so both routes are FerroEHR’s own design.

Changed

  • Breaking for clients that relied on the old leniencies: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution now read the body through the released NewContribution schema, and refuse every shape it does not define. Each refused shape answers 400: (#3595)

    • a lifecycle_state in the CODE_PHRASE shape {"terminology_id": …, "code_string": …}, as a plain string, or as a bare {"value": …}. Send a DV_CODED_TEXT: {"value": "complete", "defining_code": {"terminology_id": {"value": "openehr"}, "code_string": "532"}};
    • a description sent as a plain string, on the contribution audit, on a commit_audit or on an attestation. Send a DV_TEXT or DV_CODED_TEXT object: {"_type": "DV_TEXT", "value": "…"};
    • a commit_audit or contribution audit without change_type. The server no longer infers creation or modification from preceding_version_uid; send the code of the change, for example 249 (creation), 251 (modification), 523 (deleted) or 666 (attestation);
    • a commit_audit or contribution audit whose _type is ATTESTATION or UPDATE_ATTESTATION, or any value other than UPDATE_AUDIT or AUDIT_DETAILS. Send the audit as UPDATE_AUDIT (or with no _type) and put each attestation in the version’s attestations list as an UPDATE_ATTESTATION. A 666 (attestation) version carries its attestations there too, and one with no attestation is refused;
    • a missing audit, which answered 422 before. A missing versions list, an attestation without reason or is_pending, a system_id that is not a string, and a preceding_version_uid sent as a plain string instead of {"value": "…"} are refused the same way.
  • An attestation’s own system_id is kept, and one that states none takes the system_id of the version audit it belongs to instead of the server’s. An attestation committer is now checked against its PARTY_PROXY rules like every other committer, so an invalid one answers 422 (#3595).

  • The served OpenAPI document describes each extension route by what it does and by the ITS-REST text that governs it, and no longer names openEHR Service Model operations, interfaces or status codes. Routes, request and response fields and status codes are unchanged (#4213).

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The storage-bench lane measures the baseline and the candidate on one runner in eight interleaved rounds each and gives two verdicts. The plan verdict fails an operation whose median pages per call rose more than 10 % and one page above the baseline’s highest round, whatever its time. The time verdict divides each round by its machine factor, so a slow spell of the runner no longer reads as the code, and fails an operation whose median p50 or p99 rose more than 15 % with every candidate round slower than every baseline round, or a candidate build that ran more than 15 % slower across the board in every round. A latency is no longer compared with a record measured on another runner (#4318).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • The LANG specification text is re-vendored at f80763ab (#4377). SPECLANG-18 states how a P_BMM schema’s JSON form encodes a cardinality interval, which the JSON schema reader already followed; its documentation now cites the rule, and a test pins it. SPECLANG-16 types the P_BMM class lists as keyed collections; the generated P_BMM model takes that change with #4497. The nine openehr-* crates step to 0.0.122.

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • The viewer runs on ferrohealth-server 0.4.2 and takes its Content-Security-Policy from the family’s server-rendered console headers (FerroHEALTH/FerroSYS#202) in place of a layer of its own. The policy a deployment sees is unchanged: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; object-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none', with no nonce. FerroUI’s assets under /ferrohealth-ui/ keep the Cache-Control FerroUI sets, and a refusal under that path is no-store whatever FerroUI’s router sets, so a cache in front of the viewer keeps no error answer for an asset. The REST server takes the same ferrohealth-server release with no change in what it serves (#4439).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The published conformance record is a Veredictum 0.3.3 released-profile run, committed as docs/conformance/ferroehr/results.json with its summary.md. The landing page and the book’s conformance page show the Veredictum badges of the released and development runs on main and the record’s counts per part (BASE and OPTIONS), in place of the Veredictum 0.1.6 heat grid, chapter bars and tier pills. The docs and landing pages admit images from raw.githubusercontent.com/FerroHEALTH/FerroEHR/ for the badges (#4451).

  • The measured performance lane runs Veredictum 0.3.3’s perf. scripts/conformance.sh perf composes the conformance deployment with the measurement overlay, runs the pinned clinical-mix workload, checks the record with perf-judge and redraws the book’s charts from docs/conformance/ferroehr/perf-result.json. scripts/conformance.sh perf-aa measures the deployment against itself, so the record carries its between-run noise floor. The measurement overlay also switches response compression off, and the IXIT names the deployment’s containers for the run’s server-side readings. The Performance and Benchmarks pages describe the new instrument. They keep the earlier figures until a record measured on the reference host replaces them (#4452).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

  • The P_BMM model follows SPECLANG-16. The LANG 1.1.0 BMM that openehr-codegen reads is now the specifications-LANG file at f80763ab, so P_BMM_SCHEMA.primitive_types and class_definitions are maps keyed by class name, P_BMM_CLASS.ancestor_defs is a map of P_BMM_BASE_TYPE keyed by type signature (a simple ancestor included), and P_BMM_GENERIC_TYPE.generic_parameter_defs is an optional map keyed by formal parameter name. The formal and actual generic parameters (P_BMM_CLASS.generic_parameter_defs, P_BMM_GENERIC_TYPE.generic_parameter_defs, BMM_GENERIC_CLASS.generic_parameters) are insertion-ordered IndexMaps, because the specification fixes their declaration order, and the model builders bind each actual parameter to its formal by name. In the released v1_0 generation, P_BMM_GENERIC_TYPE.generic_parameter_defs now carries the 1..* bound its BMM states. The nine openehr-* crates step to 0.0.123 (#4497).

  • The conformance pipeline runs Veredictum 0.3.1. The Conformance workflow drives the catalogue through the Veredictum Action twice, each run on fresh volumes: under the released profile, the frozen openEHR release, which is the record and decides the job, and under the development profile, the pinned master of each openEHR repository, which is published beside it under its own job summary heading and artifact and never counted toward it. Both runs include the SMART documents: the Action now takes the SMART Platform base URL and the committed test issuer’s key. The workflow reads the result paths and the exit status from the Action’s outputs. scripts/conformance.sh takes the profile as CONF_PROFILE (released by default; a development run writes to docs/conformance/ferroehr-development/) and drives the SMART documents in the same run, and the IXIT carries the test issuer’s scoped_tokens declaration (#4503).

  • The viewer groups a template’s releases. The template list draws one group per template, the newest release first with the number of releases and the earlier releases under it, in both the ADL 1.4 and the ADL 2 tab. A template’s page names its release beside its title, where every other release is a link with the newest marked, also with FerroUI’s scripts refused; the release select and its Open button above the tabs are gone. The Overview lists the releases with the compositions committed against each, and a new Compare tab lists the paths one release adds, removes and changes against another (#4506).

  • Every write checks the committer against the login that authenticated the request where the server can compare them: a committer naming another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, is refused with a 403. A name only, or an identifier of another authority, is stored as sent. The direct writes (composition, EHR status, directory, EHR, demographic) now also refuse a committer that breaks its own rules with a 422, as the contribution path already did (#4520).

  • An audit record no longer adds a row to audit.audit_search_token for its request id or for the version uid a create wrote. ITI-81 entity.identifier finds both from the record’s own request_id and resource_id columns, through a new hash index on request_id (audit migration 0020), so the token table stays flat as requests are audited. Retention now deletes the search tokens that no remaining record names, among them the request id tokens of records written by 5.6.0-rc5 and earlier, and the database still refuses to delete a token a kept record names (#4524).

  • A clinical or party node row below the root no longer repeats its archetype header, language and encoding: an archetype root’s archetype_details and archetype_node_id are rebuilt from the archetype column and a new rm_version column, and a language in ISO_639-1 and an encoding in IANA_character-sets keep only their code, in the new language and encoding columns. Every reader puts the keys back where they stood, so the canonical JSON, every AQL result and the served key order are unchanged; a header of any other shape stays in the fragment. A vital signs version takes 17.1 KB instead of 21.6 KB and an International Patient Summary 131 KB instead of 159 KB. Migrations ext/0007, clinical/0028 and party/0020 add the columns and the readers; no data is migrated, and a row written before keeps its whole fragment and reads the same (#4525).

  • ext.storage_generation() reports generation-4, the storage generation with the node header columns, and the storage benchmark records for it are committed under docs/benchmarks/storage/generation-4/. Migration ext/0008 changes only the label; nothing in the serving path reads it (#4525).

  • Five response messages no longer name the openEHR Service Model. The admin dump route’s 400 for a missing file_sys_loc now reads “request body must carry a non-empty file_sys_loc string, a location on the server’s file system (FerroEHR’s own dump/load extension)”, and its 400 for an encoding member reads “encoding is not accepted: FerroEHR’s own dump/load extension defines no encoding, so no value names one”. The admin activity-report 400s for a_service now list the accepted service names (“query parameter a_service is required: one of Admin, Definitions, …” and “query parameter a_service names no platform service: … (expected one of …)”). The not-implemented answer of store_query_set reads “store_query_set is not implemented: no openEHR specification defines a query set”. The openehr-rm validation docs for the three PARTY and PARTY_RELATIONSHIP relationship invariants now name the separate relationship container as FerroEHR’s own design against RM demographic §Party Relationships (#4531).

  • Breaking for clients that relied on the remaining commit-audit leniencies: the audit rules of the CONTRIBUTION commit now hold on every write path (#4543).

    • A description object without _type, such as {"value": "…"}, is refused with 400 on a contribution audit, a commit_audit and an attestation, and so is an attestation reason without _type. Send {"_type": "DV_TEXT", "value": "…"} or a DV_CODED_TEXT with its _type.
    • A stated empty system_id is refused with 422 on the direct writes too (openehr-audit-details: system_id="" on a create, update or delete), where the server used to replace it with its own. Leave system_id out to get the server’s, or send a non-empty one. A system_id that is not a string stays a 400 on both CONTRIBUTION routes.
    • A commit audit of type UPDATE_ATTESTATION handed to the library’s direct write calls (create_composition, update_composition, the EHR_STATUS, DIRECTORY and party writes) is refused with 400, as on the CONTRIBUTION routes. Send the audit as UPDATE_AUDIT and put each attestation in the version’s attestations list.
  • Members that NewContribution or UPDATE_AUDIT do not declare, such as a reason or is_pending on a plain audit or an extra top-level member of a contribution, are accepted and ignored: the released schemas leave both objects open (#4543).

  • The openehr-* crates take their version when they are published. A pull request that changes the crates no longer bumps their lockstep 0.0.x version: the crates publish reads crates.io, takes the next version when packaged content changed since the newest published one, and refuses to publish changed content under a version crates.io already holds. The version in the manifests on main can therefore trail the newest version on crates.io (#4555).

  • The served Web Template of an ADL 2 operational template writes the aqlPath of an ELEMENT’s value and of a data value’s parts without an [idN] predicate, since a data value carries no archetype_node_id (ADL2 master05-paths §“Relationship with W3C Xpath”). A DV_INTERVAL and its parts read …/value, …/value/lower and …/value/upper, the form the ADL 1.4 Web Template writes, where they read …/value[id6]/lower[id7] before, and the parts’ ids are lower and upper (#4559).

  • The served Web Template gives every DV_INTERVAL<T> node its lower and upper children, each of type T, also when the operational template leaves them unconstrained (RM data_types §DV_INTERVAL). Such a node had no children before. An unconstrained part carries the open inputs of its type, and its aqlPath is …/value/lower or …/value/upper. This holds for ADL 1.4 and ADL 2 templates (#4559).

  • The template example fills such an interval: an unconstrained part takes the example value of its constrained sibling, so the limits stay consistent (#4559).

  • The served Web Template of an ADL 2 operational template writes a coded input whose at-codes the template binds to a terminology as the bound terms: its code list carries each binding’s code and label and its terminology names the bound terminology (ADL2 master08-terminology_integration §“From Constraints to Concrete Codes in Data”). COMPOSITION.category bound at1 to openehr::433 reads {"value": "433", "label": "event"} with terminology: "openehr", where it read the local at1 before, and an [acN@ttttt] constraint takes the bound terms of the terminology it names. A FLAT or STRUCTURED composition without a category, and the example composition, now carry openehr::433. The aqlPath of every data value drops its [idN] predicate, so a constrained category and context/setting read /category and /context/setting (#4574).

  • A CONTRIBUTION’s versions[] entry records its merge inputs: the other_input_version_uids it states are stored on the committed ORIGINAL_VERSION and served on every read, where the server used to refuse the member with 400. The uids may name versions the server does not hold; an empty list or an entry that is not an OBJECT_VERSION_ID is a 422 (#4577).

  • Other members UPDATE_VERSION does not declare, such as item or an extension member, are accepted and ignored, where the server used to refuse them with 400: the released schema leaves the object open, as it does NewContribution and UPDATE_AUDIT. A uid or contribution on a version entry is still refused with 400, because the server assigns both, and so is a _type other than UPDATE_VERSION or ORIGINAL_VERSION (#4577).

  • The admin dump route’s refusals name its parameters by their wire names and list the accepted values, not the Service Model class names. A bad logical_format or compression_format now reads “<field> must be a string, one of …” or “<field> names no accepted value: … (expected one of …)”, and a non-integer segment_split_size reads “segment_split_size must be an integer size in kb”. The served OpenAPI text for the route no longer names EXPORT_SPEC or EXPORT_FORMAT (#4634).

  • Five archive dump and load failures still answer 500, and their bodies are now fixed text that echoes no value from the archive: (#4657)

    • an unreadable data segment no longer names the segment, though manifest.json and demographic-commons.json are still named;
    • an unknown logical format in the manifest no longer echoes the declared value;
    • a version whose id carries a path separator, and a version whose commit audit the export lacks, no longer name the version.
  • A failed expansion of a party’s protected national identifiers still answers 500, and its body now reads “expanding the protected identifiers of a party version failed” without echoing the requested version uid: a 500 body carries curated text only (#4657).

  • Two query failures still answer 500, and their bodies now carry curated text only. A stored query whose record holds no query text reads “the stored query has no query text” without echoing its name, and a terminology server failure while a TERMINOLOGY('expand', …) operand is resolved reads “the terminology service failed while the query was being planned” instead of the upstream message (#4657).

  • The db_transactions counter (outcome="commit") now counts every committed single-object write. It used to leave out EHR_STATUS and DIRECTORY writes (#4659).

  • OPTIONS on the API base path sends conformance_profile as BASE+OPTIONS where it sent STANDARD. The build derives the value from the committed Veredictum record, docs/conformance/ferroehr/summary.md: BASE+OPTIONS when the BASE rows and the run OPTIONS rows have none failed and none errored, BASE when only the BASE rows do, and the member left out otherwise. ITS-REST 1.1.0 types the member as a free string and CNF names no profiles, so the values are FerroEHR’s own. The [rest.identity] conformance_profile configuration key is removed: a configuration that still sets it is refused at boot as an unknown key (#4672).

  • The viewer answers every request other than GET and HEAD that arrives without a session, to any path but the sign-in page, the OIDC routes, sign-out and the search answer, with the sign-in page that says the session ended, before any handler runs. A post to a page returns there after signing in; a post to the export, preference, sidebar or paging route returns to the dashboard. The export post used to answer 303 to /login, and the preference, sidebar and paging posts ran without a session (#4724).

  • The conformance instrument is Veredictum 0.4.0, and the published conformance record (docs/conformance/ferroehr/results.json with its summary.md) is a 0.4.0 released-profile run. The party statement answers the 0.4.0 catalogue: it keeps the xml-namespace-v2 arm, which now asks only whether the TRIAL http://schemas.openehr.org/v2 namespace is also accepted, and selects the new development-profile option arm adl2-id-coded-supported (#4771).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

  • The Veredictum 0.1.6 conformance documents: CONFORMANCE_REPORT.md, CONFORMANCE_STATEMENT.md, CONFORMANCE_CERTIFICATE.md, verdicts.json, run-exceptions.json, the CORE, STANDARD, OPTIONS, SEC-BASIC and performance badge files, and the heat grid and chapter bars the book and the landing page drew from them (#4451).

  • The CONF_PERF_CLASS, CONF_PERF_HOURS and CONF_STRESS settings of the measured lanes are gone, along with the class runs, the step-load stress ladder and the AQL probe, which Veredictum 0.3.1 removed (#4452).

  • The separate SMART step of the Conformance workflow, the smart stage of scripts/conformance.sh with its CONF_SMART_OUT, and docs/conformance/party/ferroehr/ixit-smart.json are gone: the one catalogue run covers the SMART documents. The party statement no longer carries the performance block, which the Veredictum 0.3.1 statement schema refuses (#4503).

  • The empty linkage storage domain is gone: its schema, its ferroehr_linkage role, the [storage.linkage] and [storage.migrate.linkage] configuration tables, its pool, the database.linkage and backup.linkage chart values, the ferroehr-backup-linkage compose job and the btree_gist extension only its table needed. A configuration that still sets [storage.linkage] or [storage.migrate.linkage] is refused at boot as an unknown key. Cohort and AQL demographic join crossings still write their linkage-domain access records. Upgrade note: remove those keys and values before you upgrade. On a co-located database, migration audit/0021_drop_linkage_domain.sql drops the schema, the role and the extension at the first boot, and raises a NOTICE for any of them it cannot drop. When the linkage domain lived in a database of its own, [storage.audit] names another database, or the cluster holds more than one FerroEHR database, run in each such database, as a superuser or a member of ferroehr_linkage: DROP SCHEMA IF EXISTS linkage CASCADE; REVOKE ALL ON SCHEMA ext FROM ferroehr_linkage; REVOKE ALL ON ALL FUNCTIONS IN SCHEMA ext FROM ferroehr_linkage; ALTER DEFAULT PRIVILEGES IN SCHEMA ext REVOKE EXECUTE ON FUNCTIONS FROM ferroehr_linkage; DROP OWNED BY ferroehr_linkage; DROP EXTENSION IF EXISTS btree_gist; then DROP ROLE ferroehr_linkage; once per cluster. The released grant files still name ferroehr_linkage, so a first install whose migrator holds no CREATEROLE still provisions that role and drops it after the first boot (#4684).

  • A party migration drops the database function party.resolve_national_identifier. No server path has called it since the linkage domain was removed (#4697).

  • The viewer no longer reads LEPTOS_SITE_ADDR: a value naming another address than server.listen no longer refuses the start. server.listen (FERROEHR_VIEWER__SERVER__LISTEN) sets the listen address (#4725).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • An AQL query aimed at named EHRs (the ITS-REST ehr_id parameter) reads those EHRs’ objects first and each object’s own content after them, whatever the planner statistics say. When an archetype’s rows fell out of the most-common-value list of the archetype statistics, the planner estimated them at one row and could start from that archetype over the whole population, evaluating every projected path for every EHR’s rows before the EHR filter: the storage bench’s wide projection over one EHR read about 10 000 pages per call in some rounds and 1 200 in others. A node contained in a versioned object is now read inside its parent on such a query; results do not change (#4487).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

  • AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri argument whole when its value carries a ?, as the QUERY master03 example url=http://snomed.info/sct?fhir_vs does. Before, the arguments were read from after the first ? anywhere in params_uri, so the value-set URL reached the terminology server truncated and the arguments before it were lost. A params_uri that starts with a path still has its arguments read from after the path’s ? (#4490).

  • An ADL 2 template whose constituent archetype the CDR does not hold is refused at upload, with a 422 whose errors carry one VARXR entry per missing archetype (AOM2 C_ARCHETYPE_ROOT validity rule VARXR); the 2013 CKM t_encounter_opt_test, built on the refused apgar, no longer loads, so the sandbox library pin is 231 accepted / 91 refused. A stored template whose constituent was deleted later answers its JSON form and example with the same VARXR list, and the viewer says so on every tab of the template, names each missing archetype and links to uploading it, instead of a raw 422 (#4505).

  • ADL 2 validation admits a constraint on an RM computed attribute. VCARM accepts an attribute “stored or computed” (AOM2 master04.5 §Validity Rules: C_ATTRIBUTE), and a function with no arguments is a computed attribute (master04.2 §Attribute Nodes), so an archetype that constrains DV_PROPORTION.is_integral, DV_QUANTITY.is_integral or EVENT.offset is no longer refused with 422. The generated RM model (openehr_rm::v1_2::model) now carries each class’s computed attributes from the BMM functions (RmClass::computed_attributes, computed_attribute). openEHR-EHR-CLUSTER.microscopy_prostate_carcinoma.v1.0.0 from the 2013 CKM ADL 2 export now uploads, and the sandbox archetype-library pin moves to 232 accepted and 90 refused. DV_QUANTITY.property, which no RM release defines, stays refused. The nine openehr-* crates step to 0.0.124 (#4513).

  • DELETE /ferroehr/rest/extensions/v1/definition/artefact/adl2/{artefact_id} refuses with 409 Conflict while another stored ADL 2 artefact does not compile without the one deleted (a slot filler, a use_archetype reference or a specialisation parent, directly or through another artefact), and the message names every one of them. Before, the delete succeeded and left each template built on the archetype unable to compile (#4522).

  • The viewer’s ADL 2 template upload names each violation the CDR found when it refuses the source: every AOM2 rule code with the CDR’s words for it, so a template whose constituent archetype the CDR does not hold names that archetype (VARXR). The upload now asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details, and reads the stored template’s id from the ETag of the 201 (#4523).

  • An ADL 2 operational template whose included archetype reuses the template’s node ids uploads again, where it was refused with 422 under VCOSU. Node-id uniqueness now holds per archetype and restarts at each C_ARCHETYPE_ROOT (AOM 2 VCOSU, “unique within the archetype”), and on an operational template the codes and interior node ids below a root are judged against that constituent’s entry in component_terminologies (OPT 2 master03-opt_raw). A duplicate node id inside one archetype is still refused. The nine openehr-* crates step to 0.0.125 (#4529).

  • The viewer’s dashboard draws the EHR or composition count as unavailable, with the reason in plain words, when the CDR answers its count query with a result set of another shape (no row, several rows, an empty row, several values, or a value that is no whole number), where it drew 0 before. The template Releases table draws such a count the same way. A real zero still draws 0 (#4533).

  • An ADL 2 operational template upload runs the terminology integrity rules (VTLC, VATCV, VTVSID, VTVSUQ, VTVSMD, VTTBK, VTCBK, and VTSD where the level is carried) on each component_terminologies entry, and each violation names the component’s archetype id (#4536).

  • A COMPOSITION committed against an ADL 2 operational template is validated against the template’s leaf constraints, where it was stored with 201 whatever it held. An inlined archetype recorded by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability) now matches its template node, so the C_INTEGER, C_REAL, C_STRING, C_BOOLEAN, tuple, value-set and cardinality constraints below it answer 422 (AOM 2 master04.5). An archetype root the template does not inline is refused as a node the template does not define, and a C_INTEGER or C_REAL list ({10, 15, 20}) admits each of its values where only the first was read. Canonical, FLAT and STRUCTURED commits take the same check. The nine openehr-* crates step to 0.0.127 (#4540).

  • A versioned read whose version path segment is no SEMVER version and no {major} or {major}.{minor} prefix, such as v1.x, 1.x or 01, now answers 400 before any lookup, where it answered 404 (parameters/path/version.yaml, responses/400.yaml). This covers the ADL 2 template read GET /definition/template/adl2/{template_id}/{version}, the stored-query read GET /definition/query/{qualified_query_name}/{version} and both stored-query executions at a version, GET and POST /query/{qualified_query_name}/{version}. A well-formed version the server does not hold still answers 404 (#4541).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the content below an inlined archetype that the composition records by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability). The read left out everything below such a root with no error; it now matches the root to its template node the way the commit does. The nine openehr-* crates step to 0.0.128 (#4551).

  • A COMPOSITION committed against an ADL 2 operational template whose C_INTEGER or C_REAL constraint is a union of ranges (for example magnitude matches {|0..5|, |10..15|}) is refused with 422 when the value lies outside every range. Such a value was stored. The served Web Template does not change: its input validation.range holds one range, so a union leaves it out, as before, and the check runs on the commit (#4552).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the value of an ELEMENT that holds a DV_INTERVAL whose lower and upper the template constrains. The read left the value out with no error (#4559).

  • A FLAT or STRUCTURED read of a COMPOSITION returns the value of an ELEMENT that holds a DV_INTERVAL whose template leaves the interval’s lower and upper unconstrained, for ADL 1.4 and ADL 2 templates. The read left the value out with no error, and a FLAT or STRUCTURED commit could not name the parts (#4559).

  • A COMPOSITION written from FLAT or STRUCTURED against an ADL 2 operational template carries what a canonical body built to that template carries. A node the template gives no term text, such as the HISTORY and ITEM_TREE of an OBSERVATION, is named for its RM type (History, Tree) instead of its bare id-code (id2). Every inlined archetype root records the complete identifier the template names (…cnf_minimal.v1.0.0, BASE master07-referencing §Reconstitutability) in its archetype_node_id and archetype_details.archetype_id, and archetype_details.rm_version is the template’s rm_release. Before this change the root recorded the interface form (…v1) and the server’s RM 1.2.0. Compositions built against an ADL 1.4 template keep their major-version archetype ids and RM 1.2.0 (#4560).

  • ferroehr report in a release build names the openehr-* crate versions that release published to crates.io: the release lanes stamp the published version into the checkout before they build the server, where the report used to name the floor the source tree holds. A new openehr_crates_origin field says published for a release build and source_tree for any other, a pre-release included (#4565).

  • An ADL 2 archetype reference resolves to one defined stored release: an interface reference (….v1) takes the newest release of its major version, a reference with a minor or patch version narrows the match to it, and an alpha or beta version matches only its full identifier (BASE Artefact Identification, Referencing). With several releases of one archetype stored, a template upload, an operational template and the delete’s dependents check used whichever release PostgreSQL returned last, and a reference to a major version no stored release carries resolved to another major version (#4567).

  • An ADL 2 upload and an ADL 2 delete no longer race. An upload validated before a concurrent delete of an archetype it uses committed could store a template whose constituent was gone. Both now hold one transaction-scoped advisory lock from the repository read to the commit, so the second write decides on what the first committed: an upload after the delete is refused with VARXR, and a delete after the upload is refused with a 409 naming the template (#4568).

  • The syslog form (DICOM PS3.15, IHE ITI-20) of an EHR-level audit record names its object beside the patient, as a URI participant object with the id its FHIR AuditEvent names. The record of placing, releasing or expiring an EHDS Art. 3(3) access delay names the held composition as access-delay:<vo>, so an Audit Record Repository fed by syslog can tell which composition a hold covers (#4569).

  • An ADL 2 template or archetype that references two major versions of one archetype (….v1 and ….v2), or one archetype under two namespaces, now compiles against both: the openehr-adl repository holds an artefact under its namespace, archetype and major version, and resolves each reference within that key (BASE Artefact Identification, Referencing). A reference without a namespace now resolves in the namespace of the artefact that makes it, and never in another namespace (#4579).

  • A read of a stored ADL 2 template by a partial id or a version (GET /definition/template/adl2/{template_id}) now resolves the way template compilation does: with v1.0.0 stored beside v1.0.0-rc.1, ….v1, ….v1.0 and version=1.0.0 answer the release in either upload order, and the release candidate is read only by its full version (BASE Artefact Identification, Referencing) (#4580).

  • A write that sends no committer, on a route that takes a bare RM body (the EHR_STATUS, COMPOSITION and DIRECTORY writes, and the demographic party and relationship writes), now records a Bearer caller’s login as issued by the token issuer, as EHR creation and every other default committer already did. It recorded the issuer ferroehr for every caller before; a Basic caller’s login is still issued by ferroehr (#4585).

  • The viewer’s ADL 1.4 template upload opens the stored template’s page, read from the ETag (else the Location) of the CDR’s 201, where it went to an address with no template id. A refused upload names each violation the CDR found, its rule code or path and the CDR’s words for it, as the ADL 2 upload does: the upload asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details (#4596).

  • A 422 validation refusal tells the client to send Prefer: return=representation only when the request did not send it; a request that sent it receives the violations in errors and a detail without the advice (#4597).

  • An EHR Extract import refusal for a received version now names that version’s position in the extract (chapters[i]/items[j]/item/versions[k]), so an importer can find the defective ORIGINAL_VERSION in a large extract (#4602).

  • The import reads a version’s other_input_version_uids with the parser the CONTRIBUTION path uses and refuses an entry that is not an OBJECT_VERSION_ID, where it used to drop the entry. On the wire the extract decode already refuses such an entry with 422, so the import stays all or nothing (#4602).

  • An AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts within the major: org.a::… no longer finds data recorded in org.b, an id without a namespace finds only data recorded without one, and .v1.4.2 no longer finds data recorded under v1.0.0 (BASE Artefact Identification §Referencing and §Supporting Archetype-based Querying; #4629, #4630).

  • A query naming a parent archetype finds the data of a specialised child release only when that release declares a parent release at or after the queried one. Before, once one stored release of the child qualified, the data of every release of it was returned (BASE Artefact Identification, Referencing §Supporting Archetype-based Querying; #4637).

  • POST /definition/template/adl1.4 and POST /definition/template/adl2 answer 400 for every invalid template, the only refusal ITS-REST 1.1.0 declares for both uploads besides the 409 duplicate. An OPT that breaks the OPT 1.4 schema, an ADL 2 source that breaks an AOM2 rule such as VCORM or VCARM, and an ADL 2 template whose constituent archetype is not held were refused with 422. The problem keeps its detail and, with Prefer: return=representation, its errors. Upgrade note: a client that read a 422 from a template upload reads a 400, with the same detail and errors, under the about:blank problem type (#4643).

  • A directory create or update whose FOLDER carries an empty folders array, at the root or in a sub-folder, is refused with 400 and commits nothing. The ITS-JSON FOLDER schema gives folders minItems: 1, and the canonical JSON reader now refuses the empty array the same way in every document that carries a FOLDER, a CONTRIBUTION included. The viewer leaves folders out of a folder with no sub-folders (#4644).

  • A canonical XML FOLDER, EHR_STATUS, PERSON, AGENT, GROUP, ORGANISATION, ROLE or PARTY_RELATIONSHIP document now has the root element items with xsi:type naming its class, the element the ITS-XML schemas publish for a LOCATABLE (Structure.xsd). The directory reads (latest, by version and at a time), the EHR_STATUS reads, the demographic reads and the demographic template examples served <folder>, <ehr_status>, <person> and the like, which no ITS-XML schema declares. XML input accepts the same form (#4645).

  • POST /definition/template/adl1.4 builds the template’s Web Template before it stores the template, so a template it cannot build is refused with 400 and nothing is stored. Once the template is stored, the upload answers 201 with its Location; with Prefer: return=representation and a Web Template Accept it serves the Web Template built before the store, and the OPT if that cannot be served. It no longer answers 422 for a template it has already stored (#4647).

  • A composition whose HISTORY carries events: [], and an AGENT, GROUP, ORGANISATION or PERSON carrying languages: [], are refused with 400 and commit nothing. The ITS-JSON schemas give both arrays minItems: 1, and the canonical JSON reader now refuses the empty array as it does an empty FOLDER folders. The FLAT builder no longer writes an empty events for a HISTORY that has no event (#4648).

  • A write to a restricted COMPOSITION (POST, PUT or DELETE on /ehr/{ehr_id}/composition) refused under a GDPR Art. 18 restriction now answers with the processing-restricted problem type, as the reads of a restricted record already did, in place of the generic forbidden type. The status stays 403 (#4657).

  • When an EXTRACT_SPEC criterion’s AQL runs past the query time budget, the 400 body now states the timeout as plain text. It used to start with a control-character marker that was only ever meant for the server itself (#4657).

  • Under a separated clinical credential, POST /ehr, PUT /ehr/{ehr_id} and an EHR_STATUS update that name a subject no longer fail with a database error: the subject pseudonym guard reads the deployment posture, and the clinical role now holds SELECT on ext.posture through the ext migration 0009 (#4675).

  • A cohort that names its subjects now finds, under [ehr_index] provider = "local", the EHR created over POST /ehr with that subject in its EHR_STATUS; it read a cross-reference table no route wrote, so it matched nothing. A cohort predicate and the AQL demographic join resolve a party to the EHRs whose EHR_STATUS.subject names it. The EHR_STATUS subject is now the one source of which EHR belongs to which subject. Upgrade note: linkage.subject_ehr and the linkage.erase_ehr function go with the linkage schema, which audit migration 0021 drops with the linkage domain. Rows written there through the service API are dropped with the table; the EHRs they named keep their subject in their own EHR_STATUS (#4675).

  • The identifier scan ([privacy.identifier_scan]) no longer refuses a DV_IDENTIFIER whose type declares an item rather than a person, such as a laboratory filler_order_number that passes the NHS Number check by chance. A DV_IDENTIFIER in FEEDER_AUDIT.originating_system_item_ids or feeder_system_item_ids is scanned only when its type names a person identifier. A type naming a person (patient, NHS, BSN, social security and similar) keeps the value scanned in every slot, and free text and untyped identifiers in clinical content are scanned as before. The default rule set is unchanged; a deployment that knows its jurisdictions still narrows rules, so a Dutch one with rules = ["nl-bsn"] never claims an NHS Number (#4702).

  • An inline primitive constraint in a parsed archetype or template now carries the node id AOM2 ADL_CODE_DEFINITIONS.Primitive_node_id defines, at9999 in an at-coded archetype and id9999 in an id-coded one, where it carried the constant’s name Primitive_node_id. An ADL 1.4 archetype converted to ADL2 gets the id-coded form (#4739).

  • The ABAC policy-enforcement point checks every item-tag operation as it checks its target: a COMPOSITION tag read after the call like the COMPOSITION read, a COMPOSITION tag write before the call with the tagged version’s template, category and origin, and the EHR tag read before the call like the EHR read. These operations were unchecked before, so a patient-scoped or template-scoped rule now refuses a tag read or write wherever it refuses the target’s read or write. A COMPOSITION tag read also falls under the EHR_ACCESS privacy ceiling and its audit record names the tagged COMPOSITION (#4768).

  • Under spec_profile = "stable", a canonical-JSON party read lists in reverse_relationships every PARTY_RELATIONSHIP another party’s current version carries with this party as its target, one LOCATABLE_REF each naming the source party’s version container and the relationship’s path, as RM 1.1.0 §Party Relationships requires. The development profile, whose RM 1.2.0 makes the attribute a function, is unchanged. The released-profile conformance run now judges the server on the stable profile (#4775).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

  • CVE-2026-84243, the glibc gettext LANGUAGE directory traversal (GLIBC-SA-2026-0025), has no Debian fix for trixie yet and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never leave the C locale, so glibc ignores LANGUAGE and opens no message catalog (security/vex/ferroehr-os.openvex.json). In ferroehr-postgres only the container environment the operator writes can set LANGUAGE; no database connection can (security/vex/postgres-os.openvex.json) (#4500).

5.6.0-rc7 - 2026-11-01

5.6.0-rc7 is the seventh release candidate of 5.6.0, published for testing. It carries the CONTRIBUTION break of 5.6.0-rc6: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution read the body through the released NewContribution schema and answer 400 for a lifecycle_state that is no DV_CODED_TEXT, a description that is a plain string or an object without _type, an audit without change_type, an audit typed ATTESTATION or UPDATE_ATTESTATION, a missing audit or versions, and a preceding_version_uid sent as a plain string, and a stated empty system_id is refused with 422 on every write path (#3595, #4543). These notes carry every change since 5.5.0, those of the 5.6.0-rc6, 5.6.0-rc5, 5.6.0-rc4 and 5.6.0-rc3 sections below included. New since 5.6.0-rc6: the EHDS Art. 10 opt-out from access through the access services, off by default until [authz.rbac] access_service_roles declares the access-service roles, with the Art. 10(2) vital-interest override marked as an emergency (#4516); every write’s committer checked against the authenticated caller (#4520) and a Bearer caller’s default committer issued by the token issuer (#4585); ADL 2 FLAT and STRUCTURED writes that carry the names, archetype references and RM release the template gives (#4560), a served Web Template that writes a bound coded input as the bound terms (#4574) and gives every DV_INTERVAL its lower and upper parts (#4559); ADL 2 validation that checks a C_INTEGER or C_REAL union of ranges (#4552), a repository keyed by namespace, archetype and major version (#4579), a read that resolves a release before its pre-releases (#4580) and an upload and a delete ordered by one lock (#4568); an AQL archetype predicate that matches the namespace exactly and the version as the oldest release (#4629, #4630); a CONTRIBUTION that records its merge inputs and accepts the UPDATE_VERSION members the schema leaves open (#4577), and an EHR Extract import that refuses a malformed merge input and names the version’s position (#4602); the syslog form of an access-delay record naming the held composition (#4569); the viewer’s ADL 2 and ADL 1.4 uploads naming each violation the CDR found (#4523, #4596); storage generation 4 and its benchmark records (#4525); ferroehr report naming the published openehr-* crate versions in a release build (#4565); and response texts without the Service Model vocabulary (#4531, #4634). The conformance run on main under the released profile fails no row against the server and still has one errored row, Veredictum’s own (Vernum-Projecten/Veredictum#964, closed upstream after the pinned 0.3.2, which still errors on it), and the triage on #4455 stays open, so this candidate does not claim a green Veredictum record. 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: a write whose committer names another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, answers 403, and a direct write whose committer breaks its own rules answers 422; a write that sends no committer on a bare RM body route records a Bearer caller’s login as issued by the token issuer, where it recorded ferroehr; an AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts, so a query may find less than before; a FLAT or STRUCTURED composition under an ADL 2 template records the complete archetype identifier and the template’s rm_release, and the served Web Template writes a bound coded input, COMPOSITION.category among them, as the bound terms (openehr::433) where it wrote the local at-code; a CONTRIBUTION commit must take the shapes above; a write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as committer and may change only what the person committed, and a service auth_failure answers 403 where it answered 500; a versioned read whose version is no SEMVER version or prefix answers 400 where it answered 404; the served Web Template of an ADL 2 template writes a data value’s aqlPath without an [idN] predicate; a composition that breaks an ADL 2 template’s leaf constraint, and an ADL 2 template whose constituent archetype is not held, answer 422; every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes, so keep it longer than the idle timeout of your clients and load balancers; a client that wrote a space as + in a query string sends %20; a party delete names the latest OBJECT_VERSION_ID in its path and answers 400 for a versioned object id; a body the RM refuses after the ITS-JSON schema admits it answers 422, and a misplaced CONTRIBUTION member, a Simplified Format commit without openehr-template-id and a flat |raw value without _type answer 400; a tag update under Prefer: return=identifier answers 200 with the tag list; with [authz.rbac] enabled a token holding only system/ SMART on openEHR scopes reaches clinical operations without a role; the viewer image serves no /pkg/ and reads no LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, and its Content-Security-Policy is script-src 'self'; style-src 'self'; Helm chart 21.0.0-rc7 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data, and ext migration 0008 names generation 4, the generation with the node header columns, and changes only the label; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; clinical 0024 and party 0015 index the version uid, ext 0006, clinical 0025 and party 0016 add the ELEMENT value search index, built over every hot node row of an existing store, party 0017 adds the archetype statistics, and clinical 0026 and party 0018 grant MAINTAIN on the partitioned tables to the serving roles; clinical 0027 and party 0019 add the access delay, and ext 0007, clinical 0028 and party 0020 add the node header columns with no data migrated; audit migration 0019 gives the search token index a pending list of 512 kB, and audit 0020 adds the hash index on request_id; clinical 0029 adds the access opt-out. Helm chart 21.0.0-rc7 carries appVersion 5.6.0-rc7. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j, and the OpenVEX judgement of CVE-2026-84243.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The operations page gains a storage sizing section, measured on 5.6.0-rc5: about 21.6 KB of clinical storage per version of a vital signs composition (2.0 times its canonical JSON, 85 % of it node rows) and 155 KB for an International Patient Summary, about 1.2 KB per audit record of a read, the effect of [audit.store] retention on growth, pg_wal up to max_wal_size, the REINDEX INDEX CONCURRENTLY clinical.node_hot_pkey step (20 % of that index after heavy concurrent writing) and a query for the size per relation, with the method to reproduce the figures (#4304).

  • EHDS Art. 5: what a person inserts in their own EHR is marked as theirs, and they cannot alter what a professional committed. A write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as its AUDIT_DETAILS.committer, its external_ref naming the login, whatever committer the request names; a COMPOSITION it commits must name PARTY_SELF or PARTY_RELATED as composer; and its update or deletion of a version whose committer is not PARTY_SELF, directly or inside a CONTRIBUTION, is refused with 403. A professional’s writes are unchanged. The contribution list extension gains inserted_by_person per row, and the viewer names a PARTY_SELF committer “The person or their representative”. A service auth_failure status now answers 403 where it answered 500 (#4350).

  • The EHDS Art. 3(3) delay of the person’s own access to newly registered data, a Member State option and off by default. The new [access_delay] section (enabled, categories, hours) holds every new COMPOSITION of a listed [audit.categories] category from requests that may be the person’s own (a bearer token with a SMART patient/ grant or the ABAC patient claim, also beside user/ or system/ grants, as ferrohealth-smart 0.8.7 decides it) for at most hours: a read answers 404 as for an unknown composition, and a query, an export and a CONTRIBUTION leave it out, while a professional reads it as before. The treating professional releases a held composition from the clinical route POST {ext}/ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/access-delay/release, with the checks of a write to that EHR, and the controller with POST /admin/access-delay/release; a professional holds one by hand within the limit with POST /admin/access-delay and reads the register with GET /admin/access-delay; every hold ends at its limit, and placing, releasing and expiry are each an access record. The subject-scoped ITI-81 read of the person’s own trail (EHDS Art. 9) leaves out the records about a held composition until it is released or expires, then serves them with their original times (ehds-logging 0.3.3) (clinical migration 0027_access_delay.sql, party migration 0019_access_delay_column.sql) (#4351).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • The Conformance workflow runs the Veredictum smart documents (smart_discovery, smart_scopes) against the same FerroEHR build after the Veredictum Action, which forwards no SMART Platform base URL and no scoped tokens. The same pinned image runs veredictum test with --platform-url and docs/conformance/party/ferroehr/ixit-smart.json, minting each access token in memory from the committed test issuer the SMART overlay trusts, and the job uploads its results, JUnit and CTRF views beside the catalogue’s and adds its summary under a heading of its own. scripts/conformance.sh gains the smart stage, and run drives it after the catalogue, so a local run covers SMART too (#4450).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

  • The EHDS Art. 10 opt-out from access through the access services of Arts. 4 and 12, a Member State option and off by default. The new [authz.rbac] access_service_roles key declares the roles of the access-service callers; while an EHR carries the opt-out, a caller holding one reads it as absent: a route naming the EHR or an object in it answers 404 as for an unknown EHR, a population query leaves its data out, a scoped query, an export and the lookup by subject find no such EHR, and a write to it is refused as absent. Every other caller, the provider’s own care included, reads it as before. The controller records the opt-out on the person’s instruction with POST /admin/access-opt-out, reverses it with POST /admin/access-opt-out/reverse (the register keeps the reversal) and reads the register with GET /admin/access-opt-out; recording is refused with 400 while no access-service role is declared. An access-service caller that declares one of [audit] emergency_purpose_codes and does not act for the person reads the EHR under the Art. 10(2) vital-interest override, with the emergency mark on its access record and an access-opt-out-override record against the EHR. Recording, reversing and reading the register are access records (clinical migration 0029_access_opt_out.sql) (#4516).

Changed

  • Breaking for clients that relied on the old leniencies: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution now read the body through the released NewContribution schema, and refuse every shape it does not define. Each refused shape answers 400: (#3595)

    • a lifecycle_state in the CODE_PHRASE shape {"terminology_id": …, "code_string": …}, as a plain string, or as a bare {"value": …}. Send a DV_CODED_TEXT: {"value": "complete", "defining_code": {"terminology_id": {"value": "openehr"}, "code_string": "532"}};
    • a description sent as a plain string, on the contribution audit, on a commit_audit or on an attestation. Send a DV_TEXT or DV_CODED_TEXT object: {"_type": "DV_TEXT", "value": "…"};
    • a commit_audit or contribution audit without change_type. The server no longer infers creation or modification from preceding_version_uid; send the code of the change, for example 249 (creation), 251 (modification), 523 (deleted) or 666 (attestation);
    • a commit_audit or contribution audit whose _type is ATTESTATION or UPDATE_ATTESTATION, or any value other than UPDATE_AUDIT or AUDIT_DETAILS. Send the audit as UPDATE_AUDIT (or with no _type) and put each attestation in the version’s attestations list as an UPDATE_ATTESTATION. A 666 (attestation) version carries its attestations there too, and one with no attestation is refused;
    • a missing audit, which answered 422 before. A missing versions list, an attestation without reason or is_pending, a system_id that is not a string, and a preceding_version_uid sent as a plain string instead of {"value": "…"} are refused the same way.
  • An attestation’s own system_id is kept, and one that states none takes the system_id of the version audit it belongs to instead of the server’s. An attestation committer is now checked against its PARTY_PROXY rules like every other committer, so an invalid one answers 422 (#3595).

  • The served OpenAPI document describes each extension route by what it does and by the ITS-REST text that governs it, and no longer names openEHR Service Model operations, interfaces or status codes. Routes, request and response fields and status codes are unchanged (#4213).

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The storage-bench lane measures the baseline and the candidate on one runner in eight interleaved rounds each and gives two verdicts. The plan verdict fails an operation whose median pages per call rose more than 10 % and one page above the baseline’s highest round, whatever its time. The time verdict divides each round by its machine factor, so a slow spell of the runner no longer reads as the code, and fails an operation whose median p50 or p99 rose more than 15 % with every candidate round slower than every baseline round, or a candidate build that ran more than 15 % slower across the board in every round. A latency is no longer compared with a record measured on another runner (#4318).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • The LANG specification text is re-vendored at f80763ab (#4377). SPECLANG-18 states how a P_BMM schema’s JSON form encodes a cardinality interval, which the JSON schema reader already followed; its documentation now cites the rule, and a test pins it. SPECLANG-16 types the P_BMM class lists as keyed collections; the generated P_BMM model takes that change with #4497. The nine openehr-* crates step to 0.0.122.

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • The viewer runs on ferrohealth-server 0.4.2 and takes its Content-Security-Policy from the family’s server-rendered console headers (FerroHEALTH/FerroSYS#202) in place of a layer of its own. The policy a deployment sees is unchanged: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; object-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none', with no nonce. FerroUI’s assets under /ferrohealth-ui/ keep the Cache-Control FerroUI sets, and a refusal under that path is no-store whatever FerroUI’s router sets, so a cache in front of the viewer keeps no error answer for an asset. The REST server takes the same ferrohealth-server release with no change in what it serves (#4439).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

  • The P_BMM model follows SPECLANG-16. The LANG 1.1.0 BMM that openehr-codegen reads is now the specifications-LANG file at f80763ab, so P_BMM_SCHEMA.primitive_types and class_definitions are maps keyed by class name, P_BMM_CLASS.ancestor_defs is a map of P_BMM_BASE_TYPE keyed by type signature (a simple ancestor included), and P_BMM_GENERIC_TYPE.generic_parameter_defs is an optional map keyed by formal parameter name. The formal and actual generic parameters (P_BMM_CLASS.generic_parameter_defs, P_BMM_GENERIC_TYPE.generic_parameter_defs, BMM_GENERIC_CLASS.generic_parameters) are insertion-ordered IndexMaps, because the specification fixes their declaration order, and the model builders bind each actual parameter to its formal by name. In the released v1_0 generation, P_BMM_GENERIC_TYPE.generic_parameter_defs now carries the 1..* bound its BMM states. The nine openehr-* crates step to 0.0.123 (#4497).

  • The conformance pipeline runs Veredictum 0.3.1. The Conformance workflow drives the catalogue through the Veredictum Action twice, each run on fresh volumes: under the released profile, the frozen openEHR release, which is the record and decides the job, and under the development profile, the pinned master of each openEHR repository, which is published beside it under its own job summary heading and artifact and never counted toward it. Both runs include the SMART documents: the Action now takes the SMART Platform base URL and the committed test issuer’s key. The workflow reads the result paths and the exit status from the Action’s outputs. scripts/conformance.sh takes the profile as CONF_PROFILE (released by default; a development run writes to docs/conformance/ferroehr-development/) and drives the SMART documents in the same run, and the IXIT carries the test issuer’s scoped_tokens declaration (#4503).

  • The viewer groups a template’s releases. The template list draws one group per template, the newest release first with the number of releases and the earlier releases under it, in both the ADL 1.4 and the ADL 2 tab. A template’s page names its release beside its title, where every other release is a link with the newest marked, also with FerroUI’s scripts refused; the release select and its Open button above the tabs are gone. The Overview lists the releases with the compositions committed against each, and a new Compare tab lists the paths one release adds, removes and changes against another (#4506).

  • Every write checks the committer against the login that authenticated the request where the server can compare them: a committer naming another login of the caller’s token issuer or of this server, or a PARTY_SELF committer from a caller who is not the record’s person, is refused with a 403. A name only, or an identifier of another authority, is stored as sent. The direct writes (composition, EHR status, directory, EHR, demographic) now also refuse a committer that breaks its own rules with a 422, as the contribution path already did (#4520).

  • An audit record no longer adds a row to audit.audit_search_token for its request id or for the version uid a create wrote. ITI-81 entity.identifier finds both from the record’s own request_id and resource_id columns, through a new hash index on request_id (audit migration 0020), so the token table stays flat as requests are audited. Retention now deletes the search tokens that no remaining record names, among them the request id tokens of records written by 5.6.0-rc5 and earlier, and the database still refuses to delete a token a kept record names (#4524).

  • A clinical or party node row below the root no longer repeats its archetype header, language and encoding: an archetype root’s archetype_details and archetype_node_id are rebuilt from the archetype column and a new rm_version column, and a language in ISO_639-1 and an encoding in IANA_character-sets keep only their code, in the new language and encoding columns. Every reader puts the keys back where they stood, so the canonical JSON, every AQL result and the served key order are unchanged; a header of any other shape stays in the fragment. A vital signs version takes 17.1 KB instead of 21.6 KB and an International Patient Summary 131 KB instead of 159 KB. Migrations ext/0007, clinical/0028 and party/0020 add the columns and the readers; no data is migrated, and a row written before keeps its whole fragment and reads the same (#4525).

  • ext.storage_generation() reports generation-4, the storage generation with the node header columns, and the storage benchmark records for it are committed under docs/benchmarks/storage/generation-4/. Migration ext/0008 changes only the label; nothing in the serving path reads it (#4525).

  • Five response messages no longer name the openEHR Service Model. The admin dump route’s 400 for a missing file_sys_loc now reads “request body must carry a non-empty file_sys_loc string, a location on the server’s file system (FerroEHR’s own dump/load extension)”, and its 400 for an encoding member reads “encoding is not accepted: FerroEHR’s own dump/load extension defines no encoding, so no value names one”. The admin activity-report 400s for a_service now list the accepted service names (“query parameter a_service is required: one of Admin, Definitions, …” and “query parameter a_service names no platform service: … (expected one of …)”). The not-implemented answer of store_query_set reads “store_query_set is not implemented: no openEHR specification defines a query set”. The openehr-rm validation docs for the three PARTY and PARTY_RELATIONSHIP relationship invariants now name the separate relationship container as FerroEHR’s own design against RM demographic §Party Relationships (#4531).

  • Breaking for clients that relied on the remaining commit-audit leniencies: the audit rules of the CONTRIBUTION commit now hold on every write path (#4543).

    • A description object without _type, such as {"value": "…"}, is refused with 400 on a contribution audit, a commit_audit and an attestation, and so is an attestation reason without _type. Send {"_type": "DV_TEXT", "value": "…"} or a DV_CODED_TEXT with its _type.
    • A stated empty system_id is refused with 422 on the direct writes too (openehr-audit-details: system_id="" on a create, update or delete), where the server used to replace it with its own. Leave system_id out to get the server’s, or send a non-empty one. A system_id that is not a string stays a 400 on both CONTRIBUTION routes.
    • A commit audit of type UPDATE_ATTESTATION handed to the library’s direct write calls (create_composition, update_composition, the EHR_STATUS, DIRECTORY and party writes) is refused with 400, as on the CONTRIBUTION routes. Send the audit as UPDATE_AUDIT and put each attestation in the version’s attestations list.
  • Members that NewContribution or UPDATE_AUDIT do not declare, such as a reason or is_pending on a plain audit or an extra top-level member of a contribution, are accepted and ignored: the released schemas leave both objects open (#4543).

  • The openehr-* crates take their version when they are published. A pull request that changes the crates no longer bumps their lockstep 0.0.x version: the crates publish reads crates.io, takes the next version when packaged content changed since the newest published one, and refuses to publish changed content under a version crates.io already holds. The version in the manifests on main can therefore trail the newest version on crates.io (#4555).

  • The served Web Template of an ADL 2 operational template writes the aqlPath of an ELEMENT’s value and of a data value’s parts without an [idN] predicate, since a data value carries no archetype_node_id (ADL2 master05-paths §“Relationship with W3C Xpath”). A DV_INTERVAL and its parts read …/value, …/value/lower and …/value/upper, the form the ADL 1.4 Web Template writes, where they read …/value[id6]/lower[id7] before, and the parts’ ids are lower and upper (#4559).

  • The served Web Template gives every DV_INTERVAL<T> node its lower and upper children, each of type T, also when the operational template leaves them unconstrained (RM data_types §DV_INTERVAL). Such a node had no children before. An unconstrained part carries the open inputs of its type, and its aqlPath is …/value/lower or …/value/upper. This holds for ADL 1.4 and ADL 2 templates (#4559).

  • The template example fills such an interval: an unconstrained part takes the example value of its constrained sibling, so the limits stay consistent (#4559).

  • The served Web Template of an ADL 2 operational template writes a coded input whose at-codes the template binds to a terminology as the bound terms: its code list carries each binding’s code and label and its terminology names the bound terminology (ADL2 master08-terminology_integration §“From Constraints to Concrete Codes in Data”). COMPOSITION.category bound at1 to openehr::433 reads {"value": "433", "label": "event"} with terminology: "openehr", where it read the local at1 before, and an [acN@ttttt] constraint takes the bound terms of the terminology it names. A FLAT or STRUCTURED composition without a category, and the example composition, now carry openehr::433. The aqlPath of every data value drops its [idN] predicate, so a constrained category and context/setting read /category and /context/setting (#4574).

  • A CONTRIBUTION’s versions[] entry records its merge inputs: the other_input_version_uids it states are stored on the committed ORIGINAL_VERSION and served on every read, where the server used to refuse the member with 400. The uids may name versions the server does not hold; an empty list or an entry that is not an OBJECT_VERSION_ID is a 422 (#4577).

  • Other members UPDATE_VERSION does not declare, such as item or an extension member, are accepted and ignored, where the server used to refuse them with 400: the released schema leaves the object open, as it does NewContribution and UPDATE_AUDIT. A uid or contribution on a version entry is still refused with 400, because the server assigns both, and so is a _type other than UPDATE_VERSION or ORIGINAL_VERSION (#4577).

  • The admin dump route’s refusals name its parameters by their wire names and list the accepted values, not the Service Model class names. A bad logical_format or compression_format now reads “<field> must be a string, one of …” or “<field> names no accepted value: … (expected one of …)”, and a non-integer segment_split_size reads “segment_split_size must be an integer size in kb”. The served OpenAPI text for the route no longer names EXPORT_SPEC or EXPORT_FORMAT (#4634).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

  • The separate SMART step of the Conformance workflow, the smart stage of scripts/conformance.sh with its CONF_SMART_OUT, and docs/conformance/party/ferroehr/ixit-smart.json are gone: the one catalogue run covers the SMART documents. The party statement no longer carries the performance block, which the Veredictum 0.3.1 statement schema refuses (#4503).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • An AQL query aimed at named EHRs (the ITS-REST ehr_id parameter) reads those EHRs’ objects first and each object’s own content after them, whatever the planner statistics say. When an archetype’s rows fell out of the most-common-value list of the archetype statistics, the planner estimated them at one row and could start from that archetype over the whole population, evaluating every projected path for every EHR’s rows before the EHR filter: the storage bench’s wide projection over one EHR read about 10 000 pages per call in some rounds and 1 200 in others. A node contained in a versioned object is now read inside its parent on such a query; results do not change (#4487).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

  • AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri argument whole when its value carries a ?, as the QUERY master03 example url=http://snomed.info/sct?fhir_vs does. Before, the arguments were read from after the first ? anywhere in params_uri, so the value-set URL reached the terminology server truncated and the arguments before it were lost. A params_uri that starts with a path still has its arguments read from after the path’s ? (#4490).

  • An ADL 2 template whose constituent archetype the CDR does not hold is refused at upload, with a 422 whose errors carry one VARXR entry per missing archetype (AOM2 C_ARCHETYPE_ROOT validity rule VARXR); the 2013 CKM t_encounter_opt_test, built on the refused apgar, no longer loads, so the sandbox library pin is 231 accepted / 91 refused. A stored template whose constituent was deleted later answers its JSON form and example with the same VARXR list, and the viewer says so on every tab of the template, names each missing archetype and links to uploading it, instead of a raw 422 (#4505).

  • ADL 2 validation admits a constraint on an RM computed attribute. VCARM accepts an attribute “stored or computed” (AOM2 master04.5 §Validity Rules: C_ATTRIBUTE), and a function with no arguments is a computed attribute (master04.2 §Attribute Nodes), so an archetype that constrains DV_PROPORTION.is_integral, DV_QUANTITY.is_integral or EVENT.offset is no longer refused with 422. The generated RM model (openehr_rm::v1_2::model) now carries each class’s computed attributes from the BMM functions (RmClass::computed_attributes, computed_attribute). openEHR-EHR-CLUSTER.microscopy_prostate_carcinoma.v1.0.0 from the 2013 CKM ADL 2 export now uploads, and the sandbox archetype-library pin moves to 232 accepted and 90 refused. DV_QUANTITY.property, which no RM release defines, stays refused. The nine openehr-* crates step to 0.0.124 (#4513).

  • DELETE /ferroehr/rest/extensions/v1/definition/artefact/adl2/{artefact_id} refuses with 409 Conflict while another stored ADL 2 artefact does not compile without the one deleted (a slot filler, a use_archetype reference or a specialisation parent, directly or through another artefact), and the message names every one of them. Before, the delete succeeded and left each template built on the archetype unable to compile (#4522).

  • The viewer’s ADL 2 template upload names each violation the CDR found when it refuses the source: every AOM2 rule code with the CDR’s words for it, so a template whose constituent archetype the CDR does not hold names that archetype (VARXR). The upload now asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details, and reads the stored template’s id from the ETag of the 201 (#4523).

  • An ADL 2 operational template whose included archetype reuses the template’s node ids uploads again, where it was refused with 422 under VCOSU. Node-id uniqueness now holds per archetype and restarts at each C_ARCHETYPE_ROOT (AOM 2 VCOSU, “unique within the archetype”), and on an operational template the codes and interior node ids below a root are judged against that constituent’s entry in component_terminologies (OPT 2 master03-opt_raw). A duplicate node id inside one archetype is still refused. The nine openehr-* crates step to 0.0.125 (#4529).

  • The viewer’s dashboard draws the EHR or composition count as unavailable, with the reason in plain words, when the CDR answers its count query with a result set of another shape (no row, several rows, an empty row, several values, or a value that is no whole number), where it drew 0 before. The template Releases table draws such a count the same way. A real zero still draws 0 (#4533).

  • An ADL 2 operational template upload runs the terminology integrity rules (VTLC, VATCV, VTVSID, VTVSUQ, VTVSMD, VTTBK, VTCBK, and VTSD where the level is carried) on each component_terminologies entry, and each violation names the component’s archetype id (#4536).

  • A COMPOSITION committed against an ADL 2 operational template is validated against the template’s leaf constraints, where it was stored with 201 whatever it held. An inlined archetype recorded by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability) now matches its template node, so the C_INTEGER, C_REAL, C_STRING, C_BOOLEAN, tuple, value-set and cardinality constraints below it answer 422 (AOM 2 master04.5). An archetype root the template does not inline is refused as a node the template does not define, and a C_INTEGER or C_REAL list ({10, 15, 20}) admits each of its values where only the first was read. Canonical, FLAT and STRUCTURED commits take the same check. The nine openehr-* crates step to 0.0.127 (#4540).

  • A versioned read whose version path segment is no SEMVER version and no {major} or {major}.{minor} prefix, such as v1.x, 1.x or 01, now answers 400 before any lookup, where it answered 404 (parameters/path/version.yaml, responses/400.yaml). This covers the ADL 2 template read GET /definition/template/adl2/{template_id}/{version}, the stored-query read GET /definition/query/{qualified_query_name}/{version} and both stored-query executions at a version, GET and POST /query/{qualified_query_name}/{version}. A well-formed version the server does not hold still answers 404 (#4541).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the content below an inlined archetype that the composition records by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability). The read left out everything below such a root with no error; it now matches the root to its template node the way the commit does. The nine openehr-* crates step to 0.0.128 (#4551).

  • A COMPOSITION committed against an ADL 2 operational template whose C_INTEGER or C_REAL constraint is a union of ranges (for example magnitude matches {|0..5|, |10..15|}) is refused with 422 when the value lies outside every range. Such a value was stored. The served Web Template does not change: its input validation.range holds one range, so a union leaves it out, as before, and the check runs on the commit (#4552).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the value of an ELEMENT that holds a DV_INTERVAL whose lower and upper the template constrains. The read left the value out with no error (#4559).

  • A FLAT or STRUCTURED read of a COMPOSITION returns the value of an ELEMENT that holds a DV_INTERVAL whose template leaves the interval’s lower and upper unconstrained, for ADL 1.4 and ADL 2 templates. The read left the value out with no error, and a FLAT or STRUCTURED commit could not name the parts (#4559).

  • A COMPOSITION written from FLAT or STRUCTURED against an ADL 2 operational template carries what a canonical body built to that template carries. A node the template gives no term text, such as the HISTORY and ITEM_TREE of an OBSERVATION, is named for its RM type (History, Tree) instead of its bare id-code (id2). Every inlined archetype root records the complete identifier the template names (…cnf_minimal.v1.0.0, BASE master07-referencing §Reconstitutability) in its archetype_node_id and archetype_details.archetype_id, and archetype_details.rm_version is the template’s rm_release. Before this change the root recorded the interface form (…v1) and the server’s RM 1.2.0. Compositions built against an ADL 1.4 template keep their major-version archetype ids and RM 1.2.0 (#4560).

  • ferroehr report in a release build names the openehr-* crate versions that release published to crates.io: the release lanes stamp the published version into the checkout before they build the server, where the report used to name the floor the source tree holds. A new openehr_crates_origin field says published for a release build and source_tree for any other, a pre-release included (#4565).

  • An ADL 2 archetype reference resolves to one defined stored release: an interface reference (….v1) takes the newest release of its major version, a reference with a minor or patch version narrows the match to it, and an alpha or beta version matches only its full identifier (BASE Artefact Identification, Referencing). With several releases of one archetype stored, a template upload, an operational template and the delete’s dependents check used whichever release PostgreSQL returned last, and a reference to a major version no stored release carries resolved to another major version (#4567).

  • An ADL 2 upload and an ADL 2 delete no longer race. An upload validated before a concurrent delete of an archetype it uses committed could store a template whose constituent was gone. Both now hold one transaction-scoped advisory lock from the repository read to the commit, so the second write decides on what the first committed: an upload after the delete is refused with VARXR, and a delete after the upload is refused with a 409 naming the template (#4568).

  • The syslog form (DICOM PS3.15, IHE ITI-20) of an EHR-level audit record names its object beside the patient, as a URI participant object with the id its FHIR AuditEvent names. The record of placing, releasing or expiring an EHDS Art. 3(3) access delay names the held composition as access-delay:<vo>, so an Audit Record Repository fed by syslog can tell which composition a hold covers (#4569).

  • An ADL 2 template or archetype that references two major versions of one archetype (….v1 and ….v2), or one archetype under two namespaces, now compiles against both: the openehr-adl repository holds an artefact under its namespace, archetype and major version, and resolves each reference within that key (BASE Artefact Identification, Referencing). A reference without a namespace now resolves in the namespace of the artefact that makes it, and never in another namespace (#4579).

  • A read of a stored ADL 2 template by a partial id or a version (GET /definition/template/adl2/{template_id}) now resolves the way template compilation does: with v1.0.0 stored beside v1.0.0-rc.1, ….v1, ….v1.0 and version=1.0.0 answer the release in either upload order, and the release candidate is read only by its full version (BASE Artefact Identification, Referencing) (#4580).

  • A write that sends no committer, on a route that takes a bare RM body (the EHR_STATUS, COMPOSITION and DIRECTORY writes, and the demographic party and relationship writes), now records a Bearer caller’s login as issued by the token issuer, as EHR creation and every other default committer already did. It recorded the issuer ferroehr for every caller before; a Basic caller’s login is still issued by ferroehr (#4585).

  • The viewer’s ADL 1.4 template upload opens the stored template’s page, read from the ETag (else the Location) of the CDR’s 201, where it went to an address with no template id. A refused upload names each violation the CDR found, its rule code or path and the CDR’s words for it, as the ADL 2 upload does: the upload asks for Prefer: return=representation, the preference under which ITS-REST lets a 4xx carry its error details (#4596).

  • An EHR Extract import refusal for a received version now names that version’s position in the extract (chapters[i]/items[j]/item/versions[k]), so an importer can find the defective ORIGINAL_VERSION in a large extract (#4602).

  • The import reads a version’s other_input_version_uids with the parser the CONTRIBUTION path uses and refuses an entry that is not an OBJECT_VERSION_ID, where it used to drop the entry. On the wire the extract decode already refuses such an entry with 422, so the import stays all or nothing (#4602).

  • An AQL archetype predicate matches the namespace exactly and reads the version it names as the oldest release it accepts within the major: org.a::… no longer finds data recorded in org.b, an id without a namespace finds only data recorded without one, and .v1.4.2 no longer finds data recorded under v1.0.0 (BASE Artefact Identification §Referencing and §Supporting Archetype-based Querying; #4629, #4630).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

  • CVE-2026-84243, the glibc gettext LANGUAGE directory traversal (GLIBC-SA-2026-0025), has no Debian fix for trixie yet and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never leave the C locale, so glibc ignores LANGUAGE and opens no message catalog (security/vex/ferroehr-os.openvex.json). In ferroehr-postgres only the container environment the operator writes can set LANGUAGE; no database connection can (security/vex/postgres-os.openvex.json) (#4500).

5.6.0-rc6 - 2026-11-01

5.6.0-rc6 is the sixth release candidate of 5.6.0, published for testing. It breaks clients that relied on the lenient CONTRIBUTION commit: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution read the body through the released NewContribution schema and answer 400 for a lifecycle_state that is no DV_CODED_TEXT, a description that is a plain string or an object without _type, an audit without change_type, an audit typed ATTESTATION or UPDATE_ATTESTATION, a missing audit or versions, and a preceding_version_uid sent as a plain string, and a stated empty system_id is refused with 422 on every write path (#3595, #4543). These notes carry every change since 5.5.0, those of the 5.6.0-rc5, 5.6.0-rc4 and 5.6.0-rc3 sections below included. New since 5.6.0-rc5: the EHDS Art. 3(3) delay of the person’s own access to newly registered data, off by default under the new [access_delay] section, with release by the treating professional or the controller (#4351); the EHDS Art. 5 mark of what a person inserts in their own EHR, with their update or deletion of a professional’s version refused with 403 (#4350); ADL 2 validation that admits a constraint on an RM computed attribute (#4513), holds VCOSU and the code checks per constituent archetype (#4529), refuses a template whose constituent archetype is not held (#4505) and the delete of an artefact another stored artefact needs (#4522), runs the terminology integrity rules on each component terminology (#4536), enforces a template’s leaf constraints at the commit (#4540) and resolves a reference to the newest stored release of its major version (#4567); FLAT and STRUCTURED reads under an ADL 2 template that return the content below a root recorded by its full archetype reference (#4551) and a DV_INTERVAL value (#4559); a node row below the root that keeps its archetype header, language and encoding in columns (#4525); an audit record whose request id and version uid ITI-81 finds from the record’s own columns, with retention reaping the search tokens no record names (#4524); a versioned read with a malformed version segment answered 400 (#4541); the viewer’s release groups and Compare tab (#4506) and its malformed AQL count drawn as unavailable (#4533); the conformance pipeline on Veredictum 0.3.2 with the released and development profiles and their badges (#4503); and the openehr-* crates versioned at publish (#4555). The conformance run on main still has known red rows, tracked on #4455, two of them Veredictum’s (Vernum-Projecten/Veredictum#964 and Vernum-Projecten/Veredictum#965), so this candidate does not claim a green Veredictum record. 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: a CONTRIBUTION commit must take the shapes above; a write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as committer and may change only what the person committed, and a service auth_failure answers 403 where it answered 500; a versioned read whose version is no SEMVER version or prefix answers 400 where it answered 404; the served Web Template of an ADL 2 template writes a data value’s aqlPath without an [idN] predicate; a composition that breaks an ADL 2 template’s leaf constraint, and an ADL 2 template whose constituent archetype is not held, answer 422; every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes, so keep it longer than the idle timeout of your clients and load balancers; a client that wrote a space as + in a query string sends %20; a party delete names the latest OBJECT_VERSION_ID in its path and answers 400 for a versioned object id; a body the RM refuses after the ITS-JSON schema admits it answers 422, and a misplaced CONTRIBUTION member, a Simplified Format commit without openehr-template-id and a flat |raw value without _type answer 400; a tag update under Prefer: return=identifier answers 200 with the tag list; with [authz.rbac] enabled a token holding only system/ SMART on openEHR scopes reaches clinical operations without a role; the viewer image serves no /pkg/ and reads no LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, and its Content-Security-Policy is script-src 'self'; style-src 'self'; Helm chart 21.0.0-rc6 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; clinical 0024 and party 0015 index the version uid, ext 0006, clinical 0025 and party 0016 add the ELEMENT value search index, built over every hot node row of an existing store, party 0017 adds the archetype statistics, and clinical 0026 and party 0018 grant MAINTAIN on the partitioned tables to the serving roles; clinical 0027 and party 0019 add the access delay, and ext 0007, clinical 0028 and party 0020 add the node header columns with no data migrated; audit migration 0019 gives the search token index a pending list of 512 kB, and audit 0020 adds the hash index on request_id. Helm chart 21.0.0-rc6 carries appVersion 5.6.0-rc6. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j, and the OpenVEX judgement of CVE-2026-84243.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The operations page gains a storage sizing section, measured on 5.6.0-rc5: about 21.6 KB of clinical storage per version of a vital signs composition (2.0 times its canonical JSON, 85 % of it node rows) and 155 KB for an International Patient Summary, about 1.2 KB per audit record of a read, the effect of [audit.store] retention on growth, pg_wal up to max_wal_size, the REINDEX INDEX CONCURRENTLY clinical.node_hot_pkey step (20 % of that index after heavy concurrent writing) and a query for the size per relation, with the method to reproduce the figures (#4304).

  • EHDS Art. 5: what a person inserts in their own EHR is marked as theirs, and they cannot alter what a professional committed. A write under a token whose SMART resource scopes are all patient/ scopes records PARTY_SELF as its AUDIT_DETAILS.committer, its external_ref naming the login, whatever committer the request names; a COMPOSITION it commits must name PARTY_SELF or PARTY_RELATED as composer; and its update or deletion of a version whose committer is not PARTY_SELF, directly or inside a CONTRIBUTION, is refused with 403. A professional’s writes are unchanged. The contribution list extension gains inserted_by_person per row, and the viewer names a PARTY_SELF committer “The person or their representative”. A service auth_failure status now answers 403 where it answered 500 (#4350).

  • The EHDS Art. 3(3) delay of the person’s own access to newly registered data, a Member State option and off by default. The new [access_delay] section (enabled, categories, hours) holds every new COMPOSITION of a listed [audit.categories] category from requests that may be the person’s own (a bearer token with a SMART patient/ grant or the ABAC patient claim, also beside user/ or system/ grants, as ferrohealth-smart 0.8.7 decides it) for at most hours: a read answers 404 as for an unknown composition, and a query, an export and a CONTRIBUTION leave it out, while a professional reads it as before. The treating professional releases a held composition from the clinical route POST {ext}/ehr/{ehr_id}/versioned_composition/{versioned_object_uid}/access-delay/release, with the checks of a write to that EHR, and the controller with POST /admin/access-delay/release; a professional holds one by hand within the limit with POST /admin/access-delay and reads the register with GET /admin/access-delay; every hold ends at its limit, and placing, releasing and expiry are each an access record. The subject-scoped ITI-81 read of the person’s own trail (EHDS Art. 9) leaves out the records about a held composition until it is released or expires, then serves them with their original times (ehds-logging 0.3.3) (clinical migration 0027_access_delay.sql, party migration 0019_access_delay_column.sql) (#4351).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • The Conformance workflow runs the Veredictum smart documents (smart_discovery, smart_scopes) against the same FerroEHR build after the Veredictum Action, which forwards no SMART Platform base URL and no scoped tokens. The same pinned image runs veredictum test with --platform-url and docs/conformance/party/ferroehr/ixit-smart.json, minting each access token in memory from the committed test issuer the SMART overlay trusts, and the job uploads its results, JUnit and CTRF views beside the catalogue’s and adds its summary under a heading of its own. scripts/conformance.sh gains the smart stage, and run drives it after the catalogue, so a local run covers SMART too (#4450).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

Changed

  • Breaking for clients that relied on the old leniencies: POST /ehr/{ehr_id}/contribution and POST /demographic/contribution now read the body through the released NewContribution schema, and refuse every shape it does not define. Each refused shape answers 400: (#3595)

    • a lifecycle_state in the CODE_PHRASE shape {"terminology_id": …, "code_string": …}, as a plain string, or as a bare {"value": …}. Send a DV_CODED_TEXT: {"value": "complete", "defining_code": {"terminology_id": {"value": "openehr"}, "code_string": "532"}};
    • a description sent as a plain string, on the contribution audit, on a commit_audit or on an attestation. Send a DV_TEXT or DV_CODED_TEXT object: {"_type": "DV_TEXT", "value": "…"};
    • a commit_audit or contribution audit without change_type. The server no longer infers creation or modification from preceding_version_uid; send the code of the change, for example 249 (creation), 251 (modification), 523 (deleted) or 666 (attestation);
    • a commit_audit or contribution audit whose _type is ATTESTATION or UPDATE_ATTESTATION, or any value other than UPDATE_AUDIT or AUDIT_DETAILS. Send the audit as UPDATE_AUDIT (or with no _type) and put each attestation in the version’s attestations list as an UPDATE_ATTESTATION. A 666 (attestation) version carries its attestations there too, and one with no attestation is refused;
    • a missing audit, which answered 422 before. A missing versions list, an attestation without reason or is_pending, a system_id that is not a string, and a preceding_version_uid sent as a plain string instead of {"value": "…"} are refused the same way.
  • An attestation’s own system_id is kept, and one that states none takes the system_id of the version audit it belongs to instead of the server’s. An attestation committer is now checked against its PARTY_PROXY rules like every other committer, so an invalid one answers 422 (#3595).

  • The served OpenAPI document describes each extension route by what it does and by the ITS-REST text that governs it, and no longer names openEHR Service Model operations, interfaces or status codes. Routes, request and response fields and status codes are unchanged (#4213).

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The storage-bench lane measures the baseline and the candidate on one runner in eight interleaved rounds each and gives two verdicts. The plan verdict fails an operation whose median pages per call rose more than 10 % and one page above the baseline’s highest round, whatever its time. The time verdict divides each round by its machine factor, so a slow spell of the runner no longer reads as the code, and fails an operation whose median p50 or p99 rose more than 15 % with every candidate round slower than every baseline round, or a candidate build that ran more than 15 % slower across the board in every round. A latency is no longer compared with a record measured on another runner (#4318).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • The LANG specification text is re-vendored at f80763ab (#4377). SPECLANG-18 states how a P_BMM schema’s JSON form encodes a cardinality interval, which the JSON schema reader already followed; its documentation now cites the rule, and a test pins it. SPECLANG-16 types the P_BMM class lists as keyed collections; the generated P_BMM model takes that change with #4497. The nine openehr-* crates step to 0.0.122.

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • The viewer runs on ferrohealth-server 0.4.2 and takes its Content-Security-Policy from the family’s server-rendered console headers (FerroHEALTH/FerroSYS#202) in place of a layer of its own. The policy a deployment sees is unchanged: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; object-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none', with no nonce. FerroUI’s assets under /ferrohealth-ui/ keep the Cache-Control FerroUI sets, and a refusal under that path is no-store whatever FerroUI’s router sets, so a cache in front of the viewer keeps no error answer for an asset. The REST server takes the same ferrohealth-server release with no change in what it serves (#4439).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

  • The P_BMM model follows SPECLANG-16. The LANG 1.1.0 BMM that openehr-codegen reads is now the specifications-LANG file at f80763ab, so P_BMM_SCHEMA.primitive_types and class_definitions are maps keyed by class name, P_BMM_CLASS.ancestor_defs is a map of P_BMM_BASE_TYPE keyed by type signature (a simple ancestor included), and P_BMM_GENERIC_TYPE.generic_parameter_defs is an optional map keyed by formal parameter name. The formal and actual generic parameters (P_BMM_CLASS.generic_parameter_defs, P_BMM_GENERIC_TYPE.generic_parameter_defs, BMM_GENERIC_CLASS.generic_parameters) are insertion-ordered IndexMaps, because the specification fixes their declaration order, and the model builders bind each actual parameter to its formal by name. In the released v1_0 generation, P_BMM_GENERIC_TYPE.generic_parameter_defs now carries the 1..* bound its BMM states. The nine openehr-* crates step to 0.0.123 (#4497).

  • The conformance pipeline runs Veredictum 0.3.1. The Conformance workflow drives the catalogue through the Veredictum Action twice, each run on fresh volumes: under the released profile, the frozen openEHR release, which is the record and decides the job, and under the development profile, the pinned master of each openEHR repository, which is published beside it under its own job summary heading and artifact and never counted toward it. Both runs include the SMART documents: the Action now takes the SMART Platform base URL and the committed test issuer’s key. The workflow reads the result paths and the exit status from the Action’s outputs. scripts/conformance.sh takes the profile as CONF_PROFILE (released by default; a development run writes to docs/conformance/ferroehr-development/) and drives the SMART documents in the same run, and the IXIT carries the test issuer’s scoped_tokens declaration (#4503).

  • The viewer groups a template’s releases. The template list draws one group per template, the newest release first with the number of releases and the earlier releases under it, in both the ADL 1.4 and the ADL 2 tab. A template’s page names its release beside its title, where every other release is a link with the newest marked, also with FerroUI’s scripts refused; the release select and its Open button above the tabs are gone. The Overview lists the releases with the compositions committed against each, and a new Compare tab lists the paths one release adds, removes and changes against another (#4506).

  • An audit record no longer adds a row to audit.audit_search_token for its request id or for the version uid a create wrote. ITI-81 entity.identifier finds both from the record’s own request_id and resource_id columns, through a new hash index on request_id (audit migration 0020), so the token table stays flat as requests are audited. Retention now deletes the search tokens that no remaining record names, among them the request id tokens of records written by 5.6.0-rc5 and earlier, and the database still refuses to delete a token a kept record names (#4524).

  • A clinical or party node row below the root no longer repeats its archetype header, language and encoding: an archetype root’s archetype_details and archetype_node_id are rebuilt from the archetype column and a new rm_version column, and a language in ISO_639-1 and an encoding in IANA_character-sets keep only their code, in the new language and encoding columns. Every reader puts the keys back where they stood, so the canonical JSON, every AQL result and the served key order are unchanged; a header of any other shape stays in the fragment. A vital signs version takes 17.1 KB instead of 21.6 KB and an International Patient Summary 131 KB instead of 159 KB. Migrations ext/0007, clinical/0028 and party/0020 add the columns and the readers; no data is migrated, and a row written before keeps its whole fragment and reads the same (#4525).

  • Breaking for clients that relied on the remaining commit-audit leniencies: the audit rules of the CONTRIBUTION commit now hold on every write path (#4543).

    • A description object without _type, such as {"value": "…"}, is refused with 400 on a contribution audit, a commit_audit and an attestation, and so is an attestation reason without _type. Send {"_type": "DV_TEXT", "value": "…"} or a DV_CODED_TEXT with its _type.
    • A stated empty system_id is refused with 422 on the direct writes too (openehr-audit-details: system_id="" on a create, update or delete), where the server used to replace it with its own. Leave system_id out to get the server’s, or send a non-empty one. A system_id that is not a string stays a 400 on both CONTRIBUTION routes.
    • A commit audit of type UPDATE_ATTESTATION handed to the library’s direct write calls (create_composition, update_composition, the EHR_STATUS, DIRECTORY and party writes) is refused with 400, as on the CONTRIBUTION routes. Send the audit as UPDATE_AUDIT and put each attestation in the version’s attestations list.
  • Members that NewContribution or UPDATE_AUDIT do not declare, such as a reason or is_pending on a plain audit or an extra top-level member of a contribution, are accepted and ignored: the released schemas leave both objects open (#4543).

  • The openehr-* crates take their version when they are published. A pull request that changes the crates no longer bumps their lockstep 0.0.x version: the crates publish reads crates.io, takes the next version when packaged content changed since the newest published one, and refuses to publish changed content under a version crates.io already holds. The version in the manifests on main can therefore trail the newest version on crates.io (#4555).

  • The served Web Template of an ADL 2 operational template writes the aqlPath of an ELEMENT’s value and of a data value’s parts without an [idN] predicate, since a data value carries no archetype_node_id (ADL2 master05-paths §“Relationship with W3C Xpath”). A DV_INTERVAL and its parts read …/value, …/value/lower and …/value/upper, the form the ADL 1.4 Web Template writes, where they read …/value[id6]/lower[id7] before, and the parts’ ids are lower and upper (#4559).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

  • The separate SMART step of the Conformance workflow, the smart stage of scripts/conformance.sh with its CONF_SMART_OUT, and docs/conformance/party/ferroehr/ixit-smart.json are gone: the one catalogue run covers the SMART documents. The party statement no longer carries the performance block, which the Veredictum 0.3.1 statement schema refuses (#4503).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • An AQL query aimed at named EHRs (the ITS-REST ehr_id parameter) reads those EHRs’ objects first and each object’s own content after them, whatever the planner statistics say. When an archetype’s rows fell out of the most-common-value list of the archetype statistics, the planner estimated them at one row and could start from that archetype over the whole population, evaluating every projected path for every EHR’s rows before the EHR filter: the storage bench’s wide projection over one EHR read about 10 000 pages per call in some rounds and 1 200 in others. A node contained in a versioned object is now read inside its parent on such a query; results do not change (#4487).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

  • AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri argument whole when its value carries a ?, as the QUERY master03 example url=http://snomed.info/sct?fhir_vs does. Before, the arguments were read from after the first ? anywhere in params_uri, so the value-set URL reached the terminology server truncated and the arguments before it were lost. A params_uri that starts with a path still has its arguments read from after the path’s ? (#4490).

  • An ADL 2 template whose constituent archetype the CDR does not hold is refused at upload, with a 422 whose errors carry one VARXR entry per missing archetype (AOM2 C_ARCHETYPE_ROOT validity rule VARXR); the 2013 CKM t_encounter_opt_test, built on the refused apgar, no longer loads, so the sandbox library pin is 231 accepted / 91 refused. A stored template whose constituent was deleted later answers its JSON form and example with the same VARXR list, and the viewer says so on every tab of the template, names each missing archetype and links to uploading it, instead of a raw 422 (#4505).

  • ADL 2 validation admits a constraint on an RM computed attribute. VCARM accepts an attribute “stored or computed” (AOM2 master04.5 §Validity Rules: C_ATTRIBUTE), and a function with no arguments is a computed attribute (master04.2 §Attribute Nodes), so an archetype that constrains DV_PROPORTION.is_integral, DV_QUANTITY.is_integral or EVENT.offset is no longer refused with 422. The generated RM model (openehr_rm::v1_2::model) now carries each class’s computed attributes from the BMM functions (RmClass::computed_attributes, computed_attribute). openEHR-EHR-CLUSTER.microscopy_prostate_carcinoma.v1.0.0 from the 2013 CKM ADL 2 export now uploads, and the sandbox archetype-library pin moves to 232 accepted and 90 refused. DV_QUANTITY.property, which no RM release defines, stays refused. The nine openehr-* crates step to 0.0.124 (#4513).

  • DELETE /ferroehr/rest/extensions/v1/definition/artefact/adl2/{artefact_id} refuses with 409 Conflict while another stored ADL 2 artefact does not compile without the one deleted (a slot filler, a use_archetype reference or a specialisation parent, directly or through another artefact), and the message names every one of them. Before, the delete succeeded and left each template built on the archetype unable to compile (#4522).

  • An ADL 2 operational template whose included archetype reuses the template’s node ids uploads again, where it was refused with 422 under VCOSU. Node-id uniqueness now holds per archetype and restarts at each C_ARCHETYPE_ROOT (AOM 2 VCOSU, “unique within the archetype”), and on an operational template the codes and interior node ids below a root are judged against that constituent’s entry in component_terminologies (OPT 2 master03-opt_raw). A duplicate node id inside one archetype is still refused. The nine openehr-* crates step to 0.0.125 (#4529).

  • The viewer’s dashboard draws the EHR or composition count as unavailable, with the reason in plain words, when the CDR answers its count query with a result set of another shape (no row, several rows, an empty row, several values, or a value that is no whole number), where it drew 0 before. The template Releases table draws such a count the same way. A real zero still draws 0 (#4533).

  • An ADL 2 operational template upload runs the terminology integrity rules (VTLC, VATCV, VTVSID, VTVSUQ, VTVSMD, VTTBK, VTCBK, and VTSD where the level is carried) on each component_terminologies entry, and each violation names the component’s archetype id (#4536).

  • A COMPOSITION committed against an ADL 2 operational template is validated against the template’s leaf constraints, where it was stored with 201 whatever it held. An inlined archetype recorded by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability) now matches its template node, so the C_INTEGER, C_REAL, C_STRING, C_BOOLEAN, tuple, value-set and cardinality constraints below it answer 422 (AOM 2 master04.5). An archetype root the template does not inline is refused as a node the template does not define, and a C_INTEGER or C_REAL list ({10, 15, 20}) admits each of its values where only the first was read. Canonical, FLAT and STRUCTURED commits take the same check. The nine openehr-* crates step to 0.0.127 (#4540).

  • A versioned read whose version path segment is no SEMVER version and no {major} or {major}.{minor} prefix, such as v1.x, 1.x or 01, now answers 400 before any lookup, where it answered 404 (parameters/path/version.yaml, responses/400.yaml). This covers the ADL 2 template read GET /definition/template/adl2/{template_id}/{version}, the stored-query read GET /definition/query/{qualified_query_name}/{version} and both stored-query executions at a version, GET and POST /query/{qualified_query_name}/{version}. A well-formed version the server does not hold still answers 404 (#4541).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the content below an inlined archetype that the composition records by its complete identifier (openEHR-EHR-OBSERVATION.x.v1.0.0, BASE master07-referencing §Reconstitutability). The read left out everything below such a root with no error; it now matches the root to its template node the way the commit does. The nine openehr-* crates step to 0.0.128 (#4551).

  • A FLAT or STRUCTURED read of a COMPOSITION committed against an ADL 2 operational template returns the value of an ELEMENT that holds a DV_INTERVAL whose lower and upper the template constrains. The read left the value out with no error (#4559).

  • An ADL 2 archetype reference resolves to one defined stored release: an interface reference (….v1) takes the newest release of its major version, a reference with a minor or patch version narrows the match to it, and an alpha or beta version matches only its full identifier (BASE Artefact Identification, Referencing). With several releases of one archetype stored, a template upload, an operational template and the delete’s dependents check used whichever release PostgreSQL returned last, and a reference to a major version no stored release carries resolved to another major version (#4567).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

  • CVE-2026-84243, the glibc gettext LANGUAGE directory traversal (GLIBC-SA-2026-0025), has no Debian fix for trixie yet and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never leave the C locale, so glibc ignores LANGUAGE and opens no message catalog (security/vex/ferroehr-os.openvex.json). In ferroehr-postgres only the container environment the operator writes can set LANGUAGE; no database connection can (security/vex/postgres-os.openvex.json) (#4500).

5.6.0-rc5 - 2026-10-31

5.6.0-rc5 is the fifth release candidate of 5.6.0, published for testing. 5.6.0-rc4 was tagged and not published: its vulnerability record found CVE-2026-84243 in glibc with no judgement, so the release stopped as a draft, and 5.6.0-rc5 carries its section below. These notes carry every change since 5.5.0, those of the 5.6.0-rc4 and 5.6.0-rc3 sections below included. New since 5.6.0-rc4: CVE-2026-84243 is judged not_affected in every image (#4500); AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri value that carries a ? whole (#4490); an AQL query aimed at named EHRs reads those EHRs’ objects first whatever the archetype statistics say (#4487); the viewer takes its Content-Security-Policy from ferrohealth-server 0.4.2, with the same policy (#4439); the Conformance workflow runs the Veredictum SMART documents beside the catalogue (#4450); the storage bench measures baseline and candidate on one runner (#4318); and the LANG specification is re-vendored, with the openehr-* crates at 0.0.122 (#4377). New in 5.6.0-rc4 over 5.6.0-rc3: an AQL lookup by an ELEMENT value, such as a person by an identifier, reads the one object through a value index in both stores (#4437), and a lookup of an EHR by its EHR_STATUS.subject reads the EHR row (#4442); a request on a kept-alive connection idle past the head bound is answered, so a client that reuses the connection for a commit no longer sees EOF (#4381); the server keeps planner statistics on the partitioned tables, the serving roles hold MAINTAIN on them, and a version read probes its commit audit by key (#4438, #4453); an AQL path read is estimated at the rows it reads and a page ordered by the version uid reads an index with no parallel worker (#4429, #4431); EHR e NOT CONTAINS runs (#4465), a DV_DATE_TIME comparison compares instants (#4466), and the REST, party, template and Simplified Format fixes the Veredictum 0.3.0 catalogue found (#4458 to #4464, #4467 to #4475); the viewer draws every page on its server with FerroUI 0.6 (#4433); AQL TERMINOLOGY('expand', …) takes the value set from the url argument of a $expand query form (#4488); and the conformance pipeline runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment with FerroTERM beside it, in CI through the Veredictum GitHub Action (#4448, #4449, #4476). 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes, so keep it longer than the idle timeout of your clients and load balancers; a client that wrote a space as + in a query string sends %20; a party delete names the latest OBJECT_VERSION_ID in its path and answers 400 for a versioned object id; a body the RM refuses after the ITS-JSON schema admits it answers 422, and a misplaced CONTRIBUTION member, a Simplified Format commit without openehr-template-id and a flat |raw value without _type answer 400; a tag update under Prefer: return=identifier answers 200 with the tag list; with [authz.rbac] enabled a token holding only system/ SMART on openEHR scopes reaches clinical operations without a role; the viewer image serves no /pkg/ and reads no LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, and its Content-Security-Policy is script-src 'self'; style-src 'self'; Helm chart 21.0.0-rc5 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; clinical 0024 and party 0015 index the version uid, ext 0006, clinical 0025 and party 0016 add the ELEMENT value search index, built over every hot node row of an existing store, party 0017 adds the archetype statistics, and clinical 0026 and party 0018 grant MAINTAIN on the partitioned tables to the serving roles; audit migration 0019 gives the search token index a pending list of 512 kB. Helm chart 21.0.0-rc5 carries appVersion 5.6.0-rc5. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j, and the OpenVEX judgement of CVE-2026-84243.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • The Conformance workflow runs the Veredictum smart documents (smart_discovery, smart_scopes) against the same FerroEHR build after the Veredictum Action, which forwards no SMART Platform base URL and no scoped tokens. The same pinned image runs veredictum test with --platform-url and docs/conformance/party/ferroehr/ixit-smart.json, minting each access token in memory from the committed test issuer the SMART overlay trusts, and the job uploads its results, JUnit and CTRF views beside the catalogue’s and adds its summary under a heading of its own. scripts/conformance.sh gains the smart stage, and run drives it after the catalogue, so a local run covers SMART too (#4450).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

Changed

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The storage-bench lane measures the baseline and the candidate on one runner in eight interleaved rounds each and gives two verdicts. The plan verdict fails an operation whose median pages per call rose more than 10 % and one page above the baseline’s highest round, whatever its time. The time verdict divides each round by its machine factor, so a slow spell of the runner no longer reads as the code, and fails an operation whose median p50 or p99 rose more than 15 % with every candidate round slower than every baseline round, or a candidate build that ran more than 15 % slower across the board in every round. A latency is no longer compared with a record measured on another runner (#4318).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • The LANG specification text is re-vendored at f80763ab (#4377). SPECLANG-18 states how a P_BMM schema’s JSON form encodes a cardinality interval, which the JSON schema reader already followed; its documentation now cites the rule, and a test pins it. SPECLANG-16 types the P_BMM class lists as keyed collections; the generated P_BMM model takes that change with #4497. The nine openehr-* crates step to 0.0.122.

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • The viewer runs on ferrohealth-server 0.4.2 and takes its Content-Security-Policy from the family’s server-rendered console headers (FerroHEALTH/FerroSYS#202) in place of a layer of its own. The policy a deployment sees is unchanged: default-src 'self'; script-src 'self'; style-src 'self'; img-src 'self' data:; connect-src 'self'; object-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none', with no nonce. FerroUI’s assets under /ferrohealth-ui/ keep the Cache-Control FerroUI sets, and a refusal under that path is no-store whatever FerroUI’s router sets, so a cache in front of the viewer keeps no error answer for an asset. The REST server takes the same ferrohealth-server release with no change in what it serves (#4439).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • An AQL query aimed at named EHRs (the ITS-REST ehr_id parameter) reads those EHRs’ objects first and each object’s own content after them, whatever the planner statistics say. When an archetype’s rows fell out of the most-common-value list of the archetype statistics, the planner estimated them at one row and could start from that archetype over the whole population, evaluating every projected path for every EHR’s rows before the EHR filter: the storage bench’s wide projection over one EHR read about 10 000 pages per call in some rounds and 1 200 in others. A node contained in a versioned object is now read inside its parent on such a query; results do not change (#4487).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

  • AQL TERMINOLOGY('validate', …) and TERMINOLOGY('subsumes', …) keep a params_uri argument whole when its value carries a ?, as the QUERY master03 example url=http://snomed.info/sct?fhir_vs does. Before, the arguments were read from after the first ? anywhere in params_uri, so the value-set URL reached the terminology server truncated and the arguments before it were lost. A params_uri that starts with a path still has its arguments read from after the path’s ? (#4490).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

  • CVE-2026-84243, the glibc gettext LANGUAGE directory traversal (GLIBC-SA-2026-0025), has no Debian fix for trixie yet and is judged not_affected in every image. The ferroehr and ferroehr-viewer programs never leave the C locale, so glibc ignores LANGUAGE and opens no message catalog (security/vex/ferroehr-os.openvex.json). In ferroehr-postgres only the container environment the operator writes can set LANGUAGE; no database connection can (security/vex/postgres-os.openvex.json) (#4500).

5.6.0-rc4 - 2026-10-31

5.6.0-rc4 is the fourth release candidate of 5.6.0, published for testing. These notes carry every change since 5.5.0, those of the 5.6.0-rc3 section below included. New since 5.6.0-rc3: an AQL lookup by an ELEMENT value, such as a person by an identifier, reads the one object through a value index in both stores (#4437), and a lookup of an EHR by its EHR_STATUS.subject reads the EHR row (#4442); a request on a kept-alive connection idle past the head bound is answered, so a client that reuses the connection for a commit no longer sees EOF (#4381); the server keeps planner statistics on the partitioned tables, the serving roles hold MAINTAIN on them, and a version read probes its commit audit by key (#4438, #4453); an AQL path read is estimated at the rows it reads and a page ordered by the version uid reads an index with no parallel worker (#4429, #4431); EHR e NOT CONTAINS runs (#4465), a DV_DATE_TIME comparison compares instants (#4466), and the REST, party, template and Simplified Format fixes the Veredictum 0.3.0 catalogue found (#4458 to #4464, #4467 to #4475); the viewer draws every page on its server with FerroUI 0.6 (#4433); AQL TERMINOLOGY('expand', …) takes the value set from the url argument of a $expand query form (#4488); and the conformance pipeline runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment with FerroTERM beside it, in CI through the Veredictum GitHub Action (#4448, #4449, #4476). 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes, so keep it longer than the idle timeout of your clients and load balancers; a client that wrote a space as + in a query string sends %20; a party delete names the latest OBJECT_VERSION_ID in its path and answers 400 for a versioned object id; a body the RM refuses after the ITS-JSON schema admits it answers 422, and a misplaced CONTRIBUTION member, a Simplified Format commit without openehr-template-id and a flat |raw value without _type answer 400; a tag update under Prefer: return=identifier answers 200 with the tag list; with [authz.rbac] enabled a token holding only system/ SMART on openEHR scopes reaches clinical operations without a role; the viewer image serves no /pkg/ and reads no LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, and its Content-Security-Policy is script-src 'self'; style-src 'self'; Helm chart 21.0.0-rc4 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; clinical 0024 and party 0015 index the version uid, ext 0006, clinical 0025 and party 0016 add the ELEMENT value search index, built over every hot node row of an existing store, party 0017 adds the archetype statistics, and clinical 0026 and party 0018 grant MAINTAIN on the partitioned tables to the serving roles; audit migration 0019 gives the search token index a pending list of 512 kB. Helm chart 21.0.0-rc4 carries appVersion 5.6.0-rc4. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

  • The Conformance workflow runs the Veredictum 0.3.0 catalogue in CI through the Veredictum GitHub Action, pinned by commit SHA and running the release image by digest. It builds FerroEHR from the commit, composes it with FerroTERM beside it on fresh volumes, reviews the party statement with veredictum validate, fails on a failed or errored row, writes the Markdown summary to the job summary and uploads the results, JUnit and CTRF views. It runs on pull requests that touch the server, the party set, the conformance compose files or the workflow, on main, on a release tag and twice a week. scripts/conformance.sh takes a stage (run, review, up, down), so CI and a local run compose the same deployment, and scripts/checks/veredictum-action.sh fails while the Action SHA, its version comment, the image digest and VEREDICTUM_VERSION disagree (#4449).

  • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now send Last-Modified: the commit that created the EHR, or the later commit that added or deleted one of its folder hierarchies (#4462).

Changed

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

  • FerroEHR runs on ferrohealth-server 0.4.1, which keeps a product’s own fallback and the layers around it when the stack merges its routes (FerroHEALTH/FerroSYS#199). FerroEHR, v5.6.0-rc3 included, was not exposed to the regression 0.4.1 fixes: the REST API and the viewer declare their own fallback inside their layers, and the management endpoints guard each route they serve, so a path no route matches answers as it did before (#4427).

  • The viewer draws every page on its server with FerroUI 0.6, with no Leptos and no WebAssembly: a link and a form load a new page, and FerroUI’s own small script modules add behaviour on top of what the server drew, so every page works with JavaScript off. The image ships the one ferroehr-viewer binary, with no /site tree beside it and no hash.txt; FerroUI’s stylesheet, theme and scripts are served from inside it under /ferrohealth-ui/0.6.0/, and /pkg/ no longer answers. The Content-Security-Policy is now script-src 'self'; style-src 'self' with no per-response nonce and no 'wasm-unsafe-eval'. The image no longer reads LEPTOS_SITE_ROOT, LEPTOS_OUTPUT_NAME or LEPTOS_HASH_FILES, so a deployment that sets them can drop them; LEPTOS_SITE_ADDR naming another address than server.listen still refuses the start. A proxy rule written for /pkg/ (the hosted deployment’s Cache-Control for it is removed) has nothing left to match (#4433).

  • scripts/conformance.sh runs the Veredictum 0.3.0 catalogue against one FerroEHR deployment: it reviews the party statement with veredictum validate, drives veredictum test in the release image pinned by digest with Basic login, and writes results.json, the HAR recording and the JUnit, CTRF and Markdown views to docs/conformance/ferroehr/. Docker or Podman runs it (CONTAINER_ENGINE). The FerroEHR statement now claims the 15 capability folders and answers every option of the 0.3.0 catalogue, and the IXIT keeps only what 0.3.0 reads (#4448).

  • The serving roles may now analyze their store’s partitioned tables: new migrations (clinical 0026, party 0018) grant MAINTAIN on node, version and vo_attestation, and on clinical.ehr, to ferroehr_clinical and ferroehr_party, so the hourly statistics refresh works under the role split. MAINTAIN also allows VACUUM, CLUSTER, REINDEX and LOCK TABLE on exactly those tables. A role that still lacks it gets the warning naming the tables, as before (#4453).

  • With the role gate on ([authz.rbac] enabled), a bearer token whose SMART on openEHR resource scopes are all system/ scopes, such as system/template-*.r, now reaches a clinical operation without a role: it is a backend client, and its scopes decide what it reaches. A user/ or patient/ token belongs to a person and still needs at least one role, as does a token with no SMART on openEHR scope and a Basic user; an admin operation still needs the admin role (#4467).

  • A tag update (PUT …/tags on EHR_STATUS, COMPOSITION and the five party kinds) sent with Prefer: return=identifier now answers 200 with the stored tag list and Preference-Applied: return=identifier. It answered 204 (#4475).

  • The conformance deployment runs FerroTERM beside FerroEHR, pinned by digest, with [terminology.server] pointed at it, locally in scripts/conformance.sh and in the Conformance workflow, so the aql-terminology-supported arm of the party statement runs against a real terminology server (#4476).

Removed

  • The conformance pipeline no longer runs EHRbase or the extra FerroEHR deployments. docker/sut-ehrbase.yml, the EHRbase record and party set, docs/conformance/COMPARISON.md, the book’s Comparison with EHRbase page, and the openPGP-signing, fail-open and stable-spec compose overlays (docker/sut-signing-pgp.yml, sut-pgp-parallel.yml, sut-terminology-failclosed.yml, sut-terminology-failopen.yml, sut-spec-stable.yml) with the test signing keys under corpus/keys/ are gone. scripts/conformance.sh no longer takes CONF_SUT, CONF_IXIT or CONF_STATEMENT, and refuses CONF_PERF_CLASS and CONF_STRESS until the measured lanes move to the Veredictum perf/ tree (#4448).

  • The Veredictum pin watcher workflow, scripts/checks/veredictum-pin.sh with its veredictum-pin-guard job, and the no-conformance-run label are gone: a pin bump now runs the new catalogue against its own build in the Conformance workflow (#4449).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • A request sent on a kept-alive connection that sat idle longer than [server] header_read_timeout_ms (30 s by default) is answered. The server used to close such a connection, sometimes while the request was arriving, so a client that reused it got EOF with no answer and no server log line, and a client that never retries a POST on a reused connection, such as Go’s net/http, reported a composition commit as failed. An idle connection now stays open until [server] idle_timeout_ms (120 s by default) and then ends cleanly before any new request byte; keep that limit longer than the idle timeout of your clients (90 s for Go) and of any load balancer in front (#4381).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

  • An AQL query ordered by c/uid/value with a LIMIT no longer starts a parallel worker on every call, and no longer switches between that plan and a sort of the whole join depending on the planner statistics. The version uid it sorts on had no index, so the planner sorted the version rows through a parallel Gather Merge, which costs a worker’s start-up per query on a server of any size: about 19 ms on the storage bench’s CI runner, where the two flat-shape path reads measured 23 ms in some rounds and 42 ms in others. Migrations clinical/0024 and party/0015 add an index on the uid expression of the hot version partition, so the query reads versions in uid order from it in one backend and stops at the limit (#4429).

  • An AQL query that reads values along a path no longer spends up to a second compiling itself before it runs. Each path step was bound through a set-returning function that PostgreSQL estimates at ten rows, so a chain of steps multiplied the row estimate tenfold per step: the planner chose join orders for a statement millions of times larger than the real one and switched on JIT compilation for it. On the storage bench, a wide projection of 9 values over one EHR read in 1 s instead of 15 to 20 ms, and an ordered WHERE on a magnitude in about 0.5 s instead of 50 ms (#4430). Each step is now a LATERAL subquery that the planner estimates from its own plan; results do not change (#4431).

  • An AQL lookup by an ELEMENT value, such as a person by an identifier (WHERE id/items[at0001]/value/id = $v) or a composition by a text value, read every node of the archetype in both stores and tested the value last: about 350 ms at 50 000 persons. The engine now anchors such an equality on the ELEMENT rows of the same version by their search text (ext.node_search_text, ext migration 0006), which the new partial expression index idx_node_hot_value_search and its statistics object serve (clinical migration 0025, party migration 0016), so the lookup reads the one object: 26 buffers and under a millisecond at 50 000 persons. The party store also gets the archetype dependency statistics the clinical store has had since 5.6.0-rc2 (party migration 0017). The storage bench measures the lookup as person_identifier_lookup (#4437).

  • The revision history and every point read of a version now reach the version’s commit audit through a lateral subquery that probes pk_commit_audit by key once per version row, so stale statistics on the partitioned version table can no longer turn the read into a walk of the whole commit audit index (1 318 buffers against 21 for a four-version object beside 1 300 others) (#4438).

  • A revision history or a current-version read of a COMPOSITION committed after the partitioned version table last had statistics could walk the whole commit audit index: the planner held the highest commit audit id of the seed for the highest there is and rated a merge join over pk_commit_audit as cheap. On a store of 1 300 compositions the history of a four-version object read 1 318 buffers instead of 21, and the storage bench read revision_history at about 440 buffers per call in 3 of 8 rounds. The server now keeps the partitioned parents’ statistics (#4453), and each read probes the commit audit by key (#4438).

  • An AQL lookup of an EHR by the subject of its EHR_STATUS (WHERE e/ehr_status/subject/external_ref/id/value = $u AND e/ehr_status/subject/external_ref/namespace = $n) joined every EHR_STATUS version and read the subject out of its stored fragment: about 150 ms on a store of a few thousand EHRs. The two leaves now read the EHR row’s subject_id and subject_namespace, the copy the commit keeps of the current EHR_STATUS subject, so the lookup enters through uq_ehr_subject and reads one row. The storage bench measures it as ehr_by_subject (#4442).

  • The partitioned tables node, version and vo_attestation of the clinical and party stores never had planner statistics of their own, since autovacuum does not analyze a partitioned table, so the planner estimated joins over them without any (PostgreSQL 18 §24.1.3). The server now runs ANALYZE ONLY on them, and on ehr while autovacuum has never analyzed it, after applying the migrations, after load_ehrs, at boot and every hour. The role it runs as needs the MAINTAIN privilege on those tables; without it the server logs the tables it could not analyze. The population CONTAINS of an archetype now enters through idx_node_hot_arch_subsume (3 703 buffers to 1 745 on the storage bench’s seed) (#4453).

  • A party delete (DELETE /demographic/{agent,group,organisation,person,role}/{uid_based_id}) and the party relationship delete now answer 400 when the path names the versioned object instead of its latest OBJECT_VERSION_ID, and delete nothing. They answered 204 and deleted the current version. The release requires the preceding_version_uid in that path (ITS-REST agent_delete.yaml and its four siblings) (#4478) (#4458).

  • A CONTRIBUTION member outside the operation’s member set, a party in an EHR CONTRIBUTION or a FOLDER in a demographic CONTRIBUTION, now answers 400 as a body that fails the request schema. It answered 422 (#4459).

  • A party update whose body carries no uid now checks each relationship source against the party the path names and refuses a foreign source with 422, as RM Relationships_validity requires. The check ran only when the body carried a uid, so such an update committed a relationship sourced at another party (#4460).

  • An ADL 1.4 template upload whose XML ends with an element still open now answers 400, as for any document that is not well-formed XML. The well-formedness gate stopped at the end of input without checking that every element was closed, so a truncated upload answered 422 (#4461).

  • A canonical JSON body that the published ITS-JSON schema admits and the RM refuses, such as a DV_URI or DV_EHR_URI without value or a DV_INTERVAL limit DV_ORDINAL without symbol, now answers 422 naming the member, on the direct commit routes, the CONTRIBUTION members and the party commits. A body the schema refuses still answers 400. All of them answered 400 (#4463).

  • A directory create or update whose root FOLDER names an archetype in archetype_node_id and carries no archetype_details now answers 422 and commits nothing: an archetype root carries archetype_details (RM common §The LOCATABLE Class). FerroEHR created such a directory. A 553|incomplete| commit and a nested FOLDER keep it optional. The viewer’s empty directory root now carries archetype_details (#4464).

  • AQL EHR e NOT CONTAINS <class expression> now runs and returns the EHRs that hold no matching object. It was refused as unsupported with 400 (#4465).

  • An AQL comparison between a DV_DATE_TIME path that ends on the data value itself, such as c/context/start_time, and a date-time literal now compares the instants both name, so values with different offsets compare correctly. The comparison read the whole data value and matched nothing (#4466).

  • A container member whose node id names an object constraint of the template but whose RM type conforms to none of them, such as an ELEMENT where the template allows only a CLUSTER, is now refused with 422. Only the opposite direction was refused (#4468).

  • An ADL 1.4 template upload with Prefer: return=representation now answers the representation Accept names: the OPT XML, or the Web Template under application/openehr.wt+json or application/json. An Accept outside those answers 406 and stores nothing. The upload always answered XML (#4469).

  • A Simplified Format COMPOSITION commit without the openehr-template-id header, and a flat |raw value without _type, now answer 400. Both answered 422 (#4470).

  • A Simplified Flat or Structured …/action_archetype_id now sets the String ACTIVITY.action_archetype_id, and the flat output writes it back. The converter built an object from the template’s string constraint, so the commit answered 422 (#4471).

  • ctx/action_ism_transition_current_state, as a rubric or a code, now sets ACTION.ism_transition.current_state of every ACTION the document gives no state. The converter wrote 524|initial| first and the context value never applied (#4472).

  • ctx/instruction_narrative now becomes INSTRUCTION.narrative of every INSTRUCTION the document gives no narrative. A placeholder <narrative> was stored instead (#4473).

  • The Simplified Flat DV_INTERVAL suffixes |lower_unbounded, |upper_unbounded, |lower_included and |upper_included are now read on an interval whose bounds are template nodes, and the flat output writes an unbounded flag when true and an included flag when false. The suffixes were refused as unknown with 422 (#4474).

  • AQL TERMINOLOGY('expand', …) against a FHIR terminology server takes the value set from the url argument when params_uri is the $expand query form (url=http://…), as the openEHR AQL examples write it, and still takes a bare value-set URL as it is. A query form without url is a 400 that names the missing argument. Before, the whole params_uri went to the server as the URL, and the query was answered 400 because the value set was not found (#4488).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

5.6.0-rc3 - 2026-10-30

5.6.0-rc3 is the third release candidate of 5.6.0, published for testing. These notes carry every change since 5.5.0, those of the 5.6.0-rc2 section below included. New since 5.6.0-rc2: FROM PERSON p CONTAINS EHR e finds the EHR whose EHR_STATUS.subject.external_ref names the party (#4411); a projection ending on a list attribute answers the list, where it answered it wrapped in a second list, and COUNT over such a path counts its elements (#4413); every error on the openEHR, extension and management routes answers RFC 9457 problem details (#4342); a + in a query string is a literal plus (#4356); the local EHR index is a supported posture (#4379); an audit store that 5.6.0-rc2 migrated starts and upgrades in place (#4424); FerroEHR takes ferrohealth-server 0.4.0 (#4417) and FerroAUDIT 0.3.2 (#4366); and storage and AQL changes keep counts, reads and supersedes fast as a store grows (#4286, #4289, #4340, #4375, #4399, #4420). 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; an error body is application/problem+json, so a client that read message reads detail, and one that read validationErrors sends Prefer: return=representation and reads errors; audit migration 0017 refuses an audit store that holds records written by 5.5.0 or earlier, while an audit store that 5.6.0-rc2 migrated upgrades in place; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; [server] header_read_timeout_ms no longer bounds an idle kept-alive connection, which the new [server] idle_timeout_ms (120000 by default) closes; a client that wrote a space as + in a query string sends %20; Helm chart 21.0.0-rc3 refuses to render until config.deployment_profile names sandbox or production, and a values file for its FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer. Storage generation 3 adds node columns and migrates no data; clinical migration 0020 adds the origin to retention policies and 0021 statistics on the archetype columns, and clinical 0022 and 0023 and party 0013 and 0014 rebuild the version tree’s uniqueness constraint and index its version id columns, which reads every version row of an existing store once; audit migration 0019 gives the search token index a pending list of 512 kB. Helm chart 21.0.0-rc3 carries appVersion 5.6.0-rc3. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

  • The openehr-its client reads an RFC 9457 problem details refusal: ErrorBody::problem() exposes its type, title, status, detail and instance, ignoring a member of the wrong type (RFC 9457 §3.1), and ErrorBody::message() answers the problem’s detail, else its title, where it answered only the ITS-REST Error’s message. The openehr-* crates move to 0.0.115 (#4355).

  • The local EHR index ([ehr_index] provider = "local") is a supported production posture beside the IHE patient identity source: EHR_STATUS.subject names a patient by its own namespace and identifier, which GET /ehr?subject_id=&subject_namespace= matches, and with no [privacy] subject_namespaces declared the reference is stored as the client sends it. The compose quickstart selects it with FERROEHR_EHR_INDEX=local docker compose up, which runs its ferroehr-config-local block (#4379).

  • CI checks every inline configuration of the compose quickstart with the config check of the image the file pins, FerroEHR’s and FerroPIX’s, so the quickstart can no longer carry a configuration its pinned server refuses (#4380).

Changed

  • A population AQL query that only counts its root objects, such as SELECT COUNT(c) FROM EHR e CONTAINS COMPOSITION c or the same with COUNT(*), reads the current version rows and no longer probes the root node of every composition. At 545 000 compositions the count fell from about 9 s to 0.2 s (#4286).

  • A REST request checks a connection out of each database pool once and keeps it until it answers, where it took one for each statement: two to four checkouts for a read of an EHR, an EHR_STATUS or a COMPOSITION, each waiting in the pool’s queue under load and each paying the pool’s release round trip. At 48 concurrent clients and a pool of 10, throughput rose by about 18 % and the median latency fell by about 13 % (#4289).

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • The viewer takes FerroUI 0.5.8: a menu item draws a visible bar under the pointer and the keyboard focus, and the top bar’s search no longer opens its panel again over the page the router draws after a search is submitted or one of its options followed (#4319).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

  • A statement the database cancels, by its statement_timeout or a cancel request (SQLSTATE 57014), answers 503 with Retry-After instead of 500: the statement rolled back, so the request can be repeated. By default the request timeout answers before the 60 s statement timeout fires, so on a request path it is raised by an operator’s cancel or by a forward step of the database host’s clock, which fires every running statement’s timeout at once (#4338).

  • Every error on the openEHR routes (/ferroehr/rest/openehr/v1/), the extension routes (/ferroehr/rest/extensions/v1/) and the management routes answers RFC 9457 problem details, application/problem+json, in place of the { message, validationErrors, error } body shaped after the ITS-REST example. The status codes are unchanged. A problem carries type, title, status, detail and the request id as a urn:uuid: instance; the refusals a client acts on have FerroEHR types documented on the book’s Problem types page (aql-invalid, query-timeout, validation-failed, content-invalid, conflict, version-conflict, unauthenticated, forbidden, processing-restricted, unavailable, rate-limited), every other error is about:blank. A request with Prefer: return=representation receives the additional details ITS-REST allows: the violations of a validation refusal as errors ({path, detail}, where validationErrors held "<path>: <message>" strings), the byte position of an AQL parse refusal, and the current_version of a 412. The stack’s own 408, 413, 429, 503 and panic 500 are problems too, and an OperationOutcome under the FHIR base, where an authentication refusal is now an OperationOutcome as well. The served OpenAPI document declares the Problem schema for every 4xx and 5xx, and the viewer reads problem details. A client that read message reads detail; one that read validationErrors sends Prefer: return=representation and reads errors (#4342).

  • The FHIR AuditEvent of an EHDS access record names the object it reads, creates, updates or deletes by its openEHR object id as entity.what.identifier with the system urn:openehr:object-id, where it wrote a what.reference that no FHIR server resolves and HAPI FHIR refused (FHIR R4 references.html §2.3.0.3). ITI-81 finds the record by entity.identifier. FerroEHR takes FerroAUDIT 0.3.2 (ihe-atna and ehds-logging) (#4366).

  • The access log costs the request path less under load. The audit store writes a drained batch, its search tokens and its token sets in one transaction on one connection (one commit where it took up to three), stores only the tokens it does not hold yet, and the GIN index on search_tokens takes new entries through a pending list of 512 kB. On the release load test the store wrote 1.5 KB of WAL per record where it wrote 5.2 KB (#4375).

  • [ehr_index] provider = "local" beside pixm_url, assigning_authority, [ehr_index.oauth] or [ehr_index.tls] is accepted, where it was refused: the IHE keys are kept, not used and named at boot, so provider alone switches the index, in the file or as FERROEHR__EHR_INDEX__PROVIDER. Declared pseudonym namespaces stay in force under either index (#4379).

  • A build no release lane marked reports its commit beside the package version, as SemVer build metadata (5.6.0-rc2+dev.1a2b3c4d5e6f), on ferroehr --version and server_version of GET /ferroehr/rest/status (the boot banner and /management/info already name the commit beside the version), so a report from a build of main no longer reads as the release it follows; a release build reports the bare version (#4380).

  • The compose quickstart runs FerroPIX 0.0.12, pinned by digest, with its configuration moved to that release’s keys (public_base_url at the top level, the Patient Identity Source’s client under [[trust.issuer]]), and Dependabot proposes each FerroPIX release (#4380).

  • The Helm chart runs FerroPIX 0.0.12 by digest, with the FerroPIX chart 0.3.4 built for it, and the hosted sandbox runs FerroPIX 0.0.12 too; both had stayed at 0.0.7 and 0.0.10 when the quickstart moved. A values file for the chart’s FerroPIX moves ferropix.config.server.public_base_url to ferropix.config.public_base_url and each issuer’s clients to ferropix.config.trust.issuer, the keys FerroPIX reads since 0.0.8; the chart’s examples and the production overlay do. FerroPIX’s chart now also runs its pods with enableServiceLinks: false and in a user namespace (ferropix.hostUsers: false, the default). A CI guard holds the quickstart, chart and sandbox to one FerroPIX release and checks the sandbox’s and the chart’s FerroPIX configuration under it, so the Dependabot proposal for the quickstart pin moves all three (#4392).

  • openehr-query, the AQL 1.1.0 lexer, parser, typed AST and printer, is published under Apache-2.0 from 0.0.116, like the generated model crates and openehr-its, so any Rust project can parse AQL with it. Versions 0.0.60 to 0.0.115 keep the Business Source License 1.1 they were published under (#4405).

  • FerroEHR takes ferrohealth-server 0.4.0 in place of 0.3.9. The new [server] idle_timeout_ms (120000 by default) closes a kept-alive HTTP/1.1 connection that waits that long for its next request, with a clean end of stream, and [server] header_read_timeout_ms counts from the first byte of each request head on a kept-alive connection (from accept for the first one, as before). The head bound used to count from the end of the previous answer, so it also closed an idle connection after 30 s, sometimes as a client sent its next request into it, which a client such as Go’s net/http saw as EOF on a POST it does not retry (#4381). A deployment that set header_read_timeout_ms to bound idle connections sets idle_timeout_ms instead. The refusals the listener writes itself (408, 413, 429, 503) and a panic’s 500 keep naming the server’s request id as their instance when the client sends its own X-Request-Id, as every other FerroEHR refusal does (#4417).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • The measurement overlay docker/sut-measurement.yml switches the caller rate limiter off by removing the [server.caller_rate] keys, so the server logs caller rate limiting configured with enabled=false at boot, instead of setting a rate of 4294967295 per second that ferrohealth-server 0.3.5 truncated to a zero interval and that answered stray 429s under concurrency. scripts/conformance.sh refuses to start a measurement stack whose ferroehr service still sets a caller rate (#4281).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • Superseding a composition or a party no longer slows down as its history grows. The foreign-key check that every new node and attestation row makes against its version could scan all versions of the object instead of probing one, depending on the planner statistics at the time: on the storage bench, superseding an object with a few hundred versions took about 14 ms instead of 9 ms when the plan flipped. Migrations clinical/0022 and party/0013 re-key the version-tree uniqueness constraint uq_version_tree with the object id last, so the check always probes the primary key. The constraint enforces the same uniqueness as before (#4340).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

  • A + in a query string is a literal plus, as RFC 3986 §2.1 and §3.4 decode a URI query, on every ITS-REST operation and every extension route: a date-time offset such as 2026-10-28T10:00:00+01:00 or an AQL parameter sent with an unencoded + reaches the server unchanged, where it was read as a space. A space in a query value is sent as %20. This reverses the form-decoding of #3891: a client that writes a space as + now sends a plus. The openehr-* crates move to 0.0.114 (#4356).

  • An AQL EHR e[ehr_id/value='…'] predicate whose value is not a UUID ('1234', the AQL 1.1.0 example) answers an empty result, as its WHERE e/ehr_id/value = '…' spelling does, on POST and GET /query/aql; it was refused with 400. Both spellings lower the EHR id comparison through one path, and a value the database cannot cast is refused with a message that no longer claims it came from a query parameter (#4378).

  • An AQL query with an archetype predicate no longer builds the archetype index bitmap of the whole store once per version. The planner multiplied the selectivities of the archetype columns, which each archetype fixes together, and expected about 15 000 of 545 067 blood-pressure OBSERVATIONs; a query over one EHR then took 38.8 s and answered 408. Clinical migration 0021_archetype_statistics.sql gives the columns a statistics object, and the same query takes 0.3 s (#4399).

  • FROM PERSON p CONTAINS EHR e (and the ORGANISATION and ROLE roots) joins a party to the EHR whose EHR_STATUS.subject.external_ref names it, so a party created with POST /demographic/person and an EHR created with POST /ehr answer their row with no other call. Before, the join read only the party-to-EHR map, which no REST route writes, and answered no rows on a deployed server. The reference matches when its id.value is the party’s uid without ::<system>::<version> and its type is the root class or its RM ancestor (ACTOR, PARTY), in any namespace; an EHR linked both ways answers once (#4411).

  • An AQL projection whose path ends on a list attribute inside a node’s data (c/feeder_audit/originating_system_item_ids, c/links, c/context/participations, c/composer/identifiers, .../mappings) answers the list itself, [{…}, {…}], and null when the list is absent. It answered the list wrapped in a second list, [[{…}, {…}]]; a client that unwrapped that outer array reads the column directly now. A path that continues below a list (.../originating_system_item_ids/id) still answers the array of its matches, ["…", "…"]. COUNT over a path ending on a list counts its elements (#4413).

  • Reading one version of a composition, EHR_STATUS, folder or party by its version id no longer slows down as the object’s history grows. Once a connection reused its prepared read, the plan scanned every version of the object and kept the one asked for: on the storage bench’s history of 500 versions such a read touched about 520 buffers instead of 12 to 24. Migrations clinical/0023 and party/0014 add an index on the version tree id columns of both storage tiers, so the read probes the one version it returns (#4420).

  • An audit store that 5.6.0-rc2 migrated starts again. Audit migration 0017 changed after 5.6.0-rc2 shipped it, so the server refused such a store on the migration’s checksum at boot. 0017 is back to its 5.6.0-rc2 text, and the new audit migration 0019_search_tokens_pending_list.sql gives the search token index its pending list of 512 kB, so a store that 5.6.0-rc2 migrated upgrades in place (#4424).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

5.6.0-rc2 - 2026-10-29

5.6.0-rc2 is the second release candidate of 5.6.0, published for testing. 5.6.0-rc1 was tagged but not published: the release plan refused it because its section missed a change merged just before the cut (#4361), so these notes carry every change since 5.5.0, those of the 5.6.0-rc1 section below included. 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; audit migration 0017 refuses an audit store that holds records written by an earlier release, so point the release candidate at an empty audit store; a [server.tls] table with client_ca_file must set client_auth = "required" (or "optional") beside it, or the configuration is refused at load; and Helm chart 21.0.0-rc2 refuses to render until config.deployment_profile names sandbox or production. Storage generation 3 adds node columns and migrates no data, and clinical migration 0020 adds the origin to retention policies. Helm chart 21.0.0-rc2 carries appVersion 5.6.0-rc2. Security carries two advisories, GHSA-2mc2-h3jq-c5wg and GHSA-52mf-qg39-4q4j.

Added

  • The logging component’s intended performance is stated as checkable claims (EHDS Annex II 1.1): every composition create and read recorded exactly once and every AQL execution at least once, no access succeeding without its record, at most 125 % of the latency and at least 80 % of the throughput with the access log off, and a 50-record ITI-81 page within 250 ms at p95. scripts/perf/audit.sh measures them with the access log off and on (the load test’s new --audit mode counts the records through ITI-81 and times retrievals), scripts/perf/audit-verdict.sh gives the verdict, the release refuses a claim not met and attaches the summary as ferroehr-<tag>-logging-performance.md (Annex III point 2), and the weekly performance guard runs it on main. The book’s compliance section gains the page that states the claims and the method (#4293).

  • scripts/interop/atna.sh exercises the logging component’s ATNA transactions against independent implementations run from published images (EHDS Annex II 1.4): the ITI-20 FHIR feed into HAPI FHIR as the Audit Record Repository, judged by its validator with IHE BALP 1.1.4; the ITI-20 syslog feed into syslog-ng over RFC 5425 TLS, each payload read as a DICOM PS3.15 AuditMessage; and the same ITI-81 searches put to FerroEHR and to HAPI. The audit trail page documents it (#4296).

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

  • Retention periods and access rules can take the origin and the category of the data into account (EHDS Annex II 3.4). The new top-level [origins] register declares the origins a deployment recognises, each by the id its data carry and the clients whose direct commits come from it; a body without a FEEDER_AUDIT committed by a registered client is stamped with that client’s origin instead of the server’s system_id. PUT /admin/retention/policy takes an optional origin naming a registered origin, the register and GET /admin/retention/due report it, and the due list lists an origin-keyed period over the objects of that origin (clinical migration 0020_retention_by_origin.sql). The ABAC request carries category (the access record’s category tokens) and origin resource attributes on composition, contribution and query operations, the Cedar schema declares both, and a remote PDP binding may name them as parameters (#4299).

Changed

  • The Helm chart, at a new major version (21), has no default deployment profile: it refuses to render until config.deployment_profile names sandbox or production, so an installation that follows a quickstart instead of the instructions for use no longer runs the sandbox posture by omission (EHDS Annex II 1.2). A values file that relied on the old sandbox default must now set it, for example --set config.deployment_profile=sandbox; values-production.yaml sets production as before. The server binary’s own default and the Compose files, which declare sandbox explicitly, are unchanged (#4294).

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

  • The FHIR rendering of an access record ehds-logging does not classify (a template upload, a stored-query write, a template read) now carries the IHE BALP 1.1.4 entity:data slice its Create, Read, Update or Delete profile requires: type 2, role 4 and what.identifier, naming the operation when the record names no object. HAPI FHIR’s validator refused the template upload record without it (#4353).

Security

Shipped with functional changes because: both fixes exist only in ferrohealth-server releases that carry functional changes in the same serving path beside them (0.3.7 two fixes, 0.3.8 and 0.3.9 the problem details module, the kept listener TLS and the unmasked remedy), so there is no security-only release of the crate for FerroEHR to take, and the client CA refusal ships in the release candidate the advisory names.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

  • A [server.tls] table that sets client_ca_file with client_auth at none, written or by default, is refused at load, naming the setting (GHSA-52mf-qg39-4q4j). Before 5.5.0 a client CA alone meant that every client presents a certificate; 5.5.0 moved the listener onto ferrohealth-server, where client_auth defaults to none and the CA was read and ignored, so a configuration carried over passed config check and served without asking a client certificate. FerroEHR takes ferrohealth-server 0.3.9, which refuses it and names the remedy in full. A 5.5.0 deployer who sets client_ca_file sets client_auth = "required" beside it. The default template no longer pairs a client CA with none, and says what each client_auth value does (#4349).

5.6.0-rc1 - 2026-10-28

5.6.0-rc1 is the first release candidate of 5.6.0, published for testing. 5.5.0 stays the stable release: this pre-release moves neither the latest image tags, the book under /docs/latest/ nor the hosted sandbox, it publishes no crates, and it receives no fixes; the next release candidate or 5.6.0 itself replaces it. Before you try it: every route ITS-REST 1.1.0 does not define moves from /ferroehr/rest/openehr/v1/ to /ferroehr/rest/extensions/v1/, and the ITI-81 retrieval to /ferroehr/rest/fhir/r4/AuditEvent, with no redirect from the old paths; and audit migration 0017 refuses an audit store that holds records written by an earlier release, so point the release candidate at an empty audit store. Storage generation 3 adds node columns and migrates no data. Helm chart 20.0.1-rc1 carries appVersion 5.6.0-rc1. Security carries one advisory, GHSA-2mc2-h3jq-c5wg.

Added

  • FerroEHR ships an EHDS category map for the published CKM templates and COMPOSITION archetypes whose data sit in one Art. 14(1) priority category (the International Patient Summary, ePrescription, the generic laboratory result and CCTA report templates; the health_summary, prescription and transfer_summary archetypes), so an installation that declares no [audit.categories] keys records those accesses classified instead of unclassified (EHDS Annex II 3.2(c)). A deployment that declares any template or archetype replaces it whole. [audit.categories] shipped_map = false with nothing declared leaves no map, which deployment_profile = "production" refuses as the new no_category_map gap unless deployment_accepts names it (#4298).

Changed

  • The audit store keeps each record’s facts in columns and renders its FHIR R4 AuditEvent from them when the record is read over ITI-81 or delivered by the FHIR feed, instead of storing the rendered document in every record. A record takes about a quarter of the space it took before. The configuration the rendering depends on (the audit source, the store horizon, the category map and the emergency purposes) is kept once per distinct configuration in audit.audit_render_policy, and every record names the policy it was written under, so it renders the same document after the configuration changes. The tamper chain hashes every column of the record in one digest and checks each render policy against its digest; the earlier digest versions are retired. Audit migration 0017 needs an empty audit store and refuses one that holds records written by an earlier release, naming the remedy. ITI-81 parameters on an element of the document (address, the .identifier parameters, entity-type, entity-role, type, subtype) are matched against the rendered document (#4300).

  • The three audit indexes on recorded_at ascend, so their pages fill instead of splitting half full on every insert; the newest-first ITI-81 search scans them backward (#4302).

  • A clinical or party node row now stores only the part of its canonical JSON fragment that its own columns do not hold: _type, a lowercase archetype_node_id, a plain or simply coded name, and an ELEMENT value in one of the common basic shapes (a quantity, count, ordinal, coded or plain text, date, time, duration, identifier, boolean or URI) move into columns, the value into the new typed value columns. Every reader rebuilds the omitted keys, so the canonical JSON and every AQL result are unchanged, and a filter on a quantity’s magnitude or a code reads its column directly. The node heap shrinks by about a third (storage generation 3); a row written before keeps its whole fragment and reads the same. Migrations ext/0005, clinical/0019 and party/0012 add the columns and the readers; no data is migrated (#4306).

  • Every ITI-81 search parameter answers through an index again. The store takes the search tokens of a record’s rendered AuditEvent when it writes the record: each distinct token once in audit.audit_search_token, the code-system tokens (type, subtype, entity-type, entity-role, source.identifier) as one shared set in audit.audit_search_set, and their ids on the record in search_tokens under a GIN index; entity reads a hash index on resource_id. A search on address or a token parameter no longer renders every record in its date window. The tamper chain covers search_tokens, and the verifier reports a changed token or set (#4313).

  • Nothing ITS-REST 1.1.0 does not define is served under the openEHR base path any more. FerroEHR serves three namespaces beside each other under /ferroehr/rest/: /openehr/v1/ holds only the ITS-REST routes, /fhir/r4/ the FHIR and IHE interfaces, and /extensions/v1/ FerroEHR’s own extensions. The ITI-81 retrieval moved from /ferroehr/rest/openehr/v1/fhir/r4/AuditEvent to /ferroehr/rest/fhir/r4/AuditEvent. These routes moved from /ferroehr/rest/openehr/v1/... to /ferroehr/rest/extensions/v1/... with the same relative path: /terminology/..., /admin/event_subscription..., /demographic/party_relationship..., /demographic/versioned_party_relationship/..., /definition/archetype/..., /definition/artefact/..., /message/export..., /message/import..., /query/cohort, and the admin extensions /admin/config, /admin/template/{template_id}, /admin/query/{qualified_query_name}/{version}, /admin/party/..., /admin/archive/..., /admin/report/..., /admin/dump, /admin/load, /admin/integrity/..., /admin/restriction..., /admin/access-restriction..., /admin/research-objection and /admin/retention/.... The old paths answer 404, with no redirect. The served OpenAPI document and its family documents name the new paths; the admin extensions, the archetype library and the cohort query have family documents of their own (admin-extensions, archetypes, cohort, contributions). The paged CONTRIBUTION list moved too, to /ferroehr/rest/extensions/v1/ehr/{ehr_id}/contribution (#4330).

  • The server and viewer images build on the current gcr.io/distroless/cc-debian13:nonroot digest, so they carry the base image’s latest Debian package updates (#4335).

  • Canonical JSON validation against the openEHR ITS-JSON schemas runs on jsonschema 0.59.0, which refuses a value whose regex match the engine cannot finish where 0.58 accepted it under not or patternProperties. The openehr-* crates move to 0.0.113 (#4336).

Fixed

  • The hosted sandbox seeds its full demo dataset again. The v5.5.0 reseed stopped at the ADL 2 archetype library (233 stored / 89 refused against the pinned 231 / 91) because the box’s .env still forced FerroTERM 0.1.6, which reads none of the FerroTERM 0.3 settings the sandbox posture passes, so it served no LOINC and VETDF could only warn on braden_scale’s mistyped LA9605-4. The corpus, the pin and the CDR were right. The seed now asks the CDR for a LOINC code before the library step and names a posture without LOINC (SANDBOX_TERMINOLOGY=without-loinc opts a local run into 233 / 89), checks that the two files it names move together, and a corpus test holds the manifest’s pin to the adjudicated split with and without LOINC. The FerroPIX rollout script retires a stale FERROTERM_IMAGE override, checks that FerroTERM runs the pinned version and serves LOINC, and wipes before it seeds over an unfinished seed (#4278).

  • A response to a request whose body FerroEHR left unread, such as a 404 to a POST of a path it does not serve or a declared-length 413, now carries Connection: close. The connection closed after such a response without saying so, and a client that sent its next request on it met a closed socket, which a non-idempotent request cannot safely retry (RFC 9112 §9.6) (#4283).

  • A [server.caller_rate] above 1 000 000 000 requests per second no longer answers 429 to concurrent requests: such a rate now limits nothing, as its operator meant (#4283).

  • Every request leaves one request line, a timeout 408, a declared-length 413 and the 500 of a panicked handler included, and a 413 for a body with no declared length that passes [server] body_limit_bytes is answered with the openEHR error body, as the declared-length 413 is (#4283).

  • An AQL query whose WHERE clause chains thousands of NOT, AND or OR terms no longer overflows the stack and aborts the server. The parser refuses a query that nests brackets and CONTAINS deeper than 64 levels or builds an expression deeper than 512 levels with 400 Bad Request, before any recursive pass runs, on the ad-hoc, stored-query store and execution paths alike. A long OR over one path is written as MATCHES {…}, which the bound does not count. A statement PostgreSQL refuses for its own complexity limit (SQLSTATE class 54) also answers 400 instead of 500. openehr-query gains ParseError::TooDeep (#4284).

  • An AQL query (ad hoc or stored) whose client disconnects, or that the server’s own query.timeout_ms or HTTP request timeout gives up on, no longer keeps its statement running and its pooled connection busy until the statement ends. The connection is closed instead of returned to the pool, so the slot frees at once, and every pooled session now sets PostgreSQL’s client_connection_check_interval to one second, so the backend stops the abandoned statement within about a second. The composition-count statistics behave the same way (#4285).

  • The server’s resident memory no longer climbs under sustained load until the container’s memory limit kills it. The ferroehr binary now allocates through mimalloc instead of glibc malloc, whose per-thread arenas kept freed memory and never gave it back. On a load test of 10 000 mixed jobs at 48 workers the peak resident set fell from 2.3 to 3.2 GB (growing run after run) to about 600 MB, and throughput rose by about a third (#4288).

  • A request whose database connection a PostgreSQL restart, a crash or the network cuts now answers 503 Service Unavailable with Retry-After, as a request that cannot get a connection already did, instead of 500. SQLSTATE class 08 (except 08P01), 57P01, 57P02, 57P03 and 57P05, and a reset, aborted or closed socket on an open session count as such a cut (#4290).

  • While PostgreSQL is down, requests that need the database answer 503 with Retry-After about a second after the first one starts waiting, and at once after that, instead of each queueing up to db.acquire_timeout_secs (30 seconds by default). An acquire that has waited a second tries one connection of its own; when that cannot reach the database, the outage is known until a fresh attempt connects, and requests are served again as soon as one does. A pool that is only busy keeps queueing as before (#4323).

Security

Shipped with functional changes because: the fix exists only in ferrohealth-server 0.3.7, which carries two functional fixes in the same serving path beside it, so there is no security-only release of the crate for FerroEHR to take.

  • FerroEHR takes ferrohealth-server 0.3.7 in place of 0.3.5 (GHSA-2mc2-h3jq-c5wg). Every earlier release held a plain-HTTP connection that never sent a byte for as long as the client kept it open, so about 8400 silent connections from a client with no credentials exhausted the quickstart’s file descriptors and the CDR stopped answering, health checks included; the listener then retried a failed accept at once and logged every failure. Such a connection now closes once [server] header_read_timeout_ms passes (30 s by default), counted from accept, and after an EMFILE or ENFILE the listener pauses up to 1 s before it accepts again and logs one warning every 10 seconds (#4283).

5.5.0 - 2026-10-27

5.5.0 moves the server and the viewer onto the FerroHEALTH family’s shared configuration, HTTP, telemetry, authorization and audit crates, and it carries breaking changes for every deployment, each with its upgrade step under Changed or Removed. Before you upgrade: name the configuration file with FERROEHR_CONFIG (or --config), since ./ferroehr.toml and /etc/ferroehr/ferroehr.toml are no longer searched; move the REST keys from [server] to [rest] and rename server.bind to server.listen; move [authz.rbac] role_claims and [authz.rbac.basic_roles] to [auth] role_claims and each [[auth.basic.users]] roles, and remove user_role; rename token_url to token_endpoint in the outbound OAuth tables; remove [audit.fhir_feed] batch_size, max_retries and [usage_report] slow_aql_ms; list a reverse proxy in [server] trusted_proxies; and give the viewer server.listen in place of LEPTOS_SITE_ADDR. The server refuses to start without a patient identity source for its EHR index unless [ehr_index] provider = "local" is set, ITI-81 audit retrieval requires a date parameter, and Template Data Document import is gone. Helm chart 20.0.0 carries appVersion 5.5.0. Security carries two advisories, GHSA-vphj-3f8c-5hpx and GHSA-cwhv-87pq-3c3c.

Added

  • The access record names the data permit a secondary-use caller acts under (#3379). A caller declares the permit’s identifier in the header [audit] permit_header names (default x-data-permit) beside its purpose of use; every access record keeps it as permit, so the trail says under which permit a population query, an export or a cohort query ran (EHDS Art. 68). A value that is not 1 to 128 characters of letters, digits and . _ : / # - is recorded as absent. The column joins the tamper chain as digest version 4 (audit migration 0016). The FHIR rendering carries it too (#4120).

  • AQL takes VERSIONED_OBJECT as a FROM class (#3379), the shape of FerroBRIDGE’s OMOP batch query (EHR e CONTAINS VERSIONED_OBJECT vo CONTAINS VERSION v CONTAINS COMPOSITION c). The QUERY specification lets FROM name any RM class (master03 §FROM), and FerroEHR refused this one. The variable answers uid/value and owner_id/id/value (RM common master06 §VERSIONED_OBJECT) and refuses any other path. The operand sets no version scope of its own.

  • The book records the secondary-use contract with FerroBRIDGE (#3379). The new Secondary use page states the batch query FerroBRIDGE loads the OMOP CDM from (time_committed > $since, ordered by the commit instant and the version uid, paged by offset and fetch), the marks that take an EHR out of a batch, the reconciliation that removes erased compositions, and where pseudonymisation for research happens. An integration test holds the batch to it. The compliance overview no longer names the withdrawn secondary-use read model (#3160) as planned.

  • The EHR index asks an IHE patient identity source over PIXm (ITI-83) by default (#3380). [ehr_index] provider = "ihe" names a PIX Manager by its FHIR base (pixm_url) and the clinical domain (assigning_authority, one of privacy.subject_namespaces), with the client credentials grant of [ehr_index.oauth] (renewed once on a 401) and the TLS material of [ehr_index.tls] (trust_roots_file, client_identity_file, min_version). The client is ihe-iti 0.0.62’s PIXm Consumer and sends only what ITI-83 specifies, by POST unless invocation = "get", so any conformant PIX Manager serves. has_ehr_for_subject and create_ehr_for_subject ask the source for the subject’s pseudonym, and the subjects of an EHR are the pseudonym its EHR_STATUS carries; the source is never asked to resolve a pseudonym back. The server reads the source’s CapabilityStatement at boot, and a source that fails a request is an error: nothing falls back to the in-instance map. PDQm is not used (GDPR Art. 5(1)(c)).

  • The book states FerroEHR’s contract as a federation node (#3383). The new page “FerroEHR as a federation node” reads the openEHR Federation Tier with AQL 1.0.0 from the node’s side. It lists the endpoints a gateway uses and resolution by the local ehr_id alone. It gives the four admission conditions of N42a and the configuration that meets them: a unique system_id, and an import under an ehr_id the node chose. It also covers authentication by the deployment’s own authorization server and the refusals the node applies whatever the gateway did. It names FerroFED 0.0.13 as the release whose harness runs FerroEHR as its node. An integration test moves an EHR between two nodes under a node-chosen id and finds every version uid and its creating_system_id unchanged.

  • The FLAT builder builds an INTERVAL_EVENT at an abstract EVENT node when the input names its width (#3898), on the 0.0.102 crate line. RM data_structures §EVENT makes EVENT abstract over POINT_EVENT and INTERVAL_EVENT, and a /width, /math_function or |sample_count under such a node (master05 §INTERVAL_EVENT) now chooses the interval; without them the node still builds as a POINT_EVENT. The flattener writes those rows back for an interval event held at an abstract node.

  • The audit store deletes each record at its own retention end (#3899). Every record stores the first date it may be deleted on: an access record the end ehds-logging states for its categories and origins (Regulation (EU) 2025/327 Annex II 3.4), any other record the day after its date of access plus the [audit.store] horizon, and none for a record kept without end. The reaper deletes a record on or after that date and never before it, where it deleted by one store-wide horizon. Records written before this release keep the horizon. The end joins the chain digest (version 3), so moving one shows in the chain check.

  • The viewer trusts a private CA for the CDR and the OIDC issuer on every platform: [cdr.tls] and [auth.oidc.tls] take trust_roots_file, a PEM bundle trusted beside the platform’s roots, with client_identity (_file) for mutual TLS and min_version, read through ferrohealth_server::tls_client. A file that cannot be read or holds no certificate refuses the start. The Helm chart names the CDR’s CA through cdr.tls.trust_roots_file in place of SSL_CERT_DIR, and viewer.auth.oidc.caExistingSecret mounts an issuer CA; the UI journeys pass the development Keycloak CA the same way (#3909).

  • Opaque access tokens are checked by token introspection under [auth] mode = "introspection" (#4033). ferrohealth-smart’s introspection check (RFC 7662) posts each bearer token to the issuer’s introspection endpoint, named by its metadata or by [auth.introspection] endpoint, and authenticates as the client of [auth.introspection.client] or with the token of [auth.introspection.bearer]. An active response admits the caller with its scopes and roles and is reused until its exp, an inactive token is a 401, and an endpoint that cannot answer is a 503. The mode and the table were refused at boot. Helm chart 15.1.0 names the mode.

  • The ITI-81 audit retrieval applies every search parameter RESTful ATNA §3.81.4.1.2.2 says the repository shall support (#4040): address, agent.identifier, patient.identifier, entity.identifier, entity-type, entity-role, source.identifier, type, subtype and outcome, beside patient, agent, entity and action. Several comma-separated values match any of them, and a repeated parameter must match too. Tokens take the code, |code and system|code forms of FHIR R4 Search §3.1.1.4.10. These parameters were a 400 not-supported. The subject-scoped audit role still names exactly one subject, by patient or patient.identifier.

  • The openehr-* crates carry the APIs FerroBRIDGE kept local copies of (#4047, #4048), on the 0.0.99 crate line. The FLAT builder takes a value whole under |raw (Simplified Formats master04 §Raw canonical JSON) on ROLE.performer (<root>/performer|raw, a PARTY_REF), on the COMPOSITION language, territory and composer, and on the _feeder_audit, _link:i, _participation:i, _other_participation:i, _provider and _health_care_facility families. A ctx/ key still wins over a path spelling of the same attribute, and a raw value of the wrong class, or one beside another suffix, is refused as InvalidRaw. openehr-am implements FromStr for ArchetypeHrid (openehr_adl::hrid::parse_hrid now reads through it), openehr-base adds TerminologyId::new(name, version_id), and openehr-term names the audit_change_type and version_lifecycle_state concepts (codes::AuditChangeType, codes::VersionLifecycleState), held equal to the embedded terminology. Until a ferrohealth-smart release requires 0.0.99, the build holds a registry copy of openehr-sdt 0.0.98 for it beside the workspace 0.0.99.

  • An ADL 2 Web Template names a node under a template overlay by its parent archetype. openehr_adl::opt::create_opt records each overlay’s specialisation parent in the OPT’s other_meta_data (template_overlay_parents), and the ADL 2 builder uses the parent’s id for the node, its paths and its slot admissions, so data and mappings written against the published archetype match. An ADL 2 header now quotes a free metadata value that is not a bare token and reads a quoted one back (#4048).

  • TEMPLATE_ID answers its BASE 1.3.0 HRID functions (#4068), on the 0.0.104 crate line. SPECBASE-42 gives TEMPLATE_ID is_hrid, namespace, qualified_rm_class_name, concept_id, version_id and the major, minor and patch versions (BASE base_types master05 §Template Identifiers). An identifier in the namespaced HRID form of the Artefact Identification reference expression (master03 §Formal Model) yields its parts; a legacy free-text template name stays valid and answers false and empty parts.

  • openehr-its writes an OPT 1.4, on the 0.0.105 crate line: opt14_writer::build and write assemble an operational template from a template definition (the root archetype, the archetypes placed in its slots or open attributes, and the occurrences, existences and attribute constraints it narrows) and the AM 1.4 archetypes it names. Internal references are inlined and deleted nodes removed (OPT2 raw flattening), a slot filler is checked against the slot’s assertions and type, a widening is refused, and the output is byte-deterministic, valid against ITS-XML Template.xsd and accepted by the template upload (#4083).

  • The server runs the FerroHEALTH HTTP stack of ferrohealth-server 0.3.4, and [server] is its section, the same keys in every FerroHEALTH product: listen, base_path (a path every route is served under, the health routes included; empty by default), request_timeout_ms (a request still running is answered 408), header_read_timeout_ms, body_limit_bytes, drain_delay_ms and shutdown_timeout_ms (on SIGTERM readiness turns 503 at once, the listener closes after the delay, and the requests in flight get the bound to finish), max_concurrent_requests with retry_after_s, trusted_proxies with forwarded_header, and compression with compression_min_bytes. [server.caller_rate] (requests_per_second, burst) limits each authenticated caller. [server.tls] (certificate_file, key_file, client_auth, client_ca_file, min_version, reload_interval_s, healthcheck_identity_file) reads the files again on SIGHUP too, and the management listener on its own port drains with the main one under the same bound (#4087).

  • [rest] holds the openEHR REST surface: base_path, swagger_ui, cors_permissive, system_id, [rest.identity] and body_limit_bytes, the clinical body tier below server.body_limit_bytes, which the template-upload and /message routes keep (#4087).

  • ferroehr serve names the default command beside config check and healthcheck (#4087).

  • The viewer serves the family’s health routes and commands: /health/liveness, /health/readiness and /health, ferroehr-viewer config check (reads the configuration and builds the CDR and issuer clients without reaching either), --config and --set key=value, the [server] table with listener TLS, and the bounded drain on SIGTERM (server.drain_delay_ms, server.shutdown_timeout_ms). ferroehr-viewer with no subcommand still serves (#4139).

  • The OPT 1.4 writer clones nodes and narrows types, on the 0.0.106 crate line: opt14_writer::Narrowing::Clone copies an archetype node that may occur more than once under its own node id, told apart by a name constraint and carrying its own narrowings and slot fills, and refuses a clone of a node that occurs at most once, a name another copy carries, or copies the attribute’s cardinality cannot hold (AOM2 VSONCO, VACMCU, VACMCO). Narrowing::Redefine narrows an attribute’s children to conforming subtypes (a DV_TEXT value to a DV_CODED_TEXT) or drops alternatives, keeping the archetype’s other constraints, and refuses a type that does not conform (AOM2 VSONCT). An attribute the template constrains can also hold archetypes placed by a slot fill. The writer now rebuilds 143 of the 305 CKM operational templates to their upstream node tree, up from 67 (#4150).

  • The OPT 1.4 writer refuses a redefined leaf constraint that admits a value the archetype’s does not, as AOM2 Phase 2 validation (VPOV) requires: a string list or pattern, an integer or real list or range, a boolean, a date, time, date/time or duration range, pattern or timezone validity, a code phrase’s terminology or code list, a quantity’s property, units, magnitude or precision, and an ordinal list. Each refusal is a WriteError::Widening naming the archetype, the path and the constraint. The openehr crates move to 0.0.107 (#4163).

  • The Helm chart mounts a private issuer CA for the server: auth.caExistingSecret (with auth.caExistingSecretKey, ca.crt by default) mounts the Secret’s PEM bundle and points the server’s [auth] ca_file at it, as viewer.auth.oidc.caExistingSecret does for the viewer. Chart 19.3.0 (#4235).

  • The quickstart viewer signs in through the Keycloak overlay: with --profile viewer, the viewer at http://localhost:3000 takes the authorization-code flow with PKCE through the demo realm’s ferroehr-viewer client and calls the server with the user’s token. The issuer moves to https://keycloak.localhost:8443, a name the containers resolve through a network alias and a browser resolves to the loopback address with no hosts entry, and the development certificate names it (#4243).

Changed

  • A server with no patient identity source refuses to start (#3380). The binary’s default [ehr_index] provider is the IHE one, and without pixm_url and assigning_authority the refusal names both choices: configure the identity source, or opt in to the in-instance map with [ehr_index] provider = "local", which serves the GET /ehr?subject_id=… lookup over EHR_STATUS content and nothing more. Upgrade step: a deployment without an identity source adds [ehr_index] provider = "local" (FERROEHR__EHR_INDEX__PROVIDER=local).

  • The compose quickstart runs FerroPIX beside FerroEHR (#3380), pinned by tag and digest, with its own PostgreSQL and a development OpenID Provider, and points [ehr_index] at it; FerroPIX mints the pseudonym in urn:oid:2.999.7.1 on FerroEHR’s first ask. The server now runs in the issuer’s network namespace, which publishes ports 8080, 8090 (FERROEHR_PIX_PORT) and 8180 (FERROEHR_ISSUER_PORT).

  • Helm chart 19.0.0 takes FerroPIX as the chart dependency ferropix (#3380), on by default: the chart points config.ehr_index at its Service, adds the egress rule to its pods, and routes the grant’s client secret through secrets.ehrIndexClientSecret. A values file that names no identity source is refused at render. Upgrade step: supply FerroPIX’s ferropix.database.existingSecret, ferropix.dataKey.existingSecret, its issuer and pseudonymisation domain, and config.ehr_index, or set ferropix.enabled: false with config.ehr_index.provider: local.

  • The EHDS system documents live in the FerroHEALTH book, and each release attaches them at a pinned revision (#3705). The EHR system Cadasto B.V. declares is FerroEHR with FerroBRIDGE, so its information sheet (EHDS Art. 38), instructions for use (Art. 30(1)(d)), readiness record and technical documentation (Art. 37, Annex III) move to https://ferrohealth.eu/docs/ehds/, and the old addresses in this book redirect there. docs/ehds/system-documents.env pins the FerroHEALTH commit and the revision of each page. The release lane fetches the two attached pages at that commit, refuses a page whose Revised line differs from the pin, and appends this release’s facts under “This copy”: the system version (this FerroEHR release with the FerroBRIDGE release its chart runs beside), the release date and the versions of the standards, read from docs/VERSIONS.md (Art. 38(2)(b) and (e)). The release notes name the pinned revisions, and release-notes.sh refuses notes that do not.

  • Breaking: the XML readers of openehr-its refuse a document outside the schema (#3893), on the 0.0.103 crate line. A child element a type’s closed xs:sequence does not declare, or a child element inside a text element, is refused with XmlError::UndeclaredChild (XML Schema 1.1 Part 1 §3.4.4.2) where it was skipped as if absent, and opt14::from_xml and the AOM2 readers refuse a root other than the one the schema publishes (<template>, ITS-XML components/AM/Release-1.4/Template.xsd) with XmlError::UnexpectedRoot. One tool-written element no schema declares, match_negated on a C_ATTRIBUTE, is read when false and refused when true (XmlError::UnsupportedExtension): a negated match changes the constraint, and no schema says how. A CDR upload of a template outside the schema is now refused; the CNF Robot valid_templates/validation/clinical_content_validation.opt and two epSOS CKM templates are such documents (#4077, #4078).

  • [auth] ca_file is trusted under introspection mode too: with ferrohealth-smart 0.8.6 the server reads an introspecting issuer’s metadata, and calls its introspection and token endpoints, trusting the bundle beside the platform roots and the grant’s own ca_file. The e2e compose overlay names the development Keycloak CA through FERROEHR__AUTH__CA_FILE in place of SSL_CERT_FILE and SSL_CERT_DIR (#3909).

  • The shared crates move to their newest lockstep releases: ferrohealth-smart 0.7.3, ferrohealth-authz 0.1.3, ihe-atna and ehds-logging 0.2.12 and ferrobridge-term 0.1.138. Every secret in the [auth] section now renders as ***, the family mask, and [auth] ca_file names a PEM bundle of extra certificate authorities for the issuer’s metadata and key set, for an issuer behind a private PKI (#3909).

  • Breaking: the audit trail believes a forwarded client address only from a trusted proxy (#4025). The source network address of an audit record (client_ip, the DICOM NetworkAccessPointID, the FHIR AuditEvent network) is the address the family server stack names: the connection’s peer, or the client a peer listed in [server] trusted_proxies names in the [server] forwarded_header (RFC 7239 §8.1). FerroEHR read the first X-Forwarded-For hop from any caller, so a client could write any address into the trail. A deployment behind a reverse proxy lists the proxy in trusted_proxies and sets forwarded_header = "x-forwarded-for" when the proxy does not send Forwarded; until then its records name the proxy.

  • FerroEHR moves to ferrohealth-usage-report 0.2.0 (#4025). The report FerroEHR sends is unchanged.

  • The CDR takes up ferrohealth-smart 0.6.0 and ferrobridge-term 0.1.134 (#4026, #4027). Token checking, Basic users, the discovery document and every outbound client credentials token behave as before. The terminology client now refuses, before sending, a code with no system on an STU3 server; the CDR’s server is R4B, so its binding checks are unchanged. The new [auth] mode = "introspection" key of the shared section, and an [auth.introspection] table, are refused at boot: this server checks bearer tokens as JWTs only, and a setting it would ignore stops the start.

  • Every secret and every masked URL renders through ferrohealth-secret, the family’s one mask (en-route finding of #4028). EHR’s own config::secret module gives way to ferrohealth-secret’s Secret, SecretUrl and RedactedUrl. A secret still renders as *** and a URL’s userinfo as ***@, and a URL’s query is now masked as ?*** too, in ferroehr config check, GET /admin/config, /management/env and ferroehr report. ferrohealth-smart moves to 0.7.1 and ferrohealth-authz to 0.1.1, ihe-atna and ehds-logging to 0.2.11, and ferrobridge-term to 0.1.137.

  • Breaking: the outbound client credentials tables take the ferrohealth-smart client table, so token_url is now token_endpoint (#4028). [terminology.server.oauth], [audit.fhir_feed.oauth] and [authz.abac.remote.oauth] are the family’s shared grant: the old key token_url stops startup by name, and each table gains private_key_jwt (RFC 7523), mutual TLS (tls_client_auth and self_signed_tls_client_auth, RFC 8705), ca_file, refresh_margin_ms and timeout_ms. issuer, client_id, the client secret, scope, audience and resource keep their names. Upgrade: rename token_url to token_endpoint in each table (in the Helm chart, under config.terminology.server.oauth and the other two), or ..._OAUTH__TOKEN_URL to ..._OAUTH__TOKEN_ENDPOINT in the environment. ferrohealth-smart moves to 0.7.0 and ihe-atna and ehds-logging to 0.2.10, the FHIR Feed sends its bearer token through ihe-atna, and every BALP AuditEvent’s source.observer carries an identifier beside its display.

  • Breaking: ihe-atna’s forwarder delivers the audit trail, and [audit.fhir_feed] batch_size and max_retries are gone (#4029). Syslog frames are stored in ihe-atna’s spool and delivered from it with backoff, retried until the repository takes them, so a failed send is no longer lost. The local audit store keeps only the FHIR document, so the spool, not the store, is syslog’s outbox, and delivered_syslog_at is no longer written. New [audit.syslog] keys: spool_dir (a spool that survives a restart; unset, it is held in memory), spool_max_messages, spool_max_bytes and retry_max_ms. The FHIR Feed delivers from the store through an ihe-atna outbox, or with the store off from an in-memory spool. A record the repository refuses is set aside until the next start, and the records behind it flow on. retry_max_ms takes the place of batch_size and max_retries, which stop startup by name. A new audit_delivery health indicator reports each forwarder. Upgrade: remove batch_size and max_retries from [audit.fhir_feed], and set [audit.syslog] spool_dir to a directory only the server can read.

  • Breaking: the ITI-81 audit retrieval (GET /fhir/r4/AuditEvent) requires a date parameter (#4030). The query is now read by ihe-atna’s retrieve, the family’s audit crate, and the answer written by it: RESTful ATNA §3.81.4.1.2.1 has at least one date present, so a query without one is a 400 OperationOutcome. A date value takes every FHIR prefix (eq when none, so date=2026-07-12 is that whole day, read in UTC), patient, agent and entity also answer to their .identifier forms, _count=0 returns the total alone, and the Bundle carries self and next links. A parameter the local store cannot apply yet is a 400 not-supported where it was silently ignored. Upgrade: an API client, a patient portal on the subject-scoped audit role included, adds a date window to every retrieval (for example date=ge2026-07-01); the viewer’s audit page sends the last 30 days when no window is set. ihe-atna and ehds-logging move to 0.2.8.

  • Breaking: [authz.rbac] user_role is retired (#4038 follow-up). It named a role nothing required once a Basic user’s roles became its [[auth.basic.users]] roles list. The key now stops startup by name and points at that list. Upgrade: remove user_role (on Helm chart 16.0.0, config.authz.rbac.user_role), and give each Basic user the roles it needs, for example roles = ["USER"].

  • Breaking: the caller’s roles come from ferrohealth-smart, so the two role keys move to [auth], and the policy decision is ferrohealth-authz 0.1.0 (#4038). [authz.rbac] role_claims is now [auth] role_claims (same default, realm_access.roles included), and [authz.rbac.basic_roles] is now the roles list of each [[auth.basic.users]] entry; both old places stop startup by name. Roles stay upper-cased, and a Basic user whose entry lists no role holds none, where a user absent from basic_roles used to hold user_role. [authz.assurance] and [authz.professional] keep their keys and are read by the shared sections; the RFC 9470 step-up challenge on patient data no longer carries an error_description. The Cedar engine and the remote decision point are the family’s: the Cedar schema and the remote wire body are unchanged, a remote resource kind with no policy is a 500 instead of a 403 (boot validation already refuses that configuration), and authz_cedar_decisions and authz_remote_pdp_calls count one per decision where they counted one per attribute combination. Upgrade: move role_claims from [authz.rbac] to [auth] (FERROEHR__AUTH__ROLE_CLAIMS in the environment), and each user’s entry of [authz.rbac.basic_roles] to roles under its [[auth.basic.users]] table (in the Helm chart, under config.auth.basic.users), and give a user that had no entry roles = ["USER"]. Helm chart 15.0.0 takes the same keys (config.auth.role_claims, roles under config.auth.basic.users, and token_endpoint in each grant). ferrobridge-term moves to 0.1.136, whose SmartToken replaces the terminology client’s own token adapter (#4031).

  • The openehr-rm tests take the change-type and lifecycle-state codes from openehr_term::codes (#4053), in the codegen templates of both generations, so no literal code or rubric can drift from the embedded terminology.

  • SECURITY.md records the main ruleset as it stands (#4067): no approving review, the conclusion check, the merge queue, and the organization-admin bypass, with what a bypass merge skips (any lane still running, ui-e2e included) and when it is used. The weekly external link check leaves out gesetze-im-internet.de and dicom.nema.org, which refuse the GitHub runner every week while answering by hand (#4074).

  • The BASE, AM, TERM and LANG specifications are re-vendored (#4068, #4069, #4070, #4071) at BASE 93abaf9f, AM 51aa07b2, TERM b2dd077d and LANG e61a7192. The TERM languages code set restores pt-pt, adds ro-md and ru-md, and names the replacement of each deprecated code (SPECTERM-36), so the openehr-term code-set lookup accepts the two new codes. The BASE Interval.has post-condition gains its brackets, which the implementation already followed. AM’s Archetype Identification document is superseded by BASE Artefact Identification (SPECAM-103), and the citations move with it. The nine openehr-* crates step to 0.0.104.

  • The bundled PostgreSQL image is rebuilt on the 2026-10-21 respin of postgres:18.6 (#4084, sha256:edb0c08d…), a same-version respin with rebuilt base packages. The respin ships the fixed xz packages, so the TEMP-1147318-639065 statement leaves security/vex/postgres-os.openvex.json.

  • Breaking: the server reads only the configuration file it is told to read (#4087). The loader is ferrohealth-config 0.2.5, the FerroHEALTH products’ shared one, and it reads the file --config or FERROEHR_CONFIG names. ./ferroehr.toml and /etc/ferroehr/ferroehr.toml are no longer searched: a deployment that mounts a file there without naming it starts on the environment and the defaults alone. Set FERROEHR_CONFIG=/etc/ferroehr/ferroehr.toml; the Compose files, the conformance stacks and Helm chart 16.0.2 set it. The fallbacks keep their order (the libpq variables, DATABASE_URL, DATABASE_URL_UNPOOLED, RUST_LOG, PORT), and ferroehr config check prints each key beside the source that set it.

  • Breaking: the viewer reads only the configuration file FERROEHR_VIEWER_CONFIG names (#4087), through the same ferrohealth-config loader as the CDR. ./ferroehr-viewer.toml and /etc/ferroehr/viewer.toml are no longer searched. A variable under FERROEHR_VIEWER_ that is neither an override nor the Compose port or image parameter refuses the start, and a *_file secret loses only its trailing line ending, where all surrounding whitespace was trimmed; an empty file is refused.

  • Breaking: [auth] role_claims defaults to ["roles", "groups", "entitlements"] (#4087), the RFC 9068 §2.2.3.1 carriers every FerroHEALTH product reads. realm_access.roles is no longer in the default: a Keycloak issuer sets FERROEHR__AUTH__ROLE_CLAIMS=roles,groups,entitlements,realm_access.roles, as the Keycloak Compose overlay and the viewer’s journey stack now do.

  • Breaking: the server no longer refuses a FERROEHR_ variable outside the override form, and a file refusal says less (#4087). Only a FERROEHR__<SECTION>__<KEY> override that names no key refuses the start. A single-_ spelling such as FERROEHR_DB_MAX_CONNECTIONS, or a per-subsystem pointer such as FERROEHR_SIGNING_CONFIG, sets nothing and is left alone, because the viewer’s FERROEHR_VIEWER_ variables share the prefix (ferrohealth-config 0.2.5). A refusal of an unknown key in the file names the key and the closest valid one, without its line number; a syntax error still names its line and column.

  • Breaking: /health/readiness and /management/health serve the FerroHEALTH readiness document of ferrohealth-health 0.1.1 (#4087). The body is {"state", "phase", "indicators": [{"name", "state", "detail"?}]} with lowercase states (up, degraded, down), replacing {"status", "components"}; the public probe leaves out every detail. Readiness also answers 503 while the process boots and from the moment it receives SIGTERM or SIGINT, before the listener stops. The optional indicators (audit_sender, audit_delivery, audit_chain, events) read at worst degraded, so a damaged audit chain now reads degraded where it read DOWN, and an indicator that is up carries no detail. The viewer’s health tab reads the same document. Probes that test only the status code are unaffected.

  • Breaking: logging, the request log and the OpenTelemetry export move to the [telemetry] section of ferrohealth-telemetry 0.2.1 (#4087), the FerroHEALTH products’ shared one, with no aliases for the old keys. [log] is gone: log.format and log.filter become telemetry.format and telemetry.filter (FERROEHR__TELEMETRY__FORMAT, FERROEHR__TELEMETRY__FILTER, RUST_LOG still a fallback), and the filter now defaults to info,hyper=warn,tower=warn,h2=warn and refuses the start when it does not parse. The export moves to a [telemetry.otlp] table: otlp_endpoint becomes otlp.endpoint, traces_sample_ratio becomes otlp.sample_ratio, and metrics_push becomes otlp.metrics, which defaults to true, so a configured endpoint now receives the metrics as well as the spans. telemetry.flame_file moves to management.flame_file. telemetry.service_name is removed: the service.name is always ferroehr. telemetry.environment no longer defaults to dev and is left off every span and metric unless set. A file or variable that still spells an old key refuses the start, naming it. Helm chart 17.0.0 renames config.log to config.telemetry.

  • Breaking: the HTTP metrics are the ferrohealth-telemetry request log’s (#4087). http_server_request_duration_seconds is labelled http_request_method, http_route, http_response_status_code and url_scheme (OpenTelemetry HTTP semantic conventions); the status_class label is gone, so a query on status_class="5xx" becomes http_response_status_code=~"5..", as the bundled alert rules and Grafana dashboard now read. The http_server_active_requests, http_server_request_body_size and http_server_response_body_size series are removed, and no series on /management/prometheus carries the otel_scope_name label any more. The viewer’s management page loses its in-flight requests tile, and the dashboard’s in-flight panel shows the request rate.

  • Breaking: responses no longer carry x-trace-id, and a client’s traceparent is no longer read (#4087). Every request starts a trace of the server’s own, so a caller cannot join or steer the server’s traces; correlate a request through its x-request-id, which the request log writes on every line.

  • Breaking: the REST-surface keys move from [server] to [rest]: server.base_path, server.swagger_ui, server.cors_permissive, server.system_id and [server.identity] are rest.base_path, rest.swagger_ui, rest.cors_permissive, rest.system_id and [rest.identity] (FERROEHR__REST__…), with no alias; the old names are refused at startup. Upgrade: rename the keys, and move a [server.identity] table to [rest.identity] (#4087).

  • Breaking: server.bind is server.listen, an IP address and a port; a host name such as localhost:8080 refuses the start. It listens on 0.0.0.0:8080 by default, as before. Upgrade: rename the key (#4087).

  • Breaking: the request-body limit is server.body_limit_bytes, 64 MiB by default, the ceiling of every route and the limit of the template-upload and /message routes, and rest.body_limit_bytes (unset by default) narrows the other routes below it, where it was 16 MiB. The chart sets 16 MiB. Upgrade: set rest.body_limit_bytes = 16777216 to keep the old clinical limit (#4087).

  • Breaking: server.max_in_flight is server.max_concurrent_requests, unset (no bound) by default where it was 256, and the bound now covers every FerroEHR route (the status document, SMART discovery and the management surface included); only the health routes are never refused. Upgrade: set max_concurrent_requests = 256 to keep the old bound (#4087).

  • Breaking: [server.rate_limit] is [server.caller_rate] (requests_per_second, burst) keyed on the authenticated subject, and limits nobody until the table is set, where the principal tier was on by default; the per-address tier is gone. The chart and the Compose quickstart set 1024 per second with a burst of 2048. A 429 carries Retry-After and no x-ratelimit-* headers (#4087).

  • Breaking: [server.tls] takes the family keys: the table’s presence turns TLS on (enabled is gone), cert_file is certificate_file, reload_interval_secs is reload_interval_s (unset by default, where the files were read every 60 seconds; SIGHUP always reads them), and client_auth = "off" is "none". The chart’s config.server.tls follows, and an empty table means TLS off (#4087).

  • Breaking: [server.connection] is [server] header_read_timeout_ms (30000 by default, where the bound was 10 seconds, and 0 is refused); max_concurrent_streams and the HTTP/2 keep-alive keys are gone (#4087).

  • Breaking: ferroehr healthcheck asks GET /health/readiness on the server.listen port (over TLS with [server.tls] set) and exits 1 while the server drains, where it asked {rest root}/status; the reason on stderr reads healthcheck: unhealthy: …. GET /health answers the readiness status with {"state": …} and GET /health/liveness answers {"state":"up"}, where both answered a plain-text OK (#4087).

  • The request id is the server’s own, minted for every request and carried by the request log and the access record; a client’s X-Request-Id is returned to it unchanged and no longer recorded. A request past server.request_timeout_ms or over the body limit, a refused caller and a panicked handler answer the openEHR { error, message } body, and the panic message reaches neither the body nor the log. Responses are compressed at the default gzip and Brotli quality, from 1 KiB up (#4087).

  • The boot banner is the FerroHEALTH banner: the wordmark, the build line ferroehr <version> (<commit>), and the facts (manufacturer, profile, pins, support period, deployment) as one list, with the sandbox notice below (#4087).

  • Chart 19.0.0: config.server carries max_concurrent_requests and caller_rate, and config.rest the REST surface; the rendered ferroehr.toml changes accordingly (#4087).

  • Breaking: the usage report is sent by ferrohealth-usage-report 0.1.2 (#4087), the FerroHEALTH products’ shared reporter, on the same envelope v2. The status object loses spec_profile, and postgres_major is left out when the database does not answer. The daily metrics lose the aql group and the aql_executions and aql_slow counters, and gain the responses_4xx counter. Routes outside the API families (the health family, the management surface, the status and discovery documents) are counted in no group. A replica adds its counts to the shared window at its own daily attempt instead of every five minutes, and a daily report that does not reach the collector keeps its counts for the retry an hour later where it dropped them. A failed or refused send is logged at WARN where it was DEBUG, and the boot line takes the family’s wording. The clinical migration 0018_usage_report_envelope drops the slow column and the aql rows and adds usage_report_counter.

  • HTTP Basic credentials are cached for 300 seconds by default (#4088). [auth.basic] cache_ttl_seconds defaulted to 0, so every Basic request ran the Argon2id derivation, 10 to 15 ms of CPU, which capped the mixed ITS-REST and AQL workload against a release build on one host at 408 op/s; with the cache it reached 1559 op/s. ferrohealth-smart 0.7.5 makes 300 the default. Only a credential that verified is kept, under a keyed hash drawn at boot, and the cache empties at a restart. Set 0 to check every request. FerroEHR takes ferrohealth-smart 0.7.5, ferrohealth-authz 0.1.5, ihe-atna and ehds-logging 0.2.13 and ferrobridge-term 0.1.139. Helm chart 16.0.1 states the new default.

  • Three AQL shapes no longer read every composition (#4089, #4090), measured against a release build on one host holding about 51 000 compositions. A population query ordered by c/context/start_time/value with a small fetch reads the compositions in start-time order from a new index and stops at the limit: 169 ms became 7 ms at p50. A c/archetype_details/template_id/value predicate on a COMPOSITION reads the indexed template id of the version: 142 ms became 2 ms for a template nothing matches, and 29 ms became 11 ms for one of thirteen templates. A COMPOSITION, EHR_STATUS or EHR_ACCESS root is bound at its root row, so a containment with OR probes one row per version where it read every node of it. Clinical migration 0017 adds the index and the planner statistics, and analyzes the node table once.

  • The scheduled audit chain check verifies the records written since the last intact link (#4100). audit.verify_audit_chain_incremental() recomputes only the records after the checkpoint it keeps in audit.audit_chain_checkpoint, checks the chain head, and moves the checkpoint when it finds nothing: on 428,000 records a full pass took about 35 s, a run over the last 1,000 records 76 ms and one with nothing new 2 ms. A full pass runs every [audit.store] full_verify_interval_seconds (new, default one week, 0 for on demand only) and after any run that found damage, and SELECT * FROM audit.verify_audit_chain() stays the full pass on demand. With the cheaper run, verify_interval_seconds defaults to one hour, where it was one day. The chain digest is unchanged.

  • The FHIR Feed owes only the records written while it is configured (#4102). A record waiting for delivery is a row in audit.fhir_feed_outbox, which the store writes with the record while [audit.fhir_feed] is on and delivery removes. The two partial indexes over the unset delivery stamps, which held every record of a deployment without a sink (about 4.8 MB each at 428,000 records), are dropped; the syslog one served no query, since syslog delivers from its own spool. Records stored before this release and never delivered ship once a feed first runs.

  • The FHIR AuditEvent carries the data permit (#4120). The stored record, the ITI-81 retrieval and the ATX: FHIR Feed write the permit a caller declared in [audit] permit_header at agent.policy of the user agent, after the token’s jti (FHIR R4 AuditEvent.agent.policy), on an access record and on every other record. The DICOM audit message over syslog has no element for a policy (DICOM PS3.15 A.5.1.1), so it carries none.

  • The per-EHR access record of an Extract export or import carries the declared purpose and data permit (#4120), taken from the request as the other service-layer access records take them, so the subject-scoped ITI-81 retrieval answers an export of GET /message/export/{ehr_id} with its permit.

  • FerroEHR’s own [server] defaults again (#4120): with no listen and no body_limit_bytes in the file or the environment, the server listens on 0.0.0.0:8080 and reads bodies up to 64 MiB, and a partial [server] table keeps both for the keys it does not name (ferrohealth-config 0.2.6 Loader::defaults). The container image no longer sets PORT=8080, and the quickstart Compose file, the dev configuration and chart 19.0.1 no longer write the two keys. Upgrade: a bare binary that must stay on the loopback sets listen = "127.0.0.1:8080".

  • The shared crates move to their newest releases (#4120): ferrohealth-smart 0.8.3, ferrohealth-authz 0.2.2, ihe-atna and ehds-logging 0.3.1, ihe-iti 0.0.65, ferrobridge-term 0.1.140, ferrohealth-server 0.3.5 and ferrohealth-config 0.2.6.

  • A pull request records its changelog entry as a fragment under changelog.d/ and no longer edits CHANGELOG.md (#4123). The fragment is changelog.d/<issue>-<kebab-slug>.<section>.md, in the format changelog.d/README.md gives, and the release cut assembles the fragments into the new version’s section with scripts/release/changelog.sh --assemble. The changelog-guard job refuses a malformed fragment and a pull request that writes under [Unreleased], so two open pull requests no longer conflict on the changelog.

  • The chart pins the FerroPIX release beside its digest. ferropix.image.tag (0.0.7) states the release whose image ferropix.image.digest pins, and scripts/checks/chart-appversion.sh now holds the tag in the chart’s Artifact Hub images entry to it, as it holds the digest. The dependency chart still deploys the digest, so the rendered manifests do not change (#4131).

  • Breaking: the viewer listens on server.listen, not LEPTOS_SITE_ADDR. The viewer reads the family’s [server] table from ferrohealth-server 0.3.5, defaulting to 127.0.0.1:3000 (the image sets 0.0.0.0:3000 through FERROEHR_VIEWER__SERVER__LISTEN). A LEPTOS_SITE_ADDR that names another address refuses the start and names the key that replaced it; unset it. server.base_path and server.caller_rate are refused, since the viewer’s pages are served at / and it mounts no per-caller limit (#4139).

  • Breaking: ferroehr-viewer healthcheck asks /health/readiness over the loopback (over TLS when server.tls is set) and exits 0 only on 200, where it asked /login at LEPTOS_SITE_ADDR and took any 2xx or 3xx. A draining viewer is now unhealthy. The Helm chart’s viewer probes ask /health/liveness and /health/readiness, where all three asked / (#4139).

  • Breaking: the viewer’s browser security headers are the family console’s. Referrer-Policy is no-referrer (was strict-origin-when-cross-origin), Cross-Origin-Opener-Policy: same-origin is new, and the content-hashed bundle under /pkg/ is Cache-Control: no-cache, revalidated on each load, where it was public, max-age=31536000, immutable. Every document stays no-store under the nonce Content-Security-Policy (#4139).

  • Breaking: every request to the viewer that is not a safe method passes the family same-origin check, server function calls and the result export included, where only page form posts compared Origin with Host. The browser’s Sec-Fetch-Site: same-origin passes; without it Origin or Referer must name auth.oidc.public_base_url, and a request that shows neither is refused 403. A post without a session now returns to the page it was posted to after the sign-in, rather than to the page its Referer named (#4139).

  • Breaking: the viewer’s CDR and OIDC clients speak TLS 1.3 at least by default, as every family client does; cdr.tls.min_version and auth.oidc.tls.min_version set to "1.2" lower it. A request now runs under the stack’s server.request_timeout_ms (60 s) and server.body_limit_bytes (16 MiB), and a form post’s ceiling is that limit (#4139).

  • The viewer hydrates the whole page on FerroUI 0.5: the server draws the first page and the browser draws every page after it, so a link or a filter no longer reloads the document, and the browser’s back and forward walk the pages it drew. The pages load their code in six groups as they are first opened. Every page still works with no WebAssembly, and a post still answers with the page it was posted to (#4147).

  • The book and the architecture no longer claim the openEHR Service Model as the design of the service layer (#4211). The service layer is FerroEHR’s own design, judged by what goes in and comes out (BASE, RM, AM, TERM, QUERY/AQL, ITS-REST, ITS-JSON, ITS-XML and the serialisation formats); the vendored SM text stays as reference only. No behaviour changes.

  • Breaking: the viewer logs through the family’s [telemetry] section of ferrohealth-telemetry 0.2.2, as the CDR does (#4214). The console writes to stdout where it wrote to stderr, and telemetry.format defaults to auto, which renders json when stdout is not a terminal, so a container’s log is now JSON lines; set FERROEHR_VIEWER__TELEMETRY__FORMAT=pretty for the old human lines. RUST_LOG is now the lower-priority fallback of telemetry.filter (FERROEHR_VIEWER__TELEMETRY__FILTER wins over it), the filter defaults to info,hyper=warn,tower=warn,h2=warn where it was info, and a filter that does not parse refuses the start where it fell back to info. Each request writes the CDR’s request line (method, matched route, status, latency, request id), and a panic writes the CDR’s one fixed line without the panic message. [telemetry.otlp] exports the viewer’s spans and its http.server.request.duration metric as service.name ferroehr-viewer. Helm chart 19.2.0 sets the viewer’s format and filter from viewer.telemetry, and the Compose viewer services set FERROEHR_VIEWER__TELEMETRY__FORMAT.

  • The viewer takes FerroUI 0.5.4: its skip link keeps the browser’s own fragment navigation under hydration, so the viewer no longer patches the link after the page hydrates, and a code editor’s key costs the edit, not the length of the text (#4227).

  • The quickstart Keycloak overlay names its development CA through [auth] ca_file: docker-compose.keycloak.yml sets FERROEHR__AUTH__CA_FILE in place of SSL_CERT_FILE and SSL_CERT_DIR, which only the Linux platform verifier reads (#4236).

  • The development and quickstart Keycloak moves to 26.8.1, and its realms form a complete OIDC example: both compose files pin quay.io/keycloak/keycloak:26.8.1 by digest, held together by scripts/checks/compose-image-tags.sh and watched weekly by the base-image watcher. In both realms the authorization-code flow requires PKCE (S256) and accepts only the viewer origins on port 3000 as redirect targets (the quickstart realm accepted any), and a service account holding USER lets a client-credentials token reach the clinical API (#4240).

  • The hosted sandbox (sandbox.ferroehr.eu) resolves patients through FerroPIX 0.0.10, the FerroHEALTH patient identity source, as the compose quickstart does: FerroPIX runs beside the CDR on loopback with its database on the database box, each EHR_STATUS names its patient by the pseudonym FerroPIX minted in the clinical domain urn:oid:2.999.7.1, and POST /query/cohort takes subjects in the synthetic hospital domain urn:oid:2.999.1.1. The nightly reseed registers the demo patients with FerroPIX through the deploy script’s new identities verb before it seeds, so after a reset the CDR’s EHRs and FerroPIX’s identities agree. FerroTERM moves from 0.1.6 to 0.3.1 on the same index layout, the viewer’s memory limit from 1536m to 1024m, and deploy/hosted/sandbox-ferropix-rollout.sh carries the operator steps on both boxes (#4242).

  • Breaking: the quickstart Keycloak realm gives the viewer and a backend service their own clients: ferroehr-viewer (ferroehr-viewer-quickstart-secret) takes the authorization-code flow with PKCE and the password grant, and ferroehr-service (ferroehr-service-quickstart-secret) takes only the client-credentials grant, with the system/ SMART on openEHR scopes and a client_id claim. Both put ferroehr in aud. The single ferroehr client and its ferroehr-quickstart-secret are gone, so a script that fetched a token with them moves to ferroehr-viewer. The overlay names ferroehr-service in [smart] backend_clients, so its system/aql-* grant runs AQL. The dev realm makes the same split and drops the unused Keycloak Authorization Services, and scripts/ui-e2e.sh keeps its scratch files under target/ui-e2e/ (#4245).

  • The Helm chart’s usageReport.enabled reaches the other products it runs: FerroTERM (FERROTERM__USAGE_REPORT__ENABLED) and a FerroBRIDGE release from 0.0.7 on ([usage_report]) follow it, each overridable with terminology.usageReport.enabled or bridge.usageReport.enabled. The FerroPIX dependency chart reads only ferropix.usageReport.enabled, so the chart now refuses to render usageReport.enabled=false while that stays true: a release you switched off sent FerroPIX’s report regardless. Upgrading a release that sets usageReport.enabled=false with FerroPIX on takes --set ferropix.usageReport.enabled=false as well. Chart 20.0.0 (#4270).

Removed

  • FerroEHR holds no identity logic (#3380): link_as_subject, the server-side subject pseudonym (mint_subject_pseudonym), the party merge and split, and resolve_ehr_for_identity, the national-identifier-to-EHR crossing, are gone. Matching, cross-referencing identifier domains and minting pseudonyms belong to the identity source.

  • Breaking: ferroehr config default is gone with no alias; the annotated template is app/ferroehr/assets/ferroehr.default.toml in the repository. ferroehr healthcheck --url and FERROEHR_HEALTHCHECK_URL are gone too: the probe follows server.listen and server.base_path. Upgrade: copy the template from the repository, and probe another URL with an HTTP client of your own (#4087).

  • Breaking: [usage_report] slow_aql_ms is removed (#4087), with the AQL series it set the threshold of. A file or variable that still sets it refuses the start; remove the key. state_dir, the shared section’s directory for the instance id, is refused: FerroEHR keeps the instance id and the counts in the clinical database. The endpoint is now refused when the configuration is read, with the same rule (https, or http to a loopback host, never with credentials).

  • Breaking: Template Data Document (TDD) import is removed (#4199). POST /message/tdd/{ehr_id} and POST /message/tdd/{ehr_id}/batch are gone and answer 404, with the service behind them, the message-tdd extension and the Tds capability in the conformance statement. TDD and the Template Data Schema (TDS) come from the template tooling of the ADL 1.4 era and are no current openEHR format; send the composition as canonical JSON or XML, or as FLAT or STRUCTURED against its Web Template. The /message group keeps the four EHR Extract routes, and server.body_limit_bytes now covers template upload and /message/import alone.

  • Breaking: openehr-sdt drops the flat::tdd module (#4199): from_tdd, from_tdd_with and TDD_TEMPLATE_NS are removed, with the FlatError::TddXml, FlatError::TddAttribute and FlatError::TddElement variants and the crate’s quick-xml dependency. The openEHR crates move to 0.0.109 together.

  • Breaking: openehr-sdt 0.0.111 no longer carries smart_scopes. The SMART on openEHR scope grammar moved, with its API and tests, to ferrohealth-smart 0.8.5 as ferrohealth_smart::openehr::smart_scopes, so a consumer of openehr_sdt::smart_scopes names that path. FerroEHR pins ferrohealth-smart 0.8.5, its scope gate and the viewer’s scope previewer (both halves) parse with it, and the build resolves one openehr-sdt. Token checking and the scopes a token grants behave as before (#4206).

Fixed

  • A Cedar policy’s principal.patient is the caller’s patient claim ([authz.abac] patient_claim) (#4038). It was the patient of the object in play, so a rule such as principal.patient == resource.patient in the shipped consent.cedar held for every object a query touched.

  • The agent guard refuses wrapped and re-ordered spellings of a force push or a recursive delete (#4046). .claude/hooks/block_dangerous.sh let through git -C . push --force … main, /usr/bin/git push, \rm -rf, rm --recursive --force, bash -c "rm -rf …", find … -delete and similar. It now normalises the command first (a path or backslash before the binary, a launcher such as env or sudo, a sh -c or eval wrapper, git’s global options before the verb), judges every rm target and every deleting find starting point against /tmp, and refuses a push that deletes main or master. block_dangerous_test.sh runs each spelling.

  • An archetype HRID’s version status reads as the VERSION_STATUS it names (#4051), on the 0.0.102 crate line. AOM2 ARCHETYPE_HRID.version_status maps -rc, -alpha, -beta and no modifier to release_candidate, alpha, beta and released, where rc and the empty modifier read as VersionStatus::Other. openehr-adl writes the HRID back with the modifier (-rc, never -release_candidate), so parse and write are inverse.

  • A party reference whose type is not the builder’s default survives a FLAT or STRUCTURED round trip (#4052). master05 gives PARTY_REF.type no row, so a _health_care_facility that is not an ORGANISATION, or a _provider, performer, composer, feeder-audit party or PARTY_PROXY datum that is not a PERSON, is now written as |raw (master04 §Raw canonical JSON) and read back whole; it came back with the default type.

  • An ADL 2 Web Template reads the ADL 1.4 term constraints of an archetype read in the ADL 1.4 dialect (#4056). openehr-adl keeps such a constraint verbatim (openehr::433, local::at0003,at0004;at0003) for its 1.4 to 2 converter, and an at-coded ADL 2 template over unconverted archetypes carried it into the ADL 2 builder, which took the whole string for one local code. The builder now reads it as the OPT 1.4 builder reads a C_CODE_PHRASE: an openehr code carries its terminology and rubric, local codes their archetype rubrics, and the assumed code becomes the default. The nine openehr-* crates move to 0.0.100.

  • An ordinal of an archetype read in the ADL 1.4 dialect lists its at-codes (#4059). The symbol of each [value, symbol] row kept the verbatim local::at0006 spelling, which the ADL 2 builder took for the code and its label; it now reads the at-code with its archetype rubric, as #4056 does for coded text. The nine openehr-* crates move to 0.0.101.

  • A BASE 1.3 TERMINOLOGY_CODE keeps its terminology_id in XML (#4076). ITS-XML 1.0.2 types TRANSLATION_DETAILS.language and the other slots that hold one as CODE_PHRASE, whose terminology_id carries a <value> child; the codec wrote and read bare text instead, so an XSD-valid document read with an empty terminology_id. It now writes and reads the TERMINOLOGY_ID shape.

  • A configuration load or check that fails keeps its cause (#4104). Fifteen sites in the server’s commands turned the failure into a string, so the error chain ended at the message; they now carry the source error.

  • Cohort queries work under the IHE patient identity source. That provider keeps no party-to-EHR map, so a cohort selected by a demographic predicate found no EHR. POST /query/cohort now takes subjects, identifiers in any domain, each resolved through the [ehr_index] provider: under the IHE provider by a PIXm (ITI-83) translation forward into the server’s clinical domain, never back. The new [cohort] subjects = true turns the surface on without a predicate. Under the IHE provider a predicate is refused 400, and a failing identity source fails the query instead of shrinking the cohort. The cohort digest takes each subject by namespace only (#4114).

  • The chart guard holds the FerroPIX Artifact Hub image to its pinned digest (#4115). scripts/checks/chart-appversion.sh checked the FerroTERM and FerroBRIDGE artifacthub.io/images entries against their pins, but not the FerroPIX entry the ferropix dependency added, so its digest could drift from ferropix.image.digest in values.yaml and Artifact Hub would scan an image the chart does not run. The guard now compares the two, and --self-test moves the digest on either side.

  • ui-screenshot-guard judges only the pull request’s own commits (#4117). It diffed the base and head commits directly, so a branch behind main failed on main’s viewer changes and passed on main’s screenshot refresh. scripts/checks/ui-screenshot.sh now reads the paths from the merge base (base...head), and its --self-test runs both cases on a branch behind main before every verdict.

  • The usage report groups a request by the route below the openEHR API mount (server.base_path followed by rest.base_path). A REST root that contains an API family word, such as /ferroehr/ehr/v1, no longer places a query or definition route under ehr, and a route outside the mount counts in no group, as before #4128 (#4132).

  • The local EHR index holds an EHR created for a subject. Under [ehr_index] provider = "local", create_ehr_for_subject now enters the subject in the local map with the EHR’s creation, as a Primary association with no location, so subject_ehrs and a subject cohort find the EHR. Before, only EHR_STATUS named the subject and the map stayed empty until add_ehr_subject ran (#4138).

  • [smart] issuer_metadata_file is read by the same parse as metadata fetched from an issuer (#4165), so a file that repeats a member name is a boot error. It used to be read as it stood, keeping whichever value the parser took.

  • The viewer draws the CDR’s refusal where it used to draw nothing: on an EHR’s Status tab when the version’s envelope read is refused, in place of the contribution, signed and lifecycle facts, and in Edit the directory’s reference picker when the composition list is refused, in place of an empty select (#4170).

  • The access record of every route that addresses one EHR names that EHR, read from the matched route’s {ehr_id} parameter: GET /message/export/{ehr_id} and POST /message/import/{ehr_id} recorded none, so a patient’s ITI-81 search missed them, and DELETE /admin/ehr/all no longer names an EHR all. The Extract record of an export or import names the requesting user where it named UNKNOWN, and a read of GET /admin/restriction is recorded against its EHR (GDPR Art. 15(1)(c), EHDS Art. 9(2)) (#4192).

  • The compose-built viewer image builds its WebAssembly bindings again: its wasm-bindgen CLI now follows the wasm-bindgen version Cargo.lock resolves instead of a separate pin that had fallen behind it. A CI guard holds every other wasm-bindgen and cargo-leptos pin to the lock and to one another, and CI builds the viewer image from source whenever its Dockerfile changes (#4222).

  • The viewer takes FerroUI 0.5.3: the top bar’s search no longer panics when a navigation leaves the page while its answer panel is being placed, and a search that finds the session over now sends every other open viewer tab to sign-in too (#4224).

  • A composition commits when an element’s value matches any one of the alternatives its template allows (#4226), on the 0.0.110 crate line. ADL 1.4 §“Single-valued Attributes” makes the object constraints under a single-valued attribute alternatives, only one of which the data needs to match. The free text of an open coded value-set (FLAT or STRUCTURED |other, a canonical DV_TEXT) was refused as expected RM type conforming to DV_CODED_TEXT but found DV_TEXT where the template allows DV_TEXT beside DV_CODED_TEXT, and a value that matched only a later alternative of the same type, such as a second DV_QUANTITY with other units, was checked against the first alone. When no alternative matches, the refusal names every alternative with what each one found, and FLAT writes a value under one of several same-type alternatives, not under each.

  • POST /query/cohort and an AQL query rooted at a demographic class now leave out every EHR the population gate leaves out: an EHR whose current EHR_STATUS.is_queryable is false and an EHR under a research objection no longer answer rows in a cohort or a demographic join, as they already did not in an ad-hoc population query (RM EHR_STATUS.is_queryable: “True if this EHR should be included in population queries”). The research objection leak was a privacy defect. A query scoped to one EHR by ehr_id and an EXTRACT’s criteria still answer the EHR they name (#4247).

  • The hosted sandbox’s database box runs a pinned image (#4252). deploy/hosted/cloud-init-postgres.yaml defaults FERROEHR_POSTGRES_IMAGE to ghcr.io/ferrohealth/ferroehr-postgres:5.4.2 by its index digest, where the box ran the moving latest tag of the image’s pre-FerroHEALTH namespace, and compose-image-tags.sh holds the default and the hosted README to one release reference. Step 6 of sandbox-ferropix-rollout.sh moves the box onto the pin: it refuses a change of PostgreSQL major, recreates the container on its data volume and checks the CDR’s readiness and a ferropix login afterwards. Step 1 of the script no longer prints an empty line for each container that is absent.

Security

Shipped with functional changes because: the patched ferrohealth-smart arrives in the same dependency move as the family mask and the lockstep releases built on it. The zstd move is a lock-only update that a security-only 5.4.3 could have carried; the maintainer chose to ship it in 5.5.0, released the day after the fix landed, instead of cutting a separate patch.

  • ferrohealth-smart 0.7.1 masks the userinfo of every URL in its grant tables, a URL no parser accepts included. With 0.7.0 a configured token_endpoint or issuer that did not parse could reach the configuration renderings and the log in full (CWE-532, GHSA-vphj-3f8c-5hpx) (#4050).

  • The zstd bindings move to zstd 0.14.3, zstd-safe 8.2.0 and zstd-sys 2.1.2, past RUSTSEC-2026-0391, RUSTSEC-2026-0398, RUSTSEC-2026-0403 and RUSTSEC-2026-0410. Response compression and the outbound clients’ response decompression reach these crates, and the affected functions were not shown unreachable (GHSA-cwhv-87pq-3c3c) (#4095).

5.4.2 - 2026-10-25

5.4.2 is a patch release with no breaking change: a deployment upgrades in place. Helm chart 14.0.1 carries appVersion 5.4.2 and changes nothing else.

Changed

  • The viewer’s search answers under the field as you type (#4024), on FerroUI 0.4.0. From the second character the panel under the top bar’s search shows the EHRs, compositions, parties, templates and stored queries that match, five of a kind, from the server function /api/find_answer, which posts the words so they stay out of the address and the log. Arrow keys move through the hits, Enter opens one, Escape closes the panel. The find page lists every hit in the page.

  • The CDR takes up ferrohealth-smart 0.5.0 (#4002), and the build holds one copy of openehr-sdt where it held two (0.0.90 for the SMART crate beside the workspace’s 0.0.98). Reading a trusted issuer’s authorization-server metadata at boot, the server now refuses a document that repeats a member name (RFC 8259 §4), naming the location that answered it.

  • SECURITY.md states the family’s response times, with fix-time targets by severity. An acknowledgement within about 5 business days and an initial assessment within about 10, both best effort as before, and now a fixed release published within 7 days of the confirmed assessment for a Critical vulnerability, 30 days for High, 90 days for Medium and with the next release for Low. The CRA risk assessment says business days where it said working days.

Fixed

  • The audit drain sends a held syslog connection’s records without a millisecond wait per record (#3954). FerroEHR moves to ihe-atna and ehds-logging 0.2.7 (FerroAUDIT). The liveness probe the syslog sink runs before each record waited for a timer tick of about a millisecond, which held one connection to roughly 880 records per second. It now answers at once, about 47 000 records per second in FerroAUDIT’s measurement.

  • A revoked Direct Key signature no longer lets the primary key sign (#3972). The PGP signer now applies a valid Certification Revocation of a Direct Key signature, created at or after it: the revoked signature is cancelled (RFC 9580 §5.2.3.10), so neither its signing flag nor its key expiration counts any more, and a version 6 primary key left with no Direct Key signature in force is not used.

  • The viewer’s Paths tab offers Build a query only on a COMPOSITION template (#4001). A node of a demographic template (a PERSON, ORGANISATION, GROUP, AGENT or ROLE root) no longer links to the query builder, whose one query shape, FROM EHR e CONTAINS COMPOSITION c, can never match a party. The viewer moves to FerroUI 0.3.7, which makes the node’s query link optional and keeps a long URL or word in a banner, a notice or a setting row inside the screen at 320 px.

  • An access record says when the patient lookup failed (#3900). When the audit drain cannot read the subject behind an EHR (the database refused or did not answer), the EHDS access record now writes the patient lookup as unavailable instead of not-found, so a reviewer can tell a record that names no patient because the lookup failed from one whose EHR has no subject (Regulation (EU) 2025/327 Art 9(1)). A failed lookup is asked again for the next record instead of being remembered.

Security

Shipped with functional changes because: no shipped artefact is affected (zip 8.6.0 ran only in a build script, as the withdrawn VEX statements recorded), so there is no security update to separate from the other changes this release carries.

  • zip 8.6.0 leaves the build (#3799; GHSA-vv5m-jxxh-4627, GHSA-4v9h-p7m9-rh26, GHSA-4684-h2q9-2vcp). utoipa-swagger-ui 10.0.2 unpacks its vendored Swagger UI archive with zip 9, so the yanked zip 8.6.0 and its three extraction advisories (RUSTSEC-2026-0367, RUSTSEC-2026-0368, RUSTSEC-2026-0369) no longer reach the dependency graph, and deny.toml and security/vex/rust-advisories.openvex.json drop their exceptions for them.

5.4.1 - 2026-10-24

5.4.1 carries breaking changes for operators who use a terminology server, each with its upgrade step under Changed. Before you upgrade: move [terminology.external] to the one [terminology.server] (its url, fail_on_error, request_timeout_ms and optional oauth grant) and rename FERROEHR__TERMINOLOGY__EXTERNAL__ variables to FERROEHR__TERMINOLOGY__SERVER__; on Helm, chart 14.0.0 takes config.terminology.server and secrets.terminologyClientSecret. A deployment without an external terminology server upgrades unchanged.

Changed

  • The CDR reaches its terminology server through ferrobridge-term (#3085), FerroBRIDGE’s published FHIR terminology client, in place of its own HTTP client and the third-party fhir-model decoding, which leaves the build. The configuration is unchanged. On the wire, $lookup, $validate-code, $expand and $subsumes are now a POST of a FHIR R4B Parameters body instead of a GET with the inputs in the query string, so a code no longer travels in a URL; the CodeSystem?url=&_summary=count probe stays a GET. The SM at_date is sent as the date input of $lookup and $validate-code. Each call is made once, so request_timeout_ms still bounds it. The answers are read strictly against the R4B operation definitions: a $lookup answer without its name or display, an $expand answer without an expansion, or a $subsumes outcome outside its four codes is now an upstream fault (500, or the unresolvable branch of fail_on_error at commit) where the CDR used to fall back to the code as the term text, an empty value set, or “not subsumed”. A server that answers by the operation definitions sees no difference. The fhir feature of the internal ferroehr-ext crate is gone with the decoder it gated; the fhir features of ferroehr, ferroehr-rest and ferroehr-server keep their meaning.

  • The conformance statement declares unmatched-member-refused (#3988, register AMB-229). The server refuses a container member whose RM type the template does not allow, such as a CLUSTER committed where items[at0004] is constrained to ELEMENT (422, AM AOM 1.4 C_OBJECT.rm_type_name, ADL 1.4 cADL §Occurrences). The statement claimed the lenient unmatched-member-accepted arm, so the served and declared postures disagreed. The posture is the strict one by owner rule: FerroEHR holds to the specification and makes no allowance for other implementations’ mistakes. Catalogue case CONT-ITEM-unmatched_member_refused now applies and CONT-ITEM-unmatched_member_accepted is not applicable. A client that commits a mismatched member gets the same 422 as before.

  • Breaking: the CDR asks one terminology server, configured under [terminology.server] (#3085). All terminology beyond the openEHR support terminology is FerroTERM’s job: the CDR keeps the in-process openEHR terminology the RM validates against and one client to one FHIR R4B terminology server (BASE architecture_overview/master12-terminology.adoc §Binding Terminology Value-sets to Archetypes puts the resolver of a constraint binding in that server). [terminology.server] carries enabled, url, fail_on_error and request_timeout_ms, and the optional [terminology.server.oauth] grant takes the keys [audit.fhir_feed.oauth] already uses. Every key is a scalar, so the whole section is reachable through FERROEHR__TERMINOLOGY__SERVER__*. The fail-open default and the 422 for a resolved non-member are unchanged (register AMB-172). A configuration that still names [terminology.external] is refused at boot, naming the key. Upgrade: move [terminology.external] enabled and fail_on_error to [terminology.server]; put the one server’s base URL in [terminology.server] url (the URL of the provider you used as default, for FerroTERM http://ferroterm:8080/r4b) and its request_timeout_ms beside it; move an [terminology.external.oauth2_clients.<name>] table to [terminology.server.oauth], where scopes = [...] becomes one space-separated scope = "..." string, auth_method becomes client_auth, and refresh_leeway_secs goes; rename every FERROEHR__TERMINOLOGY__EXTERNAL__ environment variable to FERROEHR__TERMINOLOGY__SERVER__. A deployment that spread code systems over several servers loads them into FerroTERM, which serves SNOMED CT, LOINC, ICD and any FHIR CodeSystem or ValueSet from one index.

  • Breaking: Helm chart 14.0.0 wires the terminology server through config.terminology.server and takes its client secret as secrets.terminologyClientSecret (#3085). terminology.enabled with terminology.wireCdr injects enabled, url and fail_on_error under config.terminology.server, and refuses a values file that sets the URL itself. secrets.terminologyClientSecret is one string, mounted at /etc/ferroehr-secrets/terminology.server.oauth.client_secret, and needs the grant declared under config.terminology.server.oauth. Upgrade: move config.terminology.external to config.terminology.server as above, and secrets.terminologyOauth2ClientSecrets.<name> to secrets.terminologyClientSecret.

  • A token endpoint URL is masked in every rendering of an outbound grant (#3085). token_url under [audit.fhir_feed.oauth], [authz.abac.remote.oauth] and [terminology.server.oauth] now keeps its host and path and masks its userinfo in ferroehr config check, /management/env, GET /admin/config and ferroehr report, as the terminology client’s own token URL already did.

  • The conformance lane and the real-release terminology probe run FerroTERM (#3085, #2236). The terminology profile of docker/sut-ferroehr.yml composes FerroTERM 0.1.6, pinned by digest, with the shaped seed of docker/terminology/seed/ mounted read-only as its code-system directory, in place of HAPI FHIR and the PUT seeder docker/terminology/seed.sh, which is gone. The opt-in PROBE_ONLY=terminology family of scripts/deploy-probe.sh builds a local SNOMED CT release into an index with the quickstart overlay’s own ferroterm-build and drives FerroTERM over it, in place of Snowstorm and Elasticsearch.

  • The conformance run composes a fourth deployment for the fail-open unresolvable-binding branch (#3985). docker/sut-terminology-failopen.yml runs the same image as its own compose project on ports 8083 and 5434, with [terminology.server] fail_on_error = false and a server URL nothing answers on, and the ixit declares it as the sut_fail_open instance. The stable-profile deployment now uses the primary’s FerroTERM, which the ixit already declared for it.

Removed

  • The terminology route map, per-server credentials and caches are gone (#3085); FerroTERM is where the capability lives. Removed keys: [terminology.external.providers.<name>] with type, operation, connect_timeout_ms, cache_ttl_secs, cache_capacity, oauth2_client, client_cert_path, client_key_path and ca_bundle_path; [terminology.external.routes]; [terminology.external.oauth2_clients]. The CDR no longer selects a server per terminology, no longer presents a per-server mutual-TLS client certificate or pins a per-server trust bundle (the server is verified against the platform’s trust store), no longer caches terminology answers, and no longer validates membership through $expand ($validate-code answers it). AQL TERMINOLOGY(), the /terminology/* routes, commit-time binding resolution and the ADL 2 VETDF check all ask the one server.

Fixed

  • A method a resource does not serve is 405 with Allow under SMART (#3987). With bearer authentication on, DELETE /ehr and every other method a route does not serve was refused 403 by the SMART on openEHR scope check before the router could answer. The gate now answers 405 with the Allow field naming the served methods once the caller has authenticated, whatever the token grants, and runs the RBAC and scope checks only on a served method (ITS-REST overview/Requests_and_responses §HTTP Methods, RFC 9110 §15.5.6). A request without a credential is still 401.

  • A ROLE operational template whose single attributes state 0..* uploads (#3994). POST /definition/template/adl1.4 refused deployed ROLE templates (Role.v2, root openEHR-DEMOGRAPHIC-ROLE.Role.v1) with 422 Members_valid, because they constrain ROLE.capabilities and CAPABILITY.credentials without a cardinality while the children state occurrences 0..*. ADL 1.4 master05-cadl.adoc §Occurrences calls that “not illegal” and reads it as at most one value, so the CDR now accepts the template and applies that cap: the Web Template gives such a node max 1, and the ADL2 conversion states the cap and records it in conversion_details, next to the template’s undefined ac0001 reference. A ROLE party built against the template, and a PERSON that holds it through PERSON.relationships, commit and read back as before. The disagreement with the stricter AOM 1.4 Members_valid invariant is reported upstream (#3996).

  • The template example of a demographic template is a party (#3995). GET /definition/template/adl1.4/{template_id}/example (and the ADL 2 twin) answered 400 the template root PERSON is not a COMPOSITION for every PERSON, ORGANISATION, GROUP, AGENT or ROLE template. It now answers an example party of the root class at every detail_level, in canonical JSON or XML, with a uid for type=output; an example ROLE carries a placeholder performer, which the RM makes mandatory. FLAT and STRUCTURED stay 406 for a party example, since the Simplified Formats map no party class. The viewer’s Example tab offers JSON and XML only for a demographic template, and for a template rooted at an ENTRY or a SECTION it says that no example exists instead of showing the CDR’s refusal.

5.4.0 - 2026-10-24

5.4.0 carries breaking changes for operators, each with its upgrade step under Changed. Before you upgrade: replace an ECDSA or Ed448 OpenPGP signing key in [signing] mode = "pgp" with an Ed25519 key, and check that the certificate’s primary key or a subkey may still sign; tell viewer users that the viewer’s addresses change, so bookmarks and links into it need updating, and that the theme falls back to the system’s until it is chosen again; and serve the viewer’s whole pkg/ directory if you serve its site from your own web server. No configuration key, database migration or image name changes.

Changed

  • Breaking: an ECDSA or Ed448 OpenPGP signing key is refused at boot (#3965). In [signing] mode = "pgp" the server no longer starts when the key’s signing component is ECDSA (P-256, P-384, P-521, secp256k1, Brainpool) or Ed448, and the error names RUSTSEC-2026-0387; see Security below. P-384, P-521 and Ed448 keys already failed the boot signature, because the server signs with SHA-256 and RFC 9580 requires a longer digest for them, so the keys that change behaviour are P-256 and secp256k1. Ed25519 and RSA keys load as before, and the RSA boot warning now recommends Ed25519 alone. Such a certificate keeps verifying the versions it signed when it is listed in signing.retired_key_paths. Upgrade: before upgrading a deployment that signs with such a key, generate an Ed25519 signing key (gpg --quick-generate-key "…" ed25519 sign), export its armored secret key to signing.key_path, and add the old certificate’s public half to signing.retired_key_paths (the book’s PGP signing page and the signing section of the configuration reference). The digest mode, the default, is unaffected.
  • The viewer is rebuilt on FerroUI 0.3.6 (#3932, #3942, #3946, #3957, FerroHEALTH/FerroUI#161). Every screen follows the FerroEHR mockups of FerroUI’s design/consoles/ferroehr/ at 1440 and 390 px, light and dark, with FerroUI’s components and none of the viewer’s own copies. The viewer now runs in islands mode: the server draws every page whole, every link is a link and every action a form that posts to its page, so the whole viewer works with JavaScript off, and the browser hydrates only the parts that add behaviour (the code editor, path trees, the chart legend, date pickers, the user menu’s keyboard pattern, the subject reveal and Copy). Editors end in a docked decision bar; a refused save keeps what you typed beside the CDR’s words. The navigation is Dashboard, EHRs, Parties, Templates, Terminology, Stored queries, Query, Server, Audit log, Subscriptions, Configuration and About, with search in the top bar. The activity charts of the dashboard and of an EHR’s contributions label their first day and every tenth day back from today, and a staged change’s Apply sits beside its What changes select.
  • Breaking: the viewer’s addresses change. /demographics/… is /parties (/parties/{kind}/{uid}, /parties/relationships/{uid}, /parties/contributions/{uid}); an EHR’s tabs are paths (/ehrs/{id}/status, /directory, /contributions, /tags) instead of ?tab=; /queries/builder and /queries/aql are /query and /query/aql; /queries/stored?load=… is /queries/{name}/{version}/run; /system and /operations are /server with its tabs, the conformance manifest moves to /about, the runtime configuration to /configuration; an ADL 2 release is /templates/adl2/{stem}/{release}. Upgrade: update bookmarks and links into the viewer; an old address answers 404.
  • Breaking: display preferences are cookies of their own. The theme is set in the user menu and on Your preferences (/preferences), and each preference is its own ferroehr-viewer-* cookie; the ferroehr-theme cookie and the /theme route are gone, so a theme chosen before the upgrade falls back to the system’s until it is chosen again.
  • Breaking: a listing that can show patient data keeps its page out of the address (FerroUI leptos-ui.md §0): the EHRs, an EHR’s compositions and contributions and the audit log page through a session cookie (/paging); ?offset= and ?page= on them are ignored. Filters stay in the address.
  • An open tab now idles out (#3057). The viewer polls nothing in the background, so session.idle_minutes measures the requests the browser makes, and a tab left untouched signs out like a closed one; the next page asked for is the sign-in page saying the session ended, and signing in returns to it. A revoked session is noticed on the next request.
  • The viewer’s site is built with cargo leptos build --split (scripts/cargo-leptos.sh adds it) and FerroUI’s split feature, so FerroUI’s lazy islands (the date pickers and the tree) load from chunks of their own on a page that draws them (FerroUI #331). A deployment that serves the site from its own web server serves the whole pkg/ directory, chunks included; the image already does.
  • The dashboard’s compositions per day count commits, read from each version’s commit_audit/time_committed, as its note says; it counted context start times before.
  • A query’s results are FerroUI’s results pane (#3942). Table and Chart sit in the Results heading and run the query again in the view chosen, which the next Run keeps; the separate Show as choice is gone. Export in the heading opens a menu of CSV and JSON (#3946), whose items post to /export/aql/csv and /export/aql/json, and /export/aql with a format field no longer answers.

Added

  • Your preferences: the theme, density, reduced motion, the language, the first day of the week, how long a revealed subject id stays shown, and the keyboard shortcut hints, each kept in its own cookie.
  • Older versions on an EHR’s directory reads further back than the newest ten versions, and a version reached that way restores.
  • The viewer’s journeys run at a phone’s 390 px in CI beside 1440 px, and the screenshot pass captures every route in both widths and both themes as the ui-route-shots artifact (#3910). Keyboard journeys cover where the focus lands after a route change, the query builder by keyboard, and the user menu’s keyboard pattern (#2793).
  • The parts the FerroEHR mockups draw and FerroUI 0.3.4 now carries (#3942): Your access in the user menu; Details on the sandbox band, leading to Server’s Overview; Copy the id (or the name) in the More menus of an EHR, a directory, a composition, a party, a relationship, a template and a stored query; Save to the CDR beside Run in Query’s header; relative times in FerroUI’s words; one editor per staged change of a new contribution, with no cap of twelve; the contribution glyph; Discard that leaves a form with an empty required field, so the relationship and subscription editors mark their required fields again; the time of each point in a chart’s figures table; Your access’s grants built from the CDR’s own reading of each scope; and the How configuration works dialog on Configuration.
  • The parts the FerroEHR mockups draw and FerroUI 0.3.5 now carries (#3946): Your access reads the OpenID Connect profile scope as the request for your profile, where it drew it refused although the CDR accepts it; a query’s exports are one Export menu; and Stored queries takes its scroll glyph in the navigation, on the dashboard and in the search’s answers. The activity charts of the dashboard and of an EHR’s contributions label their first and last day as the mockups do (22 Sep).
  • A template’s Download in its More menu (#3942): the stored ADL 1.4 operational template (.opt) or ADL2 template (.adlt) as a file, read with your session from /download/template/{family}/{id}.
  • New contribution follows its mockup more closely (#3942): the audit section is headed Contribution audit, the committer’s placeholder names who commits when it is left empty, and Terminology takes its book glyph.

Removed

  • The viewer’s Tailwind build and its own stylesheet. The viewer serves FerroUI’s stylesheet and its FerroEHR theme only; the image build no longer runs the Tailwind standalone binary.
  • The EHR status’s own tag panel. A tag on the EHR status is still listed on the EHR’s Tags tab and opens the Status tab; setting one is done over the API, as the mockup has no panel for it.
  • The stored-query listing’s pager: every stored query is listed, grouped by namespace.

Fixed

  • The OpenPGP signing subkey is the newest one that may sign (#3967). In [signing] mode = "pgp" the server took the first subkey in the file that any binding signature had ever flagged for signing, so a subkey added for rotation never signed, and a revoked or expired subkey kept signing. It now signs with the subkey created last among those whose newest valid binding grants signing, that have not expired, and that carry no valid revocation (RFC 9580 §5.2.1.12, §5.2.3.10, §5.2.3.13), and with the primary key when none qualifies. A certificate with one signing subkey signs as before. Upgrade: a deployment whose certificate holds several signing subkeys signs with a different one after the upgrade; every version signed before keeps verifying, since the certificate still carries the old subkey.
  • The OpenPGP primary key signs only when the certificate lets it (#3969). In [signing] mode = "pgp", when no subkey qualified to sign, the server signed with the primary key without checking it. It now signs with the primary key only when the primary key’s governing self-signature (the newest valid Direct Key signature when it carries key flags, else the newest valid self-signature on the primary User ID) carries the signing flag 0x02, neither that self-signature nor the key has expired, and no valid Key Revocation covers it (RFC 9580 §5.2.1.11, §5.2.3.10, §5.2.3.13, §5.2.3.29). A self-signature with no key flags grants no signing. When neither a subkey nor the primary key may sign, the server refuses to start and says what to change. Upgrade: a certificate whose primary key lacks the signing flag, or has expired or been revoked, and which has no usable signing subkey, no longer boots; add an Ed25519 signing subkey (gpg --quick-add-key <FINGERPRINT> ed25519 sign) and export the armored secret key to signing.key_path again. Every version signed before keeps verifying, since the certificate still carries the primary key. A certificate generated with signing on its primary key, or with a valid signing subkey, loads as before.
  • A signing subkey stops signing when its primary key is revoked or expires (#3971). In [signing] mode = "pgp" the server checked a signing subkey’s own binding, expiry and revocation but not the primary key’s, so the subkey of a certificate whose primary key carried a valid Key Revocation, or had passed its Key Expiration Time, kept signing. Now no subkey signs while the primary key is revoked or expired (RFC 9580 §5.2.1.11, §5.2.3.13), and with nothing left that may sign the server refuses to start. Upgrade: a deployment whose certificate’s primary key is revoked or expired no longer boots; generate a new Ed25519 certificate (gpg --quick-generate-key "…" ed25519 sign), export its armored secret key to signing.key_path, and list the old certificate’s public half in signing.retired_key_paths so the versions it signed keep verifying. To keep an expired certificate, extend its primary key’s expiry (gpg --quick-set-expire <FINGERPRINT> 1y) and export it again.
  • The audit drain no longer loses the records it sent last when it exits (#3959). It dropped its syslog connection without a TLS close_notify. With the repository’s session tickets still unread, the kernel answered with a reset, which could discard the last records, with no warning. The drain now closes the connection as RFC 5425 §4.4 asks, through FerroAUDIT 0.2.5’s Connection::close (ihe-atna and ehds-logging =0.2.5), and logs and counts a close that fails. The syslog connection also sends each record at once (TCP_NODELAY).

Security

Shipped with functional changes because: only the newest release is supported (SECURITY.md), so a security-only patch cut from 5.3.3 beside the next release would be out of support the moment that release lands and give no one a release to stay on; the fix is itself a behaviour change, refusing ECDSA and Ed448 signing keys, so there is no dependency bump to ship alone.

  • RUSTSEC-2026-0387 in crypto-bigint 0.5.5 (#3965, security/vex/rust-advisories.openvex.json). Its operations meant to be constant-time can branch on their operands, and the 0.5 line, which pgp and openidconnect still use, has no fixed release. With ECDSA and Ed448 signing keys refused, the server runs those operations only over public keys and signatures, in verification.

5.3.3 - 2026-10-23

Fixed

  • An element whose value the template leaves open keeps its FLAT path (#3933). The web template dropped an ELEMENT that carries no value constraint, or a value constrained to {*}, so the vcf_report template’s “Test method” was missing from the FLAT example and a FLAT write to …/laboratory_test_result:0/test_method/text_value was refused as an unknown simplified path. Such an element admits any data value (AOM 1.4 C_COMPLEX_OBJECT.any_allowed, ADL 1.4 “Any” constraints), so it now keeps its node with one alternative per concrete data value type, keyed by type as in the Simplified Formats CLUSTER example (…/test_method/text_value, …/test_method/quantity_value|magnitude); DV_STATE is not among them, since it has no simplified form (#3934). Web templates, FLAT and STRUCTURED examples and FLAT paths change for templates with such elements. The openehr-* crates move to 0.0.96.
  • A DV_STATE is no longer built as a coded text from FLAT or STRUCTURED (#3934). The converter read and wrote a DV_STATE with the DV_CODED_TEXT suffixes, so a write built a value the RM refuses (RM data_types §DV_STATE: value: DV_CODED_TEXT, is_terminal: Boolean) and a read emitted keys no table defines. The Simplified Formats RM mapping chapter has no DV_STATE table, so a write that decomposes a DV_STATE node is now refused with a 422 naming the node and the missing table; the node is written and read as |raw canonical JSON, and the FLAT and STRUCTURED examples leave it out. A canonical commit’s DV_STATE is now checked against the template’s code list on its value. The gap is reported upstream in #3936. The openehr-* crates move to 0.0.97.

Security

Shipped with functional changes because: only the newest release is supported (SECURITY.md), so a separate 5.3.2 patch published beside 5.3.3 would be out of support the moment 5.3.3 lands and give no one a release to stay on; the fix is a dependency bump with no behaviour change of its own.

  • ctutils 0.4.5 (RUSTSEC-2026-0385, GHSA-h46f-99p9-gwvf). Its Choice mask generation and its u128 comparisons could compile to branches on secret values; the crate sits under the aes-gcm, hmac and digest implementations the server uses.

5.3.2 - 2026-10-22

The v5.3.0 and v5.3.1 tags were pushed but neither release published: the v5.3.0 scan refused the container base’s ca-certificates (DSA-6572-1), and the v5.3.1 plan refused a CRA risk assessment with no revision for that base change. 5.3.2 is the same release on the rebuilt base, and no 5.3.0 or 5.3.1 artifacts exist.

Changed

  • The Rust toolchain moves to 1.99.1. The pinned channel, both container base images (digest re-resolved to the rust:1.99.1-slim-trixie index) and the CI job containers move together; the MSRV stays at 1.98, the toolchain minus one, the rule docs/VERSIONS.md now states for the whole FerroHEALTH family. Clippy 1.99’s new assert_is_empty lint is answered at every site with an assert_eq! or assert_ne! that prints the value on failure, two float_cmp expectations it no longer needs are removed, and the unit tests under the published crates’ src/ (and the codegen templates they are emitted from) change with it, so the openehr-* crates move to 0.0.94 with no API change.

  • The access log moves to FerroAUDIT 0.2.4 (ihe-atna and ehds-logging =0.2.4), and the record writes again what #3897 had to drop: the token’s jti at agent[user].policy, the ITS-REST operation id as a second subtype, outcomeDesc, source.type, and in the DICOM message the optional EventOutcomeDescription, AuditSourceTypeCode and ParticipantObjectDataLifeCycle. A Basic user is recorded with no issuer and no client instead of FerroEHR’s own URN and UNKNOWN, an acting mode the deployment did not establish adds no person role, the caller’s roles are written by the crate, an origin names its originating system with no invented endpoint or status, and FerroEHR’s own agent is addressed by its server_host. A record kept without end (retention_days = 0) says so and names no end date, a day-counted horizon is stated in days, and a horizon under three years, which only a deployment under no Union jurisdiction passes, names that horizon as the ground of its floor instead of claiming Art. 9(2). The category map takes every archetype id openEHR BASE 1.3.0 admits, a CKM draft’s .v0 and the namespaced ADL 2 form among them, and compares ids as openEHR BASE asks, so the evidence keeps each id as the access named it: a template id keeps its case (records written before keep the folded form), and a stored composition’s root archetype stays in the node table’s comparison form, and the map’s digest moves once with the upgrade: records keep the digest of the map that classified them (#3902).

  • The minimum supported Rust version is 1.98 (was 1.97), the version fhir-types, pulled by ihe-atna’s FHIR writer, declares. The nine openehr-* crates move to 0.0.93 with it.

  • ferrohealth-smart 0.4.0 (#3895), the version every FerroHEALTH product pins. The refusal statuses and challenges do not change.

  • The access log is written through the FerroHEALTH products’ shared audit crates (#3887). The EHDS access record, the Art. 14(1) categories and their classification, the Art. 9(2) retention floor and the Art. 11(5) emergency mark come from ehds-logging 0.1.1, and the BALP AuditEvent, the DICOM audit message, the syslog sender and the ATX: FHIR Feed client from ihe-atna 0.1.1 (FerroAUDIT). FerroEHR still decides which accesses are logged, what each record says and where it is stored.

  • [audit.categories] names categories by their HL7 Europe codes. A priority category is Patient-Summaries, Electronic-Prescriptions, Electronic-Dispensations, Medical-Imaging, Laboratory-Reports or Discharge-Reports, or <system>|<code>; a national category is declared in the new national list as <system>|<code>; a key of no category maps to the word "none" instead of ["none"]. The old spellings (patient-summary, test-results, national:<code> and the others) refuse boot. The Helm chart’s config.audit.categories takes the same values (chart 13.0.0).

  • A stored access record names each category by its <system>|<code> token, and category_basis says what the classification was read from: returned, written, queried or resource-kind (audit migration 0012). Records written before keep their spellings. Evidence ids are stored in the case-folded form the map is read in. A query whose constraints name an id the map does not hold, or bind no id to a content operand, is now recorded unclassified beside the categories that did map.

  • The retention register keys a period on the category token (clinical migration 0016). PUT {base}/admin/retention/policy takes a category as [audit.categories] names it and stores the token; periods keyed on a priority spelling move to its token, and a period keyed on a national:<code> spelling refuses the migration until it is declared again.

  • The FHIR AuditEvent follows the BALP 1.1.4 patterns ehds-logging and ihe-atna write. The person is the agent:user slice (the token’s issuer and subject, the professional’s identifier as ihe-otherId, the assurance level as ihe-assuranceLevel), the client application an agent of its own, the organisation a logical Organization reference; the categories, the retention, the security labels and the origins ride in ehds-categories and origin entities; the emergency mark is the ehds-emergency-access entity; a query’s subtype is search. The token’s jti agent, the ITS-REST operation-id subtype, the outcomeDesc and the audit source type are no longer written.

  • The DICOM audit message is written without a byte order mark, EventOutcomeDescription, AuditSourceTypeCode or ParticipantObjectDataLifeCycle (ITI TF-2 §3.20.4.1.2).

Removed

  • The UDP syslog transport ([audit.syslog] transport). The syslog feed sends RFC 5424 over RFC 5425 TLS only, on port 6514 by default, and a transport key refuses boot. The audit_syslog_udp deployment gap is gone with it.

Fixed

  • A + in a query-string parameter is read as a space (#3891). Every query parameter the REST API reads, q of GET /query/aql included, is decoded as application/x-www-form-urlencoded (WHATWG URL Standard §5.1): + and %20 are a space and %2B is a literal plus. A form-encoded AQL query that wrote its spaces as + was refused with a 400 before. A client that sends an unencoded + meaning a plus, such as a time-zone offset, encodes it as %2B, as RFC 6570 form expansion and form encoders already do.

  • An xsi:type the XML schemas do not define is refused with an error that names them (#3892). An OPT 1.4 upload whose constraint node carries a type outside the ITS-XML OPT 1.4 schemas (Template.xsd, OpenehrProfile.xsd, Archetype.xsd), such as T_ARCHETYPE_ROOT in place of C_ARCHETYPE_ROOT, answers 422 with a message saying that those schemas define no such subtype. The RM and AOM2 XML readers name their schemas the same way. The refusal itself is unchanged.

  • A Basic request the verified-credential cache admits writes no ATNA login record (#3895). Since the move to ferrohealth-smart 0.3.1 every Basic request was marked a fresh authentication, so with suppress_login_events = false each one wrote a login record (DICOM PS3.15 §A.5.1, EventID 110114). The crate’s Caller::cached now tells the cache hit apart: one login record per credential verification, as [auth.basic] cache_ttl_seconds documents.

  • A lookup by subject no longer echoes the subject id (#3906). The 404 of GET /ehr?subject_id=…&subject_namespace=… for a subject with no EHR, and the message extension’s refusal for an unknown subject_id, named the identifier, a national patient number in many deployments. Both now name only the namespace or the parameter. A test drives the lookup under a subscriber that records every log line and span attribute and asserts the identifier reaches none of them; the request log and traces already recorded the path and route without the query string. The book’s Security page says that the identifier travels in the URL and that a proxy in front of FerroEHR must not log query strings.

  • An archetype that fills two slots under different names gets the right path id in each (#3920). The web template took a slot fill’s name from whichever fill of the same archetype came first, so in a template that fills openEHR-EHR-CLUSTER.reference_sequence.v1 once as “Genome assembly” and once as “Reference sequence”, both fills were named genome_assembly and a FLAT path through the second, such as …/genomic_substitution_variant/reference_sequence/source_name, was refused as an unknown simplified path. Each fill now takes its node id from its own root term (ITS-REST Simplified Formats, Node ID Generation Rules). Web templates, FLAT and STRUCTURED examples, and FLAT paths change for templates that fill one archetype under two names. The openehr-* crates move to 0.0.95.

  • <root>/context/start_time sets the history origin and event times (#3919). A FLAT composition that gave its start time by path, without ctx/time, stored that start time but timed every untimed observation history and event with the commit time. The path spellings …/context/start_time, …/context/_end_time, …/context/_location and …/context/setting now feed the same defaults as ctx/time, ctx/end_time, ctx/location and ctx/setting (Simplified Formats, ctx/ vocabulary and RM mappings, EVENT_CONTEXT). An explicit ctx/ key still wins for the defaults.

  • A date or time pattern refusal says which parts are required, optional or not allowed (#3918). A value that breaks its template’s date, time or date-time pattern used to be refused with only the pattern, such as yyyy-mm-ddTXX:XX:XX, which reads as a match to anyone who takes XX for two digits. The refusal now names the parts in words (“the template allows a date only; a time is not allowed here”, “the template requires the second, which the value lacks”) and ends with a key to the pattern symbols. The book’s Templates & validation page explains them.

  • A stored composition’s archetype id keeps its case in the access log (#3908). The access record named a stored composition’s root archetype in lowercase, so a read of an openEHR-EHR-COMPOSITION.report.v1 composition recorded openehr-ehr-composition.report.v1 as its evidence, while template ids and ids taken from AQL text kept theirs. openEHR BASE asks identifiers to be case-preserving through persistence and case-insensitive in comparison (Composite Identifiers and Case). The node table keeps the lowercase id only as its comparison key beside the stored fragment, which has always held the id as written; the access log now reads the id from the fragment, so the evidence names the archetype as the composition wrote it, and AQL predicates and the category map still match in any case. Records written before keep the lowercase form.

Security

Shipped with functional changes because: only the newest release is supported (SECURITY.md), so a separate 5.2.1 published beside 5.3.2 would be out of support the moment 5.3.2 lands and give no one a release to stay on; the fix is a dependency and base-image bump with no behaviour change of its own.

  • rustls-webpki 0.103.16 and webpki-roots / webpki-root-certs 1.0.10 (RUSTSEC-2026-0381, RUSTSEC-2026-0382, RUSTSEC-2026-0383, GHSA-pmpg-vfwr-99qv). The outbound TLS client no longer accepts certificate paths through the distrusted emSign intermediates.
  • The ferroehr and ferroehr-viewer images move to the distroless base with ca-certificates 20250419+deb13u1 (DSA-6572-1, GHSA-pmpg-vfwr-99qv, #3926). The system trust store no longer holds the compromised eMudhra (emSign) intermediates; it is the only package that differs from the previous base.

5.2.0 - 2026-10-20

Added

  • [smart] issuer_metadata_file supplies the metadata of a jwks_file issuer (#3875). It names a JSON file with the RFC 8414 §2 metadata of the first [auth] issuer, read at boot, so a server whose key set is a file can serve the SMART discovery document with a token_endpoint. The metadata must name that issuer and every endpoint in it must be https (a loopback host excepted); a file set for an issuer whose metadata the server fetches refuses the start.

  • An authorised, audited read of a party’s protected national identifiers (#3776). GET {base}/admin/party/identifiers?version_uid=… serves a party version with the identifiers [demographic.identifier_protection] seals put back to their values, a historical version included. It needs the admin role and the new [authz.rbac] identifier_reader_role; unset, nobody may call it. Every read writes one access record per scheme expanded, naming the reader, the party and the scheme, never the value.

  • The EHDS Art. 8 restriction of access by health professionals (#3682). On the person’s instruction the controller records a restriction with POST {base}/admin/access-restriction, for a whole EHR or one versioned object (GET reads the register, POST …/lift lifts it). Restricted data then reads as absent to every professional route, query and export: a restricted composition or EHR answers 404 as an unknown one does, queries and exports leave it out, a CONTRIBUTION lists none of its versions, the event stream withholds it, and a write to it is refused as a write to an unknown target. A whole-EHR restriction covers data recorded after it. The admin routes still see everything, and nothing is deleted. A caller holding the new [authz.rbac] emergency_access_role that declares one of [audit] emergency_purpose_codes reads restricted data (EHDS Art. 11(5)), and its access record carries the emergency mark. Two migrations add the register and the marks (clinical/0015_access_restriction.sql, party/0011_access_restriction_column.sql).

Changed

  • Authentication runs on ferrohealth-smart 0.3.1 (#3875). The crate now holds what FerroEHR kept around it: the cache of verified Basic credentials, the RFC 6750 §2.1 b64token check, the refusal of a blank sub and the redacted rendering of [auth]. What changes for you:

    • [authz] basic_cache_ttl_seconds is now [auth.basic] cache_ttl_seconds (FERROEHR__AUTH__BASIC__CACHE_TTL_SECONDS), and its default is 0, which checks every request. Set 60 to keep the former behaviour. The old key is an unknown key and refuses the start.
    • [auth] scopeless_subjects is accepted: a listed subject whose token carries no resource scope meets every SMART on openEHR requirement, and a token that carries scopes gets only its scopes. It refused the start before.
    • A bearer token outside the b64token alphabet answers 401 with error="invalid_token" (RFC 6750 §3.1: a malformed token). It answered 400 invalid_request before. A header that is no scheme and credential, and two Authorization headers, still answer 400.
    • A rendered configuration (ferroehr config check, /management/env, ferroehr report) writes a Basic password hash as [redacted].
    • With [audit] suppress_login_events = false, every request a Basic credential authenticates writes a login record, since the credential is checked on each request. Before, only a request that missed the cache did.
    • The Helm chart is 12.1.2; its values document the two keys.
  • An EHR extract never carries a protected national identifier between deployments, and the refusal says what to do instead (#3777). Importing a party whose identifier another deployment sealed answers 422 naming the remedy: import the extract without the party and register the person through the receiving deployment’s party API, which seals the identifier under its own key. The book’s EHR Extract page records the decision and the steps.

Fixed

  • A FLAT OBSERVATION without a history origin takes the earliest event’s time (#3888, openehr-sdt 0.0.91). Simplified Formats master05 §OBSERVATION says /history_origin “will be set to the time of the earliest event if not explicitly set”; the builder used the composition’s context time. Times are compared as instants across offsets, ctx/history_origin still wins, and with no event time the context time stands.

5.1.0 - 2026-10-20

Added

  • The Helm chart mounts a database CA for sslmode=verify-full (#3788, chart 12.1.0). database.caConfigMap names a ConfigMap holding the CA the database certificates chain to (key database.caConfigMapKey, ca.crt by default); the server, the migration Job and the backup CronJobs mount it at /etc/ferroehr-db-ca, where a DSN’s sslrootcert names it. values-production.yaml sets it to ferroehr-db-ca and tells you to end every DSN in ?sslmode=verify-full&sslrootcert=/etc/ferroehr-db-ca/ca.crt.

  • A demographic query joins a party to the content of its EHR (#3831). FROM PERSON p CONTAINS EHR e CONTAINS COMPOSITION c (with ORGANISATION or ROLE as the root too, and party classes beside the EHR in an AND) answers a row per party and clinical row of an EHR that party is the subject of, through the party-to-EHR map. It runs as one statement per database credential, joined, ordered and paged by the server; a column, WHERE condition or ORDER BY key reading both sides, an aggregate, DISTINCT, and a half above query.max_result_rows each answer 400. The party read and the map read each leave an access record in their own domain.

  • The audit feed and the remote policy service authenticate with OAuth 2.0 client credentials (#3829). [audit.fhir_feed.oauth] and [authz.abac.remote.oauth] take a client-credentials grant (token_url or issuer, client_id, client_secret or client_secret_file, client_auth, scope, audience, resource). The server sends the token as a bearer credential, renews it before it expires, and fetches a new one once when the upstream refuses it with 401. A policy service that cannot be reached with a token fails the request closed. HTTP Basic through the URL stays, and setting both is a startup error.

  • AQL rooted at PERSON, ORGANISATION or ROLE (#3818), the subset of the experimental AQL demographic extension a client needs to find a party by what it holds. POST /query/aql answers rows over the latest version of each party, leaving out logically deleted ones, with paths inside the party in SELECT, WHERE and ORDER BY (its details tree, its identities, contacts and capabilities, p/uid/value), CONTAINS CLUSTER and the other content classes inside a party, named parameters, and the offset and fetch paging. It runs under the authorisation of a Demographic API read, on the demographic database credential. GROUP, AGENT, PARTY_RELATIONSHIP, a demographic class below the root, EHR or clinical content anywhere but directly under the root, and an ehr_id scope each answer 400 naming the construct.

  • openehr-sdt validates a party against its template (#3832). rm_instance::validate_party runs the RM-invariant and terminology passes with the template’s party class (PERSON, ORGANISATION, GROUP, AGENT or ROLE) as the root type, then the archetype-conformance pass over the template, so a party missing a mandatory element is refused at that element’s path. A template rooted at another class is refused at /. The nine openehr-* crates move to 0.0.90.

Changed

  • Authentication and SMART on openEHR run on ferrohealth-smart 0.2.0 (#3872). This is a breaking change to the configuration with no deprecation period. [auth.oidc] is gone: the token issuer is configured in [auth] itself (issuer or [[auth.issuer]], audience, algorithms, token_profile, clock_skew_seconds, jwks_max_age_seconds, jwks_refresh_interval_seconds, jwks_file, timeout_ms, realm). The server reads each issuer’s metadata and key set at boot, accepts only asymmetric algorithms, requires typ: at+jwt under the default token_profile (RFC 9068 §4), and answers every invalid token with one 401 that names no reason. An issuer on plain http is accepted only on a loopback address. A Basic user is user, password_hash or password_hash_file, and scopes; its roles move to [authz.rbac.basic_roles], and the verified-credential cache to [authz] basic_cache_ttl_seconds. [auth.oidc.assurance] and [auth.oidc.professional] become [authz.assurance] and [authz.professional].

  • Every route enforces its SMART on openEHR scope whenever [auth] is on (#3872). There is no switch any more: an EHR, composition, directory or contribution route needs a composition- scope, a template route a template- scope, a query an aql- scope, the ADMIN API the [auth] product_scope (default __ferroehr.platform), and the DEMOGRAPHIC API a client listed in [smart] demographic_clients or the product scope. A credential without the scope gets 403 with the insufficient_scope challenge. A patient/ grant counts only for an issuer in [smart] patient_issuers and is bound to the EHR the token’s ehrId claim names; a system/aql-* grant counts only for a client in [smart] backend_clients.

  • The SMART configuration document is built from the issuer’s metadata (#3872). [smart] enabled serves it, needs an issuer whose metadata names a token_endpoint, and lists only the authorization_code and client_credentials grants and the S256 challenge method. Its capabilities and scopes_supported come from [auth] capabilities and [auth] scopes_supported beside what the server enforces.

  • The terminology client fetches its client-credentials token through ferrohealth-smart (#3872). The [terminology.external.oauth2_clients] keys are unchanged.

  • The Helm chart and the compose files follow the new keys (#3872, chart 12.0.0). The chart’s Basic users are keyed by user, and the CDR credential FerroBRIDGE uses is HTTP Basic only.

  • The Keycloak overlay and the development stack serve Keycloak’s issuer over TLS (#3873). A one-shot keycloak-tls service generates a development CA and Keycloak’s certificate on the first up, Keycloak serves the issuer at https://keycloak:8443/auth/realms/ferroehr, and the CDR trusts that CA through SSL_CERT_DIR with token_profile = "jwt", so a Keycloak token is accepted again. Both realms put the SMART on openEHR scopes and the product scope into every token. The token call on port 8081 is unchanged.

  • The viewer takes FerroUI 0.2.0 and reads every word it draws from its catalog (#300, #3860). The labels it hands FerroUI (table captions, column headers, pager nouns with their plurals, navigation, tabs, fields, dialogs, toasts) are looked up in locales/en/ferroehr-viewer.ftl at run time, and the stray-literal baseline drops from 1918 to 505 literals. English remains the only locale.

Removed

  • The HMAC token key and the static bearer token (#3872). hmac_secret and hmac_secret_file are unknown keys, as are the inline jwks_json key set and allow_insecure_issuer; the Helm chart drops secrets.authOidcHmacSecret and bridge.cdr.auth. Machine-to-machine callers authenticate with HTTP Basic or an OAuth 2.0 client-credentials token from the issuer. Only the asymmetric algorithms RFC 9068 §4 names verify a bearer token, so a token signed with a shared secret is refused.
  • [smart] ehr_id_claim, patient_claim, require_smart_scopes and [smart.endpoints] (#3872). The launch context is the token’s ehrId claim, enforcement is always on, and the discovery document’s endpoints come from the issuer’s metadata.
  • [auth] scopeless_subjects is refused at boot (#3872), because a SMART on openEHR requirement admits no caller without scopes.

Fixed

  • A TDD keeps two same-archetype siblings told apart by name apart (#3833). Converting a TDD of a template whose events (or other compacted nodes) share an archetype node id and differ by name, such as a Baseline and a Follow up event, now builds two nodes, each with the content the TDD spells out under its own element, where it used to merge them into one.

  • A rotated mTLS client CA takes effect without a restart (#3769). [server.tls] client_ca_file is re-read every reload_interval_secs, as the certificate and key already were, and a changed CA verifies every new handshake; a file that fails to load is logged and the previous CA stays in force.

  • An unmounted path outside the API answers 404 (#3834). With server.swagger_ui at private or admin_only, the documentation guard wrapped every unmatched path, so /.well-known/smart-configuration on a server with SMART off answered 401, and the viewer’s System screen said the CDR refused the session. The guard now covers the documentation routes only, and the viewer shows SMART as not enabled.

  • Four viewer layout defects (#3841). The Create EHR id field’s placeholder reads “UUID, or blank” and fits; the audit log’s From and To filters take a wider share of the row so an instant shows whole, its Event column keeps “REST · E” on one line, and an opened record scrolls inside a narrow box instead of widening the page. The book captures are taken with reduced motion, so the results Table/Chart toggle no longer shows mid-fade.

  • The OpenAPI family titles carry no em dash (#3865). The served documents and the viewer’s family selector read “FerroEHR: Messaging”, “openEHR: EHR” and so on.

  • No viewer string carries an em dash (#3860). The query builder’s template choice reads “Any template”, an absent value reads “(no value)”, and the hints, notes and form errors that used a dash use a colon or a full stop.

5.0.0 - 2026-10-19

Added

  • openehr-sdt builds a canonical PARTY from a party Web Template (#3810). flat::build::build_party (and flat::convert::party_from_flat / party_from_structured) turns a FLAT or STRUCTURED document of a PERSON, ORGANISATION, GROUP, AGENT or ROLE template into the party, with every list attribute of the demographic package (identities, contacts, addresses, capabilities, roles, languages, relationships) built as an array, no composition attribute, and a refusal for a ctx/ key. A ROLE takes its performer from <root>/performer|id, |namespace, |type and an optional |id_scheme. The result is decoded as the root class before it is returned, so a party the RM refuses is reported as an error. Simplified Formats define FLAT for compositions only, so this is FerroEHR’s own extension over the same template walk. flat::build::build_root is the same walk for any non-COMPOSITION root, and flat::example::example_root_instance generates an example instance of such a root class for validation and tests.
  • The openEHR terminology serves interoperability (672) as a term mapping purpose (#3838, SPECTERM-31): TERM_MAPPING.purpose validation accepts it, and the terminology endpoints list it in all five languages.
  • The openehr-sdt builders stamp the ARCHETYPED.rm_version the caller chooses (#3809). flat::build::BuildOptions carries it into build_composition_with, build_party and flat::tdd::from_tdd_with; build_composition and from_tdd keep stamping 1.2.0.

Changed

  • The viewer draws with FerroUI 0.1.7 and is ready for more languages, English the one that ships (#300). FerroUI’s words and the viewer’s come from Fluent catalogs, the viewer’s under app/ferroehr-viewer/locales/, so a language joins as a catalog file with no code change. The locale is chosen once per request, the best language of Accept-Language the viewer’s catalog holds or else the new [ui] locale (default en), and the page writes it as <html lang>. scripts/checks/stray-literals.sh fails CI on a viewer string outside the catalog beyond the per-file baseline the remaining strings hold while they move.
  • A structure the Web Template compacts away (an ITEM_TREE under a party identity or an ENTRY, for example) is rebuilt with the name its term definition gives instead of its at-code (#3810).
  • The nine openehr-* crates move to 0.0.89.
  • Only the newest release is supported (#3853). When a new release is published, every earlier release stops receiving fixes: there are no maintenance branches, no long-term-support line and no backports. If you are not on the newest release, upgrade; upgrading is free. Under the Cyber Resilience Act, Cadasto B.V. commits to keep publishing security fixes for at least five years after each release (Art 13(8)) and ships them only in the newest release (Art 13(10)). SECURITY.md, the release notes and the compliance pages say so in place of a five-year support period per release. The end month of that commitment stays on the boot banner, in ferroehr --version and in the management documents; the banner line now reads newest release only; fixes published until the end of <month>.
  • Archetype identifiers follow the BASE 1.3.0 grammar under the development profile (#3835, #3845). The namespaced ADL 2 form (org.openehr::openEHR-EHR-OBSERVATION.blood_pressure.v2.1.4, with an optional -rc, -alpha or -beta modifier and build count) is accepted wherever an ARCHETYPE_ID is read, and openehr_base::v1_3::ArchetypeId gains namespace(), minor_version() and patch_version(). A namespaceless identifier with a full version stays accepted, because AM 2.4.0 artefacts carry it (the conflict is reported upstream, #3844); every other identifier the grammar refuses, such as a malformed namespace or version modifier, is refused naming it. The stable profile keeps BASE 1.2.0’s grammar.

Fixed

  • The viewer serves its hydration bundle on a route of its own (#3849). A /pkg/ file the bundle lacks answers 404 and a read failure 500, each logged with its path and status; neither reaches the rendered app or the sign-in redirect any more, which had answered a browser 302 for its own WebAssembly and failed the ui-e2e journeys.
  • The template example endpoints answer 400 Bad Request naming the root for a template whose root is not a COMPOSITION (an OBSERVATION or a SECTION), instead of an object carrying COMPOSITION members its class does not have (#3780). The 200 body the specification defines is an instance of a versioned-object class, which such a template does not describe.
  • Same-archetype siblings told apart only by name (two EVENTs Baseline and Follow up of one archetype node, for example) stay two nodes when a FLAT or STRUCTURED document builds, so the example of such a template passes the server’s own validation (#3779).

4.3.11 - 2026-10-19

Changed

  • The viewer draws in FerroUI, the shared FerroHEALTH component library (#3802). The shell, the navigation, the theme toggle, the tables and their pager, the dialogs, the access side panel, the upload field, the charts, the document panes with their copy button, the tabs, the tree, the toasts, the alerts, the loading skeletons, the buttons, fields, cards and empty states, the icons and the document highlighter come from ferrohealth-ui, the library FerroBRIDGE, FerroFED and FerroTERM share; thaw, leptos-use and leptos-chartistry are gone. The viewer keeps its teal and slate. The moon and sun button in the top bar switches between dark and light, the viewer follows the operating system until you press it, and the choice is a server cookie the server writes on the page, so the first paint is already in your theme and nothing is kept in browser storage. The Content Security Policy drops 'unsafe-inline' from style-src: the stylesheets are the viewer’s own files. A listing that can show patient data (the demographic tags, the EHR tags) keeps its page in the screen and out of the URL; the other listings keep ?page= (now counted from 1) and ?size= in the URL. Fields in a row share its width, every key and value list is one definition grid, a page’s delete action sits on its title row, the query builder’s controls, labels and output choice are FerroUI’s, an EHR’s subject shows once in its header, a tag table drops an empty target-path column, and a refusal no longer repeats “forbidden”. Every JSON or AQL input (the party documents, the EHR status other_details, the composition editor, the directory JSON, the staged change, the relationship details, the AQL editor and its parameters, and the template upload) is a highlighted, numbered editor that opens on an indented document, and on a phone every listing shows each row as a card with every column named. Every timestamp reads as “2026-10-18 23:18:37 UTC” on one line, with the exact value the CDR sent in its tooltip.
  • The viewer moves to FerroUI 0.1.4 (#3811). Every timestamp, the build date on the Operations screen included, is drawn by FerroUI’s shared timestamp, the same in every FerroHEALTH console. On a phone the dashboard and Operations figures are compact tiles with the glyph, the number and the label on one line, a field and its button stack with the button on its own line, and a disabled field keeps its edge in the dark theme.
  • The viewer moves to FerroUI 0.1.5 (#3815), in step with the other FerroHEALTH consoles. The screens look the same: 0.1.5 adds a per-row form field and a header title the viewer does not use yet, and its browser half no longer asks for a wasm-bindgen of its own.
  • The viewer moves to FerroUI 0.1.6 (#3820), in step with the other FerroHEALTH consoles. The screens look the same: 0.1.6 adds the shared field shapes the other consoles needed, and every viewer field already draws through the library since #3819.

Fixed

  • On a phone the CDR status pill no longer covers the FerroEHR mark in the top bar (#3812). The pill shows from a 64rem window up, where the top bar has room for its full words.
  • The viewer shows the FerroEHR mark as its tab icon and answers /favicon.ico with it, instead of a 404 on every page load (#3808).

Security

Shipped with functional changes because: no shipped artefact is affected (the VEX statements below record why), so there is no security update to separate from the viewer changes this release carries.

  • cargo deny passes again: the viewer’s WebDriver test client no longer reaches zip 8.6.0, and the three zip extraction advisories (RUSTSEC-2026-0367, RUSTSEC-2026-0368, RUSTSEC-2026-0369) are recorded as not affecting FerroEHR in the Rust-dependency VEX document, since the one remaining carrier, utoipa-swagger-ui’s build script, never calls the extraction APIs they concern (#3806; security/vex/rust-advisories.openvex.json).

4.3.10 - 2026-10-16

Changed

  • The usage report moves to FerroPULSE report format 2 (#3795). schema is 2; spec_profile, db_ok, migrations_ok and postgres_major move into a status object on both reports; the daily metrics.window_24h holds groups (the eight endpoint groups and aql, each with p50/p95/p99, 5xx count and histogram) and counters (aql_executions, aql_slow). The report carries no new information. ferroehr usage-report --print prints the new body, and the book’s usage report page lists every field.

  • The admin dump and load archive is written and read with zip 9.0.4 (#3798). The zip maintainers yanked every 8.x release, including the 8.6.0 that POST {base}/admin/dump and POST {base}/admin/load used. Nothing changes for you: the routes, the archive layout and the compression are as before. The build still unpacks the vendored Swagger UI with zip 8.6.0, because no utoipa-swagger-ui release accepts zip 9 yet; that copy runs only at build time and is not part of the server binary.

Fixed

  • The usage report is off in CI (#3794). A server started inside a CI job reported itself to FerroPULSE like a deployment. When GITHUB_ACTIONS or CI is true or 1 in the server’s environment, the server now sends no usage report and stores no instance id, whatever usage_report.enabled says, and the boot line reads usage report OFF: CI detected (<variable>). Outside CI nothing changes: the report stays on by default and the switch turns it off. ferroehr usage-report --print still prints in CI and says on stderr why the instance sends nothing.

4.3.9 - 2026-10-15

Added

  • openehr-query renders a column path with its variable (#3782). IdentifiedPath::column_path_with_variable gives c1/uid/value where column_path_text gives /uid/value, so a federation gateway can keep two variables of one class apart in RESULT_SET.columns[].path. FerroEHR’s own result sets are unchanged.

  • The Helm chart deploys FerroBRIDGE beside the CDR (#3713, chart 11.1.0). bridge.enabled=true installs FerroBRIDGE 0.0.5 (pinned by digest) as its own workload under the same Restricted security context, with its config.toml pointed at the release’s CDR Service, its CDR credential read from the Secret named in bridge.cdr.existingSecret, its mapping files from bridge.mappings.existingConfigMap, and its terminology server set to the chart’s FerroTERM when terminology.enabled is on. FerroBRIDGE authenticates no caller yet (FerroHEALTH/FerroBRIDGE#400), so the chart keeps it cluster-internal: a ClusterIP Service, no Ingress, and a NetworkPolicy that admits only the peers in bridge.networkPolicy.ingressFrom; an empty list or an admit-everything peer is refused. The image is listed in artifacthub.io/images, so Artifact Hub scans it.

  • AQL reads variables declared inside an OR containment branch (#3736). CONTAINS (OBSERVATION o[...] OR EVALUATION ev[...]) with paths through o and ev in SELECT, WHERE or ORDER BY answers one row per match, with null for the branch that did not match, and both branches matching pair up; it was refused with “not supported yet”. A branch that is its own versioned object (EHR e CONTAINS (COMPOSITION c1 OR COMPOSITION c2), or a composition a FOLDER references) reads the current, unrestricted, undeleted version with its uid (#3745). Reading a variable declared under NOT CONTAINS is refused with a named 400.

  • An admin route physically deletes one demographic party (#3683). DELETE {base}/admin/party/{party_id} removes the party and its relationships (SM physical_party_delete), admin-gated and recorded like the EHR physical delete. It is our own extension; no ITS-REST operation governs it.

  • A live probe of values-production.yaml (#3692). scripts/deploy-probe-k8s.sh --production-only installs the production overlay against four PostgreSQL clusters, one per storage domain, and checks that the server boots and that it refuses when domains share a cluster; --render-check runs without a cluster. The probe runs on kind with podman as well as Docker.

  • openehr-its: a bound on the bytes the REST client reads of one answer (#3600). ReqwestTransport::with_max_body(limit) refuses a body whose declared length is past the limit without reading it and cuts a chunked body off at the limit; the call fails with the new ClientError::AnswerTooLarge, which carries the answer’s status and is never retried.

  • [authz.abac] patient_namespace (FERROEHR__AUTHZ__ABAC__PATIENT_NAMESPACE, #3771): the subject namespace patient-claim ids are issued in. When set, a claim names the EHR subject with that id in that namespace; when unset, an id held under two namespaces matches no EHR.

Changed

  • The release pipeline checks the linkage the OS VEX statements rest on (#3694). Each statement that argues from what the binary loads now carries that linkage as data (the needed libraries, the loader, the imports it rules out), and the release lane refuses a ferroehr or ferroehr-viewer binary, on either architecture, that needs another library or imports a function a statement rules out.

  • openehr-its: the template-example response body is typed (#3549). definition_template_adl1.4_example_get and definition_template_adl2_example_get answer a _type-dispatched TemplateExampleRetrievedBody over their eight RM classes instead of untyped JSON. The duplicate XSD tree schemas/xml/components/, which no code read, is removed (#3753).

  • The vendored openEHR specifications move to current upstream (#3716, #3717, #3718, #3719, #3720, #3721, #3723). BASE, RM, AM, TERM, LANG, CNF and the ITS-JSON docs follow upstream master, as the development generations are pinned; upstream agent files (CLAUDE.md) and build tooling (development/) are no longer vendored. ITS-XML stays on its release tags, and the spec-update watcher now reads a tag pin as one (#3722). The generator’s inputs (BMM, ITS-JSON schemas, ITS-XML XSDs, ITS-REST OAS) are fetched by scripts/vendor/codegen-inputs.sh, which CI runs with --check.

  • TERM 3.1.0 amendment 3.1.4 (SPECPR-450) (#3717, #3727). Template upload accepts a DV_QUANTITY constrained to the properties 761–765 (flow rate per area, volume per area, mitotic count, length per area, concentration per volume), and FLAT/STRUCTURED quantity validation knows their units (mg/m², mm/m², ng/mL/cm³, μg/L/cm³); a unit filed under a different property is refused under 764 and 765.

  • AQL result row counts change for paths through repeated nodes (#3733, #3734). A SELECT path through a repeated step returns one row per node where it returned one row with an arbitrary node’s value; a whole-object projection of an absent path keeps its row with a null cell; a path step no longer reaches past its direct child, so c/content[...] does not find an entry inside a SECTION. Clients that counted rows per composition need to group them. The rules are under Fixed.

  • An AQL text that does not tokenise names the pass in its 400 (#3684). The body reads invalid AQL: the query could not be tokenised: <lexer diagnostic>, and a stored query the same after its own prefix; the diagnostic is unchanged. Bearer-token refusals log which check refused the token and its cause, each once; the 401 body is unchanged.

  • The conformance pipeline fetches Veredictum from its new home (#3594), https://github.com/Vernum-Projecten/Veredictum.

Security

Shipped with functional changes because: the patient-scope fix (#3771) is built on this release’s rewritten AQL row bindings, the per-version identifier rows (#3772) and the audit digest version 2 (#3686) need this release’s party and audit migrations, and the runtime-role grants need its new bookkeeping migrations, so a security-only patch from the v4.3.6 tag cannot carry them.

  • Protected national identifiers stay protected on every path (#3772, GHSA-4w9c-h7v6-xmh2). A party create or update response no longer echoes the plaintext identifier: with protection on it is the stored version, carrying the reference a later read serves (relationship write responses also serve the stored form under multimedia offload). An EHR extract or an admin load whose party carries a protected identifier in plaintext is refused (422, or a per-container failure for a load), because an imported ORIGINAL_VERSION stays a faithful copy (RM common §Copying). Admin dump archives carry the sealed rows as ciphertext only, and a load restores them under the same deployment key. Physically deleting a party removes every sealed row of the party and its deleted relationships. A scheme in [demographic.identifier_protection].schemes that the database registry does not hold is refused at boot. The served OpenAPI declares the 409 and the identifier 422 cases on the party create and update operations.

  • The runtime roles can no longer write their migration bookkeeping (GHSA-jf68-mmp7-grjf). The shipped grants on every table of a domain schema covered its _sqlx_migrations; new migrations (clinical 0014, party 0009, linkage 0005) leave the runtime roles SELECT only, and the server refuses to start while a runtime role can write a bookkeeping table. The boot check of database grants now runs in every database a domain pool reaches, so a domain in its own database is no longer left unchecked.

  • ABAC patient scoping covers queries that read only the EHR row (#3771, GHSA-xqc2-gwmg-jhmg). With [authz.abac] on, SELECT e/ehr_id/value FROM EHR e, COUNT(*) FROM EHR e, e/ehr_status/... paths, EHR e CONTAINS (A OR B) and an ehr_id-scoped query naming another patient’s EHR returned EHRs of other patients. The patient scope now applies to the EHR row and to every versioned object, on every query plan; the policy-decision post-check runs for those queries too, and a subject id held under two namespaces admits neither EHR, since the patient claim carries no namespace. A touched EHR or template id that cannot be read fails the query instead of being left out. The direct EHR, composition and status routes apply the same rule: an EHR that records no subject, or whose subject id another namespace also holds, answers 403 to a caller with a patient claim. The query policy check now receives the subjects of the touched EHRs as its patient attribute, so a patient-keyed rule applies even without a configured claim, and it sees the EHR_STATUS template on e/ehr_status/... paths.

  • The viewer’s Basic sign-in follows only a same-origin next (GHSA-23g2-35q8-43rq). After a successful or refused sign-in, an absolute, scheme-relative or backslash destination lands on the dashboard instead (an open redirect).

  • rustls 0.23.46 (GHSA-wjr5-jv48-6hf2) closes RUSTSEC-2026-0346, -0347, -0348 and -0349. Of the four, RUSTSEC-2026-0349 reached FerroEHR’s TLS clients; the other three need a custom session store, Encrypted Client Hello or the brotli feature, none of which FerroEHR uses.

  • An unresolvable signing key set no longer discloses the identity provider’s URLs (#3666, GHSA-rch2-2r9h-cjwv). The 503 names only the failed step (issuer URL, OIDC discovery, key-set fetch or read); it quoted the transport error, which named the provider’s discovery and JWKS URLs. The cause is still logged.

  • ehr_access_default = "restricted" holds with RBAC off (#3742, GHSA-rhf2-6x3j-6qrh). [authz.rbac] ehr_access_default = "restricted" was ignored when [authz.rbac] enabled = false, leaving every EHR without ACCESS_CONTROL_SETTINGS open on its path and in AQL, while the deployment posture check stopped reporting the open default. The default is now enforced with RBAC on or off. With [auth] enabled = false the combination is refused at boot, since no caller carries a role. A new deployment gap, rbac_off, is reported when authentication is on and the RBAC gate is off; production refuses it unless it is accepted by name. [authz.rbac] role_claims now applies to role extraction with RBAC off as well.

  • The audit hash chain covers every column of an access record (#3686, GHSA-29xh-548g-g2mq). Digest version 2 hashes every column (purpose, legal basis, categories, assurance level and value, professional, the emergency mark and the rest), so a change to any of them shows in audit.verify_audit_chain(). Records written before the upgrade keep verifying under version 1, and the verifier reports a record moved back to an older digest version. A test fails when a migration adds a column the digest does not cover.

Fixed

  • The release pipeline accepts an arm64 program as built. The VEX linkage check required the dynamic loader in DT_NEEDED, where glibc lists it on amd64 and not on arm64, so it refused both arm64 binaries. It now matches the declared loader against the program interpreter (PT_INTERP) and accepts it in DT_NEEDED without requiring it. The v4.3.7 tag carries this content without the fix, and the v4.3.8 tag without the OS VEX statements below; neither release was published, and a tag cannot move, so the release ships as 4.3.9.

  • Three OS findings in the database image carry a VEX judgement. Debian has no fix yet for CVE-2026-107161 (Cyrus SASL DIGEST-MD5 client), CVE-2026-107708 (the krb5 KDC) or CVE-2026-107778 (libkrb5 rd_cred.c) in ferroehr-postgres. The first two are not reached: PostgreSQL never starts a SASL bind and the image ships no KDC. CVE-2026-107778 is reached only through a GSSAPI configuration FerroEHR does not ship; it is affected with an action statement in security/vex/postgres-os.openvex.json and joins #3708.

  • The chart’s SBOM is attested (#3710). The CycloneDX SBOM the chart lane writes carried no serialNumber, which actions/attest requires of a CycloneDX document, so chart 11.0.1 was published without its SBOM attestation. The generator now writes a urn:uuid: serial number and refuses to hand on a document without bomFormat, serialNumber and specVersion. Chart 11.0.2 carries the attestation.

  • The POST /query/aql example in the served OpenAPI runs (#3737). Its request and response examples sorted on the SELECT alias (ORDER BY temperature DESC), which the server refuses with 400: AQL 1.1.0 §ORDER BY sorts on an identified path. The examples now sort on o/data[at0002]/events[at0003]/data[at0001]/items[at0004]/value/magnitude.

  • Stale statements corrected (#3743). CITATION.cff and the Zenodo record no longer say a built-in runner verifies conformance at every release; they name Veredictum, the separate CNF 2.0 runner, and the committed run records. The Helm chart description drops “hardened-by-default” and states that the default NetworkPolicy admits every source until networkPolicy.ingressFrom is set and that the default-deny egress policy is off. SECURITY.md names GET /management/status as the endpoint that reports support.status, and the README calls Veredictum separate rather than independent.

  • More stale statements corrected (#3743). The operations guide and the CRA user information and risk assessment pages name GET /management/status as the endpoint that reports support.status; the public GET /ferroehr/rest/status carries only status, server_version, openehr_rest_api_version, timestamp and deployment.profile. The Conformance and Benchmarks pages call Veredictum a separate runner instead of an independent or built-in one. MAINTAINERS.md states the support period as SECURITY.md does: five years per release, with every fix in the newest release. The conformance statement no longer declares the tenancy extension, which was removed in 4.3.1. The Zenodo record’s references and related identifiers cite the pinned specification releases (RM 1.2.0 development, BASE 1.3.0, TERM 3.1.0, AM 1.4 and 2.4.0, QUERY 1.1.0, ITS-REST 1.1.0), and its notes name Veredictum. The provenance record of Delegated Regulation (EU) 2026/881 names the FerroHEALTH post-market page as the page that cites it (#3706).

  • AQL no longer pairs values from different nodes in one row (#3733). Two paths in SELECT under one repeated node (the systolic and diastolic of an EVENT) read the same node: each structure step a SELECT path descends is bound once per row, one row per node, and paths through different repeated steps give every combination of their nodes. A path that reaches nothing reads null and keeps its row; a whole-object projection used to drop that row. In WHERE, every step a path descends that SELECT has not bound is bound once for the whole clause, and a row is kept when some assignment of those bindings makes the clause true under SQL three-valued logic, so sys >= 140 AND dia >= 90 needs both readings on one event, De Morgan’s laws hold, and NOT (x > 140) over readings 120 and 150 holds. Aggregates read the items of their path that a projection under the same WHERE would keep, without adding rows, so COUNT(*) next to MAX, SUM or AVG is unchanged; COUNT over a list attribute counts its elements and ignores a stored JSON null. EXISTS is path-level (AQL 1.1.0 §EXISTS): it holds when data at the path exists, read per node only where SELECT binds the path’s prefix. A test of a list attribute inside one node’s data (links, participations) holds when any element satisfies it and is unknown when the list is empty, so NOT (c/links/target/value = 'x') drops a composition without links. AQL 1.1.0 does not define the grain over a repeated attribute; the book states the rules.

  • AQL CONTAINS and path steps no longer match the node itself (#3734). CLUSTER c1 CONTAINS CLUSTER c2 returned each cluster as its own child, NOT CONTAINS of the node’s own class never held, and a path step could bind the node it started from. A path step now follows exactly one attribute to a direct child (BASE paths §Basic Syntax), so c/content[openEHR-EHR-OBSERVATION.x.v1] no longer finds an observation nested inside a SECTION and items[at0004] no longer finds an element in a sub-cluster; CONTAINS still finds both.

  • A deleted composition stays out of unscoped AQL queries (#3735). A query with no EHR scope that read nothing from the composition’s content (SELECT c/uid/value, SELECT e/ehr_id/value, SELECT 1) returned a row for the deleting version.

  • A CONTRIBUTION cannot delete an already deleted version (#3738). It was accepted with 201 and added another deleted version, while the DELETE route answered 400. Both now answer 400 with the same message (ITS-REST 400_already_deleted), and nothing is committed. A second DELETE of an already deleted directory answers 400 as well.

  • Example compositions pass the server’s own validation (#3739, #3740). GET /definition/template/adl1.4/{template_id}/example now carries every attribute the template makes mandatory, an EVENT state or ENTRY protocol without data points included, every SECTION the template requires but names only, a container’s minimum number of children, and the archetyped other_context content a template makes mandatory, and never sets value and null_flavour on one ELEMENT (RM Inv_null_flavour_indicated). The cases it still cannot cover are #3779 and #3780.

  • A quoted query parameter on a GET query binds the string it names (#3781). On GET /query/aql?q=… and GET /query/{qualified_query_name}, uid="123" now binds the string 123, as {"uid": "123"} does in a POST body; the quotes were kept as part of the value. An unquoted value reads as before.

  • The web template keeps an ELEMENT’s value beside its null_flavour. An ELEMENT constraining a text value and a coded null_flavour was folded into one coded node with an other input, which dropped the value. That fold now applies only to two alternatives of one attribute (Simplified Formats master04 §“Open Value-Sets and the |other Suffix”).

  • XML without the required id of a key/value element is refused (#3741). An OPT 1.4, archetype or RM XML document whose original_author, other_details or other StringDictionaryItem element has no id attribute is refused with 400, as the ITS-XML schema requires, instead of being stored under an empty key. id="" is still accepted.

  • [events] tls = true upgrades a broker URL in any case (#3690). AMQP:// and Amqp:// now become amqps://, as amqp:// did; the URL scheme is case-insensitive (RFC 3986 §3.1).

  • A database the server cannot connect to answers 503 (#3659). A refused connection, a TLS setup failure, an authentication or unknown-database refusal, too many connections and a database still starting answer 503 with Retry-After instead of 500. A failure on an established connection stays 500.

  • Concurrent imports into one container are serialised (#3658). The second EHR Extract import waits for the first and lands after it, or is refused as a re-import with the specific 409, instead of failing on a unique-index collision.

  • Errors print their cause once (#3684). A chained error report repeated the cause in the wrapper’s own message for database, storage, service, AQL, terminology, multimedia, eventing, XML, Simplified Format and BMM errors; each now carries the cause as its source only, and response bodies are unchanged. A version-signing failure and a RESULT_SET assembly failure now log their cause, and a failed terminology OAuth2 token grant logs its transport cause. scripts/checks/error-source-chain.sh refuses the shape.

  • A CONTRIBUTION and an import can no longer deadlock (#3658). A CONTRIBUTION that changes several objects locks them in object-id order, the order an import uses, and an import takes the EHR row lock first. The check that an EHR has no second EHR_STATUS or EHR_ACCESS now runs inside the import’s transaction, so two concurrent imports cannot both add one.

  • ferroehr report separates an unreachable database from a failed read (#3659). It says “unreachable” only when no connection could be opened, and “unreadable” when a connected read failed.

  • The operations guide no longer calls a partial schema drop a supported reset (#3685). Dropping clinical, ext and audit also drops the btree_gist extension the linkage constraint depends on, and the next boot does not recreate it; the guide now says to drop the database.

  • ferroehr-ext compiles cleanly in every feature combination (#3703), and CI runs clippy over each one, and over the server’s multimedia lane.

  • The viewer’s /system page links the Messaging OpenAPI family (#3704). The viewer and the server now read one shared family list, and a test on each side fails when they disagree; the deployment gap codes the sandbox notice names, rbac_off included, follow the same shared list.

  • The operations panel shows every object of /management/info (#3632). The build card listed only top-level values, so the manufacturer, the support period and the audit posture never appeared; each now has its own heading.

  • A JWK Set that fails to parse is reported without its content (#3666). The error carries the parser’s error class, line and column only, for an inline jwks_json and for a set fetched from the issuer, so a 503 body can no longer quote key material.

  • The DICOM audit message and the FHIR AuditEvent agree on the data flow (#3687). On a read the server is the source and the client the destination; on a write or a query the client is the source (DICOM PS3.15 §A.5.2.7 and §A.5.3.10, IHE BALP).

  • The served OpenAPI document repeats no operationId (#3691). The demographic CONTRIBUTION operations are served as demographic_contribution_create and demographic_contribution_get, the names of the released operation files; the access record keeps the released ids contribution_create and contribution_get.

  • [authz.rbac] role_claims applies with every gate off (#3742). Bearer role extraction read the default claims when RBAC, ABAC and a restricted per-EHR default were all off; the boot log now also says when only a restricted per-EHR default switches authorization on.

  • The Helm chart’s in-chart clients follow the server’s own TLS (#3714, chart 11.1.4). With config.server.tls.enabled, the viewer, FerroBRIDGE, the ServiceMonitor scrape and the kubelet probes use https, and the viewer and FerroBRIDGE trust the CA named by the new service.tls.caExistingSecret; the chart refuses to render a TLS listener beside an in-chart client without one. The migration Job runs under its own hook ServiceAccount, so a first install no longer waits on an account Helm creates only after the hooks.

  • Storing a query runs the analysis an execution runs (#3748). PUT /definition/query/{name}[/{version}] parses, lowers and generates the SQL of the query without executing it, so a query the engine would refuse answers 400 instead of being stored and failing on every execution. Terminology resolution still happens at execution (AQL 1.1.0 §TERMINOLOGY).

  • FHIR AuditEvents for deletes carry the agent types BALP Delete fixes (#3750): client DCM 110150 Application, server custodian.

  • OPT 1.4 conversion keeps the type of a primitive node without an inner constraint (#3752). An Integer, Date or Boolean node no longer becomes a String constraint.

  • A composition built from FLAT, STRUCTURED or TDD input without a category takes it from the template (#3652): its single admitted code, 433 when the template leaves the category open, and a refusal naming …/category|code when several codes are admitted and none is assumed (an assumed category value is taken). It was always 433, even where the template admits only 431.

  • The Web Template carries defaultValue from archetype assumed values. Coded text, quantity, ordinal, count, boolean, text, date/time and duration inputs take the assumed value of their constraint, as a string (ITS-REST Input.yaml), and only when the constraint admits it (AOM Assumed_value_valid), for OPT 1.4 and ADL2/OPT2 templates alike.

  • OPT 1.4 conversion keeps enumerated integer and real lists (#3752); only the range was carried.

  • The server serves a renewed TLS certificate without a restart (#3765). It re-reads [server.tls] cert_file and key_file every reload_interval_secs (new key, default 60; 0 reads them once at boot); a pair that fails to load is logged with its cause and the previous pair stays in service. Pods stayed Ready on an expired certificate because the kubelet’s HTTPS probes do not verify it.

  • An EC private key loads for [server.tls] in SEC1 or PKCS#8 (#3761), and a certificate whose EC key carries explicit curve parameters (RFC 5480 §2.1.1) is refused with the cause and the named-curve remedy instead of a bare key mismatch.

  • The viewer trusts an identity provider behind a private CA (#3767). Its OIDC client verifies against the platform trust store (SSL_CERT_DIR, SSL_CERT_FILE) instead of the bundled web roots only.

  • A wrong password at the viewer’s Basic sign-in returns to the sign-in card with the refusal (#3759) instead of answering 500.

  • The CRA provenance record lists the eighth corrigendum (#3732). 32024R2847R(08) corrects the Slovenian version only (Art. 3(17), Art. 14(7) and the titles of Articles 21 and 22); it has no English version and changes no text FerroEHR cites.

  • The four-cluster production layout prepares and boots (#3755, #3756, #3757, #3762). A database that holds the linkage domain without the party domain also receives the party migration set (schema and grants only), and a relocated domain’s database is prepared and verified on its own migration credential, the new [storage.migrate.<domain>] url/url_file (chart: database.<domain>.migrateExistingSecret). The chart’s migration Job prepares every cluster, and scripts/deploy-probe-k8s.sh --production-only proves the boot and the shared-cluster refusal live.

  • Chart: TLS and migration-hook corrections (#3760, #3766). The hook ServiceAccount survives a failed hook and is created only when the chart creates the release’s account; client_auth: required with server TLS is refused at render, since kubelet probes present no client certificate; an Ingress in front of a TLS listener needs ingress.backendTlsAnnotations; prometheus.io/scheme follows the listener.

  • The CDR trusts an identity provider behind a private CA (#3767). OIDC discovery and JWKS fetches verify against the platform trust store (SSL_CERT_DIR, SSL_CERT_FILE), as the viewer now does; reqwest 0.12 and its bundled roots leave the dependency graph.

  • openehr-lang reads BMM packages without packages (#3768). The generated BMM JSON reader (and its AM 2.4 re-emission) reads a package or schema without packages as an empty package map and leaves an empty map out on write, as the LANG BMM persistence syntax chapter’s example and every published openEHR BMM do; the class table’s 1..1 is reported upstream (#3770). openehr-lang also reads P_BMM schemas from their JSON serialisation (json_reader::read_schema_json), producing the same P_BMM_SCHEMA as the ODIN reader, uid and source_schema_id stamping included; its BMM tests now read the JSON files the generator reads, and the stale ITS-BMM ODIN copies of BASE 1.3.0, RM 1.2.0 and TERM 3.1.0 are no longer vendored.

  • A database that holds some of the storage domains gets only their schemas (plus ext); schemas only a shipped grant file names are created for the run and dropped afterwards, and ferroehr db erase removes empty schemas of other domains that earlier releases left in such a database.

  • PostgreSQL healthchecks check over TCP in the compose files, the hosted sandbox and CI, so a container is no longer reported healthy while the image’s initialisation server still runs.

  • Every connect-time database failure answers 503 (#3749). Each pool checkout goes through one typed acquire, so an unresolvable or unreachable host and a protocol failure while connecting answer 503 with Retry-After; a failure on an established session stays 500. The db_pool_acquire_duration metric is now recorded for every checkout.

  • Chart: the viewer reaches a management surface on its own port (#3758). It is given FERROEHR_VIEWER__CDR__MANAGEMENT_BASE_URL, and its NetworkPolicy admits the management port when config.management.port is set.

  • AQL SQL generation is deterministic with several EHR sources, and an AQL execution that cannot obtain a database connection answers 503 with Retry-After (#3749).

  • Protected national identifiers are sealed inside the party commit (#3772). A failed demographic commit no longer leaves a sealed identifier behind, and a failed update no longer replaces the party’s stored one. Each version of a party keeps the identifier it was committed with: a changed identifier is sealed into a new row instead of overwriting the old one, so an earlier version shows the value it held, and lookup by identifier answers the latest version only (party migration 0010_identifier_versions). A party whose identifier another party already holds answers 409; two values in one scheme, or a reference to an identifier the party does not hold, answer 422.

  • A patient-scope check that cannot reach the database answers 503 with Retry-After instead of 500 (#3749).

  • Error texts without stray space runs. The 400 for an AQL parameter that cannot be coerced, the event-subscription update 400, the served OpenAPI description of the demographic party_id parameter and an ADL parser refusal had lost their line continuations.

  • The compose image-tag guard covers the terminology overlay (#3605). scripts/checks/compose-image-tags.sh checks every FerroEHR image tag in every root and docker/ compose file, docker-compose.terminology.yml included, and proves it with a --self-test.

4.3.6 - 2026-10-08

Added

  • The server verifies the audit hash chain on a schedule (#3640, CRA Annex I Part I(2)(f)): [audit.store] verify_interval_seconds (default one day, 0 turns it off). A finding is logged at ERROR, counted in atna_audit_chain_findings and reported as the audit_chain indicator.

  • The running server states its support period (#3639): five years from the release month (CRA Art. 13(8)) on the boot banner, in ferroehr --version, on GET /management/info and on the REST status document (support.status is supported, ended or unreleased). Once the period has ended it logs a WARN at boot and daily (CRA Art. 13(19)).

  • More access records carry EHDS priority categories (#3653): logical composition deletes, CONTRIBUTION reads with or without resolve_refs, VERSIONED_COMPOSITION reads, and EHR Extract export and import records; EHR_ACCESS is classified none.

  • FerroEHR’s intended purpose is stated (#3641). The new book page Intended purpose is Cadasto B.V.’s statement of who FerroEHR is for, the data it is designed to process (with the EHDS Art. 14(1) priority categories), the environment it runs in, the security environment it assumes (TLS termination, an identity provider, database protection) and the misuse it foresees (CRA Art. 3(23), Annex II point 4; EHDS Annex III 1(a)).

  • A CRA cybersecurity risk assessment (#3635). The new book page CRA risk assessment rates FerroEHR’s cybersecurity risks, states for each point of CRA Annex I Part I(2) whether it applies, how it is implemented, the evidence and the open issue, justifies the limbs that do not apply (automatic updates, application-level encryption at rest, configuration integrity), and lists every outbound data flow, the usage report included (CRA Art. 13(2) to (4)).

  • A hazard log for the access log (#3618). The new book page Hazard log lists nine patient-safety hazards of the European logging software component, each with its control, the tests that show it holds and what remains (EHDS Annex II 1.1).

  • A claims review (#3619). The new book page Claims review records the reading of every public text against EHDS Art. 28, release by release, and re-assesses EHDS Annex II 2.5 against the authorisation defaults.

  • An information sheet and instructions for use accompany every release (#3617). The new book pages Information sheet (every point of EHDS Art. 38(2): the manufacturer, the name, version and release date, the intended purpose, the Art. 14(1) data categories, and the standards and their versions) and Instructions for use (what FerroEHR runs on, an EU deployment’s configuration with deployment_profile = "production", [audit.categories] and retention_years, the maintenance it needs and how often, and the known limitations EHDS Art. 28(b) asks to be stated, with the CRA Annex II user information mapped point by point). The release lane attaches both, cut from the tag, to each GitHub release (scripts/release/accompanying-documents.sh).

  • The CRA information and instructions to the user, linked from every release’s notes (#3650). The new book page CRA information and instructions to the user carries each point of CRA Annex II in order (the manufacturer, the vulnerability contact, identification, intended purpose and security properties, the risks, the declaration of conformity, the support period, the instructions of point 8(a) to (f) and the SBOMs), each with its source. The release notes link the copy frozen for the tag, and the release lane refuses notes without that link or the support period (scripts/checks/release-notes.sh, CRA Art. 13(18)). Rust crates gains the information for integrators of the openehr-* crates that point 8(f) asks for.

  • The technical documentation of EHDS Art. 37 and CRA Annex VII (#3616) is kept in docs/technical-documentation/, one file per Annex III element and per CRA Annex VII point that adds one, each citing its evidence. A release that changes the logging component updates it (scripts/checks/technical-documentation.sh --since <previous tag>, EHDS Art. 30(2)), and the tree states how each release’s documentation is kept for ten years, where the declaration of conformity will be published, and how a translation is requested.

  • SECURITY.md names how to hear of new releases (#3639): subscribe to the release feed (https://github.com/FerroHEALTH/FerroEHR/releases.atom) and poll the published advisories (GET https://api.github.com/repos/FerroHEALTH/FerroEHR/security-advisories), since the server does not look for newer releases itself.

  • A Cyber Resilience Act page (#3626) states FerroEHR’s position under the CRA: Cadasto B.V. as manufacturer, why BUSL-1.1 is not free and open-source software under CRA Art. 3(48), the reporting duty in force since 11 September 2026, the five-year support period, the conformity route through the EHDS (CRA Art. 32(5a)), the default-product reading beside the competing Annex III class I reading, the hosted-service case, and the questions for counsel.

  • Vulnerabilities can be reported by email (#3646): SECURITY.md, security.txt (a second Contact: mailto:info@cadasto.com) and the Complaints, incidents and vulnerabilities page name info@cadasto.com beside GitHub private vulnerability reporting, so a report needs no GitHub account (CRA Art. 13(17)). An emailed report runs through the same procedure and starts the same CRA Art. 14 clock.

  • Component vulnerabilities are reported upstream (#3648). Cadasto B.V. reports a vulnerability it finds in an integrated component to that component’s maintainer and offers its fix (CRA Art. 13(6)), and records each report in the new public register docs/registers/upstream-reports.tsv.

  • What happens if Cadasto B.V. ceases operations (#3651): SECURITY.md and the Complaints, incidents and vulnerabilities page say whom the manufacturer tells, in which order, that the support periods end on the cessation date and that every published release stays published (CRA Art. 13(23)).

  • Every access record carries the EHDS priority categories of the data it served or wrote (#3621, Annex II 3.2(c)). A new [audit.categories] map classifies an access from the template id, then the root archetype id, as patient-summary, eprescription, edispensation, imaging, test-results, discharge-report, national:<code> or none. An access the map cannot classify is recorded unclassified with its ids and is never refused. The map’s digest is on the boot line, on GET /management/info and on every record, and the FHIR AuditEvent carries the classification as entity details.

  • The retention register can key a period on an EHDS priority category (#3621, Annex II 3.4): category on PUT /admin/retention/policy, echoed by the policy and due lists. The category map is mirrored into the register at boot, and an unclassified object takes the longest period configured for its kind.

  • [audit.store] retention_years (#3625) states the access-log horizon in calendar years as an alternative to retention_days. It is compared exactly against the retention floors and ceilings, which are now registered in years, and reaped with calendar arithmetic; setting both keys is a boot error.

  • The regulation corpus gains the EHDS implementing acts and the market-surveillance acts (#3607). docs/law/eu/ now carries Commission Implementing Regulations (EU) 2026/2083 (MyHealth@EU) and 2026/2099 (cross-border identification and authentication), both applying from 26 March 2027; Commission Recommendation (EU) 2019/243 on the European electronic health record exchange format, which EHDS recital 26 names; Regulation (EU) 2019/1020 on market surveillance, whose Article 3 definitions EHDS Article 2(1)(d) takes over, at its consolidation of 12 August 2026; Regulation (EC) No 765/2008, whose Article 30 sets the general principles of the CE marking, at its consolidation of 16 July 2021; and CRA Implementing Regulation (EU) 2025/2392 on the important and critical product categories. Each is fetched by scripts/vendor/law-eu.sh with its own provenance record and digests.

  • The CRA record carries the Regulation’s three English corrigenda (#3607). They correct the title, Article 64(10) (“paragraphs 2 to 9”, where the published text says “3 to 9”) and the point number Article 67 adds to Directive (EU) 2020/1828. The vendored OJ text includes none of them, and docs/law/eu/cra/PROVENANCE.md says which file to cite for which provision.

  • Complaints, serious incidents and exploited vulnerabilities have a written procedure (#3613). The new book page Complaints, incidents and vulnerabilities names FerroEHR’s manufacturer, Cadasto B.V., and its single point of contact, the channels (GitHub private vulnerability reporting; info@cadasto.com with “FerroEHR incident” or “FerroEHR complaint” in the subject; a public issue), the public registers of complaints and of non-conforming versions under docs/registers/, and what happens on a non-conformity, an EHDS serious incident (reported within three days of awareness, EHDS Art. 44(7)) and an actively exploited vulnerability or severe incident (the CRA Art. 14 early warning within 24 hours, notification within 72 hours and final report), with NIS2 kept separate. docs/post-market.md is the maintainers’ checklist; SUPPORT.md and the issue chooser point to the page.

  • A non-conforming release can be withdrawn (#3615). scripts/release/withdraw.sh <version> --to <version> refuses a version the register of non-conforming versions does not name, moves the <major>.<minor> and latest tags of the server, viewer and PostgreSQL images that still point at it to the correcting release, and prints the advisory, the SECURITY.md line, the notices to authorities and users, and the register columns to fill. The version’s own tag, its digests and its chart version stay published, so a deployment pinned to them keeps running it until it is moved; an OCI chart registry cannot mark one chart version deprecated, and the notice says so. --dry-run prints every command and changes nothing.

  • The regulation corpus gains the CRA delegated acts and the NIS2 corrigendum (#3627). docs/law/eu/ now carries Commission Delegated Regulation (EU) 2026/881 (when a CSIRT may delay passing on a CRA Art. 14 notification, CRA Art. 14(9) and 16(2)) and (EU) 2025/1535 (the CRA Art. 2(5) exclusion of products within Regulation (EU) No 168/2013), and the English corrigendum to NIS2 Art. 19(1) beside the NIS2 text, whose record now says which file to cite. docs/law/README.md records Implementing Regulations (EU) 2026/771 (the EHDS Board) and 2026/2098 (dataset metadata for secondary use) as adopted and not vendored, with the reason.

  • Every release states its support period (#3610). Cadasto B.V. is the manufacturer of each tagged release under the Cyber Resilience Act, and SECURITY.md now gives each release a support period of five years from the month it is published (CRA Art. 13(8)), with the end date printed in a new “Support period” section of the release notes (Art. 13(19), Annex II point 7). The security update for a release in its support period ships in the newest release (Art. 13(10)).

  • SECURITY.md describes the CRA Art. 14 notifications (#3609): the early warning within 24 hours, the notification within 72 hours and the final report Cadasto B.V. sends through ENISA’s single reporting platform to the Dutch CSIRT for an actively exploited vulnerability or a severe incident, and how users are told through the security advisory.

  • A weekly EUR-Lex watcher (#3608) files one issue per new act adopted under, amending or correcting the EHDS (Regulation (EU) 2025/327) or the CRA (Regulation (EU) 2024/2847), naming the article it implements, against a committed baseline of the 18 acts known on 2026-10-06.

  • The control matrix knows the CRA (#3606) as a legal source, so a control can declare a CRA article.

  • The running system names its manufacturer (#3612), as EHDS Art. 30(1)(g) asks: Cadasto B.V., Comeniusstraat 2d, 1817 MS Alkmaar, The Netherlands, info@cadasto.com, https://www.cadasto.com/contact/. It appears on the startup banner, in ferroehr --version, as a manufacturer object on GET /management/info, in the viewer’s footer and a Manufacturer card on /system, and in an eu.ferroehr.image.manufacturer label on the three container images, all read from one source file.

  • ferroehr report (#3614) writes one redacted JSON document naming what a deployment runs: version, commit, build date, enabled features, specification pins, openehr-* crate versions, the migration level of each schema set, the deployment posture, the licence summary, the manufacturer and the effective configuration. A manifest names any part it could not read and why. No credential, URL userinfo or patient data reaches it: its configuration passes the same userinfo masker as ferroehr config check (#3629). Attach it to a complaint or a serious-incident report.

  • The Helm chart carries the access-log category map and the calendar-year horizon (#3655): config.audit.categories (empty templates and archetypes, as in the binary) and config.audit.store.retention_years (unset) in values.yaml, typed by shape in values.schema.json. Chart 11.0.1.

  • The Helm chart and the Compose files declare the deployment profile (#3637): config.deployment_profile: sandbox, with a comment listing what production refuses under the chart defaults, and deployment_profile = "sandbox" in the quickstart, dev and hosted-sandbox configurations. The new book page Returning to the original state resets an instance’s configuration and database, and the deploy-probe harness runs the Compose procedure.

  • Every published artefact has an SBOM (#3643): each release attaches a CycloneDX SBOM per openehr-* crate, attested against the .crate archive crates.io serves, and each chart version attests a CycloneDX SBOM naming every image it deploys by digest.

  • Every release publishes an exploitability record (#3636): every image and server binary is scanned with no severity floor and with unfixed findings included, each finding is joined with its OpenVEX statement in ferroehr-<tag>-vulnerability-record.json, and the release is refused, before any image is tagged, while a finding has none.

  • The CRA risk assessment declares its security-relevant paths (#3649). A release that changes one of them since the previous release is refused until the assessment gains a revision row.

  • ferroehr db erase permanently removes all data and settings of an instance (#3642; CRA Annex I Part I(2)(m), Annex II point 8(d)). It deletes every multimedia blob FerroEHR wrote, then drops the ext, clinical, party, linkage and audit schemas, one transaction per database. Without --confirm it is a dry run that prints what it would remove and the value that confirms it (the stored instance id, or the clinical database’s name). A wrong value or an unreachable blob store refuses before anything is touched, and a rerun after success is a no-op. The new Decommissioning page covers the retention duties before erasing and what the command cannot reach: backups, WAL archives, replicas, the chart’s Secrets and claims, database roles, configuration files and media remanence.

  • Patient-data requests can require an authentication assurance level and a natural person (#3622; EHDS Annex II 3.1 and 3.2). [auth.oidc.assurance] maps the token’s acr (or another claim) to the eIDAS levels low, substantial and high; with minimum set, a patient-data request below it is refused with 401 and an RFC 9470 insufficient_user_authentication challenge naming the accepted acr_values. [auth.oidc.professional] names the claim identifying the natural person; a client token is then refused on patient data with 403 unless acting_for_claim names the professional it acts for. Basic credentials are refused on patient data when either block is set. The access record gains the acting mode, the assurance level and its raw value, and the professional’s identifier (audit migration 0009); the FHIR AuditEvent and the DICOM message render the professional as the requesting participant and the client application as a second one, and neither format can carry the assurance level. Both blocks are off by default.

  • An emergency access is marked in the access record (#3624; EHDS Art. 11(5)). [audit] emergency_purpose_codes names the agreed purpose-of-use codes that declare an access necessary to protect the vital interests of the patient; an access declaring one carries emergency_access in its record (audit migration 0010) and, on the FHIR AuditEvent, a purposeOfEvent with HL7 v3 ETREAT and a plain-words text, so the patient’s subject-scoped ITI-81 retrieval shows it. Each code must also be on purpose_codes, or the server refuses to start. The mark lifts no restriction: a GDPR Art. 18 restricted object is refused with an emergency purpose exactly as without one. The DICOM syslog message has no purpose element and does not carry it.

  • production refuses three more plaintext paths (#3676, #3689; CRA Annex I Part I(2)(e)). An enabled change-event outbox on a plain amqp:// broker (no amqps://, no tls = true) opens plaintext_broker; an enabled audit FHIR feed on an http:// URL opens plaintext_audit_feed; an enabled multimedia store with allow_http = true opens plaintext_object_store. Each refuses boot unless accepted by name in deployment_accepts, and the viewer’s deployment notice names each in plain words.

  • The Helm chart ships values-production.yaml (#3678), an overlay that boots under deployment_profile = "production": one database Secret per storage domain, a migration Job under migrate = verify, the audit trail failing closed, a restricted EHR_ACCESS default, declared subject namespaces, OIDC, and a TLS Ingress with a NetworkPolicy admitting only the ingress controller. It accepts plaintext_listener by name, because TLS ends at the Ingress. Chart validation and the boot check cover it.

  • Every binary in the published images lists its Rust dependencies (#3675): the server and viewer binaries inside the images, and the viewer image built from source, are built with cargo auditable, so an image scan and the release’s exploitability record see the Rust graph. The OpenVEX statement for RUSTSEC-2023-0071 now covers the viewer image.

  • Every operating-system finding in the three images carries an OpenVEX judgement (#3673): ferroehr-os.openvex.json (25 statements for the distroless ferroehr and ferroehr-viewer images) and postgres-os.openvex.json (101 for ferroehr-postgres), each argued from what the shipped programs load and import. Seven are affected, with an action statement: CVE-2026-8674 (a long resolver search domain) in all three images, and three libxml2 and two glibc JISX0213 findings any database login reaches through PostgreSQL’s XML functions, which FerroEHR itself never calls. The release’s exploitability record now has a judgement for every finding.

Changed

  • The family-level documentation moved to the FerroHEALTH book (#3701), at https://ferrohealth.eu/docs/: the manufacturer, the security policy shared by every FerroHEALTH product (response times, disclosure, advisories, security-only patches, CRA Art. 14 reporting, cessation of operations), the post-market procedure, the manufacturer side of the CRA, the intended purpose and the shared responsibility of the EHR system FerroEHR forms with FerroBRIDGE, licensing and trademarks, and how the products fit together. The book pages Complaints, incidents and vulnerabilities, Cyber Resilience Act, Intended purpose, Shared responsibility and Licensing & legal keep their addresses and now carry only FerroEHR’s own facts, with a link to the family page. SECURITY.md keeps FerroEHR’s supported versions, its reporting routes and scope, and the repository’s security settings. The information sheet, the instructions for use, the EHDS readiness and technical documentation pages, the CRA risk assessment and the Annex II user information stay in this book.

  • deployment_profile = "production" refuses two more plaintext paths unless each is accepted by name. A syslog audit feed with transport = "udp" opens the new audit_syslog_udp gap (#3669, EHDS Annex II 1.4): set [audit.syslog] transport = "tls", or accept the gap where the UDP path stays on a trusted segment. A separate management listener ([management] port) now opens plaintext_listener (#3668), because it binds every interface in plain HTTP: unset the port so the surface shares the main listener, or accept the gap where the port is unreachable off the host.

  • What FerroEHR answers without a credential says only what a client or probe needs (#3671). /health/readiness keeps its 200/503 semantics but its body carries the aggregate status and each indicator’s name and status only; the detail moves to GET /management/health. The public {rest root}/status keeps status, versions, timestamp and the deployment profile; the licence, the deployment gaps and the support period move to GET /management/status. Both new endpoints are off until [management.endpoints] names a level for them. The Security page lists every route reachable without a credential and why, and a test drives every mounted route to keep that list closed.

  • Licence tokens are verified against the licensor’s two master keys. A [licence] file token is accepted when it carries exactly one text signature, made directly by one of the two embedded Cadasto B.V. keys (F2D214FC30698BA65A846B2BF2EA48D4C316406D, 0BCE848225EFAECA912175292DBFB014CA506311); whether a key may sign is read from the embedded certificate, never from the copy the token carries. The previous licensing key is no longer trusted, and the embedded non-commercial grant is re-signed (licence id 01a1175c-cd67-777d-bdc0-c95bf1c900d3). A token signed under the previous key is reported as refused on /management/status, and the embedded grant stays in force; the server behaves identically either way.

  • deployment_profile = "production" refuses two more postures (#3638): auth.enabled = false (auth_off) and a plaintext listener on a routable address (plaintext_listener), unless the gap is named in deployment_accepts. A loopback bind is exempt; a deployment behind a TLS-terminating ingress accepts plaintext_listener by name.

  • A security fix ships in a security-only patch release (#3644), unless the release notes record why that was technically infeasible (CRA Annex I Part II(2)); SECURITY.md no longer leaves it open.

  • Every fixed vulnerability gets a GitHub security advisory (#3645), for FerroEHR’s own code and for a dependency or base-image fix that changes a shipped artefact, with a CVSS severity, the affected and fixed versions and the remediation; the changelog’s ### Security entry carries the GHSA- identifier. SECURITY.md states when the technical details of a fixed vulnerability may be published later (CRA Annex I Part II(4)).

  • The README, the landing page and the book no longer claim multi-tenancy, “spec-compliant” or “hardened by default” (#3619). FerroEHR is single-tenant; the openEHR conformance record is described as measured, with the version it measured; the README states that TLS is off by default, that the shipped profile is sandbox, and that FHIR R4 mapping is FerroBRIDGE’s.

  • The EHDS readiness and technical-documentation pages follow the manufacturer decision (#3626). Every Annex II row was re-read on 2026-10-06; the open questions are replaced by the decisions of #3606 (Cadasto B.V. is the manufacturer, FerroEHR with FerroBRIDGE is the EHR system, FerroBRIDGE ships the interoperability component) and the questions for counsel; rows 3.1 and 1.3 now say what is missing (the assurance level, the emergency-override mark), row 3.4 reflects the category-keyed retention and retention_years, and each partial row links its open issue. The shared-responsibility page and the compliance overview state both cases: Cadasto B.V. is not the processor of a self-hosted deployment, and is the customer’s processor where it hosts FerroEHR. Verifying releases counts nine published openehr-* crates, not eight.

  • Every EU Member State carries an access-log retention floor of three years (#3625, EHDS Art. 9(2), applying from 26 March 2029 and 2031 under Art. 105). A longer national floor still wins (the Netherlands, five years), Switzerland keeps one year, and the boot refusal names the provision the floor rests on. A declared SGB V § 309 controller in Germany sets retention_years = 3, because no day count meets both the three-year floor and the three-year § 309 ceiling.

  • auth.oidc.issuer and every smart URL are refused at boot when they carry a user name or password (#3629; OpenID Connect Core 1.0 §1.2). The SMART check applies even with SMART disabled, because the discovery document publishes those URLs.

  • The system-options vendor defaults to Cadasto B.V. (#3612), read from the manufacturer constant (it was FerroEHR project); [server.identity] vendor still overrides it. The OCI vendor and authors labels of the three images, and the author of the two OpenVEX documents, name Cadasto B.V. too.

  • The quickstart docker-compose.yml forwards FERROEHR__USAGE_REPORT__ENABLED from your shell (#3623). The report stays on by default; exporting the variable as false before docker compose up now switches it off, which it did not before.

  • The ferroehr-viewer image is scanned before it is tagged (#3674): the release and main image lanes scan it by its pushed digest on both platforms, so a fixable high or critical finding stops the viewer’s tags as it stops the server’s.

  • The ferroehr-postgres image drops GnuPG and the Go gosu (#3673). GnuPG and the libraries only it uses (gnupg, gpg, gpg-agent, gpgsm, gpgconf, dirmngr, pinentry-curses, libgnutls30t64, libp11-kit0, libtasn1-6, libksba8, libassuan9, libnpth0t64) served upstream’s build-time key import alone. gosu is now a shell wrapper over util-linux setpriv under the same name, so the upstream entrypoint runs unchanged and the image carries no Go binary and no Go standard-library findings; the 22 gosu exceptions in .trivyignore.yaml and postgres-gosu.openvex.json are gone. The base is re-pinned to the 2026-10-06 respin of postgres:18.6.

Removed

  • The in-tree FHIR connector is removed; FHIR R4 mapping is FerroBRIDGE’s (#3080). This is a breaking change with no deprecation period. Mapping between FHIR and openEHR is served by FerroBRIDGE (BUSL-1.1), a separate server that keeps its mappings as FHIRconnect documents and reaches FerroEHR over ITS-REST. A stored fhir_mapping row does not carry over to it. What is gone:

    • the routes POST /fhir/r4/{resource_type}, POST /fhir/r4/{resource_type}/$validate, GET /fhir/r4/{resource_type} and the /admin/fhir_mapping CRUD, with their fhir tag and the ferroehr-fhir.openapi.json document of the served OpenAPI (the ITI-81 route now has its own ferroehr-audit.openapi.json document);
    • the outbound FHIR emitter on the ferroehr.fhir AMQP exchange and its fhir_outbound health indicator;
    • the configuration sections [fhir] (api_enabled) and [fhir.outbound] (enabled, url, url_file, exchange, tls, batch_size, poll_interval_ms, publish_max_retries) and their FERROEHR__FHIR__… environment forms, which now fail at boot as unknown keys;
    • the fhir_mapping table, dropped by the new clinical migration 0013_retire_fhir_connector.sql, which also deletes the emitter’s fhir-outbound cursor row so it no longer holds the outbox prune floor;
    • the org.fhir.rest service of the SMART discovery document;
    • the viewer’s FHIR screen and its sidebar entry;
    • the Helm values config.fhir.* and secrets.fhirOutboundUrl (chart version 11.0.1), the fhir deploy-probe family, the fhir-r4-connector and fhir-mapping-store rows of the conformance statement, and the book pages FHIR connectors and FHIR connector (the EHDS priority-category table moves to EHDS readiness).

    The plaintext_broker deployment gap now concerns [events] alone. Not affected: the FHIR AuditEvent rendering and the ITI-81 GET /fhir/r4/AuditEvent retrieval ([audit.store], [audit.fhir_feed]), the FHIR terminology client ([terminology.external]), the change-event outbox and its AMQP publisher ([events]), and the Simplified Formats. The fhir cargo feature stays and now covers the audit rendering and the terminology client only; it no longer pulls in events.

Fixed

  • An AQL query over EHR_ACCESS and a CONTRIBUTION that deletes a version are classified like the object path (#3667). The query records the category none, as an EHR_ACCESS read does, and the contribution records the categories of the newest version with content before the delete, as a single logical delete does.
  • The viewer names every deployment gap in plain words (#3670), including audit_fails_open, open_ehr_access_default, auth_off, plaintext_listener and audit_syslog_udp, instead of showing the raw code.
  • The FHIR AuditEvent of an AQL query carries the origins of the data it served (#3654), one entity per origin with the capped-count marker, as a document read does.
  • The test harness no longer connects a database socket to itself (#3630). On macOS localhost resolves to ::1 first while the test container publishes its port on IPv4 only, and a port in the ephemeral range could occasionally connect to itself (a TCP simultaneous open), which surfaced as an internal database error in a test. testkit now uses the host’s IPv4 address.
  • CI and the test harnesses no longer send usage reports (#3623). The compose smoke test, the viewer login smoke, the deploy probes, the Helm probe, the browser end-to-end runs and the conformance runs switch the report off, so throwaway instances on GitHub-hosted runners stop appearing in FerroPULSE. A CI check fails any new test launcher that leaves it on.
  • Every access record carries its request id (#3665). The audit trail’s request_id column was empty on every record, because the id was read from the response before the request-id layer copied it there; it is now read from the request and equals the x-request-id the client received.
  • FERROEHR__AUDIT__PURPOSE_CODES is read as a comma-separated list (#3624). The environment form was refused as a string, so the purpose codes could only be set in the TOML file.
  • The served OpenAPI document uses the released operationIds with a dot (#3679): the ADL 1.4 template operations (definition_template_adl1.4_*) and the stored-query store operations (definition_query_store.yaml, definition_query_version_store.yaml) were served in an underscore form that did not match the id in the access record.

4.3.5 - 2026-10-08

Security

  • auth.oidc.jwks_json refuses a symmetric key or private key material at boot (#3664, GHSA-r86f-f3vw-j4jr; RFC 7517 §4.1, RFC 7518 §6.4). The inline key set is shown in clear by ferroehr config check and GET {base}/admin/config, so a set carrying a kty: "oct" key or a private member (d, p, q, dp, dq, qi, oth, k) is now a boot error that names the key’s position and kid, never its value. A set that does not parse as a JWK Set is refused with the line and column only. Public key sets boot as before; a symmetric key belongs in auth.oidc.hmac_secret.
  • The S3 store’s endpoint refusals no longer print URL userinfo (#3656, GHSA-h2gv-2qpx-4fg8). With [multimedia] configured, an endpoint such as https://user:password@s3.example that failed validation was quoted in full in the boot error and the log. The endpoint is now validated once, by the configuration, and every refusal quotes it with the userinfo masked.
  • The configuration views no longer print URL userinfo (#3629, GHSA-4jp3-xxj8-vv86). A URL written with credentials (https://user:password@host/…) in auth.oidc.issuer, a terminology provider url or OAuth2 token_url, authz.abac.remote.server, telemetry.otlp_endpoint or multimedia.endpoint was shown in clear by GET {base}/admin/config and ferroehr config check. Every string value in those views and in GET /management/env now passes one masker, so each URL keeps its host and path with the userinfo shown as ***.
  • hickory-resolver 0.26.3 (#3700, GHSA-rjpj-g667-cmvm): the DNS resolver the server binary links through the AMQP client is updated from 0.26.1. It resolves the broker’s host name when [events] or [fhir.outbound] is enabled (both off by default), where GHSA-6w6g-hm98-mhgm and GHSA-6f2x-v7q7-m7m5 reached it for an attacker who controls that name’s DNS answers.

4.3.4 - 2026-10-06

Added

  • A usage report to FerroPULSE, on by default (#3578). At each process start and once a day, an instance sends one small JSON report to https://report.ferropulse.eu/v1/report, the collector Cadasto B.V. runs in the EU. It carries a random instance id (created once and stored in the database, so every replica shares it), the version and commit, the specification profile, the licence grant type (non-commercial or commercial, never the licence id), the deployment kind (helm, compose, binary or unknown), uptime, whether the database answers and carries its migrations, the PostgreSQL major version, coarse CPU and memory buckets, and in the daily report a bucketed request count with latency histograms, percentiles and 5xx counts per route-template group and for AQL execution. It never carries patient data, request paths, AQL text, licence ids or host names. FerroPULSE stores the source IP address of each report and the country derived from it. Several replicas send one daily report between them, and a restart loop sends at most one start report per ten minutes. A slow or unreachable collector affects neither boot nor readiness; a failed send is logged once at DEBUG and not retried. Every boot logs one INFO line saying whether the report is on. Switch it off with [usage_report] enabled = false, FERROEHR__USAGE_REPORT__ENABLED=false, or the chart value usageReport.enabled=false. The new [usage_report] section also takes endpoint and slow_aql_ms (default 1000), and ferroehr usage-report --print [--event start|daily] prints the exact JSON the instance would send without sending it. The Helm chart (10.2.1) and the compose files set the deployment kind; the conformance compose files switch the report off. A new clinical migration adds the usage_report_instance and usage_report_window tables.

Changed

  • Commercial licensing goes to Cadasto B.V. (#3572). Cadasto B.V., the Licensor, handles the commercial licence and every other business or licensing question, at info@cadasto.com or https://www.cadasto.com/contact/. The BUSL LICENSE copies name that contact for alternative licensing arrangements in place of the maintainer, and the README, the book, the landing page, the BUSL crate READMEs, MAINTAINERS.md, GOVERNANCE.md and SUPPORT.md say the same. The maintainer keeps the technical side. The licence terms do not change. The openehr-* crates step to 0.0.86 in lockstep, because three packaged LICENSE files change.
  • The Licensor and copyright holder is Cadasto B.V. (#3567). Every holder statement moves from Vernum Projecten B.V. to Cadasto B.V.: LICENSE and the crate licence files, REUSE.toml, every SPDX-FileCopyrightText header, the licensing chapter of the book, the landing page footer and the conformance statement, whose signatory is now “Ruben Talstra, for Cadasto B.V.” with the date of the change. The licence terms do not change.
  • ClientError::DpopProof says whether a send went out (#3565). Its new sent field is false when the first proof failed and nothing was sent, and true when the proof for a retry or for the re-send answering a DPoP use_dpop_nonce challenge failed, after the service had received the first request.
  • A CONTRIBUTION version no longer borrows the change set’s committer (#3550). A versions[i].commit_audit, or an attestation, that names no committer used to be committed under the CONTRIBUTION audit.committer. ITS-REST UpdateAudit.yaml requires committer on every commit audit, so such a version is now refused with 400. A CONTRIBUTION without an audit, or whose audit has no committer or no change_type, is a 400 too (it was a 422), and so is a committer that is not a PARTY_PROXY object at all; a PARTY_PROXY that breaks its own invariants stays a 422. A version that names its own committer keeps it. The server still fills one time_committed for every audit in the set, and copies the CONTRIBUTION audit’s system_id into a version that omits one. This applies to the EHR and the demographic CONTRIBUTION alike. Migration: send committer on every audit and attestation, and change_type on the CONTRIBUTION audit.
  • An attestation member of a CONTRIBUTION must state its lifecycle_state (#3590). A 666|attestation| member without one was accepted, although SM master03 §Version Update Semantics says the lifecycle state “must be supplied in all cases”. It is now refused with 400 like every other member, and a stated code outside the openEHR version_lifecycle_state group is a 422. Migration: send lifecycle_state on every 666|attestation| member.

Fixed

  • The terminology overlay switches on the /terminology/* API it describes (#3476). docker-compose.terminology.yml said the CDR’s /terminology/* routes answer from FerroTERM, but it left [terminology] api_enabled at its false default, so every route answered 404. The overlay now sets FERROEHR__TERMINOLOGY__API_ENABLED=true; a term, subsumption or value-set call for a terminology the bundled openEHR terminology does not carry is answered by FerroTERM.
  • deploy/helm/validate.sh --update regenerates the chart README (#3476). It used to restore the stale README and then fail on it, so the drift it reported could only be fixed by running helm-docs by hand.

4.3.3 - 2026-10-05

Added

  • The openehr-its client sends DPoP-bound tokens (#3558). Credentials::Dpop writes Authorization: DPoP <token> (RFC 9449 §7.1), and a DpopProver given with Client::with_dpop_prover signs the DPoP proof over each request’s final method, URL and access token. A use_dpop_nonce challenge (RFC 9449 §8 and §9) is answered with one re-send carrying the supplied nonce.
  • openehr-sdt prints a SMART scope back in the grammar (#3557). SmartScope and ResourceScope implement Display, writing the canonical form (permissions in c/r/u/d/s order), and SmartScope::format_all prints a scope list as the space-delimited string SmartScope::parse_all reads.

Changed

  • ClientError::DeadlineElapsed says whether a send went out (#3560). Its new sent field is false when the deadline passed before anything was sent, and true when it passed before a retry or the re-send answering a DPoP use_dpop_nonce challenge, after the service had received the request.
  • The container images name FerroHEALTH as vendor and author (#3561). org.opencontainers.image.vendor is FerroHEALTH and org.opencontainers.image.authors is FerroHEALTH <https://github.com/FerroHEALTH> on every image.
  • openehr-its types the Prefer-dependent response bodies (#3548). A response the OAS declares as oneOf a resource and Identifier (the 201 and 200 answers of every create and update: EHR, EHR_STATUS, COMPOSITION, DIRECTORY, CONTRIBUTION and the demographic parties) is now rest::runtime::Representation<T>, either Full(T) or Identifier(Identifier), in place of serde_json::Value. A generator test pins the bodies that remain untyped, with their reason.
  • The generated CONTRIBUTION envelope reads a deleted version and an AUDIT_DETAILS audit (#3547). UpdateVersion.data is optional, because a logical deletion commits a version with no data (RM common master06 §Logical Deletion; ORIGINAL_VERSION.data is 0..1), and UpdateAudit reads _type: "AUDIT_DETAILS" as its base, as the contribution_create docs text tells servers to.
  • ferroehr-rest decodes every ITS-REST request through the generated contract (#3545). Each handler reads its parameters with the generated *Params::from_request, the CDR’s own parameter decoder is gone, and the Simplified Formats CONTRIBUTION path reads the typed NewContribution envelope.
  • Every ITS-REST request body has a public decoder (#3551). Each operation with a body gets {operation}_request_body(content_type, bytes) in its API group, which returns the type the server trait receives (for contribution_create, a Payload of the canonical or the Simplified Formats envelope) and is the decoding the generated router itself runs. The body decoding moved to rest::decode, so it needs only the rest feature.
  • ROUTE_REQUEST_MEDIA admits every media type the Content-Type parameter declares (#3552). The released OAS lists only application/json under requestBody.content where the operation’s Content-Type parameter and the docs text also admit XML and both Simplified Formats; request_media now carries the union, and a test holds it equal to the parameter.
  • The openehr-* crates step to 0.0.81.
  • openehr-its names the openEHR identifier class of every path parameter (#3539). rest::routes::Param gains identifier, an IdentifierClass whose as_str() is the BASE class name: the class the ITS-REST OAS states only in the parameter’s description, as the BMM type of the RM attribute the value is “taken from”. ehr_id, versioned_object_uid and contribution_uid carry HIER_OBJECT_ID, version_uid carries OBJECT_VERSION_ID, and uid_based_id keeps its per-operation class: UID_BASED_ID on a read, HIER_OBJECT_ID on an update, OBJECT_VERSION_ID on a delete (IdentifierClass::UidBased, HierObject, ObjectVersion). The table is generated from a decision map that a codegen test holds to the vendored OAS descriptions.
  • Every generated *Params struct decodes a matched request (#3540). from_request(matched, query, headers) takes the RouteMatch from routes::lookup, the query string and the headers, and runs the decoding the generated router runs: the declared keys into their fields, the members of a form-exploded object such as the AQL query_parameters read as JSON where they parse, + as a literal plus, and a 400 naming the parameter for a missing, repeated or invalid one. The generated handlers call the same decoding, so an intermediary and the router cannot read a request differently. The decoding moved from rest::server to the new rest::decode, which needs only the rest feature, not the axum server half.
  • The ITS-REST route table names each operation’s request-body media types (#3543). Each API group carries ROUTE_REQUEST_MEDIA, aligned with ROUTES, holding the OAS requestBody.content keys, and RouteMatch carries them as request_media. definition_query_version_store, which declares no Content-Type parameter, now names text/plain, so a forwarding intermediary can set the header.
  • A Simplified Formats CONTRIBUTION has a typed envelope (#3541). The generated NewContribution is generic over its version content, NewContribution<T = Versionable>, so the bare name stays the canonical body and NewContribution<serde_json::Value> reads the envelope of a FLAT or STRUCTURED CONTRIBUTION, whose envelope stays canonical while each versions[i].data is simplified (ITS-REST contribution_create §Simplified Formats). The generated server admits application/openehr.wt.flat+json and application/openehr.wt.structured+json on every operation whose Content-Type parameter declares them, and hands the implementation a rest::runtime::Payload (Canonical, Flat, Structured); XML and any other type stay 415.
  • The openehr-* crates step to 0.0.80.
  • The openehr-rm attribute model covers the BASE foundation types and names reference targets (#3537). The model (v1_1::model, v1_2::model) now holds every class of the BASE + RM BMM, including the primitives (Boolean, Character, Integer, Integer64, Octet, Real, Double, String, Uri), the Ordered and Numeric markers and the containers. Each class carries its BMM package, whether the generated crates carry a type for it (generated_type), and its full BMM ancestry as ancestors, so DV_ORDERED carries Ordered. The new is_primitive, is_foundation_type and conforms_to_ordered lookups answer the AQL ORDER BY question (“primitives and Ordered types”, QUERY master03 §ORDER BY). The AQL engine expands a declared type to its descendants for RM classes only; a foundation type stays a value type. A reference-typed attribute (OBJECT_REF, PARTY_REF, LOCATABLE_REF) carries ref_target, the class the RM text says it points to (EHR.ehr_status → EHR_STATUS, PARTY_RELATIONSHIP.target → VERSIONED_PARTY), or None where the text names no single class. A consumer that read “not in the model” as “primitive” now asks is_primitive.
  • The openehr-* crates step to 0.0.79.
  • openehr-its exposes the Authorization value it sends (#3535). Credentials::header_value is public, so a consumer can validate a credential at configuration load with exactly the value the client sends. Basic credentials are checked against RFC 7617 §2 (no colon in the user-id, no control character in either part) and bearer tokens against the b64token syntax of RFC 6750 §2.1. A refusal is a typed InvalidCredentials, which never carries the secret; a request whose credentials break these rules fails with ClientError::InvalidCredentials before anything is sent.
  • The openehr-* crates step to 0.0.78.
  • openehr-query classifies AQL function calls (#3529). A call naming one of the built-in functions in QUERY master03 §Functions parses to FunctionCall::Builtin with a typed BuiltinFunction (StringFunction, NumericFunction, DateTimeFunction); any other name parses to FunctionCall::Other. Both keep the name as written, so printer::to_aql returns it unchanged, CONTAINS included. FunctionCall::Named is gone, and the Token::Contains lexer token now carries its spelling. The CDR’s AQL engine uses this classification instead of a function list of its own.
  • openehr-its names every operation’s declared parameters (#3530). Each generated *Params struct carries a PARAMS table of the path, query and header parameters it holds (name as the OAS spells it, location, required, explode, and the value kind its schema states), written in the same pass as the struct. Each API group carries ROUTE_PARAMS, aligned with ROUTES, and routes::lookup returns the matched operation’s parameters with query_param, query_key and header_param lookups, so an intermediary can pass on exactly what an operation declares.
  • openehr-its ReqwestTransport constructors take a reqwest::ClientBuilder (#3531). ReqwestTransport::new and the new with_builder_timeout (replacing with_client_timeout) build the client themselves, with redirects switched off.
  • The openehr-* crates step to 0.0.77.
  • FerroTERM moves to 0.1.6 at ghcr.io/ferrohealth/ferroterm (#3521). FerroTERM moved to the FerroHEALTH organization, and its releases from 0.1.6 on publish there. The terminology compose overlay, the hosted sandbox, the Helm chart (terminology.image, chart 10.1.6) and the FerroTERM integration test pin 0.1.6@sha256:7ec4c962…, and the book links https://github.com/FerroHEALTH/FerroTERM.

Fixed

  • AQL honours the per-EHR EHR_ACCESS decision (#3562). A query could return rows from an EHR whose path routes refused the caller: the gate ran only on routes carrying an ehr_id in the path. Ad-hoc, stored and cohort queries now withhold every EHR the caller may not read, whether the query scopes it by the WHERE clause, an EHR predicate, the ehr_id parameter, or not at all.
  • A + in a query string is a literal plus (#3545). The CDR decoded query strings as HTML forms do, where + means a space, so an unencoded offset such as version_at_time=2024-01-01T00:00:00+01:00 lost its sign. Query strings now decode per RFC 3986, as the OAS style: form parameters (RFC 6570) define; a space is sent as %20.
  • Client::forward returns a redirect as received (#3531). The client built by ReqwestTransport::with_timeout used reqwest’s default redirect policy and followed up to ten redirects, re-sending the request and its credentials to whatever host a Location named. Every ReqwestTransport constructor now switches redirects off, so a 3xx reaches the caller and nothing is sent further. No ITS-REST 1.1.0 operation declares a 3xx, so a generated client call that receives one refuses it as an undocumented status.
  • A FerroTERM pin the sandbox’s index cannot serve is refused on the pull request (#3520). FerroTERM 0.1.5 moved its artifact layout to 7, and the v4.3.2 sandbox leg failed on the layout-6 index the box carried. A ferroterm-index-layout CI job reads the layout at the pinned FerroTERM tag and refuses a pin that differs from deploy/hosted/ferroterm-index.env, the record of the box’s index. The box now runs FerroTERM 0.1.6 on a rebuilt layout-7 index, and deploy/hosted/README.md documents the rebuild.
  • The generated ITS-REST types keep the members an open schema allows (#3526). OpenAPI 3.0.3 makes additionalProperties default to true, so every ITS-REST object schema that does not set it to false admits members beyond its declared properties. The openehr-its REST types for those schemas now carry them in an additional_properties map, written back at the object’s own level, where they used to be dropped on decode. Error is one of them, and the server now builds its error body from the generated type with the error member in that map. The body carries the same three members, with error now written after message and validationErrors.

4.3.2 - 2026-10-01

Added

  • openehr-query can rewrite, locate and bind a query (#3505, #3506, #3507, #3508). visit::Visit and visit::VisitMut walk every node of the AST, so a consumer finds, replaces or removes a WHERE condition without its own recursive walk. Every IdentifiedPath and every WHERE condition carries an ast::Span, the byte range it was written at, so a diagnostic can point at a predicate without repeating its literal; spans take no part in equality. bind::bind substitutes ITS-REST query_parameters into the AST as typed literals the printer escapes, at every position the grammar admits a parameter, and reports an unbound, unknown, duplicate or unbindable parameter by name, never by value. The opt-in federation feature adds federation::parse_federated, which lifts a FROM ENDPOINT p [ … ] or ORGANISATION [ … ] directive out of a query and parses the rest as strict AQL. Breaking: WhereExpr::Identified and IdentifiedPath gained the span (WhereExpr::identified and IdentifiedPath::new build them without one).

  • openehr-its publishes the server router and the intermediary seams (#3509, #3510, #3511, #3512, #3514). Under rest-server, each API group’s server module carries the trait, one success enum per operation (the documented 2xx answers with their Location, ETag and other headers) and router(api), an axum router binding every ROUTES entry to its trait method. A method refuses with rest::runtime::Refusal, an ApiError plus the headers its answer needs: the current ETag on a 412, or the OpenAPI-declared headers of a demographic 409/412 through the generated headers structs (Refusal::with_headers). Either renders as the ITS-REST Error JSON body. rest::server::router(api) mounts all six groups at once, and rest::server::with_fallbacks makes an unknown path answer 404 and an undeclared method 405 with Allow, both with the Error body. rest::routes::lookup matches a method and path to its operation and path parameters without reading the body, and Client::forward sends a request once and returns every answer as received, so an intermediary passes bodies, Location and ETag through byte for byte. CredentialsProvider resolves the credential before every attempt and hears about a 401, for tokens that refresh. CallOptions gives a generated client a deadline and extra headers per call (with_options); the remaining budget reaches the engine as RequestTimeout, and an elapsed deadline is ClientError::DeadlineElapsed. Breaking: the server traits moved into <group>::server and return Result<success enum, Refusal>, and the response headers structs moved from client to the group module. The nine openehr-* crates step to 0.0.74.

  • The openehr-its ITS-REST client carries what a consumer meets on first adoption (#3487). The commit operations (COMPOSITION, EHR_STATUS, the directory, EHR creation, the demographic parties and relationships) take the committal-metadata headers the ITS-REST overview defines and the OpenAPI declares no parameter for: openehr-version, openehr-audit-details (one field line per value) and, on the COMPOSITION and CONTRIBUTION commits, openehr-template-id. Every documented 4xx outcome carries the answer body as an ErrorBody: the bytes as received, decoded as the ITS-REST Error when they are one, so a 422’s validationErrors reach the caller; ClientError::Unauthorized, Forbidden, ServiceFailure and UndocumentedStatus carry the same body. A 400 whose body carries message alone decodes as the Error with an empty validationErrors. A path parameter keeps : literal, so a version uid is sent as the specification writes it. Credentials hold their secret in a secrecy::SecretString (redacted Debug, zeroed on drop), with Credentials::basic and Credentials::bearer constructors. The nine openehr-* crates step to 0.0.72.

  • openehr-its ships an ITS-REST client (#3485). The new rest-client feature adds a generated client per ITS-REST 1.1.0 API group (rest::generated::<group>::client) over a hand-written runtime (rest::client: the Transport engine trait, the shipped reqwest engine, Basic and Bearer credentials, the retry budget and ClientError). Each operation takes its parameter struct, with Prefer, If-Match and Accept as fields, and answers an outcome with one variant per status the OpenAPI documents for it; declared response headers such as ETag and Location come back in a typed struct. Any other status is ClientError::UndocumentedStatus, and 401, 403 and 5xx are their own errors. Idempotent requests are retried after a transport failure or a 5xx; a POST is sent once. A create operation also answers NoContent, the 204 the ITS-REST overview prescribes under Prefer: return=minimal although the OpenAPI lists 201 alone. FerroBRIDGE keeps its hand-written client until it moves to this one (its issue #276). The DTOs, parameter structs and route tables move to a new rest feature over serde and http alone, which both rest-server and rest-client build on, so a client never compiles axum. full now includes rest-client. The nine openehr-* crates step to 0.0.71.

  • A CI guard keeps the Apache-2.0 crates free of BUSL-1.1 code: scripts/checks/licence-boundary.sh refuses a normal or build dependency, or a feature, from openehr-base, openehr-lang, openehr-term, openehr-rm, openehr-am or openehr-its on openehr-query, openehr-adl or openehr-sdt, and a license field that changes side.

Changed

  • The repository moved to the FerroHEALTH organization (https://github.com/FerroHEALTH/FerroEHR, #3516); the old rubentalstra/FerroEHR URLs redirect. Images and the Helm chart now publish to ghcr.io/ferrohealth/ferroehr, ghcr.io/ferrohealth/ferroehr-viewer, ghcr.io/ferrohealth/ferroehr-postgres and oci://ghcr.io/ferrohealth/charts/ferroehr, and the compose files, the chart and the book name those references. The v4.3.1 images and chart are copied there by digest. Everything published up to v4.3.1 stays available under ghcr.io/rubentalstra, and its signatures name rubentalstra/FerroEHR as the repository that built it. The roadmap board is now https://github.com/orgs/FerroHEALTH/projects/2. The nine openehr-* crates step to 0.0.75 for their new repository field.

  • The dependency set moves forward in one step. OpenTelemetry goes to 0.33 (opentelemetry, opentelemetry_sdk, opentelemetry-otlp, opentelemetry-prometheus) with tracing-opentelemetry 0.34; the OpenAPI stack to utoipa 6, utoipa-axum 0.3 and utoipa-swagger-ui 10; jsonschema to 0.58, sevenz-rust2 to 0.23, lapin to 4.12 and config to 0.15.27. The bundled PostgreSQL image is rebuilt on the current postgres:18.6 upstream digest (sha256:5a5a84b1…), a same-version respin with rebuilt base packages. The nine openehr-* crates step to 0.0.73.

  • FerroTERM moves to 0.1.5 everywhere the product pins it. The quickstart overlay docker-compose.terminology.yml, the hosted sandbox compose file, the Helm chart (terminology.image.digest and the pinned version its tag default and Artifact Hub annotation read) and the integration test that starts the real server all name ghcr.io/rubentalstra/ferroterm:0.1.5@sha256:61a2d4aefc4a48e9e90b9a698a64bdfeb7ef6443744a1fe50972099456891fe9, the index digest read from GHCR. That release adds the history interaction, _summary and _elements on read, a 406 on an unacceptable Accept, refuses a write that would duplicate a canonical, and states the position of a refused expression constraint on the OperationOutcome. The chart version becomes 10.1.3; an install that pins the chart keeps running the image it pinned.

  • openehr-its is split in two, and the wire layer is Apache-2.0 again (#3482). The new crate openehr-sdt (BUSL-1.1) holds the hand-written Simplified Formats (FLAT, STRUCTURED, Web Template, TDD import), the template-independent RM-instance validation and the SMART on openEHR scope grammar. openehr-its keeps the ITS wire layer: the generated canonical JSON dispatch, the canonical XML codecs, the ITS-REST contract, the OPT 1.4 and AOM2 archetype XML codecs, and the hand-written runtimes, canonical JSON entry points and wire-validation dispatcher they need. From 0.0.69 it is published under Apache-2.0, so a consumer of canonical JSON takes on no BUSL-1.1 code. Rust callers import openehr_sdt::flat, openehr_sdt::rm_instance and openehr_sdt::smart_scopes in place of openehr_its::flat, openehr_its::rm_instance and openehr_its::rest::smart_scopes. The flat and cache features left openehr-its for openehr-sdt; a wasm32-unknown-unknown consumer of openehr-its now selects default-features = false, features = ["opt14"], and openehr-sdt with default-features = false compiles to the scope grammar alone, with no dependency. Published versions keep their licence: openehr-its 0.0.60 to 0.0.67 stay BUSL-1.1 AND Apache-2.0.

  • The published openehr-* crates are nine, and all step to 0.0.69 (#3482). openehr-sdt was first published alone at 0.0.68, built against the 0.0.67 siblings; it joins the lockstep line at 0.0.69 and publishes after openehr-its and before openehr-adl.

  • FerroTERM moves to 0.1.4 everywhere the product pins it (#3474). The quickstart overlay docker-compose.terminology.yml, the hosted sandbox compose file, the Helm chart (terminology.image.digest and the pinned version its tag default and Artifact Hub annotation read) and the integration test that starts the real server all name ghcr.io/rubentalstra/ferroterm:0.1.4@sha256:f3b5f35a60d99d5c837125e5f8cf39d8eeecf9f69dbc79b9a526cc45aa706470, the index digest read from GHCR. That release reloads its served set without a restart, treats an index root’s child directories as releases, ships the ferroterm-sync service, and carries the standard concept status properties on locally authored code systems. The chart version becomes 10.1.2; an install that pins the chart keeps running the image it pinned.

Fixed

  • Two 500 answers from the openehr-its REST server runtime carry a fixed message (#3518). An illegal response header and an answer body that does not serialize put the header name or the serde error into the body the client received; both now answer with a fixed text.

  • The app and viewer images rebase onto a distroless base with the fixed OpenSSL (#3504). gcr.io/distroless/cc-debian13:nonroot moves to sha256:e792ab3d…, which carries the libssl3t64 update from DSA-6531-1 (CVE-2026-75804, CVE-2026-84782). The image scan refused the previous base with both findings at HIGH, so the main images went untagged until this change.

  • An AQL path predicate whose node boolean opens with a parameter or an archetype id parses (#3513): CLUSTER k[$node and name/value='x'] and [openEHR-EHR-CLUSTER.device.v1 or at0001] were refused, though the grammar in QUERY/docs/AQL/master07-grammar.adoc admits both.

  • The no attribution in commits CI job no longer refuses the Co-authored-by: dependabot[bot] trailer GitHub writes when it squash-merges a Dependabot pull request, which had failed the last three pushes to main. The commit-msg hook keeps that line too.

4.3.1 - 2026-09-16

Added

  • A fourth backup CronJob for the audit domain (#3399). The Helm chart renders backup.audit as soon as database.audit.existingSecret gives the audit trail a database of its own, with its own claim, credential and schedule, and drops --schema=audit from the clinical dump in the same render. Until then nothing changes: the trail shares the clinical database and travels with that dump. Before this, an audit database placed elsewhere was backed up by nothing and the operations page told the operator to dump it by hand.
  • An erasure tombstone on the party change-event stream (#3417). A physically deleted party now leaves one contribution-less row on the party domain’s own outbox before the delete commits, naming the party and every PARTY_RELATIONSHIP erased with it, so a consumer of the party stream is told to delete what it derived. GDPR Art. 19 makes the controller communicate an erasure to each recipient the data reached, and the clinical outbox has carried the same row since #3416.
  • A dispatch-only storage benchmark lane (#3370). It runs the storage benchmark against a PostgreSQL 18 service with durability on, compares every operation with the record committed for the same schema generation, and fails the run when a p50 or p99 rose further above it than the dispatched tolerance. The summary lists each operation’s baseline, measurement, delta and verdict and names both machines; the fresh record leaves as a run artifact and the committed baseline changes only in a pull request. Exploration, never a conformance record: it earns no performance class.
  • Restriction of processing, at whole-EHR or single-object grain (#3324). A restriction register records who asked, on what ground and when, and the mark it drives stops every path that would process the object: a point read, a versioned read and a revision history answer 403 with a body naming the restriction, AQL omits it at every scope, an EHR Extract skips it and refuses outright for a restricted EHR, the change-event stream neither emits nor delivers an event about it, and a write to it is refused. The stored rows are untouched — storage is the only processing GDPR Art. 18(2) leaves. Lifting stamps the register rather than erasing it, so the sequence Art. 18(3) turns on survives, and a whole-record lift leaves an object-scoped restriction standing. POST/GET {base}/admin/restriction and POST {base}/admin/restriction/lift drive it, each act an access record of its own.
  • A per-EHR objection to research processing (#3325). While it stands, the record leaves every full-population AQL query, every export and the event stream, and a query naming the ehr_id or a read for care is untouched: GDPR Art. 21(6) reaches research processing, not the care record. The public-interest override the article admits is recorded beside the objection and says on whose authority. POST {base}/admin/research-objection.
  • A retention register, and a list of what has fallen due (#3346). The period per content category and jurisdiction with the legal citation it rests on, a per-EHR anchor, a whole-record hold and a per-object hold, and a view that lists the EHRs whose period has run with the counts of objects due and objects exempted. Nothing is deleted on a timer: an openEHR record is indelible, so acting on the list is the controller’s decision, taken through the physical delete. Five routes under {base}/admin/retention/, and a new book page, Retention, restriction and objection.
  • An access-log retention ceiling beside the floor (#3346). Set [audit.store] sgb_v_309_controller = true where the deploying organisation is, or acts for, one of the controllers SGB V § 307 names for a German telematics-infrastructure application, and boot refuses any retention_days above the three-year period § 309 Abs. 1 sets — including 0, keep forever. A floor and a ceiling that contradict each other are refused outright rather than silently resolved in favour of one.

Fixed

  • A date with text after it compares as no date (#3436). openehr_date_days read the leading four, six or eight characters of a value and ignored the rest, so 20210102XYZ, 2021-01-02 BC, 2021-01-02 Europe/Amsterdam and 2021a compared and ordered as dates in AQL. BASE foundation_types master06-time_types.adoc §Iso8601_date admits none of them, and each now reads as NULL, which is a comparison miss rather than an error. 2021-01-02T10:30:45 and its space-separated and lowercase-t variants are unchanged.

Changed

  • The release profile links with thin LTO (#3098). [profile.release] gains lto = "thin", the whole-binary optimisation the Cargo book describes for a CPU-bound path; the overflow-checks, panic = "unwind" and line-table posture is unchanged. codegen-units = 1 was tried in the same cycle and withdrawn: the arm64 release build exhausted its runner twice. A release build takes longer; the throughput before/after is measured under #3452 once the box can be idle.

  • LATEST_VERSION under CONTAINS is one key probe per object (#3453). The AQL emitter used to express “this row is the current trunk version” as two correlated EXISTS subqueries over the head table, which let the planner reach the version row by vo_id alone through the commit-time index and walk every trunk version of the object for every node row: 75 601 of the 93 255 buffers a class-s population CONTAINS read. The head row is now joined on its primary key and the version row reached through its own, sys_version bound; the plan-shape tests pin the index and a ceiling of eight buffers per probe. Results are unchanged.

  • The server needs only btree_gist, which it installs itself (#3433). The second-generation schema calls nothing from uuid-ossp, pgcrypto or pg_trgm, so the PostgreSQL image no longer installs them and the CI step no longer creates them. btree_gist backs the temporal key on linkage.subject_ehr and the server creates it at boot, so a plain PostgreSQL 18 database with a migration credential that may create it serves with nothing preinstalled.

  • The conformance statement is signed by a named person for the company (#3438). A company signs through someone: the party statement’s signatory is now “Ruben Talstra, for Vernum Projecten B.V.” with the date of the change, and the rendered statement follows. AI_STATEMENT.md names Ruben Talstra as author and maintainer, since the owner of the work is the company.

  • The Licensor and copyright holder is Vernum Projecten B.V. (#3435). Every statement of the holder names the company: LICENSE and the per-crate licence texts, REUSE.toml, the SPDX header of every first-party file (emitted for the generated crates, so they were regenerated), the licensing page, the landing footer and the conformance party statement, whose rendered statement and certificate were regenerated with it. The licences themselves are unchanged — the application and the three hand-written engines stay Business Source License 1.1 with the same Additional Use Grant, Change Date and Change License, the five generated model crates stay Apache-2.0, and releases up to v4.0.17 and crate versions up to 0.0.56 stay MIT as published, under the holder they were published with. Ruben Talstra remains the maintainer credit the banner, MAINTAINERS.md, the OCI authors label and the chart maintainers carry, and the urn:rubentalstra:ferroehr conformance identifier is an identifier rather than a holder statement and stays.

  • The production deployment profile refuses the two compatibility defaults (#3323). The per-EHR EHR_ACCESS default is open and the audit trail fails open, both chosen so the conformance instrument and the development stacks work out of the box. Neither default changes; what changes is that deployment_profile = "production" now refuses to start while either stands, naming the key, the value and the remedy, unless the deployment accepts it by name in deployment_accepts — the mechanism the profile already uses. The two new gap tokens are open_ehr_access_default and audit_fails_open. GDPR Art. 25(2) asks that by default personal data not be made accessible without the individual’s intervention to an indefinite number of natural persons; the obligation binds the controller’s production posture, which is what the profile expresses.

  • The deployment posture reads the audit pool and every domain database (#3398). shared_cluster now includes the audit pool’s own cluster identity, so an audit database co-located with a pseudonymisation domain is reported instead of being invisible (Swiss DSV Art. 4 Abs. 5 keeps the log on a system separate from the one processing the data). migrate_on_runtime_credential is evaluated per domain database rather than only from [db] migrate_url: a domain relocated to a database of its own is prepared on that domain’s runtime DSN, which is the same gap under a different key.

  • ext.openehr_timestamp refuses a zone name and a BC era (#3427). PostgreSQL’s own date/time input reads 2021-01-02T10:30:45 Europe/Amsterdam and 2021-01-02T10:30:45 BC, and the separator-position gate in front of it let both through. BASE foundation_types master06-time_types.adoc §Iso8601_date_time ends the value at the offset, so both now read as NULL: an AQL comparison against such a stored value misses rather than matching a date the record does not carry. The check costs the helper 93 ms per 50 000 readings against 65 ms without it, measured on the same corpus where the first generation took 117 ms.

  • The archive export lists multimedia blobs from the reference index (#3420). It used to walk every exported body in process to find the blob keys; it now reads blob_ref, the index the node write path maintains in the same transaction as the nodes it describes, once per domain. The archive it writes is unchanged.

  • The ext value helpers are each in the form that measured fastest, and read fewer things than PostgreSQL’s own date parser did (#3351). Five of the seven no longer open a subtransaction per row: they guard their casts ahead of them instead of trapping an error. The two whose whole body is parsing a date or a time keep the trap, because measurement makes it the cheapest validator available — over 50 000 readings they cost 58 ms and 70 ms, where the same bodies validating by hand so the cast can run untrapped cost 65 ms and 235 ms — and each carries that measurement in its function comment. Every helper reads faster than the one it replaces, and a value any of them refuses reads as NULL. Three readings tighten as a result, all of them values openEHR does not define: a date whose year carried leading whitespace or a sign is now unreadable, a date handed to the time parser no longer yields a number, and ext.openehr_timestamp refuses a date that is not zero-padded (2021-1-2), an ordinal date (2021-002), and the words PostgreSQL accepts (now, today, yesterday, epoch, infinity). The first two made a stored value read as the current time. In the other direction it now reads the lowercase t and space separators on a reduced-precision date, which it used to read only at full precision.

  • A physical EHR delete reaches everything the erasure has to reach (#3347, #3403). DELETE {base}/admin/ehr/{ehr_id} runs one clinical transaction that removes the EHR and, through the foreign-key graph, its versions over both storage tiers with their nodes, attestations, contributions, commit audits, item tags, folder memberships, restriction and retention marks, multimedia blob references and pending change events; the subject proxies of a subject whose only record this was go with it, and an erasure tombstone is appended to the change-event stream, which each registered reader receives and applies by deleting what it derived (GDPR Art. 19). The cross-reference row naming the subject is then erased in the linkage domain, on the bulk route as well as the single one, and the externalized blobs no surviving version references leave the object store. The audit records naming the ehr_id stay, because Art. 17(3)(b) withholds erasure where a legal obligation requires the processing and the access-logging periods are that obligation.

  • Multimedia blob collection is an index lookup rather than a scan (#3347). clinical.blob_ref and its party twin record which stored version references which externalized blob, written by the node write path, carried across the archival tier move by their foreign key and removed with the version. The collector reads that index instead of pulling every node body of the deleted record into the process and substring-matching it against every node of both domains.

  • An OPT 1.4 template whose ordinal symbol has no <value> element is refused with the fix in the message (#3395). The ITS-XML profile schema types C_DV_ORDINAL.list as the RM DV_ORDINAL, whose symbol is a DV_CODED_TEXT carrying a mandatory value, so such a template is not a document the format admits and the refusal stands. Beside the element and line it already reported, the 422 now says to add <value/>, that an empty element is valid, and that the rubric is resolved from term_definitions. CKM’s exporter writes the element; Archetype Designer’s omits it.

  • The migration-immutability rule binds SHIPPED files only (owner ruling 2026-09-15). A migration is immutable once it is in a release (present at the latest vX.Y.Z tag); a file on main that no release carries has been applied by no installation and is fixed in place. The guard (scripts/checks/migration-immutability.sh) judges the pull request’s diff against the release tag’s tree and re-arms at every cut; nothing changes for an installation on a released version.

  • BREAKING — the storage schema is rewritten, and a database created by an earlier release must be recreated (#3342, #3344). The rewrite is greenfield (owner ruling 2026-09-14): the new migration sets replace the old ones outright, nothing is upgraded in place, and the server refuses at boot a database whose migration record predates the rewrite. Take a dump of anything worth keeping before upgrading, recreate the database, and let the server migrate it. What changed, and why:

    • The ehr schema becomes clinical and demographic becomes party, each laid out as a sequence of natural migration files with one concern apiece rather than one squashed baseline.
    • version is append-only: it carries no validity interval, so a supersession writes no close-out statement, and the validity of a version is derived from its committed_at. One mutable vo_head row per versioned object answers “what is current”, and none of the columns a commit updates appears in an index, so that update is heap-only.
    • version, node and vo_attestation are partitioned by storage tier, so archiving an EHR is row movement inside one relation set rather than a copy into mirror relations: the cold and cold_demographic schemas, every *_all union view and the vo_archive marker table are gone, the archive marker is two columns of vo_head, and a foreign key holds across the tier.
    • node gains the promoted name_code and name_terminology the AQL node predicate needs, and drops the unread citem_num.
    • The clinical and party change-control and node relations are rendered from one shared DDL template, so the two domains cannot drift.
    • The party domain gains party_relationship_target, the target-side index behind PARTY.reverse_relationships.
    • The archive file format carries a version’s committed_at instead of a validity interval. An archive written by an earlier release does not load.
  • BREAKING — the EHR id / subject cross-reference moves to the linkage domain, and erasure reaches it (#3345). linkage.party_ehr becomes linkage.subject_ehr and absorbs ehr.ehr_index: one temporal relation holding both the party-to-EHR map and the openEHR Service Model’s EHR Index associations (I_EHR_INDEX), which is what that service is for — an EHR persisted with only an EHR id, and a separate index to obtain the subject identifier. The EHR_INDEX operations and the subject-proxy subject resolution now run on the linkage pool; the clinical schema keeps only the promoted ehr.subject_id/subject_namespace pair, under the pseudonym guard, because the wire binds ehr_get_by_subject and the one-EHR-per-subject 409 to EHR_STATUS content. The subject-proxy relations (sp_subject, sp_variable, sp_data_set, sp_sample) key on a derived opaque subject_key instead of the caller’s own subject identifier, so no unguarded subject identifier remains in the clinical domain (GDPR Art. 4(5)), and a catalog test asserts it. Deleting an EHR now removes the cross-reference rows naming it, through linkage.erase_ehr — a SECURITY DEFINER function the ferroehr_linkage role may execute while still holding no DELETE on the table — so an erased record leaves no row behind saying whose it was (Art. 17(1)). Subject-proxy registrations and index associations made under an earlier release do not carry over; the storage rewrite above requires a fresh database in any case.

  • BREAKING — one pool and one DSN per storage domain, and the runtime roles are named after them (#3343). Each of the four storage domains — clinical, party, linkage, audit — now has its own configuration table, [storage.<domain>] with url/url_file, and its own connection pool carrying only that domain’s schema on its search_path. Every domain defaults to the shared [db].url, so a configuration file that names no [storage] table at all, and the compose quickstart, behave exactly as before. What changed for a deployment that had already separated its credentials:

    • [db].demographic_url, [db].demographic_url_file, [db].linkage_url and [db].linkage_url_file are gone. Move them to [storage.party] and [storage.linkage]; the old keys are refused at boot by name rather than ignored.
    • The chart’s database.demographicExistingSecret, database.linkageExistingSecret and their *Key/*Url siblings become database.<domain>.{existingSecret,existingSecretKey,url} for clinical, party, linkage and audit. The mounted files are storage.<domain>.url and the environment names are FERROEHR__STORAGE__<DOMAIN>__URL_FILE.
    • The runtime roles are the domain roles and nothing else: ferroehr_clinical, ferroehr_clinical_reader, ferroehr_party, ferroehr_party_reader and ferroehr_linkage, beside the DDL-only ferroehr_migrator. The first generation’s ferroehr_ehr, ferroehr_ehr_reader, ferroehr_demographic, ferroehr_demographic_reader and the generic ferroehr_app/ferroehr_reader pair are gone: no migration creates or grants to them, the compose database image no longer provisions them, and a DSN that authenticates as one of them holds nothing. A login role is granted the domain role of the pool it serves (GRANT ferroehr_clinical TO <login>); the clinical role also writes the local Audit Record Repository.
    • The compose backup service ferroehr-backup-demographic is ferroehr-backup-party, its directory variable FERROEHR_BACKUP_DEMOGRAPHIC_DIR is FERROEHR_BACKUP_PARTY_DIR, and the chart’s backup.demographic.* values are backup.party.*.

    A DSN naming another host moves that domain to a database or cluster of its own, which is the separation a schema inside one database cannot make: a base backup, WAL archiving and physical replication carry every schema of a database together (PostgreSQL 18, Backup and Restore). Schema preparation follows: the ext set is applied per database by whichever domain reaches it first, so two co-located domains share one copy and a relocated domain carries its own. One layout is refused rather than half-applied — the linkage migration set revokes a function the party set creates, so those two are prepared in the same database.

    The boot gate grew two refusals. Two domains configured on DIFFERENT DSNs that authenticate as the SAME database role are refused, because a separation that exists only in the configuration reads as one that holds. And a domain role that does not exist is a warning under deployment_profile = "sandbox" and a refusal under production, where absent roles mean absent grants and the check would otherwise pass by having nothing to measure. ferroehr config check prints the resolved domain layout, naming no DSN.

  • The Helm chart is 10.1.1, a major bump (#3342, #3343, #3344, #3378). The values contract changed in three ways a values file can notice: the per-domain database.<domain> blocks above replace the flat demographicExistingSecret and linkageExistingSecret keys, config.tenancy no longer exists and is refused at render with the reason, and the per-domain backup CronJobs dump the generation-2 schemas (clinical, ext, audit and party) instead of the retired mirror schemas. The docker-compose backup services dump the same sets, and all six jobs now pass --strict-names, so a --schema pattern matching nothing fails the dump instead of writing an artefact that silently lacks that half (PostgreSQL 18, pg_dump §Options).

  • A newer runtime base image (#3439). The server and the viewer image build on gcr.io/distroless/cc-debian13:nonroot at digest sha256:54df941e, the digest that tag resolves to today. The org.opencontainers.image.base.digest label moves with it in both Dockerfiles and in the two publishing lanes, which is what the image-labels guard checks.

  • Four runtime dependencies move to their current patch (#3442). lapin 4.11.0 retries the initial AMQP connection properly when recovery is on, which is the mode the events transport runs in; reqwest 0.13.5 fixes proxy authentication picking the wrong credentials when several proxies match a request, and a blocking-client timeout panic; tokio-rustls 0.26.5 returns more data per poll_read on a TLS stream; fancy-regex 0.19.2 cuts pattern build time. No API a caller here uses changed.

  • jsonschema 0.55.1 (#3443). It compiles the vendored ITS-JSON schema behind openehr-its’s schema-validation feature. The releases since 0.53.0 correct the keyword location an error names and several schema canonicalisation defects; the 626-test openehr-its suite, the ITS-JSON corpus gate among it, passes unchanged. The eight published openehr-* crates step to 0.0.67 with it, since the requirement is part of their packaged manifest.

Removed

  • The Subject Proxy service (#3431). It realised SM I_SUBJECT_PROXY_SERVICE and I_DATA_BINDING over the sp_* tables, had no REST route and no consumer outside its own configuration. Gone with it: the [subject_proxy] configuration section and its FHIR-frame executor, the sp_* tables, the erasure reach into them, the Helm config.subject_proxy values and the Subject Proxy book page. No openEHR specification obliges a CDR to realise every SM component; SM master02 presents the platform as a set of services a deployment composes. docs/architecture.md lists the component as not realised.
  • Multi-tenancy (#3378). Multi-tenancy is achieved by running separate instances: one instance, one database, one set of domain roles per organisation. That is where openEHR puts it. An openEHR system is “a distinct logical repository corresponding to an organisational entity that is legally responsible” for the data, and it is “distinct from any underlying virtualisation infrastructure or cloud computing facility, which may house multiple logical EHR systems in a multi-tenant fashion” (BASE architecture_overview/master06-design_of_the_ehr.adoc §The EHR System) — so tenancy belongs to the layer that hosts several systems, not inside one. The same chapter’s §System Identity is why housing several organisations in one system was wrong rather than merely unnecessary: system_id names that one legally responsible system, it “becomes embedded in the version identifiers of committed – and possibly signed – content”, and it “cannot easily be changed afterwards”. No relation carries a tenant column, no row policy or session GUC scopes a read, and the [tenancy] configuration, the tenant middleware, the /admin/tenant routes and the Helm and compose tenancy keys are gone. The pseudonymisation domains and their role barriers are unrelated to tenancy and stay.

Security

  • rustls 0.23.45 (RUSTSEC-2026-0285): the TLS stack no longer accepts TLS 1.3 handshake messages sent at the wrong encryption level after a key change (RFC 8446 §5.1). Lockfile bump only; no configuration changes.

4.3.0 - 2026-09-14

Added

  • The pre-rewrite storage baseline is committed (#3368). The storage benchmark’s class-s record for the first-generation schema lives at docs/benchmarks/storage/generation-1/record.json and renders on the performance page, so the storage rewrite (#3337) is compared against a committed record rather than a claim. The harness’s end-of-run relation table gains dead_tuple_percent, the exact heap figure pgstattuple reads (null where the extension cannot be created), beside the planner’s estimates.
  • A storage benchmark harness (#3367). cargo bench -p ferroehr --bench storage seeds a corpus through the ordinary write path, then times the storage layer’s hot paths one at a time: a composition commit, a supersession, the If-Match supersession, point reads by version uid and by versioned-object uid, the version at an instant, the revision history, an AQL CONTAINS chain over one EHR and over the population, archive and restore of an EHR, and one retention prune. Beside each operation’s wall-clock it records what the database did for it: the per-relation tuple counters and relation sizes from pg_stat_user_tables, buffer hits and misses from pg_stat_database, the write-ahead-log bytes either side of the measurement, one representative commit probed on its own, and EXPLAIN (ANALYZE, BUFFERS, WAL) of the population statement inside a rolled-back transaction. Everything runs through the service and the public storage API, and the relations are discovered from the catalogue, so the same harness measures a rewritten schema. The JSON record lands under docs/benchmarks/storage/<schema-generation>/ and reaches the performance page through scripts/render/storage-bench.sh. It is a benchmark, never a conformance record: conformance is the CNF suite Veredictum runs.
  • The Kubernetes deployment probe observes the FerroTERM workload (#3328). scripts/deploy-probe-k8s.sh gains a terminology stage (P-K8S-TERM-*) that reads the chart’s terminology.enabled claims off a running cluster: the absence of every terminology object while the gate is off, the rollout, the applied runtime posture from the container runtime’s own spec, the shaped seed reaching the process through the ConfigMap volume, the NetworkPolicy refusing a pod the CDR’s selector does not name in this namespace and in another while the CDR itself is admitted, and a value-set binding resolved through the Service. A mounted index is probed when K8S_TERM_INDEX_CLAIM names an existing claim, and is declared not exercised by probe id when it does not.
  • Identifier-scanner rules for Germany and Switzerland, and the Swiss access-log retention floor (#3318, #3319). de-kvnr refuses the unchangeable part of a Krankenversichertennummer (a letter, eight digits and the GKV-Spitzenverband’s check digit), ch-ahvn13 the AHV number in its plain and dotted display forms (AHVV Art. 133; the BSV’s Wegleitung check digit) and ch-epd-pid the EPD patient identification number (EPDV-EDI Anhang 1, the fixed 76133761 prefix and its check digit), each transcribed from the issuing authority’s own publication and tested against its worked examples. With a Swiss rule active, the server refuses to boot with an audit retention_days under the one year DSV Art. 4 Abs. 5 requires. The AHVV joins the vendored corpus at docs/law/ch/ahvv/.
  • The regulation corpus gains the German and Swiss layers (#3293, #3294). docs/law/de/ carries the BDSG, SGB V (the ePA and telematics provisions, §§ 341–355 and 360–363), the GDNG and the StGB (§ 203) as the XML gesetze-im-internet.de publishes, pinned by the publisher’s build stamp, plus the Digital-Gesetz as its Bundesgesetzblatt issue; docs/law/ch/ carries the DSG/FADP and DSV/DPO in German and the non-binding English, and the EPDG, EPDV and EPDV-EDI in German, each at its current Fedlex consolidation. Two vendor scripts, provenance and digests per act, the publishers’ no-copyright articles quoted in LICENSES/, and the compliance overview gains a Germany and a Switzerland section stating, provision by provision, what the software ships and what the deploying organisation must do.
  • The Helm chart runs FerroTERM beside the CDR (#3305, chart 8.3.0). terminology.enabled renders the terminology server as its own workload — Deployment, ClusterIP Service, ServiceAccount, NetworkPolicy — under the same Restricted profile as the CDR, and points the CDR’s config.terminology.external at its Service, so archetype bindings resolve at commit and AQL TERMINOLOGY() expands through it (the CDR’s own /terminology/* extension routes stay a separate switch, config.terminology.api_enabled, and are off by default). Every values file that contradicts that injection is refused at render rather than silently overwritten: a default provider of your own, an external.enabled: false, a fail_on_error disagreeing with terminology.failOnError, and a provider with no external.routes entry naming it. The licence-free shaped seed ships with the chart and is mounted as a ConfigMap, so an install from the registry serves terminology with no checkout; a release you hold a licence for is an index built off-cluster and mounted from an existing claim named in terminology.index.persistentVolumeClaim — which also switches the workload to strategy: Recreate, because a rolling update’s surge pod cannot attach a ReadWriteOnce volume the outgoing pod still holds. No Ingress is rendered for it and no value renders one: the CDR is its only caller, and its NetworkPolicy admits the CDR’s pods alone. No boolean opens that port; terminology.networkPolicy.extraIngressFrom admits a peer you write, and the values schema refuses an empty one (a bare namespaceSelector: {} selects every namespace). Above one replica, the pods spread one per node and terminology.podDisruptionBudget guards them through a drain.

Fixed

  • Under multi-tenancy an unset ferroehr.tenant_id no longer resolves to the reserved default tenant: the server stamps the tenancy posture at boot, every connection it opens declares its tenant explicitly, and the tenant reader refuses a connection that declares none. Single-tenant deployments are unchanged (#3341).
  • Under multi-tenancy the AMQP outbox drainer and the FHIR outbound emitter drain every registered tenant’s rows in turn, with a delivery cursor per tenant, instead of the default tenant’s alone (#3355).
  • The event-outbox retention prune stops at the lowest cursor of every active reader instead of deleting on age alone: the FHIR outbound emitter (and any future cursor reader) now owns a row in the new event_outbox_reader registry, reconciled from the configuration at boot, and a published row it has not reached survives the retention window. The singleton fhir_outbound_cursor table moves into the registry with its watermark (#3330).

4.2.5 - 2026-09-12

Changed

  • The compliance overview’s GDPR section covers the data-subject rights, the processor terms, breach notification and research (#3326). From the audit recorded on #3322: rows for Art. 12(3), 15, 16, 17, 18, 19, 20, 21, 28(3)(e) to (h), 33(3)(a) with 34(3)(a), and 89(1), each naming what the product ships, the tracker status and what the deploying organisation must still do, with the same split mirrored on the shared-responsibility page. Where the text is silent, on encryption at rest and on audit-log retention, the rows say so instead of implying a requirement.

  • The sandbox dataset pins 231 accepted / 91 refused ADL 2 archetypes (#3315). Beside a terminology server that serves LOINC, two archetypes of the 2013 CKM corpus are refused by the VETDF check for LOINC answer codes that do not exist: apgar binds fifteen codes without their check digit (LA6713 … LA6727) and braden_scale binds the mistyped LA9605-4. The sources stay verbatim, the manifest records the adjudication, and the getting-started page counts 233 archetypes.

4.2.4 - 2026-09-12

Fixed

  • ADL 2 archetypes with SNOMED CT or LOINC term bindings upload against a terminology server that does not serve those code systems (#3311). The VETDF check asked the server about system=SNOMED-CT&code=<the binding URI> and read every miss as “the term does not exist”, so with FerroTERM beside the CDR six archetypes of the sandbox dataset were refused and the seed failed. A binding target in the IHTSDO URI model (http://snomed.info/id/…, http://loinc.org/id/…) is now taken apart into the FHIR system and code the server is asked about, and the server’s OperationOutcome decides the outcome: a code system it does not serve leaves the binding unverified and the archetype accepted with a warning, as AOM2’s VETDF prescribes; a code absent from a served system is still refused with 422. Verified against a real FerroTERM in the test suite.

4.2.3 - 2026-09-12

Added

  • FerroTERM ships with the quickstart and runs beside the sandbox CDR (#3304). docker-compose.terminology.yml starts the Ferro family’s FHIR terminology server next to the CDR and switches [terminology.external] on, serving the licence-free shaped seed out of the ferroehr image by default and any edition you build into its index volume (ferroterm-build, the terminology-build profile). The hosted sandbox runs the same pair on one machine with the terminology server reachable only from inside the stack, which is the posture the SNOMED CT Affiliate Licence asks of a public system, and its dataset gains a composition whose coded text resolves through it. The deployment probe observes the overlay off, on and resolving, and the server down under both fail postures. FerroTERM is pinned at 0.1.3, whose image carries its own HEALTHCHECK (ferroterm healthcheck), so both compose files start the CDR only once the terminology server reports healthy (#3307).

4.2.2 - 2026-09-12

Fixed

  • A wiped clinical schema comes back with its cold-tier alias views (#3298). The three ehr.cold_* views the storage layer reads through were created by the demographic migration set, so a database whose ehr schema was dropped and rebuilt (the hosted sandbox after every nightly reset since 4.2.0) came back without them, and every composition or EHR_STATUS update answered 500 while creates and reads worked. A new clinical migration recreates the views whenever the demographic set has run, and widens a view created before vo_version.origins existed. The sandbox reset now drops all seven schemas, so a reset is a fresh start rather than a partial one.
  • The compose quickstart role can reach the linkage domain. The database init script granted the application login the clinical and demographic domain roles but not ferroehr_linkage; it worked only because that login owns the database.

4.2.1 - 2026-09-12

Fixed

  • Identifiers carry the licence stamp the design specifies (#3282). The stamp keyed its HMAC with a hash of the key material and used a fail-safe constant of its own, so no identifier a 4.2.0 server minted could be attributed to its licence by the attestation tooling; the key is now the raw material and the unlicensed constant is the design’s. Identifiers minted by 4.2.0 stay valid and simply carry no recognisable stamp.

4.2.0 - 2026-09-12

Added

  • The storage-parity sweep reports rows an older release decomposed (#3273). A fifth defect value, stale_decomposition, on POST {base}/admin/integrity/verify and its NDJSON stream: the rows reassemble into exactly the stored document, so nothing is damaged, but they are not the rows this version writes for it. The sweep now decomposes each version’s document a second time and compares, which is the only way that case is visible; POST {base}/admin/integrity/rebuild-nodes repairs it like any other finding. A release that changes how content is decomposed is therefore an upgrade step an operator runs and verifies, instead of a shape that waits for each object’s next commit.

  • Cohort queries across the pseudonymisation boundary (#3159). A new POST /query/cohort runs an AQL query over a population selected in the demographic domain: a predicate from a deployment-declared allow-list ([cohort.predicates] — city, postcode_area, sex, age_band, organisation, each naming an archetype, an at-code and how a value is matched) selects parties, the linkage domain resolves those parties to the EHRs they are the subject of, and the caller’s AQL runs scoped to exactly those EHRs. The three steps run on three separated database credentials and carry identifiers only — no name, address or party id reaches the caller. A result set serving fewer distinct EHRs than cohort.small_cell_threshold (default 5) has its rows withheld and is marked suppressed rather than looking empty; a predicate matching more than cohort.max_cohort_size parties (default 100000) is refused 422 rather than truncated. Every execution, suppressed and failed ones included, records one linkage-domain access event naming the cohort by a digest of its predicate list and never by its values. The surface is off until a predicate is bound, and the route answers 404 until then.

  • AQL EHR scoping binds one array instead of one parameter per id. A query scoped to a set of EHRs now emits ehr_id = ANY($n) with a single uuid-array bind, lifting the PostgreSQL 65535-parameter wire ceiling that previously bounded how many EHRs one query could name. Behaviour is otherwise identical. The per-EHR served-row count behind the access records and the existence check of a scoped id list are now keyed rather than probed, so a population page of a hundred thousand rows from as many EHRs is served in linear time (the 100 000-EHR cohort measured at fifty seconds before, under a second in the database itself).

  • The viewer shows the connected CDR’s deployment profile (#3264). The header’s status chip carries the declared profile (· production or · sandbox), and a sandbox deployment raises a persistent notice under the header on every authenticated screen stating that it has not made the production separations and must not hold real patient data, naming the open separations the server reported. Production renders quietly; a server that declares no profile changes nothing.

  • Deployer-facing privacy documentation (#3161). Three pages under Security: a data protection impact assessment (processing description, data categories per schema, roles, retention including the Dutch five-year access-log floor, a risk register, and the shipped controls mapped to GDPR, EHDS, EDPB 01/2025 and NEN by closed issue), an Article 30 records-of-processing template pre-filled with the activities FerroEHR performs by design, and a go-live checklist linked from the production checklist.

  • The threat model, architecture and storage pages describe the pseudonymisation boundary as shipped (#3166): what each database credential and each backup artefact can reach, the re-identification paths and the control against each, which service module runs on which pool and role, and all seven schemas including linkage.

  • The FHIR page records the priority-category mapping decision (#3206): profile mappings are not authored here and belong to FerroBRIDGE once the EHDS Article 36 implementing acts fix the exchange format; the trigger that reopens the question is stated, and the two priority categories with no committed template are recorded as an adjudicated corpus boundary.

  • The access log records the origin of the data it served (#3212, EHDS Annex II 3.2(e)). Every committed version now carries the distinct set of FEEDER_AUDIT.originating_system_audit.system_id values found in its body, or this server’s own system id when it carries none (vo_version.origins, derived at commit; an unstamped row is assessed at read). A composition, status or folder read records the origins of the version it served, an AQL execution the union over the version rows its result page came from, each with the true distinct count beside a set capped at 32; the FHIR AuditEvent carries one named entity per origin and ITI-81 serves it.

  • The server mints the subject pseudonym (#3232). Where FerroEHR itself makes a party the subject of an EHR, service::linkage::link_as_subject derives the pseudonym (a keyed, tenant-bound derivation over the party under the linkage key, an RFC 9562 custom UUID), writes it as the EHR’s EHR_STATUS.subject.external_ref in the first declared namespace, and opens the mapping; no caller-supplied value enters that path. Minting refuses, typed, when no namespace is declared or no identifier-protection key is configured.

  • The access log has a retention floor per jurisdiction (#3242). A non-zero audit.store.retention_days below the floor a jurisdiction the active identifier rules name publishes is a boot error naming both numbers. Registered: NL, five years (1830 days) under the Besluit vaststelling bewaartermijn logging. 0 keeps forever and always passes; the effective retention is on the boot line and on /management/info.

  • A declared deployment profile (#3226). The new top-level deployment_profile key (sandbox, the default, or production) says what a deployment may hold. production refuses to start while a separation it asserts is missing and not accepted by name in deployment_accepts: separated database credentials, separated clusters (read from pg_control_system().system_identifier on each pool, never from the DSN text), a declared subject pseudonym namespace, an audit trail with a durable sink, schema preparation on its own credential. sandbox names every missing separation on the banner, at warn with a structured posture = "deployment" field, and on GET /rest/status under deployment. Our own posture, not a legal requirement; the documentation says so and why.

  • The database guards the subject pseudonym (#3241). Once privacy.subject_namespaces is declared, a trigger on ehr refuses a subject reference that is not a UUID whichever code path or session writes it; the server stamps the posture on every boot and reports how many stored EHRs a newly declared namespace finds already carrying a non-pseudonym subject.

  • A subject-scoped read of the access log (#3240). GET /fhir/r4/AuditEvent was admin-only, so a portal serving a person’s right to know who accessed their record (GDPR Art. 15, EHDS Art. 9) had to hold an admin credential over every patient’s log. A new authz.rbac.subject_audit_role reads the log for one subject at a time, patient required and refused with a 403 without it; the admin keeps the unscoped retrieval, and reading a subject’s log is itself recorded as an access naming that subject. Unset, nothing changes.

  • The access record carries the caller’s roles (#3239). NEN 7513 asks under which role or authority a person accessed a record; the RBAC layer held the caller’s roles and the record dropped them. Every access event, the request-level ones and the domain-level linkage, subject, identifier and EHR-Extract records alike, now stores the roles as a JSON array (audit.audit_event.roles, NULL for an unauthenticated refusal), renders them as FHIR agent.role on the requestor and as a RoleIDCode per role on the DICOM source participant, and serves them through ITI-81.

  • Auditing off and fail_mode = "open" are said where an operator looks (#3238). Auditing disabled used to be one info-level line; the fail-open default was silent. Both are now a warn at boot with a structured posture = "audit" field, the audit_sender readiness indicator is registered whether or not auditing is on (DEGRADED with the consequence stated when it is off, UP with fail_mode, the local store and retention otherwise), and GET /management/info carries the audit posture. The audit page states at its top what each posture means.

  • The licence in force is explicit, verified and reported. Every build embeds the licensor’s non-commercial grant (what BUSL-1.1 gives everyone); a commercial licence is installed with the new [licence] file key (FERROEHR__LICENCE__FILE), an OpenPGP-signed token verified at boot against the licensor’s embedded public certificate. The server behaves identically under either grant: the outcome is logged once at boot and served on GET /rest/status as licence: {state, use, licensee, not_after, configured_token}. Every identifier the server mints (EHR ids, versioned object ids, contribution ids) carries sixteen keyed bits derived from the licence id in the trailing bytes of its UUIDv7, so a record says which grant it was written under; the bits replace random bits only, never the timestamp or counter, and carry no patient data. A configured token that does not verify leaves the embedded grant in force and is reported as refused.

  • The linkage resolution service: one path from a person to a record, over two credentials, under audit (#3158). ferroehr::service::linkage opens, resolves, merges and splits the party-to-EHR mappings the previous increment created the schema for. resolve_ehr_for_identity is the only path that crosses from an external identity to an EHR, and it crosses in the APPLICATION: the sealed identifier map answers with a party on the demographic pool, the linkage map answers with an EHR on the linkage pool, and no statement on either could perform the join, because no credential holds both schemas. Merge and split are period-closing writes in one transaction — the mapping in force gets an end date and its successor opens at the same instant — so no mapping is ever deleted and which party was the subject of an EHR when a composition was written stays answerable. Every operation records a linkage-domain access event naming the actor, the purpose of use the caller declared and whether anything matched; the resolved EHR is deliberately counted and not named, because the audit trail lives outside this domain’s role and a party/EHR pair there would be a second copy of the map. GET /ehr?subject_id&subject_namespace is unchanged on the wire and still resolves through the clinical columns — it matches an opaque pseudonym, not an identity — but now records a linkage-domain access event of its own, because who resolved a subject to a record is worth knowing wherever it happened.

  • [db] linkage_url, the third runtime credential (#3158). Alongside url and demographic_url (and migrate_url for schema preparation), a deployment can now point the linkage pool at its own DSN, with the linkage_url_file mounted-secret sibling every other credential has; setting both is a boot error, and leaving it unset shares the clinical DSN exactly as before. The Helm chart carries it as database.linkageExistingSecret / linkageExistingSecretKey / linkageUrl, mounted as a file so only its path reaches the pod’s environment (chart 8.2.0). With all three set, no credential the server uses can re-identify a record in SQL.

  • A third pseudonymisation domain: the linkage schema holds the party-to-EHR map (#3158). A migration adds the linkage schema, the linkage.party_ehr table and the ferroehr_linkage role. The table records which demographic party is the subject of which EHR, temporally: a PostgreSQL 18 WITHOUT OVERLAPS primary key admits one open mapping per party per tenant at any instant, so a merge or a split closes a row and opens another instead of deleting history. It carries identifiers only — no name, no address, no plaintext identifier — because a row is already the additional information that re-joins a pseudonymised record to a person (GDPR Art. 4(5)). The grants run in both directions: the clinical and demographic roles are revoked from linkage, and ferroehr_linkage is revoked from ehr, cold, demographic and cold_demographic, so no one credential holds the map and either side of it. The boot gate (verify_domain_isolation) now covers all five runtime roles and refuses to serve when any of them can read across. Nothing reads the new schema yet; the resolution service is still to come. Upgrading installations get an empty schema and no data movement.

  • A third backup: the linkage domain is dumped on its own, and the deployment probes cover its boundary (#3220). The Compose backup profile gains ferroehr-backup-linkage (FERROEHR_BACKUP_LINKAGE_DIR, defaulting to ./backups/linkage) and the chart gains a third CronJob under backup.linkage, with its own schedule, claim and BYPASSRLS credential. It is deliberately a separate artefact rather than a place in the demographic dump: linkage.party_ehr is the map that re-identifies a pseudonymised record, so a file carrying it beside either side of that map rebuilds the join the split exists to withhold (GDPR Art. 4(5)). The chart’s render guards are now pairwise — any two domains naming the same claim is refused, not just the original pair — and the operations page documents three dumps and a restore that supplies all three. The linkage dump carries --extension=btree_gist, because a --schema dump carries no extension and linkage.party_ehr’s temporal primary key is a GiST index over that extension’s operator classes: without it the table restores with its rows and without the key, and pg_restore reports that as an ignored error. Enabling backup therefore now requires a claim and a Secret for the linkage domain too; a values file that sets only the first two is refused at render, naming what is missing — a values contract change, so the chart goes to 8.0.0.

  • The audit trail records the accessing organisation, and the declared purpose reaches the FHIR export (#3204). An access record now carries an organisation column: the organisation the caller acted for, read from the access token’s [authz.abac] organization_claim — the same claim the ABAC layer decides on, so no second setting can name a different one, and recorded whether or not that layer is switched on. It answers EHDS Annex II 3.2(a) beside the natural person of 3.2(b), and where no claim is configured or the caller used Basic authentication it stays empty rather than being guessed. The FHIR AuditEvent export renders it as a second agent referencing an Organization, and the purpose of use as agent.purposeOfUse on the requesting person’s agent, coded with the deployment’s own vocabulary and no code system. The DICOM syslog form carries neither: PS3.15 §A.5 defines no organisation attribute and no purpose element. Existing records keep a NULL organisation — the column is added by a migration and nothing is backfilled.

  • Which EHDS priority categories the FHIR façade carries, measured rather than asserted (#3171). The FHIR chapter now carries a generated matrix of the six Annex I priority categories: which has a committed CKM template, which FHIR resource a profile mapping would target, and whether the transform is proven end to end. Four of the six — patient summaries, electronic prescriptions, imaging reports and laboratory results — have a real committed template whose example composition flattens and reverse-maps into a FHIR resource carrying its own values, asserted per category with the mapped leaf derived from the flat map rather than written into the test. Electronic dispensations and discharge reports have no committed template, and the table says so. It also says the three things it does not establish: no profile mapping ships for any category, the test exercises the transform rather than the endpoint, and none of it is conformance to an exchange format that has not been adopted. Those gaps are #3206.

  • The EHDS logging elements are mapped onto the audit trail, gaps included (#3170). Annex II 3.2 of Regulation (EU) 2025/327 lists five things the European logging software component must record on every access event. The audit page now maps each of them onto an access-event field and onto both renderings the trail leaves the server in, and says plainly where there is nothing: the accessing organisation (a) and the origin of the data (e) have no field, the category (c) is an openEHR resource class rather than an Annex I priority category, and the declared purpose of use is stored but reaches neither export. Every row is asserted by a test — including the gaps, which fail the day the field arrives, so the table and the code cannot drift apart in either direction. The retention default and the two export routes are documented beside it, with the reason retention_days = 0 keeps records forever rather than choosing a horizon for an operator. Closing the gaps is #3204.

  • Two EHDS readiness pages, generated from one source (#3169). Chapter III of Regulation (EU) 2025/327 requires an EHR system to include two harmonised software components, meet the essential requirements of Annex II, and carry technical documentation and an EU declaration of conformity. The documentation site now states, requirement by requirement, what FerroEHR provides today: an EHDS readiness page with a status and evidence for each of the 14 Annex II requirements, and a technical-documentation page mapping each Annex III element to material that exists. Both are rendered from a committed YAML by scripts/render/ehds-pages.sh, and a CI job re-renders and diffs, so a status typed into a page fails the build. Both pages say plainly that no conformity assessment has been carried out, no technical documentation has been drawn up and no declaration of conformity exists — a “shipped” row is a statement about the software, never a claim of conformity.

  • openehr-its compiles to wasm32-unknown-unknown with the flat and OPT code in it (#3149). The crate’s one full feature dragged axum, moka and jsonschema in with the parsers, so a browser consumer that wanted flat::build_web_template or the OPT 1.4 reader had to take an HTTP server and a cache too — and default-features = false left only the SMART scope grammar. The features are layered now: json → xml → opt14 → flat is the parsing spine, and cache (moka), schema-validation (jsonschema) and rest-server (axum, http, async-trait) are the layers a browser declines. full is still the default and still means all of them, so no existing consumer sees a change. A CI lane builds the wasm selection, and a second one builds the dependency-free rest::smart_scopes island, so both are checked rather than claimed.

  • Backups are taken per pseudonymisation domain (#3157). A single pg_dump of the whole database produces one file holding both the pseudonymised clinical record and the identities of its subjects, and whoever can read that file can re-identify every record in it. The Compose stack gained an opt-in backup profile with one job per domain, writing to two separate directories (FERROEHR_BACKUP_CLINICAL_DIR, FERROEHR_BACKUP_DEMOGRAPHIC_DIR), and the operations page carries the pg_dump --schema recipe, the restore procedure, and the two properties no configuration can enforce for you: different access control on the two targets, and a credential per job that reaches one domain. Point-in-time recovery stays instance-wide and the page says so. A backup credential needs BYPASSRLS: every tenant-scoped table carries FORCE ROW LEVEL SECURITY, so pg_dump refuses a table it would read through a policy and the application role cannot take a backup at all — and --enable-row-security, which makes the dump succeed, quietly writes a single tenant’s rows. The chart therefore requires a backup Secret per domain and refuses to render without one. ferroehr db verify is the restore gate — it issues no DDL and refuses a database whose grants let a runtime role read across the boundary, which is the same check the server runs at boot.

  • The deployment artifacts can reach the demographic role split (#3179). The chart gained database.demographicExistingSecret (and its demographicExistingSecretKey / inline demographicUrl siblings), mounted as a file exactly like the clinical DSN so neither credential enters the pod’s environment. Set it and the schema separation becomes a credential separation: one leaked connection string then reaches one domain rather than both. Leave it unset and both pools share the clinical DSN, which is the schema-only posture the boot self-check still verifies. The compose stacks now provision all four NOINHERIT domain roles — previously the migrations skipped them with a NOTICE, because the compose migrator holds no CREATEROLE, so those stacks ran with no domain grants at all — while deliberately keeping one login credential, which the init script says in place. scripts/deploy-probe.sh gained a family that reads the posture out of the database catalogue rather than the compose file: the roles exist, each is NOINHERIT, no clinical role can read a demographic relation, and the stack is single-credential on purpose. What it cannot show — a two-DSN deployment, and role provisioning on a managed database — it declares as not exercised.

  • National identifiers in the demographic domain can be held sealed (#3155). With the new [demographic.identifier_protection] section on, an identifier of a configured scheme never sits in the versioned body: the value moves to demographic.national_identifier, sealed with AES-256-GCM under a key derived per tenant from one configured root key, and the body keeps a reference in its place. Beside the ciphertext sits an HMAC-SHA-256 digest, so “which party holds this identifier” is answerable without decrypting anything — and, being keyed, it cannot be reversed by enumerating a nine-digit space the way an unkeyed hash could. Sealing runs before the body is decomposed and signed, so the stored, signed and served body stay one form. Only the demographic writer reaches the sealed value; the read-only twin is refused the ciphertext and the digest by column-level grant and the clinical roles are refused the table outright. Resolution goes through a SECURITY DEFINER function that takes the digest rather than the value, and every call is recorded as a linkage-domain access naming the scheme and whether it matched, never the value. Off by default, and enabling it without a key — or with a scheme list that would protect nothing — is a boot error. See Privacy & data minimisation and the key-rotation runbook in Operations.

  • Reads are logged per record, not just per operation (#3156). Every retrieval in the EHR, Query and Demographic APIs already produced an audit record; each one now also says which pseudonymisation domain it read (ehr, demographic, linkage, or system for an operation that touches no domain), the purpose of use the caller declared, the legal basis the deployment processes under, how many records were served, and the x-request-id of the request it belonged to. An AQL statement additionally produces one access record per EHR it served, each with that EHR’s served-row count, because one execute record over a population query cannot answer whose record was read. The served set is read off the rows the caller actually received rather than the rows the statement matched: a legal record of access must not name an EHR whose content was never disclosed. SELECT DISTINCT and aggregate projections carry no per-EHR breakdown — the shape makes it underivable — and record the statement and its row count. Two new keys configure the declaration surface: [audit] purpose_header (default x-purpose-of-use) and purpose_codes, an allow-list that keeps unagreed free text out of the trail, plus [audit] legal_basis. A test walks the generated route tables so a newly generated GET that emits nothing fails the build, and the trail’s append-only property — no rewrite, no delete outside the retention reaper, no truncation — is now asserted rather than assumed. See Audit trail, which maps the NEN 7513 event content onto the recorded fields.

  • The clinical side refuses identifying data, and the subject reference is bound to a pseudonym. A new [privacy] configuration section carries three write-path rules, all reported as 422 with one entry per finding naming the RM path and never the offending value. privacy.subject_namespaces names the pseudonymisation domains this deployment issues subject pseudonyms in; once one is declared, EHR_STATUS.subject.external_ref must name a listed namespace and carry a UUID. privacy.allow_identified_parties_in_ehr (false by default, announced at boot when on) decides whether a PARTY_IDENTIFIED or PARTY_RELATED in clinical content may carry name or identifiers — with it off they carry external_ref only, which still satisfies the RM’s own validity rule. [privacy.identifier_scan] checks every string leaf of every clinical write against a jurisdiction-keyed ruleset, refusing in strict (the default) and recording in warn. The shipped rules are fi-hetu, gb-nhs-number, nl-bsn, no-fodselsnummer and se-personnummer, each transcribing the checksum its own issuing register publishes, and all are active by default; a deployment narrows the list to its own jurisdictions and adds patterns for local medical-record numbers and address forms. Every rule publishes how often it claims a value that is not one of its identifiers, and those figures are measured rather than estimated: one in 11 for nl-bsn and gb-nhs-number, one in 31 for fi-hetu, one in 120 for no-fodselsnummer, one in 139 for se-personnummer, and zero findings across the 258 clinical documents in this repository’s vendored corpora. See Privacy & data minimisation.

  • Demographic parties live in their own schema, behind their own database role (#3153). Parties (PERSON, ORGANISATION, GROUP, AGENT, ROLE, PARTY_RELATIONSHIP) and their change control leave the clinical ehr schema for a new demographic schema with its own cold_demographic archival tier, and the database refuses the mix in both directions: a version with no owning EHR cannot enter ehr, and one that has an owner cannot enter demographic. Four runtime roles replace the single ferroehr_app pair: ferroehr_ehr, ferroehr_demographic, and a read-only twin of each. Each holds explicit grants on its own domain and carries an explicit revoke on the other, all four are NOINHERIT, and none is a member of another. The schema separation is unconditional. The new [db] demographic_url (or demographic_url_file) turns it into a credential separation as well, by pointing the demographic pool at its own DSN. The server now refuses to boot when the grants are wrong: a self-check enumerates every table, view, sequence and function in each domain and fails, naming the role and the object it can reach. ferroehr db verify runs the same check, and it passes when the roles do not exist, which is the development, compose and test-harness case. The ITS-REST Demographic API, the RM change-control semantics and every version identifier are unchanged. GDPR Art. 4(5) and Art. 32(1)(a), EDPB Guidelines 01/2025; no openEHR spec governs storage layout or database roles. Existing installations migrate in place: the upgrade moves the parties by kind, and refuses before moving anything if it finds an EHR-less row of a kind it does not classify.

  • The documentation site carries a compliance control matrix, generated from the tracker (#3164). A hand-kept compliance table is wrong the first time a control ships, so this one is not kept by hand: an issue declares the control it delivers as a Control: <legal source> <article or clause> line in its body, and scripts/render/control-matrix.sh renders website/book/src/compliance/control-matrix.md from the tracker. Each row shows the legal source as a link to its official publisher, the article or clause, the issue, whether the control is shipped, in progress or planned, and the pull request that closed it. Short names resolve through a registry inside the generator, so a citation on the page can never be free text: an undeclared source stops the render and names the issue. A CI job re-renders and diffs, so a control that shipped since the last render fails the build instead of sitting on the public site as “planned”.

  • The documentation site has a compliance section (#3163, #3165). Two pages join the generated control matrix. The compliance overview states the pseudonymisation boundary, then walks GDPR, the EDPB Guidelines 01/2025 on pseudonymisation, the EHDS, the Dutch UAVG and Wabvpz, and NEN 7510, 7512 and 7513, saying per requirement what the product ships, what is planned and under which issue, and what the deploying organisation still has to do. The shared-responsibility page draws the same line obligation by obligation, because most duties in these texts rest on the controller and the processor and cannot be met by software. Both pages carry the wording policy of the compliance programme: FerroEHR aims to be the first openly developed, source-available openEHR CDR with a published, tracker-backed EU compliance posture and an EHDS conformity self-assessment on its roadmap, and it claims no certification, no conformity and no compliant deployment. The schema and role split is described as shipped; the resolve map, the identifier constraints and the per-domain logging are described as planned, with their issue numbers. Every legal source is linked to its official publisher, and every openEHR statement to the published specification section it comes from.

  • The storage-parity sweep has a repair to go with it (#3143). POST /admin/integrity/verify reported that a version’s decomposed rows and its stored document disagree, and then there was nothing to run: dump and load was the only re-decomposition path, and its dump side reads each version through the node rows themselves, so it reproduced the damage it was asked to fix. POST /admin/integrity/rebuild-nodes re-derives the rows of every damaged version from that version’s stored document, one transaction per version. It runs the same sweep first and writes only what the sweep reports damaged, so a run over healthy data writes nothing. Repair goes in one direction only: the stored document is the canonical serialized form a point read serves and a digest was taken over, the rows are an index derived from it, so a version whose document is the damaged copy is refused rather than having the damage copied into the index. Each version is read under a row lock, decomposed, its rows replaced, and the version re-derived from the new rows and compared with the document before the commit; anything that fails rolls that version back untouched and the run continues. An archived object is thawed and re-archived in the same transaction, marker included. A logically deleted version rebuilds to no rows.

  • Contributor rules for personal data, and a guard that holds them (#3167). CONTRIBUTING.md and the contributing page of the book gained a “Personal data” section: synthetic data only in tests and fixtures, no identifiers in logs, traces or Debug output, no database grant spanning the clinical and demographic domains, and a review step for any change at the privacy boundary. The pull request template carries the matching “Privacy boundary” checklist, and the new privacy-boundary-guard CI job fails a pull request that touches the boundary paths without a ticked checklist. The same job reads the added lines of every diff and refuses a nine-digit value passing the BSN eleven-test or a Dutch postcode followed by a house number, proving both detectors against passing and failing fixtures before it judges anything. .claude/rules/personal-data.md carries the rules for agents.

Changed

  • The control matrix states the control, not the defect (#3249). An issue titled as the defect it fixed may carry a Control-text: line beside its Control: declaration, and the generator renders that text in the Control column; seven shipped rows that read as defects beside “Shipped” now read as controls. The Helm chart’s linkage backup comment and the deployment probe’s coverage note describe the three server pools as shipped (#3268).

  • The library’s default privacy policy is the shipped one (#3243). PrivacyPolicy::default() used to permit identified parties and scan nothing, so a service built without an explicit policy ran a more permissive posture than the binary; it now compiles to the same refusing posture the default configuration does. Embedding hosts and tests that mean to accept more install the policy that says so.

  • The control matrix lists a control’s closing pull request only once it has merged (#3253). An open pull request that says Closes #N no longer moves a row, and the pull request that closes a control regenerates the page as it will read after the merge (--closing <pr>), so the published matrix never lags a shipped control and unrelated pull requests stop failing its check.

  • The two-DSN posture is exercised, and the deployment probe says exactly what still is not (#3222). Pointing [db] demographic_url at a role of its own turns the schema separation between the clinical and demographic pseudonymisation domains into a credential separation, and nothing ran it: every test drove both domains through one login role. A new integration suite assembles the server the way the binary does — two DSNs, two pools — on one login role per domain against a real PostgreSQL, serves an EHR and a party through the wire, and asserts that each of the running server’s own pools is refused the other domain’s relations for want of privilege. The single-DSN fallback is covered beside it, so the split stays a deployment choice. The Compose deployment probe keeps its honest “not exercised” entry for credential separation, now narrowed to what genuinely remains: a real deployment booting on two mounted DSNs. No behaviour changes.

  • A default deployment accepts a named composer again (#3190). With [privacy] allow_identified_parties_in_ehr off, the clinical side refused both name and identifiers on both party proxy classes, so a server running the shipped configuration rejected the very composition its own /definition/template/adl1.4/{id}/example endpoint hands out — anything built from ctx/composer_name included. The default now follows what the two Reference Model classes actually mean. A PARTY_IDENTIFIED name is accepted: that class is the proxy for a party “other than the subject of the record”, “Typically for health care providers”, so a composer or performer name is provider identity. Still refused: identifiers on either class, which is the national-identifier slot, and name on a PARTY_RELATED, which is defined by its relationship to the subject and may be the subject itself. The configuration key is unchanged and means what it always did when set to true, so a deployment that opted in keeps working, and nothing that committed before this change is refused now.

  • The published model crates no longer carry a random-number generator. The workspace pinned uuid with v4/v7/fast-rng for everyone, but only the server and the test harness ever mint an identifier; openehr-base, openehr-rm and their siblings parse and serialize them. The unused generator is what made those crates refuse to compile for wasm32-unknown-unknown, which has no randomness source unless one is configured — so a browser consumer had to supply a backend for a capability none of the crates uses. The generator features now sit on the two crates that generate.

  • A key derived for one pseudonymisation domain opens nothing in another (#3157). The per-tenant subkeys protecting national identifiers now take the domain as part of their derivation context, so the clinical domain’s key material cannot decrypt a demographic record or reproduce its lookup digest even when both derive from the same configured root key. That is what makes a per-schema backup a separate artefact under a separate key rather than two files under one. [demographic.identifier_protection] has not shipped in a release yet, so no stored data is affected; a deployment that enabled it on an unreleased build re-seals its identifiers.

  • The Rust toolchain moves to 1.98.1 and the MSRV to 1.97. 1.98.1 is a point release for a vtable-generation miscompilation, which is reason enough on its own; the toolchain step also brings the 1.98 lint set, and the declared minimum supported version moves with it, one minor behind as before. Container builds follow: the builder base is the digest-pinned rust:1.98.1-slim-trixie, resolved fresh rather than copied. Nothing here uses a 1.98 language feature, and the msrv job verifies the declared minimum independently.

  • Dependencies bumped, and one advisory left the tree with them. argon2 0.6, jsonschema 0.53, tower-http 0.7.1, rust_decimal 1.43, futures 0.3.34, toml 1.1.5, indexmap 2.14.2, leptos-use 0.19.2, wasm-bindgen 0.2.128. rust_decimal 1.43 dropped its optional rkyv ^0.7 requirement, so RUSTSEC-2026-0235 — an out-of-bounds read no feature of ours ever compiled, but which lock-file scanners reported — is gone from Cargo.lock and its VEX statement is deleted rather than renewed. The three remaining advisory exceptions were re-checked against their upstreams on 2026-09-10 and all three still have no fixed release.

  • Database migrations are append-only, so an existing installation upgrades in place. A released migration is never edited again. sqlx records a checksum of every applied migration and refuses a database whose recorded checksum no longer matches the file, so editing one would not revise history: it would stop every server that had already run it from starting, reporting a checksum rather than the change that caused it. A correction now arrives as a new migration that carries the schema forward, and a CI guard fails any pull request that modifies, renames or deletes a migration its base branch already carries. This replaces the earlier greenfield policy, under which baselines were edited in place and deployments were expected to recreate their volume.

  • Both storage-integrity routes can be pointed at a single version (#3143). vo_id and sys_version join ehr_id and committed_since as scope parameters on POST /admin/integrity/verify and the new rebuild, so checking or repairing one record after a support incident no longer means reading the repository. The ordinal is per-object, so sys_version without vo_id is a 400 rather than a filter that would match one version of every object.

Security

  • Three tenant-scoped tables enforce tenant isolation that they carried a column for but no policy (#3219). demographic.national_identifier holds sealed national identifiers and the keyed digests that resolve them; demographic.contribution and demographic.audit hold the change-control trail over the demographic domain. All three carried a tenant_id, and the equality lookups on them were tenant-scoped, so no resolve could ever cross a tenant. What was exposed is the unscoped read: a plain SELECT by the demographic role returned every tenant’s rows. Each now enables and forces row-level security with the same tenant_isolation policy its sibling relations have carried since the baseline. Deployments that run a single tenant are unaffected, because an unset session resolves to the reserved default tenant exactly as before.

    The cause is worth stating, because it will recur otherwise: a baseline migration applies row-level security by looping over a list of relation names, and a table added by a later migration joins nothing. The omission is invisible in the table’s own definition. A test now asks the catalog instead of a reviewer — any relation carrying a tenant_id must enable and force row-level security — so the next table added after a baseline is caught by the build rather than by an audit. audit.audit_event is the one named exception: its drain batches many tenants’ records in a single insert from a task outside any request’s tenant session, so a policy there would reject audit rows rather than isolate them.

Fixed

  • A party’s contacts, addresses, identities and capabilities are decomposed into node rows (#3273). CONTACT, ADDRESS, PARTY_IDENTITY and CAPABILITY were stored inline on the party root, so nothing under them carried a row and no row-level predicate could reach it — a cohort predicate could only be bound to an archetype under the party’s own details, never to the address a party actually lives at. Each container now gets its own row, along with the ITEM_STRUCTURE and ELEMENTs beneath it, and an archetyped container scopes its own leaves. openEHR-DEMOGRAPHIC-ADDRESS.address.v1 is bindable in [cohort.predicates]; the served body is unchanged.

    Upgrade step. After deploying this version, run the storage-parity sweep (POST {base}/admin/integrity/verify). Every party version stored before it reports the new stale_decomposition defect: the rows hold the right content, in the shape the previous release wrote. Run the node rebuild once, unscoped (POST {base}/admin/integrity/rebuild-nodes), and the next sweep is clean. Until then a predicate bound to a container archetype matches a party only from its next commit onward.

  • Every documented least-privilege posture names a credential that can do what the page asks (#3228). db.migrate = "verify" reads all five _sqlx_migrations bookkeeping tables, which no least-privilege runtime role can do, ferroehr_app included; the Compose, Kubernetes, CLI, server configuration and installation pages now name [db] migrate_url, and the operations page says which postures a test exercises and which none does.

  • The split clinical role can write the audit trail (#3267). The local Audit Record Repository was granted to ferroehr_app and ferroehr_reader only, so a deployment whose clinical credential is a member of ferroehr_ehr alone wrote no access record at all (dropped under fail_mode = "open", every auditable operation refused under "closed"); ferroehr_ehr and ferroehr_ehr_reader now hold in audit what the single-domain pair holds.

  • A CONTRIBUTION into a deactivated EHR reports the state conflict first (#3257). The privacy and content validation ran before the EHR_STATUS.is_modifiable gate, so a body defect on a record that cannot be written was reported as 422 where the 409 was the answer; the state is now checked before any member’s content, and the locked in-transaction gate stays authoritative.

  • Documented: under [smart] require_smart_scopes = true a Basic caller cannot reach the template family (#3255). The conformance stack runs that posture, which is why its Basic dev users answer 403 on template upload; the configuration reference and the templates page say so, and a test pins that fail-closed refuses a Basic principal while advisory defers.

  • The management surface is inside the audit trail (#3244). /management/* was mounted outside the audit layer, so a runtime filter change through /management/loggers left no record. Every management request is now recorded as a system-domain access under its own operation id (management_info, management_env, management_loggers_set, management_loggers_reset, …).

  • Formal identifiers on a provider proxy are accepted (#3254). The identified-party rule refused identifiers on every party proxy, which turned away a clinician’s registration number on the composer and the simplified-format ctx/participation_identifiers forms, the Reference Model’s own paradigm case for PARTY_IDENTIFIED. The slot is now refused only on a PARTY_RELATED whose relationship is self; a national-identifier value is still refused by the identifier scanner wherever it sits.

  • A named PARTY_RELATED is accepted unless it is the subject (#3252). The identified-party rule refused every PARTY_RELATED carrying a name in clinical content, which turned away a composition whose consenting mother or guardian is named, content the Reference Model models on purpose and the openEHR REST API obliges a server to accept; the conformance suite went from 1104/1104 to 917/1104 on it. The rule now refuses the name only where the relationship codes self, the party being the patient; identifiers stay refused on both classes, and allow_identified_parties_in_ehr still opens everything.

  • The identifier scan covers the verbatim-replay writes (#3237). EHR-Extract import and the admin archive load stored clinical content exactly as received and skipped the data-minimisation pass every other clinical write takes, while the configuration reference said every clinical write was scanned. Both paths now refuse a record whose body carries a national identifier, a non-pseudonym subject reference or an identified party, with the same 422 and RM path the commit path reports; nothing is rewritten.

  • The control matrix is populated (#3236). The generated compliance page rendered an empty table because no issue declared a Control: line, and the staleness check passed on the empty set. Every shipped and planned control in the pseudonymisation and EHDS programme now declares the article or clause it supports, the page lists all of them with their status, and the generator refuses an empty matrix while the compliance pages link to it.

  • fail_mode = "closed" now covers the domain-level access records too (#3235). The linkage resolutions and merges, the subject-to-EHR lookup behind GET /ehr?subject_id, the national-identifier resolution and the EHR-Extract export and import each record their own access event, and each used to discard the sender’s answer, so a record the audit trail could not take under fail_mode = "closed" left the operation completing unrecorded. A rejected record now withholds the result with the same 503 and Retry-After the request-level record already produces. Under fail_mode = "open" nothing changes.

  • The documented two-DSN posture boots: schema preparation has its own credential (#3224). The server prepared the schema on the CLINICAL runtime pool, and preparation spans every schema — the DDL of all five migration sets under db.migrate = "apply", all five _sqlx_migrations bookkeeping tables under "verify". A role that is a member of ferroehr_ehr and nothing else — what the book recommends and the chart configures — holds neither, so it was refused on the first set and the deployment could not start; ferroehr_app could not run verify either. The new [db] migrate_url (or migrate_url_file, and database.migrateExistingSecret on the chart) names the credential that prepares the schema, used for that one boot step on a connection that is opened and closed again — no pool is held on it and no request is served through it. Unset, it falls back to [db] url, so a single-credential deployment is unchanged. The pseudonymisation boot gate deliberately stays on the RUNTIME pool: it measures what the serving credential can reach, and asking the migration credential — which holds every domain by design — would stop measuring anything. A credential that cannot read a set is now refused with the schema and the role it authenticates as, and the remedy, instead of a bare 42501 about a table no operator has heard of. ferroehr db verify splits the same way. Chart values contract grows three optional keys, so the chart goes to 8.1.0.

  • The template upload dialog opens empty (#3200). It cleared the source it was holding only when the CDR accepted it, so after an upload that was refused — or one whose click never reached the handler — re-opening the dialog showed the previous attempt’s source with the send button already live, offering to send something the reader had not just chosen. Opening the dialog is now a fresh attempt: the editor starts empty, and the button stays inert until a file has actually been read into it.

  • The migration hook’s pod is no longer selected by the server’s Service and PodDisruptionBudget (#3197). A selector is a subset match, and the hook pod carried the server’s selector pair plus a component label, so for as long as it existed it satisfied both — during exactly the install and upgrade when a rollout or a drain reads them. It exposed no port named http, so the Service had no endpoint to route to it, but the PodDisruptionBudget match was real and the port was a property of that pod rather than a promise about the next one. The hook now carries its own app.kubernetes.io/name, as the viewer and the backup jobs already do. The chart’s selector gate was looking only at Deployments and could not have seen this; it now judges every workload that carries a pod template, and the migration Job renders in a committed value set so the gate has something to judge.

  • A physically deleted party takes its multimedia with it (#3180). physical_delete_party removed the party’s rows without collecting the externalized blob keys they referenced, so a blob held only by that party stayed in the object store forever — a storage leak, and on a party also a deletion that did not delete, since a blob is content and an operator who ran a physical delete has reason to believe it is gone. It now collects the keys inside the same transaction and hands them to the same garbage collection delete_ehr uses, which scans both pseudonymisation domains, so a blob a clinical record still references survives.

  • The ADL 2 corpus provenance claimed a licence the files do not state (#3150). corpus/archetypes/adl2/PROVENANCE.md said the archetypes were “predominantly CC-BY-SA 3.0 where stated” and pointed at the root CC-BY-SA 3.0 reference copy. Measured over the vendored tree: of 652 archetypes exactly one states a licence, and it states CC-BY 4.0. The other 651 carry an openEHR Foundation copyright and no grant, and upstream carries no LICENSE file at the pinned commit either. The record now says that, the subtree is declared LicenseRef-openEHR-unstated in REUSE.toml with its own text in LICENSES/, and the licensing chapter carries the same statement. The vendor script that generates the provenance file was repeating the claim and is corrected too, so re-running it cannot restore the old wording. The openehr-adl regression library from the same upstream measured differently and is declared separately: 111 of 302 state a licence, predominantly CC-BY-SA 3.0, with CC-BY 4.0 and CC-BY 3.0 present and CC-BY-SA 4.0 absent — which the previous glob had asserted over it. Whether material whose terms are unstated may stay committed is a decision, tracked separately.

  • The storage-integrity sweep and its repair cover both pseudonymisation domains (#3178). Moving the parties into their own schema left the sweep reading only the clinical one, so damage to the copy no read-path check recomputes went undetected on the domain that holds the identifying data. A sweep with no scope now walks both in one pass, and every mismatch, every streamed line and every rebuild record names the domain it came from, so a report cannot describe half the store while looking like it described all of it. The repair reaches the damaged version through that same domain.

  • The compliance control matrix says which jurisdiction each legal source belongs to, and its links work (#3177). The NEN 7510 entry pointed at a page that returns 404, and NEN 7513 pointed at the publisher’s homepage rather than the standard. Both are corrected, NEN 7512 and the EDPB pseudonymisation guidelines are added, and every URL in the registry was checked by hand on the day it landed, which is the only way it can be checked: the site’s link gate runs offline on purpose, so an external URL that rots is invisible to it. The table now carries an “Applies to” column. EU and INT apply to every deployment; a two-letter code is national law and applies to a deployment in that country. openEHR is not a Dutch standard, and a page listing Dutch law beside EU law without saying which is which reads as though every deployment answers to both.

  • A FLAT or STRUCTURED read no longer drops the content of a template-named event (#3142). An operational template can fix the name of a node the simplified formats collapse away, such as the single event of a HISTORY. The constrained name then survives only inside the leaf’s aqlPath, and a composition written through the FLAT surface came back short: the reporter’s body weight and comment were present in the canonical JSON and absent from application/openehr.wt.flat+json, with no error to say so. Two halves are fixed. Building a composition named that collapsed node after its archetype node id rather than the name its own template path constrains, so the node did not satisfy the path it had just been created for, and every datum beneath it became unaddressable. Reading now tolerates it either way: a node whose runtime name differs from the template’s is still reached by archetype node id, exactly as the archetype-conformance validator already did. That applies only where the name identifies nothing the node id does not. A template that distinguishes same-node-id siblings by name keeps matching strictly, so one sibling can never claim another’s content. Compositions already stored with the wrong name read correctly without being rewritten.

4.1.1 - 2026-09-05

Added

  • A whole large repository can now be checked for storage damage in one pass (#3122). POST /admin/integrity/verify computed its whole report before answering, so it had to finish inside the server’s 30-second request timeout and a large enough repository lost the report to a 408. Sending Accept: application/x-ndjson streams the same sweep instead: one JSON object per line, a mismatch as each is found, a progress tick per page read, and a closing summary with the counts. Nothing bounds that response, and the stream carries every mismatch rather than the first thousand. The status code is committed before the work is, so a sweep that fails part-way ends with an error line instead of a summary — read to the end to tell the two apart. A request that does not name the media type, */* included, keeps the aggregated document exactly as before.

Changed

  • A refused operational template lists every violation, not the first one (#3129). Repairing a template meant one upload per defect: a reporter with a 1.7 MB template uploaded it to learn use was empty, fixed it, uploaded again to learn misuse was too, and a code list with seventeen duplicated codes named one of them at a time. validationErrors now carries an entry per violation, the way a refused composition already did. A violation that makes the tree below it meaningless, a constrained type the reference model does not have or an attribute its parent does not declare, says that its subtree went unchecked, so a short list is never mistaken for a clean one; past 200 violations the response says it stopped counting.

4.1.0 - 2026-09-05

Added

  • A terminology server that answers with the wrong FHIR resource is refused by name (#3099). The R4B Parameters resource has no mandatory member, so any JSON object decoded as an empty view: a $validate-code answered with a ValueSet or an OperationOutcome read as “no result” instead of the wrong answer it was. The decoders now check the resourceType first and report unexpected FHIR resource: expected Parameters, got OperationOutcome, which terminology.external.fail_on_error then decides on like any other upstream fault.

  • A refused OPT or canonical-XML document says where the defect is (#3067). A mandatory element or attribute that is absent used to be reported as xml parse error: missing element id, with no position and a token that matches hundreds of unrelated attributes. The refusal now names the element that should hold it with its xsi:type, its line and column, its path from the root with sibling ordinals, and the class attribute the child realises: element <default_value xsi:type="DV_IDENTIFIER"> at line 4270, column 13 (/template/definition[1]/…/children[1]/default_value[1]) is missing mandatory child <id> (DV_IDENTIFIER.id). A value a class constructor rejects, and an empty 1..* container, carry the same location. The wording no longer calls a cardinality refusal a parse error.

  • tenancy.header is refused at boot when authentication is enabled (#3093). The header override lets a request name its own tenant and wins over the JWT claim, so on a deployment with real users any authenticated caller could read and write any tenant. The server now refuses to start on that pair, naming the key and the remedy, unless the new tenancy.insecure_header_override = true accepts it explicitly for a development deployment. With authentication off, or tenancy off, nothing changes.

Changed

  • The storage-parity sweep reads a page of versions per round trip instead of one, and can be scoped (#3110). POST /admin/integrity/verify compared each stored version’s node rows with its materialized body using two sequential statements per version, so a store of 60 000 versions cost about 120 000 round trips and the request died at the 30 s timeout with the report lost. Content is now read 32 versions at a time, two statements per chunk: the same store costs under 4 000 round trips. Two optional query parameters narrow what a sweep covers, ehr_id and committed_since, so verifying one record or everything committed since an incident no longer means reading the whole repository. The report, its defect vocabulary and its status codes are unchanged.

  • A tenant key that names no registered tenant is refused instead of running unscoped (#3111). It used to fall through to the reserved default tenant, on the argument that a cross-tenant access should look like an empty result set rather than a 403 confirming another tenant exists. That argument holds only while the default tenant is empty, and the default tenant owns every row written while tenancy was off: on a deployment that enabled tenancy after going live, a misspelled claim read and wrote the entire pre-tenancy store. Such a request now gets a 403. Set FERROEHR__TENANCY__UNKNOWN_TENANT=default_tenant to restore the old behaviour where the existence-hiding is worth more. A request carrying no tenant key at all is unchanged. With tenancy enabled, boot now counts the default tenant’s stored versions and warns when it holds any, because both paths lead there.

  • A composition commit checks out one pooled connection instead of three (#3097). The EHR-writability gate, the persistent-duplicate gate and the commit each took their own connection from the pool. With multi-tenancy on every checkout costs a set_config round trip that stamps the tenant GUC, so the acquire count was the cost: measured over 100 commits, the path went from 9.3 to 6.3 connection checkouts each. Behaviour is unchanged — the gates run in the same order, answer the same statuses, and the commit is still one folded statement (or one transaction when attestations or the outbox are in play).

  • A large composition no longer holds a request worker through its whole validation (#3097). The RM, terminology and archetype-conformance passes are CPU work that ran inline on the async worker, costing about 260 ns per JSON node: the largest form the CKM publishes takes 2.5 ms, and at the default 16 MiB body limit a single commit could hold one worker for tens of milliseconds while every other request on it waited. Above 400 nodes the passes now tell the runtime to relieve the worker first. Validation itself is unchanged: the same passes run in the same order and refuse the same content.

  • openehr-query, openehr-adl and openehr-its move to the Business Source License 1.1 (owner decision 2026-09-04). The three hand-written engines (the AQL parser, the ADL 2 engine, and the canonical codecs, REST contract and Simplified Formats) now carry the same licence as the application, each with its own LICENSE naming the crate as the Licensed Work; openehr-its declares BUSL-1.1 AND Apache-2.0 and keeps the Apache-2.0 text for the openEHR-derived codecs, contract and ITS-JSON schema it embeds. The five generated model crates (openehr-base, openehr-rm, openehr-am, openehr-lang, openehr-term) stay Apache-2.0. Published versions keep the licence they were published with: the three engines are Apache-2.0 up to 0.0.59 and BUSL-1.1 from 0.0.60.

  • server.swagger_ui is an access level, and the Swagger UI needs a credential by default (#3094). The key used to be a boolean that mounted the UI and the OpenAPI documents outside the authentication layer, on by default, so an unauthenticated client received the complete operation surface, admin and message groups included. It now takes the management vocabulary off, admin_only, private (any authenticated principal; the new default) or public, and the non-public levels sit behind the same access guard the management endpoints use: unauthenticated is 401 with the WWW-Authenticate challenge, so a browser prompts for the Basic credential. An existing swagger_ui = true or false is a boot error naming the new spellings; the hosted sandbox keeps the UI public explicitly. The Helm chart default follows (config.server.swagger_ui: private).

Fixed

  • Signing in again no longer lands back on the login card (#3066). After the viewer signed a user out, an answer to a request issued under the old session could still arrive while they were signing back in. The transport reads every such answer, and the shell treated one as proof the session was over, so a freshly signed-in user was sent straight back to the login screen and only a full page reload got them past it. The shell now re-reads the session instead of trusting the detection, because a detection names no session, so a late answer to a dead one is ignored and a real expiry still signs out. The visible behaviour on a genuine expiry is unchanged.

  • A large operational template uploads in milliseconds again (#3114). The XML reader resolved every closing element’s line and column by scanning the document from byte zero, which made a parse quadratic in document size: the International Patient Summary template, 1.9 MB, took 10.08 s, and the largest form the CKM publishes was on track for about 59 s. A template upload holds that CPU on the request, and large uploads were answering 408 Request Timeout. The position and the element path are now resolved when a refusal is raised rather than on every close, so a refused document names exactly what it named before. The same two templates parse in 10.2 ms and 25.1 ms.

  • An explicit fetch or AQL LIMIT can no longer ask for a page larger than query.max_result_rows (#3092). The ceiling used to apply only to a query that carried neither bound, so fetch=10000000 materialised the whole matched set into one RESULT_SET. A page above the ceiling is now refused with 400 naming the requested size and the ceiling; the effective page (the smaller of LIMIT and fetch) is what is checked, and a page at or below the ceiling is served as written. The page is refused rather than shortened because a client paging with its own fetch as the stride would silently skip rows.

  • An ADL 2 upload can no longer take the server down through unbounded recursion (#3062). The ADL engine now carries one nesting bound of 512 levels: the cADL, ODIN and rule-expression readers refuse an artefact nested past it with a 400 that names the bound (the SUNK syntax bucket), the flattener refuses a specialisation lineage longer than the bound or a flat form that composes deeper than it, and the operational-template transform refuses fillers that reference each other in a cycle (naming the cycle) or inline past the bound, as a 422. Every remaining engine call, including compiling a stored template to its operational form and loading the stored repository, runs on the dedicated engine thread. Published archetypes and templates stay far below the bound.

4.0.18 - 2026-09-03

Changed

  • The licence of the project’s own code is the Business Source License 1.1 (LICENSE, SPDX BUSL-1.1; owner decision 2026-09-03). All non-production use is free, and production use is free for Non-Commercial Purposes only (personal use, academic or scientific research, teaching, and non-profit or public bodies outside the course of a business); any other production use, including the delivery of health care or another service for payment, and in every case offering FerroEHR or a derived work to third parties as a hosted, managed or embedded service or distributing it for a fee, needs a commercial licence from the Licensor. Each version becomes Apache License 2.0 four years after its publication. The copyright holder named everywhere is Ruben Talstra. Every SPDX header, manifest, badge, OCI image label, the Helm chart’s Artifact Hub licence, the citation metadata and every page that named MIT follow, and a CI guard now fails on a stale MIT claim in a first-party file. Releases v3.0.0 through v4.0.17 stay under the MIT terms they were published with. The eight openehr-* spec crates on crates.io are NOT under the BUSL: they are Apache-2.0 (openehr-term also CC-BY-SA-3.0 for the terminology XML it carries) from 0.0.58, the licence of the openEHR inputs they are generated from, so any Rust project can consume them; 0.0.56 and earlier stay MIT AND Apache-2.0 as published.
  • Contributions now carry an inbound relicensing grant (CONTRIBUTING.md § Licensing of contributions): contributors keep their copyright and license their work under the project licence, and additionally grant the Licensor the right to sublicense and relicense it, so the work stays one work under one licensor.
  • Every migration file carries a licence header, which changes its recorded checksum: a database created by 4.0.17 or earlier refuses to start this version with migration 1 was previously applied but has been modified. The schema is greenfield (see the operations chapter, “Upgrades”), so the remedy is to recreate the database, or drop the ehr, ext, audit and cold schemas, and reload; the hosted sandbox was reset the same way.

4.0.17 - 2026-09-02

Added

  • server.base_path (FERROEHR__SERVER__BASE_PATH) is a supported, validated knob: a deployment behind a path-prefixed reverse proxy can shorten the REST base path to as little as /ferroehr/v1 instead of stacking the proxy prefix on top of the full default. The API nest, Location headers, the System Options manifest, the served OpenAPI documents, the Swagger UI, the status document and SMART discovery all follow the configured value; the health family stays at the process root. The default is unchanged.
  • The FerroEHR Viewer gained the mirror key cdr.base_path (FERROEHR_VIEWER__CDR__BASE_PATH, default /ferroehr/rest/openehr/v1), so a viewer can drive a CDR that shortened its base path.

Changed

  • server.base_path is now checked at boot, and a value that breaks a rule stops the server with an error naming the key and every rule it broke. The first segment must be ferroehr, the last must be the openEHR API version segment v1, and the value must carry no trailing slash, no empty segment, and only unreserved URL characters. Values that previously booted and served a broken surface are now refused: a trailing slash, a missing /ferroehr segment, a path not ending in v1. The default value is unaffected.
  • ferroehr healthcheck no longer hard-codes the default REST root: with no --url/FERROEHR_HEALTHCHECK_URL it loads the same configuration as the server and probes http://127.0.0.1:<server.bind port><REST root>/status, so the container health check keeps working when server.base_path is shortened. The URL it probes under the defaults is unchanged.
  • The hosted sandbox (sandbox.ferroehr.eu) now holds a complete demo dataset instead of four templates in three EHRs. Each nightly reset loads 16 ADL 1.4 operational templates from the openEHR CKM pack, 235 ADL 2 archetypes and 5 ADL 2 templates (two of them source templates the server flattens against that archetype library), 8 EHRs carrying 182 compositions with mixed records, 6 compositions with more than one version and 2 EHRs with a second EHR_STATUS version (one not queryable, one not modifiable), 11 demographic parties, a FOLDER directory per EHR referencing real compositions, and 5 stored AQL queries under eu.ferroehr.sandbox that all return rows. The seeder (scripts/sandbox/reseed.sh) still drives only the public REST API and is now a walker over scripts/sandbox/seed/manifest.json, so adding demo content is editing data. The ADL 2 corpus is uploaded parents before children, and the split it pins (233 stored, 89 refused by AOM2 validation) is the one the corpus gate in openehr-adl adjudicates file by file. A run takes 39–45 s against a local stack and leaves a 24 MB database; a repeat run detects the marker EHR it wrote and exits without changing anything.

Fixed

  • An ADL 2 upload can no longer take the server down. The AOM engine’s validation now runs on a dedicated thread with a stack sized for its recursive walks over the stored repository, and every runtime thread starts with a larger stack; previously a deep enough artefact overflowed a worker’s 2 MiB stack and the process aborted with every in-flight request.

  • ADL 2 validation no longer reports VCATU (duplicate sibling attribute) for a specialised archetype whose root-level differential paths end in the same attribute name (/items beside /items[id9]/items): the two address different nodes of the flat parent. Twenty-nine archetypes of the vendored 2013 CKM export were refused on that false positive.

  • An ADL 2 upload whose specialise clause names a parent that is not in the repository is refused with 422 carrying VASID and the missing parent’s id, on both validation paths. Previously the parent-conformance checks were skipped and the archetype was stored as valid, so a child uploaded before its parent was never checked against it.

  • An AQL SELECT that mixes an aggregate function with a non-aggregated column (SELECT e/ehr_id/value, COUNT(c/uid/value) …) is refused with a typed 400 naming the rule. AQL 1.1 defines the aggregate functions over the selected rows and no grouping construct, so the shape has no defined result; previously the ungrouped SQL reached the database and its error surfaced as a 500.

  • The FerroEHR Viewer no longer leaves an interactive shell over a session that has ended. When a session expires or is revoked, the viewer moves the whole UI to the signed-out state on its own — no manual refresh — and lands on the sign-in screen with “Your session ended. Sign in again to continue.” plus the screen the user was on, so signing back in returns them to it. Two things drive it: any request the CDR or the viewer refuses with a 401/no-session answer signs the UI out immediately (this covers a revocation the browser could not have predicted), and the browser also re-checks the session on the session’s own deadline and on the connection indicator’s poll, so an expiry nobody clicks through is caught as well. The indicator itself now reads “Session ended” rather than blaming the CDR for a refusal that is not its. The OIDC sign-in path carries the destination too: /auth/oidc/login accepts a next parameter and the callback lands on it, so an OIDC user is returned to the screen they were on instead of the dashboard. Only a same-origin relative path is honoured — an absolute URL or a protocol-relative value falls back to the dashboard, so the login route cannot be used as an open redirect.

  • The hosted-sandbox CDR no longer crashloops when the box’s .env carries a FERROEHR_-prefixed Compose image variable the server does not recognise. The CDR service in deploy/hosted/docker-compose.yml stops loading the whole .env into the container and now receives only DATABASE_URL (interpolated, and a boot error if unset) plus its config-file path. Compose image variables stay substitution-only and never reach the server’s strict configuration check, so a renamed or stale image variable in the operator’s .env cannot stop the server from starting.

4.0.16 - 2026-09-02

Changed

  • The canonical-XML reader and the TDD/OPT XML readers refuse a non-UTF-8 document at the first read (quick-xml 0.42 validates encoding while constructing events) instead of tolerating invalid bytes into text values. Well-formed UTF-8 documents — the entire wire corpus — are unaffected; the full serialization fidelity battery passes unchanged.
  • The generic noun “console” is retired from everything a human reads: the Helm chart’s viewer.* value descriptions (which Artifact Hub publishes), the viewer’s own on-screen copy, the book, and the doc comments all say “viewer” now. The hosted sandbox’s Compose service is renamed console to viewer and its container ferroehr-sandbox-console to ferroehr-sandbox-viewer, which the box picks up at the next release deploy. The Helm chart is version 7.0.2. The name of the entry the viewer keeps inside its encrypted session cookie changes too, so open sessions need one fresh sign-in after the upgrade.
  • The hosted sandbox (sandbox.ferroehr.eu) runs on a Hetzner CX33 (4 shared vCPU, 8 GB RAM, 80 GB NVMe), resized in place from the CPX22. The hosted compose memory limits that ship inside the server image move with it: the CDR container to 4096m and the FerroEHR Viewer container to 1536m, so the services actually use the larger box.
  • The sandbox database moved from Neon to a second, dedicated CX33 running the published ferroehr-postgres image (PostgreSQL 18), reachable only over a Hetzner private network — no public database port exists. The committed posture gains deploy/hosted/cloud-init-postgres.yaml; the nightly reseed’s schema wipe now runs on the app box through the deploy script’s new wipe verb instead of a psql from the CI runner, so CI no longer holds any database credential (the SANDBOX_DATABASE_URL secret is retired).
  • The admin console is now FerroEHR Viewer, and every name it carries changes with it. Deployments must switch their references: the OCI image is ghcr.io/rubentalstra/ferroehr-viewer (was ferroehr-admin-ui), the Compose service and profile are both viewer (docker compose --profile viewer up, was --profile admin-ui), the Compose image and port overrides are FERROEHR_VIEWER_IMAGE and FERROEHR_VIEWER_PORT, the Helm values key is viewer (was adminUi) and its OIDC client secret mounts at /etc/ferroehr-viewer-secrets, the config file the viewer looks for is ./ferroehr-viewer.toml then /etc/ferroehr/viewer.toml (was ferroehr-admin-ui.toml / /etc/ferroehr/admin-ui.toml), the binary and crate are ferroehr-viewer, and the session cookie is ferroehr_viewer_session, so open sessions need one fresh sign-in. Documentation moved from /docs/*/admin-ui/ to /docs/*/viewer/, with redirects from the old pages. This rename is the chart’s 7.0.0 major bump.
  • The viewer’s own configuration environment grammar is renamed with it: every FERROEHR_ADMIN__<SECTION>__<KEY> variable becomes FERROEHR_VIEWER__<SECTION>__<KEY> (so FERROEHR_ADMIN__CDR__BASE_URL is now FERROEHR_VIEWER__CDR__BASE_URL), and the config-file pointer FERROEHR_ADMIN_CONFIG becomes FERROEHR_VIEWER_CONFIG. The chart’s default viewer.existingSecretKey moves from FERROEHR_ADMIN__AUTH__OIDC__CLIENT_SECRET to FERROEHR_VIEWER__AUTH__OIDC__CLIENT_SECRET, so an existing Secret needs its key renamed or viewer.existingSecretKey set to the old spelling. A variable left under the old prefix is not read and not reported: the viewer’s strict deserialization only sees what arrives under its own prefix.
  • The Helm chart renames the viewer’s Kubernetes objects from <release>-admin-ui to <release>-viewer and its app.kubernetes.io/name label from <name>-admin-ui to <name>-viewer. A Deployment’s selector is immutable, so a release that already runs the viewer with adminUi.enabled: true must have that Deployment, Service, ServiceAccount and NetworkPolicy deleted before the upgrade, or be reinstalled. The CDR’s own objects are untouched.

Security

  • The ferroehr and ferroehr-viewer images build from the rebuilt gcr.io/distroless/cc-debian13:nonroot base carrying openssl 3.5.7-1~deb13u2, which fixes CVE-2026-14456 (QUIC denial of service) and nine sibling advisories in libssl3t64. The library was never linked by the shipped binaries (the TLS stack is rustls/aws-lc throughout); the bounded scanner exception and its VEX statement are removed now that the fixed base exists.

Fixed

  • Composition validation now enforces a template’s decimal-precision constraint on DV_QUANTITY and DV_PROPORTION leaves: an instance whose optional precision attribute falls outside the constrained interval is refused (422) instead of silently accepted. Temporal-range constraints were already enforced; the module documentation claiming otherwise was stale and now states the real contract.
  • The FerroEHR Viewer’s session cookie sets Secure by default (session.cookie_secure now defaults to true, fail closed): a TLS-fronted deployment needs nothing, and plain-HTTP contexts (the compose quickstart on localhost, local development, the e2e harness) opt out explicitly. If you serve the viewer over plain HTTP without one of those postures, set FERROEHR_VIEWER__SESSION__COOKIE_SECURE=false or logins will not stick.
  • TDD import no longer discards instance data a TDD spells out on wrappers the WebTemplate compacted: HISTORY.origin, an event’s time and name, and the other LOCATABLE metadata (uid, links, feeder_audit) now land on the re-materialised nodes instead of RM-mandatory defaults, each value parsed as its model-declared type. Spelled-out wrapper data that cannot legally sit on the corresponding node is refused with a named error rather than silently dropped.
  • The platform validity checker (definitions_valid) now checks archetype identifiers, not just template identifiers, per the SM clause it realizes: every archetype_details declaration in the checked content contributes its archetype id and template id, archetype ids resolve through a declared template’s inlined nodes or against the stored ADL 1.4 and ADL 2 repositories, and an unknown identifier at any depth now answers false.
  • The ETag served with a query RESULT_SET now derives from SHA-256, a pinned published algorithm, instead of the standard library’s default hasher, whose algorithm may change between Rust releases. Query ETags are now stable across server builds; every served query ETag changes value once at this upgrade.

4.0.15 - 2026-09-01

This release carries everything prepared for 4.0.14, which was tagged but never published: its release pipeline failed before any artifact was built, and the v4.0.14 tag is immutable, so the content ships here under the next patch number. No 4.0.14 artifacts (release, images, chart, crates) exist.

Added

  • The admin console’s template Example tab lets you pick how the example is generated. Beside the format selector, both template families (ADL 1.4 and ADL 2) now offer a detail level (Required, Medium, Complete) and the form the example is shaped for (Input as submitted, Output as retrieved). The pane opens on Required/Input, as before, and changing either control asks the CDR for a fresh example.

Security

  • The ferroehr-postgres image builds from the respun upstream postgres:18.6 base (the tag was re-pointed upstream with rebuilt base packages and bundled binaries); the digest pin follows it everywhere it appears. The rebuilt candidate scans clean at the published-image gate.

Changed

  • The hosted sandbox (sandbox.ferroehr.eu) moved off Vercel onto a dedicated Hetzner server (deploy/hosted/ is the whole committed posture: cloud-init, compose, Caddy TLS, a key-restricted deploy script). No more scale-to-zero cold starts or free-compute slowdowns; the same URL, credentials, nightly reset and released images. The database is a standalone Neon project on its direct endpoint, no longer a platform integration.
  • The hosted sandbox (sandbox.ferroehr.eu) now serves the Admin API: the nightly-reset demo is a development/testing deployment in the released spec’s own terms, and the conformance statement’s AdminApi claims become exercisable against it. The public demo credential reaches the whole admin group, bulk delete included — it wipes what the nightly reset wipes anyway.
  • The repository’s commit history on main now starts at a single labelled import commit holding the tree FerroEHR was forked from, followed only by this project’s own commits. The 4864 inherited upstream commits and their authors are no longer part of main. Nothing published changed: all release tags still point at the commits they were cut from, release assets, Sigstore bundles, image attestations and archived deposits verify exactly as before, and the pre-rewrite lineage stays reachable in this repository through those tags. Existing clones must be re-cloned rather than pulled, and comparisons spanning the rewrite (for example git describe against main, or a tag range that crosses it) no longer share ancestry.
  • The admin console’s Template Manager uploads both template families the same way: one button in the page header (Upload OPT / Upload ADL2) opens a dialog offering a file picker and a paste area for either family, so ADL 1.4 gains the paste path and ADL 2 no longer carries a permanent card above its listing. A refusal keeps the dialog open with the server’s diagnostic beside the source, and a successful upload closes it and refreshes the list.

Fixed

  • The admin console’s two template Example panes report a failed example read the same way: an inline error in the pane, in both template families. The ADL 1.4 pane used to render the whole-screen error with a back link even though the template itself had loaded.
  • The hosted sandbox serves the console’s content-hashed script and wasm bundle with immutable caching, so a repeat visit stops re-downloading it.
  • A template example request with a present but empty detail_level or type query value is refused with 400. The declared defaults apply only to an absent parameter; an empty or whitespace-padded value is outside the closed enums and was silently read as the default.
  • The admin console’s System screen reports SMART correctly. It probed the discovery document at the origin instead of under the platform base path the CDR serves it from, and rendered whatever came back as a raw status echo (CDR answered 302: HTTP 302 on the hosted sandbox, where the origin path belongs to the console itself). It now probes /ferroehr/rest/.well-known/smart-configuration, states plainly that SMART is not enabled when the CDR serves no document, and gives an unexpected answer actionable copy that never claims SMART is disabled when the probe could not tell.

4.0.13 - 2026-08-30

Fixed

  • AQL LIKE honors the escaped single-character wildcard: \? in a pattern now matches a literal ? (QUERY master03 §Operators/LIKE), symmetrically with the already-correct \*. The string reader was consuming \? as a string escape before the pattern layer could see it, so the escaped form still matched as the wildcard.
  • Console sessions survive multi-instance deployments. The session moved from an in-process store to a sealed cookie (AES-256-GCM): any replica holding the configured session.secret (base64, at least 64 bytes; new, with session.secret_file) can serve any signed-in visitor — on serverless platforms the in-process store made every request that landed on another instance fail as 500 and the console report the CDR as down. With no secret configured the console behaves as before (one replica, ephemeral key) and says so at startup. The cookie is encrypted and authenticated, HttpOnly, SameSite=Lax, Secure per configuration; idle expiry is unchanged and rides inside the sealed payload.

4.0.12 - 2026-08-30

Added

  • The admin console’s sign-in card takes a deployment-configured notice and links: login.notice (informational text, line breaks preserved — a demo states its public credentials and usage expectations there) and login.links (label + href pairs, e.g. an API reference). Both default empty and render nothing when unset.

Changed

  • The hosted sandbox (sandbox.ferroehr.eu) leads with the admin console: the deployment now runs both published container images, the console is the landing surface (a visitor gets its sign-in screen, demo credentials ferroehr / ferroehr), and the CDR keeps its path families — /ferroehr/* (including the Swagger UI, which stays the API reference), /health*, and /management*. The console drives the sandbox CDR over the public REST base as the same demo user; the CDR’s admin and management surfaces stay off.

4.0.11 - 2026-08-29

Added

  • Validation enforces the one testable rule of DV_TEXT.formatting: a value formatted "plain_no_newlines" is refused when the text contains a newline (RM: “newlines are not allowed”). Every other formatting value stays unvalidated — markdown, "plain", the deprecated CSS form, and unknown names all remain legal, with or without newlines.

Fixed

  • The is_modifiable content-write gate is evaluated inside the commit transaction, under a row lock, instead of a separate read before it — a deactivation committed concurrently with a content write is now always observed (one or the other strictly wins; previously the content commit could land on a just-deactivated EHR). The gate’s semantics are unchanged: the atomic change set is judged as a whole — a deactivating set may carry its final content, a reactivating set enables its own content, and EHR_STATUS itself is always writable.
  • AQL date/time comparisons work on reduced-precision values. openEHR admits partial date/times (2019, 1985-06), but one stored anywhere in the scanned data made every temporal comparison on that path fail as a 400 (“your request is malformed” — for a data property). Both sides of a temporal comparison now floor a partial value to the first instant it contains (first month/day, zero time), the same completion the promoted timestamp column uses, so stored partials compare, order, and match NOW() instead of erroring. A comparison value that is not an ISO 8601 date/time at all is still the caller’s 400, now named at plan time; non-ISO forms PostgreSQL happened to parse (e.g. June 15, 2020) are refused — AQL temporal literals are ISO 8601.
  • AQL predicates and projections now see every element of a list-valued attribute (links, context/participations, composer identifiers, mappings, …). Previously the extraction took only the first match, so WHERE c/links/target/value = ... missed a match on the second link and SELECT over such a path served one value where the data held several — silently. Comparisons, LIKE, MATCHES and EXISTS are any-match; projection returns every match as one JSON array cell (null when nothing matches).
  • An AQL node predicate written on the ROOT of an identified path (c[openEHR-EHR-COMPOSITION.report.v1]/name/value) now constrains the query — it was silently discarded, so the path answered as if the predicate were not written. A predicate on a non-structure path step (links[meaning/value = ...]) lowers as a SQL/JSON path filter selecting the matching elements, and a root predicate on a whole-object projection serves the object where the source matches and a null cell where it does not.
  • Reads serve the exact canonical bytes the commit accepted: _type first and every field at its spec-declared position — including the server-stamped uid, which now lands at its place in the document instead of the end. Previously the stored body passed through a jsonb column and PostgreSQL’s own key ordering leaked onto the wire (reported by the community). Dump archives carry the same faithful bytes, so an export/load round-trip is byte-equal again.

4.0.10 - 2026-08-29

Added

  • The metrics surface gains process_resident_memory_bytes (the standard Prometheus name), sampled from procfs every five seconds beside the pool and runtime gauges — so “is the server’s memory growing?” is a dashboard read instead of a shell into the host. Absent on platforms without procfs.

Changed

  • AQL FOLDER f1 CONTAINS FOLDER f2 now also matches the folders of a VERSIONED_FOLDER that a folder’s items reference — the same reference-resolution ground as FOLDER CONTAINS COMPOSITION, one reference hop (chain CONTAINS to follow further hops). Strict sub-folder matching is unchanged.
  • The Helm chart’s committed appVersion now always equals the product version, bumped in every release PR beside the compose image tags (chart 6.0.27). Published charts are unchanged — the release pipeline keeps injecting the released version at package time — but a from-tree helm install now defaults to the current version instead of a release that could lag by one.
  • The repository’s default branch is now main (renamed from develop; full history preserved, old GitHub URLs redirect). The moving development image tag follows the branch: pull ghcr.io/rubentalstra/ferroehr:main (and the admin-ui/postgres siblings) instead of :develop, which is frozen tags are deleted. :latest keeps its meaning — the newest stable release — and release tags are unchanged.

Fixed

  • The v4.0.9 Helm chart publishes as 6.0.26 (appVersion 4.0.9). The release’s own chart leg failed on a CI guard defect before publishing, so chart 6.0.25 never existed on the registry; 6.0.26 is the same chart with the committed appVersion default refreshed to the released 4.0.9.

4.0.9 - 2026-08-28

Fixed

  • FOLDER f CONTAINS COMPOSITION c now resolves the folder’s items references (transitively over its sub-folders) instead of returning every composition in the EHR — the containment edge was silently dropped for any versioned object under a non-EHR parent, and the same defect made NOT CONTAINS vacuously false under a folder and let ORDER BY change the row count of EHR e CONTAINS FOLDER f. FOLDER CONTAINS FOLDER is now a strict sub-folder match (no self-pairs), and a CONTAINS pair the RM defines no containment relationship for (such as COMPOSITION CONTAINS COMPOSITION) is a typed refusal instead of a cartesian product (#2880).

  • The quickstart no longer publishes PostgreSQL to a host port, so a natively installed PostgreSQL (the classic Windows collision: the installer’s auto-started postgresql-x64-* service holding 5432) can no longer fail docker compose up. Host access for psql/GUI clients is the new docker-compose.db-publish.yml overlay, attached to releases beside the base file; in-stack access needs no port (docker compose exec ferroehr-postgres psql) (#2879).

  • The admin console stops re-downloading its whole WebAssembly bundle on every page load. Its /pkg/ filenames now carry a content hash and are served public, max-age=31536000, immutable, so a browser reuses them until the console is rebuilt — measured on the shipped image, a second page load transfers 0 bytes of /pkg instead of 7.6 MB. Documents keep no-store unchanged: they carry patient data and a per-request CSP nonce (#2875).

  • A bare docker compose up in a repository checkout now runs the same published-images quickstart as the downloaded standalone file. The development overlay no longer carries Compose’s auto-merged override name (it is docker-compose.dev.yml, applied only by an explicit -f), so a checkout can no longer fail on unpublished :local tags or missing build contexts (#2868).

  • A port conflict never requires editing the compose file: every published port was already a variable (FERROEHR_PORT, FERROEHR_DB_PORT, FERROEHR_ADMIN_UI_PORT, FERROEHR_S3_PORT), and the quickstart and the book now say so at the point of failure (#2869).

Changed

  • Console builds no longer touch Cargo.lock. cargo leptos resolves the workspace through an unlocked cargo metadata call of its own, which had silently re-resolved a transitive dependency during a build; every invocation now runs through a wrapper that refuses a stale lockfile up front and passes --locked to both compile legs, matching every other build in the repository (#2877).
  • The full browser e2e battery now runs against the PUBLISHED console image, in a weekly lane that drives the latest release’s three images at that release’s own tag. The battery’s shipped-artifact mode had no caller at all; the release pipeline keeps gating the login journey through the image it pushes (#2876).
  • The openehr-* spec crates step to 0.0.43. The change is internal structure only: over-complex functions in the ADL engine, the Simplified Formats web-template builder and the ISO 8601 duration parser were split into named, documented helpers. The wire behaviour, the generated output and every gate are unchanged, and nothing an API consumer calls moved.

4.0.8 - 2026-08-27

Changed

  • The container images’ distroless base (gcr.io/distroless/cc-debian13:nonroot) moves to the current upstream digest, picking up Debian package updates for the runtime layer of the server and admin-console images.

4.0.7 - 2026-08-27

Added

  • The release pipeline publishes the openehr-* spec crates itself, as an approval-gated leg: the run pauses at the crates-io environment for an explicit human approval, then publishes any unpublished crate versions in dependency order with registry read-back (a release with no crate changes passes the leg as a no-op). A cut can no longer forget the crates; the dispatch lane remains as the dry-run/recovery path.

Fixed

  • The chart packaging lane installs the pinned helm-docs, so the chart README drift check runs where the chart is actually packaged. The v4.0.6 chart publish initially failed on exactly this and was recovered by dispatch; from this release the leg carries the tool.

4.0.6 - 2026-08-27

The first stable cut of the 4.0.6 line. The bulk of the line’s changes landed in the two release candidates below: the CI/release-pipeline redesign (4.0.6-rc2) and the conformance-instrument split to the standalone Veredictum project (4.0.6-rc3).

Changed

  • The openehr-* spec crates step to 0.0.42. The change is internal structure only: long functions across the spec crates, the code generator and the test harnesses were split into named, documented helpers, with the wire behaviour, the generated output and every gate unchanged. Nothing an API consumer calls moved.

4.0.6-rc3 - 2026-08-26

Changed

  • The conformance instrument is no longer built from this repository. It is Veredictum, an independent Apache-2.0 project with its own release line, and the pipeline consumes it at a version pinned in one place (scripts/lib/veredictum.sh): the runner comes from crates.io (cargo install veredictum --version <pin> --locked) and the catalogue with its vendored specification oracle comes from a cached checkout of the matching tag. scripts/conformance.sh and the render scripts take the same arguments as before and produce the same artifacts, so the committed conformance record and every published chart, badge and document derived from it are unchanged. An instrument built separately from the system it judges can no longer be adjusted to suit that system.
  • The party set — each SUT’s ixit.json and its party statement, plus the static test issuer the SMART conformance posture trusts — moved to docs/conformance/party/<sut>/, beside the record it belongs to. A vendor running the pipeline against their own deployment points CONF_IXIT and CONF_STATEMENT at their own files exactly as before.
  • The vendored clinical-model and fixture corpus this repository’s own test suites read moved to corpus/ at the repository root. It carries what FerroEHR’s gates exercise — the CKM ADL 1.4 and AM XML archetype packs, the CKM operational template library, the fixture bodies and the version-signing test key. The catalogue-side records (the per-file corpus manifest, the generated-set recipe contracts and the synthesized-template generator) stayed with the instrument.

Removed

  • docs/conformance/coverage-report.md and the CI step that regenerated it. The report describes the catalogue’s coverage of the openEHR wire surface, so it is a claim about the instrument and is now published by Veredictum, which is the only repository that can regenerate it.
  • The CNF 2.0 design record left with the instrument; it is ARCHITECTURE.md at the root of the Veredictum repository.

4.0.6-rc2 - 2026-08-26

Added

  • A weekly report of container manifests in the registry that nothing can reach any more, the residue of a publish that pushed an image by digest and then failed before its tags were applied. It files one tracking issue listing the digests and the command that removes each. Deletion stays manual: a miscomputed reachability would destroy published signatures and attestations.

Changed

  • A published Helm chart now takes its appVersion and its artifacthub.io/images tags from the release being cut, injected when the chart is packaged, instead of from values edited into Chart.yaml beforehand. What the published chart says does not change; it can no longer say the previous release. The committed values are the default for a chart packaged between releases and for helm template against a checkout, and they may lag.
  • The conformance instrument (tools/cnf-runner — the CNF 2.0 runner, its machine-readable catalogue, schemas, and party artifacts) is now licensed under Apache-2.0 instead of MIT: attribution travels with every copy and derivative (license + NOTICE retention), and the license carries an explicit patent grant. The rest of the project stays MIT; the vendored test corpora keep their upstream terms unchanged.

Fixed

  • The chart’s generated README.md, the page Artifact Hub renders, was a release behind: it advertised chart 6.0.19 and image tag 4.0.4 while the chart was 6.0.20 with an appVersion of 4.0.5. It is regenerated, and a guard now refuses a README that disagrees with the chart it ships in.
  • Two gh attestation verify examples in Verifying releases still pinned image tag 4.0.4 at the v4.0.5 cut, and that page was then frozen into the 4.0.5 documentation. The docs guard now checks release-asset filenames and the substitute-this-tag note on that page, and it reads the workspace version rather than the chart’s appVersion.
  • Three checks in the documentation-claims guard could never fire: they tested membership of a newline-separated file list against a space-delimited pattern, so the quoted-wire-evidence check and the Rust-version check on the from-source page had never run once.
  • The documentation-claims guard now recognises an image.tag=X.Y.Z pin on a chart page. That spelling is the defect the check was written for, and it was the one spelling the pattern did not match.

4.0.5 - 2026-08-26

Changed

  • Refusing a supplied EHR_STATUS without archetype_details (a 422 the Reference Model requires — EHR_STATUS is always an archetype root) now names the remedy in the error message, and the API book documents both create-EHR branches: the content of the server-minted default, and the RM-completeness a supplied status must have.

Fixed

  • AQL parsing no longer leaks memory. The parser’s internal recursive combinators formed reference cycles that outlived every parse, so each parsed query grew the process permanently. Found by the fuzzing lane, which now runs the AQL target under LeakSanitizer without findings.
  • Printing a parsed AQL query re-escapes its string literals (backslashes, quotes, and the decoded control escapes). The parser decodes escape sequences into the AST, and four printer sites emitted the decoded value verbatim, so a printed query re-decoded them on the next parse and drifted. A stored query now round-trips to the identical AST.
  • Refusing an invalid AQL query with nested [ predicate openers took exponential time (a 222-byte input parsed for seconds to minutes, doubling per nesting level). The node-predicate grammar now parses its shared path prefix once, so refusal time stays near-linear in input length.

Security

  • Container image tags (:latest and the version tags) now move only after the vulnerability scan passes. The images are pushed by digest, scanned, and tagged last, so a failing scan leaves every previously published tag untouched (the v4.0.4 lane moved :latest before its scan verdict).
  • CVE-2026-14456 (libssl in the distroless base image) is adjudicated not affected: no FerroEHR binary links OpenSSL (rustls only). The finding is suppressed by an id-scoped, expiring ignore plus a published OpenVEX statement, pending a rebuilt upstream base.

4.0.4 - 2026-08-26

Fixed

  • An unauthenticated request to a guarded admin-console URL now answers its 302 → /login with an empty body. Previously the redirect carried a fully rendered console document — chrome, screen markup, and the serialized failure diagnostics of every server function the screen runs — visible to any client that does not follow redirects, and each anonymous hit paid a full server render. A pre-render session guard now refuses the request before any rendering happens; the sign-in screen, OIDC handshake, assets, and server-function endpoints are unaffected, and the in-view guard stays as the client-side navigation gate.

Changed

  • Composition updates got faster: the whole write pre-check (ownership, If-Match, lifecycle, is_modifiable, template stability, the cross-version invariants) now rides the write transaction’s own placement statement under the per-object lock, instead of a separate pre-read round trip. The refusal outcomes and their order are unchanged.
  • GET …/composition/{uid} with a JSON Accept now serves the stored canonical body verbatim instead of parsing and re-serializing it. The JSON is semantically identical; insignificant whitespace now follows PostgreSQL’s jsonb rendering (a space after : and ,), so byte-exact comparisons against previous responses will differ while every JSON parser sees the same document. XML, FLAT and STRUCTURED representations, and expand_multimedia=true reads, are parsed exactly as before.
  • The performance page documents the durability floor: why a single durable commit cannot beat the WAL flush, how group commit scales concurrent writers past it, and the synchronous_commit trade-off as an explicit operator choice (never a FerroEHR default).
  • The Helm chart’s default container resources now bound ephemeral storage (requests 128Mi, limits 1Gi) alongside CPU and memory: the root filesystem is read-only and the container writes only logs and tmp, so an unbounded local-write default could evict a node’s neighbouring pods. Chart 6.0.18.
  • The landing page’s Content-Security-Policy drops 'unsafe-inline' from script-src: the page’s only script element is a JSON-LD data block, which browsers never execute, so the allowance permitted inline script injection for nothing in return.
  • The hosted sandbox’s delivery pipeline is redesigned end to end (deploy/vercel/README.md): one trigger owner (the Sandbox deploy workflow, firing after a release image publishes, on sandbox-posture changes, or manually), Dockerfile.vercel tracking the :latest release pointer instead of a hand-bumped version pin, an automatic wipe-and-reseed after every successful redeploy, and no ignore script — the ordering race it guarded cannot occur when the only trigger fires after the image exists. Sandbox failures now surface only in the sandbox-named workflows. The published :latest container tag no longer moves on prerelease tags.

4.0.3 - 2026-08-25

Added

  • The hosted sandbox resets every night around midnight UTC: a scheduled job wipes the database (fenced so it can only ever run against a Neon endpoint), the next boot re-runs the migrations, and scripts/sandbox/reseed.sh seeds demo EHRs with example compositions from published CKM templates through the public API. Visiting the sandbox root or its favicon now lands on the Swagger UI instead of a 404.

  • The server understands two more deployment-platform conventions. PORT (Vercel, Cloud Run, Heroku inject it) binds 0.0.0.0:<PORT>, below FERROEHR__SERVER__BIND. The libpq environment set (PGHOST, PGUSER, PGPASSWORD, PGDATABASE, PGPORT, PGSSLMODE — what managed-Postgres integrations such as Neon inject) assembles the database DSN when no URL form is set; DATABASE_URL beats the assembled form and FERROEHR__DB__URL beats both. Direct endpoints beat pooled ones inside that layer (DATABASE_URL_UNPOOLED over DATABASE_URL, PGHOST_UNPOOLED over PGHOST): a transaction-pooled endpoint hands statements to backend connections without the session search_path, which surfaced as intermittent “relation does not exist” errors on the hosted sandbox. A Dockerfile.vercel plus vercel.json ship for Vercel’s Container preset (the hosted sandbox); the Dockerfile references the published release image, so Vercel’s build step is a pull measured in seconds instead of a half-hour Rust compile, and the sandbox runs the exact bytes CI built, signed and tested. The release-cut guard now holds its FROM tag to the workspace version alongside the compose defaults. Vercel deploys only from the develop branch (no per-PR previews), and because the Dockerfile pins the release tag, a develop push is a seconds-long re-pull of the same release image: the sandbox always runs the latest release, and its content changes only at a release cut.

  • A one-click tester sandbox: opening the repository in a GitHub Codespace boots the published quickstart stack (server, PostgreSQL 18, admin console) automatically and forwards the API and console ports. The committed .devcontainer/ pins every compose command to the standalone docker-compose.yml, so the sandbox runs release images rather than building the checkout. Documented at Installation → Try it in Codespaces and linked from the README quick start.

  • POST {base}/admin/integrity/verify sweeps the stored data for content-copy disagreement. Every version’s content is stored twice — as the materialized document a point read serves and as the decomposed rows the AQL engine queries — and read-time signature verification only ever recomputed the first, so a row edited behind the server’s back could reach a client through an AQL scalar result with no integrity signal. The sweep re-derives every stored version from its decomposed rows, compares, and reports every mismatch by identifier with a defect of content_differs, nodes_missing, nodes_unreadable, or unexpected_nodes. It reads the archived tier too, takes no lock, runs off the request path, and never logs or returns content. A finding is reported in the 200 body, not as a failed request. Behind the existing admin switch and admin role.

Changed

  • The documentation site no longer carries its own OpenAPI reference at /api/. Every “API reference” link now opens the live sandbox’s Swagger UI at https://sandbox.ferroehr.eu/ferroehr/rest/swagger-ui, served by the running server from its own handlers with the public demo credentials ferroehr / ferroehr. The reader gets the current release rather than a copy assembled at site-build time, and “Try it out” issues real requests against real data. Every deployment still serves the same UI at /ferroehr/rest/swagger-ui when swagger_ui is on.

  • A composition CREATE skips the explicit transaction when the commit is the one folded statement (no accompanying attestations, event outbox off — the common case): a single SQL statement is atomic on its own, so the BEGIN/COMMIT round trips are gone and a create is two database statements end to end (the merged pre-check read and the folded commit).

  • A version UPDATE gets three round trips faster. Closing the superseded lineage tip now rides the one folded commit statement (its leading CTE, at the same bound instant, so the close boundary and the new version’s open bound are one value by construction); the archived-object thaw rides the version-tree placement read instead of its own statement; and the first-version invariant read (archetype and category stability across versions) rides the update’s merged pre-read instead of a transaction-time statement. Measured on a 14k-version store: composition update p50 drops from 10.6 ms to 5.6 ms, within ~1 ms of create.

  • The first-version invariant read now consults both storage tiers, so updating an archived composition enforces the cross-version archetype and category invariants instead of silently skipping them (the read used to run before the thaw and saw only the primary tier).

  • The write path sheds most of another btree: the node CONTAINS-anchor index narrows from (rm_type, archetype) to (rm_type) alone — 83% of node rows carry at-code archetype text whose key bytes dominated the index’s maintenance, while every measured anchor plan either used the subsumption index (full-HRID anchors) or filtered archetype after the rm_type probe at identical latency. Measured: the one-statement commit drops a further ~0.3 ms; the read probes (EHR-scoped, archetype-anchored, rm_type-only, and the no-match worst case) are unchanged.

  • A versioned-object commit is ONE SQL statement: the decomposed node rows ride the folded commit CTE (ordered after the version row, which also satisfies their foreign key in-statement), removing the separate node insert round trip from every local write — composition, EHR_STATUS, directory, demographic, and the CONTRIBUTION route alike. A logical delete folds through with empty arrays.

  • Every versioned-object write sheds one btree: EHR-scoped AQL routes its scoping through the version spine (where the engine already mirrored the predicate), so the per-node-row ehr_id index — maintained for every row of every commit and, measured over the generated plans, serving no read the spine route does not — is gone from the baseline. Measured at a 100k-composition seed: the node insert drops ~0.5 ms per commit (~13%); EHR-scoped reads are unchanged. Deployments recreate to pick the baseline up (greenfield policy).

  • Commit latency’s tail shrinks: the large stored JSON columns (vo_version.body, vo_version.wrapped_original; the node fragments and audit columns already had it) compress with lz4 instead of PostgreSQL’s default pglz, and the bundled compose database enables wal_compression=lz4 — measured together, per-commit WAL volume roughly halves (~52 KB → ~24 KB) and the post-checkpoint p99 drops from ~180 ms to ~19 ms on the reference workload, with the median commit ~0.5–1 ms faster. Greenfield deployments pick the column compression up by recreating the database; the WAL setting is the compose default and an operator knob (BENCH_PG_WAL_COMPRESSION) elsewhere.

4.0.2 - 2026-08-24

Added

  • A mixed CONTRIBUTION now honours its own EHR_STATUS member when the content-write gate runs: content members beside an EHR_STATUS member setting is_modifiable = true are accepted against a deactivated EHR (reactivate-and-write in one atomic commit), and a deactivating CONTRIBUTION may carry its final content updates. Content against a deactivated EHR without a reactivating member stays a 409. The rule is order-independent over the change set (community report; adjudication with citations on the tracker).

  • The book gained a Version signing chapter: a digest-signing page and a PGP-signing page, each with a flow diagram (commit-time signing and read-time verification), covering what is signed, the client-supplied signature rules, key configuration and rotation, and the IMPORTED_VERSION wrapper signature.

  • Admin console: the ADL 2 template detail shows the path catalog — the same expandable tree and node inspector the ADL 1.4 detail has — built by the console from the served AOM2 JSON. The CDR’s REST surface is unchanged (the released API defines no Web Template representation on the ADL 2 resource).

  • Helm chart 6.0.16: metrics.grafanaDashboard.enabled ships the default “FerroEHR — service overview” Grafana dashboard as a ConfigMap the Grafana dashboard sidecar (kube-prometheus-stack) auto-imports; the compose observability overlay provisions the same dashboard, now extended to AQL rate and phase latency, plan-cache hit ratio, database pool, Tokio runtime, and audit throughput — every panel query written against the served metric names.

  • Docs: the compose installation page states the supported container engines — the quickstart, both profiles and both overlays are verified on Docker and Podman.

Changed

  • Current-version lookups gained dedicated partial indexes: EHR-scoped current-row reads (EHR_STATUS/FOLDER resolution, EHR summaries, the directory join) and the persistent-COMPOSITION duplicate probe now hit an index that holds exactly one entry per live object instead of walking an EHR’s version history. Deployments recreate their database to pick the baseline change up (greenfield migration policy).

  • The composition-create write transaction no longer spends a round trip reading the clock: the commit instant rides the EHR writability gate’s own statement, and the folded commit statements now BIND that instant as the audit time and version validity open bound — the stored commit time and the signed one are a single value by construction.

  • A signed commit stopped copying the composition body twice: the signature’s canonical form is produced over a shallow reference view (the data subtree and the signature drop are joined/filtered by reference, never a deep copy of the version body).

  • The admin archive load batches per relation — one unnest statement each for version rows, node rows, audits, contributions, attestations, item tags, folder ranks and archive markers — instead of a statement per row (a 1,000-version EHR record previously cost ~2,000+ round trips); the EHR-Extract export’s demographics chapter batches its party reads and version counts the same way.

  • ABAC-checked AQL queries run the scope collection concurrently with the main query under the one execution budget (wall-clock is the maximum of the two, not their sum).

  • Template existence checks on the TDD prepare and template-example surfaces probe with EXISTS instead of moving the stored OPT XML; commit decomposition and content negotiation shed their per-node and per-request allocations.

  • Version point reads (composition, EHR_STATUS, directory, party — every .../{uid} and version_at_time read) execute as ONE SQL statement, and the served canonical body is now MATERIALIZED at commit (vo_version.body, written from the same value the node rows are decomposed from) instead of being reassembled from the node subtree on every read. Measured: point-read latency drops under 1 ms (mean ~0.7 ms at c=1 on reference hardware) with the database-side read cost down ~8×; the node rows remain the AQL source of truth. Storage grows by roughly the size of each stored version’s canonical JSON.

  • The EHR_STATUS reads (current and version_at_time) and the current directory read resolve their container and read the version in ONE statement each (previously a container-resolution query followed by the version read); a revision history folds its attestations into the one metadata statement; the versioned-object container reads carry their ownership check inside the bounds statement; and a tag-collection replace returns the stored collection from its own insert instead of re-reading it after commit.

  • Two per-item read loops became single statements: a resolved CONTRIBUTION read (Prefer: resolve_refs) loads all its member versions in one batched query instead of one read per member, and the persistent-COMPOSITION uniqueness pre-check is one EXISTS probe instead of reading every live composition body in the EHR.

  • Response compression now runs at the fastest level for every negotiated encoding (it previously used the libraries’ default levels — real clients negotiating brotli paid visible per-response compression CPU for a marginal ratio gain), and a bare object read no longer re-stamps a uid the stored body already carries.

  • Every object-addressed read (full version reads and the metadata/existence lookups behind ETags, revision histories, and 404 checks) queries the two-tier union view in ONE statement instead of retrying the cold archival tier in its own transaction on a primary miss — a miss (an unknown id, a not-yet-created resource, a probing client) no longer costs four extra database round trips, and archived objects stay retrievable exactly as before.

  • AQL whole-object projections got cheaper end to end: a root-anchored projection (SELECT c FROM … CONTAINS COMPOSITION c) serves the materialized version body directly instead of re-aggregating and reassembling the node subtree, and the RESULT_SET assembly moves each document into its cell and into the response envelope instead of deep-copying the page three times on the way out.

  • Commit and update paths do less per-request work: the version body is assembled once per commit (it previously reassembled twice — once for the signature, once for storage), the composition-update pre-read and the cross-version invariant checks read two text scalars off the materialized body instead of fetching and parsing JSON fragments, and the per-request EHR_ACCESS gate shares the cached settings instead of cloning the access list.

Fixed

  • A composition decomposing to more than 4,095 node rows now commits: the node insert previously bound 16 parameters per row against PostgreSQL’s 65,535-parameter statement cap, so a sufficiently large document failed the write outright. The insert is now one fixed-text unnest statement with one array bind per column at any row count (which also ends the per-commit statement-text churn in the prepared-statement cache).

  • Multi-tenant deployments carry db.statement_timeout_ms again: the tenant-scoped pool’s connection hook replaced the base hook and silently dropped the timeout, leaving those deployments without the database-side runaway-query guard.

  • Compose: docker compose --profile s3 up -d --wait no longer fails after everything succeeded. Current Compose (v5.4, Docker and Podman alike) treats any exited container as a --wait failure, including the bucket initializer’s successful exit; the server now declares a service_completed_successfully dependency on the initializer (inert without the s3 profile), which both orders startup after bucket creation and marks the initializer as a one-shot. The documented invocation drops the initializer from the service list.

  • Compose: the observability overlay boots under Podman. Every component of the LGTM container creates its state under /data, which the image cannot create under Podman (its root directory ships mode 555 and the overlay drops all capabilities); a tmpfs mount at /data fixes it on both engines.

  • Helm chart 6.0.15: the chart-signing step retries transient Sigstore failures instead of stranding a published version unsigned. Chart 6.0.14 was pushed but lost its signature and provenance attestation to a single Fulcio connection reset; it stays published as-is (an OCI push is immutable under this repository’s lanes) and 6.0.15 is the signed, attested build of the same content. Kubernetes users should install 6.0.15.

4.0.1 - 2026-08-24

Added

  • Every published artifact class now builds at SLSA Build Level 3: the container images and the Helm chart moved into reusable build workflows (build-image.yml, build-chart.yml) alongside the release binaries’ existing lane, so the attestation-signing material is out of reach of any caller-defined step and a consumer can pin the exact signer workflow. The README carries the SLSA Build L3 badge, linked to the verification guide.
  • Release binaries and the server container image are built with cargo auditable: the shipped binary carries its own compressed dependency list in a .dep-v0 section (it survives stripping — verified first-hand), so binary-reading scanners such as syft, trivy, grype and osv-scanner recover the crate graph from the artifact itself.
  • The verifying-releases guide now covers the SBOM-attestation verify command (with its required predicate type), digest-form image verification, the registry-only and fully-offline verification paths, and a plain statement of what crates.io Trusted Publishing does and does not provide.

Fixed

  • The container lane’s change detection no longer degrades with a warning on tag pushes and manual dispatches: building everything in those cases is now declared instead of falling out of a missing event field.

  • The Helm chart publishes again as 6.0.13, carrying the v4.0.0 server as its appVersion: the v4.0.0 cut changed packaged chart content (appVersion, the generated README, the image annotations) without bumping the chart’s own version, so the tag-triggered publish was correctly refused by the immutability guard and no chart for v4.0.0 existed until this one.

4.0.0 - 2026-08-24

Added

  • Admin console: a Subscriptions screen (shown only when the CDR’s event-subscription admin API is enabled) — list, create, edit, and two-step delete of change-event subscriptions.

  • Admin console: the EHR detail’s Compositions tab gains filters — template, date range, and composer, all AQL-backed and kept in the URL so a filtered view can be shared — and its rows open the composition’s rendered clinical view directly. The EHR header now shows the subject’s identity and the queryable/modifiable badges.

  • Admin console: a Commit tab on the EHR detail — stage several changes (create a composition, amend one, modify the EHR status), set the contribution’s audit, and commit them as ONE contribution, all-or-nothing: the openEHR-native way to make correlated changes. A refused commit keeps the staging intact and shows the per-version diagnostics verbatim; staged changes live only in the open session.

  • Admin console: a FHIR screen (shown only when the CDR’s FHIR API is enabled) — the mapping-store editor (list, create, edit, and two-step delete of the connector’s mapping definitions, edited as JSON documents with the CDR’s validation diagnostics verbatim), a read-path viewer showing what a mapping produces on GET /fhir/r4/{type}, and a validate-only dry-run panel over $validate that commits nothing. The console deliberately has no path to the committing FHIR ingest door.

  • The tenancy extension gains GET /admin/tenant/current — the tenant the calling credential resolves to ({"default": bool, "tenant": record|null}; the reserved default tenant when the request runs unscoped). Multi-tenancy stays credential-derived: the read reports the middleware’s own resolution, and nothing selects a tenant.

  • Admin console: a Tenants screen (shown only when the CDR serves the tenancy extension) — the tenant registry with create, edit, and two-step delete, plus a read-only card naming the tenant this session’s credential resolves to. There is deliberately no tenant switcher.

  • Admin console: the ADL2 template rows gain the same two-step delete the ADL 1.4 rows have, driving the CDR’s existing ADL2 artefact delete (with its never-orphan refusal surfaced verbatim).

  • Admin console: the ADL2 template family — the Templates screen gains a family switch (ADL 1.4 | ADL 2, kept in the URL): list and upload ADL2 operational-template sources (validation diagnostics surface verbatim, like the OPT upload), and a per-template detail screen serving the stored source, the AOM2 canonical JSON, an example composition (JSON or XML), and version navigation across a template’s stored versions. ADL2 templates carry no Web Template, so the detail states that instead of a path catalog.

  • Admin console: a Terminology section — browse the terminologies the CDR serves, define a code, test subsumption, and expand or validate a value set — plus a terminology-aware code picker on the query builder’s coded criterion: the terminology field offers the served ids, a code can be looked up (rendered code — term) or pulled from a value-set expansion, and free-text entry keeps working for terminologies the server does not host. The read-only terminology extension API is now enabled in the dev compose stack’s configuration so these screens work out of the box.

  • Admin console: the openEHR item-tag surfaces — a tag panel on the composition viewer (following the version the viewer shows) and on the EHR detail’s Status tab (the versioned container, so tags survive new versions), plus a Tags tab listing every tag in an EHR grouped by the object it sits on, filterable by key, value and target path in the URL. One shared tag kit serves these and the demographic tag editor.

  • Admin console: a full Demographics section — a per-kind party browser and editor for all five party kinds (people, organisations, groups, agents, roles) with create, view, verbatim-merge update and logical delete; versioned-party history with time-travel; a relationships index and detail (both linked ends, edit, history, delete) plus a relationships tab on every party; demographic tag editing per party and a repository-wide tag index; and a demographic contribution viewer. The openEHR demographic API is published in the development state within the implemented REST release, and the section’s book page says so.

  • FHIR mapping dry run: POST /fhir/r4/{resource_type}/$validate (the HL7 FHIR R4 validation-operation convention) runs the whole ingest pipeline — mapping resolution, the FLAT build, the provenance stamp, and the same validation the real commit runs — and commits nothing. The OperationOutcome carries the verdict: the validator’s rejections verbatim, or the valid verdict plus the EHR disposition (the target EHR is resolved and reported, never created). Same starter-set scope, config gate and access class as the ingest door; the mapping-definition contract and a worked blood-pressure example are now documented in the book’s FHIR chapter.

Changed

  • Every JSON error body now has ONE uniform shape: {"error", "message", "validationErrors"}. Semantic-validation failures populate validationErrors with their "<path>: <message>" entries (as before); every other error now carries the member as an empty list, and validation refusals additionally gain the machine-readable error reason phrase. This makes the 400 bodies satisfy the released Error.yaml schema’s required member list; clients matching on the previous two-shape split keep working (the change is additive on both shapes).

  • The FHIR connector’s release identity is recorded as FHIR R4: every resource and element it touches is unchanged in R4B (HL7’s own R4B scope statement), so the wire’s /fhir/r4 is truthful and connector-side text and citations now say R4 consistently. The terminology integration keeps its deliberate R4B identity — the two vocabularies are intentionally different, and the documentation says which is which.

  • An event-subscription update is an explicit full replace: enabled is now required on the body — omitting it previously defaulted to true and silently re-enabled a deliberately disabled subscription — and any unknown body key is refused instead of silently dropped (echoed read-only members from a prior read stay tolerated).

Fixed

  • Admin console: the directory history panel loads a bounded window of the newest versions (with a “Load older versions” affordance) instead of one request per version back to v1, and opening the history of a logically deleted directory shows the empty state instead of a bogus row; the template detail screen fetches and parses its OPT once instead of three times; the /system repository-usage card runs its per-template counts concurrently and says when it is showing a truncated list.

  • Admin console: a CDR-refused session (401) and a wrong-role refusal (403) now read as two different situations with two different next actions — previously both collapsed into the same “forbidden” copy. The update flows send back the very ETag the CDR served instead of re-deriving the precondition from the body, and an already-zoned date-time entered in a time-travel picker (e.g. …+02:00) is no longer double-stamped with Z.

  • A composition DELETE that volunteers an If-Match precondition now honours it: a stale condition is refused with 412 Precondition Failed carrying the latest version’s ETag, evaluated only when the delete would otherwise have proceeded (RFC 9110 precondition precedence). Previously the received header was silently ignored and the delete performed.

  • Admin console, from the close-out wire audit: a refused commit’s per-path validation details (validationErrors) now render line by line instead of collapsing to one generic message; opening a deleted composition version shows a first-class “deleted at this instant” state instead of a blank pane; the /system repository-usage card sends the template id as a query-parameter binding (an apostrophe-bearing template id previously broke the card’s AQL — and could escape the string literal); every time-travel picker states that the entered instant is interpreted as UTC; the conformance-manifest reader tolerates a peer CDR omitting optional members instead of hiding the admin affordances; the composition delete no longer sends a self-satisfying If-Match; tag updates state Prefer: return=minimal explicitly; and a FHIR connector answer now counts any success status as completed, not only 200.

  • An AQL query parameter whose value cannot be coerced to the type its predicate compares against (an invalid date/time or text representation) is now the caller’s 400, naming the defect class — previously the database’s coercion failure surfaced as an opaque internal 500.

  • Every member-scoped refusal on the contribution commit now names the offending member (versions[i]) — the data-parse, template-resolution, change-type and audit refusals previously carried no index, so a client staging several changes could not tell which one was rejected.

  • Admin console: a FHIR-shaped refusal (an OperationOutcome) now renders its human diagnostics wherever an error surfaces — the shared diagnostic reader speaks both error vocabularies, so no screen can hand raw JSON to a toast.

  • The FHIR read facade serves one Bundle entry per stored composition — with several enabled mappings over one template it previously served the same composition once per mapping, duplicating fullUrls within a Bundle (HL7 FHIR R4 bdl-7) and over-counting total. The mapping that wins a composition follows the same precedence ingest uses.

  • The published conformance statement’s product version had sat at 3.6.0 since that release while the record beside it moved on — it now matches the workspace version, and the release-cut guard fails whenever the two diverge again.

  • The product is spelled FerroEHR everywhere a human reads it as a name: the admin console’s sidebar wordmark and every browser-tab title (both previously ferroehr-admin), the three published container images’ display titles, and the conformance report’s opening line (which now leads with the product name and keeps the machine sut key beside it). Technical identifiers — package names, URLs, paths, environment variables, registry references — deliberately stay lowercase.

  • Admin console: the sidebar is grouped — the domain sections (Dashboard through Terminology, plus FHIR and Tenants when served) sit above a divider, and the platform group (Operations, Audit log, System) sits below it, with System last.

  • Admin console: the query-results chart draws with linear interpolation — the smoothing default emitted invalid SVG (a NaN control point) whenever two result rows shared a timestamp on a time axis, and it drew smoothed values between samples that were never measured.

  • The unknown-configuration-key diagnostic now names the key’s full section path (auth.oidc.enabled) and attributes a file line only when the key is really defined under that section — a key arriving from the environment is named by its FERROEHR__… variable instead of pointing at an unrelated, valid line of the configuration file.

  • The template lists’ version query filter now matches the template id’s version axis, as the API documents (“taken from template_id”) — an exact version or a prefix pattern like 1.2.* previously matched nothing on either the ADL 1.4 or the ADL2 list, leaving * as the only working filter.

  • Admin console: the EHR_STATUS editor and the demographics party editor are now inert (fields and save disabled) until the served document has seeded the form, so a keystroke can no longer race the seed, be silently overwritten, and then be committed as the overwritten text.

  • Admin console: the composition links on the EHR detail’s Compositions tab now percent-encode the EHR id and the versioned-object id, like every other link the console builds.

  • The COMPOSITION and EHR_STATUS resource reads now refuse a uid_based_id of another RM kind with 404, as the REST specification requires — an EHR_STATUS id addressed on the composition route (or vice versa) previously answered 200 with the other kind’s body. Every kind-scoped read (latest, at-version, at-time, revision history, and the directory family) now discriminates the stored kind beside the owning EHR.

  • Importing an EHR Extract that advances an already-held BRANCH lineage (a later branch version arriving while the stored branch head is still open) now lands as an append: the stored tip is superseded and the new version becomes the lineage’s head, mirroring the trunk behaviour — RM common master06 §Copying / §Semantics in Distributed Systems. Previously the import was refused with a bare 409 Conflict. A stale receipt (a branch version at or below the stored tip) is now refused with a message naming the stored tip, exactly like a stale trunk re-import.

Removed

  • The event subscription’s archetype predicate: it was stored and served but consulted by nothing — the AMQP routing key carries no archetype segment, so the predicate never filtered anything in any combination. Filter by template instead (an operational template names its root archetype).

3.20.0 - 2026-08-21

Fixed

  • Committing a CONTRIBUTION whose EHR_STATUS member carries lifecycle_state = incomplete is now accepted, with the incomplete relaxation (existence and cardinality lower bounds lifted) applied to the status body exactly as for every other content type — RM common master06 defines the incomplete state generically and no released text excludes EHR_STATUS. The direct EHR_STATUS update route likewise honours an incomplete lifecycle it previously ignored. Previously such commits were refused with a 422.
  • The version lifecycle transition table now follows the formal state machine the RM designates (the RM-version_lifecycle diagram), in both directions: complete → incomplete (the drawn update edge) is accepted, while same-state re-commits of inactive or abandoned content — permissions the machine does not draw — are now refused with a 422 (resume editing via reactivate/retrieve first).

Changed

  • The from-source server image build (docker compose --build, the conformance stack, the repo-dev posture) now keeps the whole cargo target directory in a persistent BuildKit cache mount — the official Docker Rust pattern — instead of the cargo-chef dependency layer. Cargo’s own freshness now decides what recompiles, so an app-only edit no longer recompiles the eight generated openehr-* spec crates (previously every edit paid the full 20+ minute workspace compile; cargo-chef documents that local workspace crates cannot ride its layer). The first build on a machine seeds the cache; every later build compiles only what changed.
  • The entire outbound upstream-report corpus — all 215 reports of defects, contradictions, and silences in the released openEHR specifications — was re-verified first-hand against the vendored spec text in one adversarial pass. Three reports were refuted and withdrawn, one rewritten to its provable narrow claim, and every correction the pass surfaced travels on the closed reports for the future submission to openEHR. Verification is now terminal: a verified report closes as the standing record, so the open tracker stays near zero by design.
  • The conformance catalogue follows the pass’s refutations: committing a CONTRIBUTION whose EHR_STATUS member carries lifecycle_state = incomplete is now a gating acceptance case (RM common master06 defines the incomplete state generically; the reject lean existed only in the stalled CNF guide), and the empty-directory 404 cases gate unconditionally (the released OAS defines the branch — the former empty-vs-error option selection is gone).

3.19.0 - 2026-08-21

Fixed

  • Error responses name the SM call status the service model actually declares. Several refusals previously reported a neighbouring generic status: a duplicate EHR id now reports ehr_create_fail_duplicate_id, a subject another EHR already holds reports ehr_for_subject_already_exists (on create, status update, EHR-Extract import and archive load alike), an invalid uploaded template reports invalid_template, an invalid uploaded archetype reports invalid_archetype, and conflicts the service model names nothing more precise for (duplicate templates or stored queries, referenced-artefact deletes, directory and modifiability conflicts) report the honest generic conflict instead of composition_already_exists. No HTTP status moves — only the error token in the body becomes accurate.

3.18.0 - 2026-08-20

Added

  • FHIR mapping definitions gained where()/first() paths and cross-terminology code translation. The connector’s FHIRPath subset now supports single-condition where(path = literal) filters and first() (picking, say, a component by its code instead of its position), and a coded entry can declare translate to convert codes between terminologies at ingest time via a configured FHIR terminology server’s ConceptMap/$translate. Only strictly equivalent matches are taken; an untranslatable required entry refuses the ingest, an optional one writes nothing, and a translate mapping without a configured terminology server is a configuration error — the untranslated code is never passed through under the target terminology.

  • The whole-repository dump archives standalone demographic containers. POST {base}/admin/dump now writes a demographic wave beside the EHR wave: parties and party relationships that live outside any EHR land in demographic-commons.json (their shared audits and contributions) plus demographic-NNNN.json segments, in every logical format and container (loose, zip, 7z; canonical-XML exports externalize their versions as versions/*.xml like the EHR wave). POST {base}/admin/load restores them verbatim, reports an already-present container under its own kind (PERSON, ORGANISATION, GROUP, AGENT, ROLE, PARTY_RELATIONSHIP) instead of failing, and still reads archives written before the wave existed.

  • Template and archetype uploads now validate the artefact’s own meta-data against the RM resource-package rules. An OPT 1.4 whose header violates the RM class invariants — a language outside the openEHR languages code set, is_controlled without a revision history, a description without an author or details, duplicate detail languages, an empty lifecycle_state or purpose — is refused with a 422 naming the violated invariant, and validate answers with the same catalogue as upload, so the two can no longer disagree. ADL 1.4 source uploads refuse the structural and language-code rows the same way; the empty-prose rows (purpose/use/misuse as <"">) are reported as named warnings there, because the empty string is how real-world 1.4 authoring spells absence (adjudicated against the full vendored CKM library, which uploads unchanged).

Fixed

  • DV_URI accepts scheme-less and plain-text values. The validator previously refused any DV_URI.value without an RFC 3986 scheme — a floor the class does not declare: its only invariant is Value_valid: not value.is_empty, and the RM’s URI chapter explicitly allows “plain-text URIs” (encoding is the consumer’s duty at the point of use). Relative references and plain text now commit; DV_EHR_URI is unchanged (its own Scheme_valid invariant still requires the ehr scheme). The conformance catalogue’s URI cases carry the re-derived expectations.

  • ITEM_TABLE.as_hierarchy produces the specified row encoding. The openehr-rm function previously emitted a column transpose (following a contradictory one-line summary in the class table); it now encodes one CLUSTER per row renamed to the stringified row number, as the RM’s own ISO 13606 encoding rules, class description and instance figure define (the contradiction is reported upstream).

  • validate and upload can no longer disagree about a template. The ADL 1.4 OPT validation endpoint now answers with the same artefact-validity catalogue the upload enforces, so a template that would be refused on upload is reported invalid by validation too.

  • A CONTRIBUTION delete member with a contradictory lifecycle_state is refused. Declaring deleted as the change type while stating a non-deleted lifecycle previously committed with the instruction silently dropped; it is now a 400 naming the contradiction, matching the header-seam behaviour.

  • The openehr-its crate’s default XML lineage is v2. The convenience serializer previously emitted the v1 namespace, whose published schema bundle cannot describe every RM 1.2.0 class this model emits; the v1 lineage stays available by explicit selection.

3.17.8 - 2026-08-19

Fixed

  • The quickstart’s S3 bucket initializer no longer races the gateway. seaweedfs-init retried nothing: the gateway’s healthcheck only proves the S3 port answers, and a single-shot CreateBucket against a still-warming filer could fail once and never run again, leaving every multimedia commit against a missing bucket. The initializer now retries until the bucket is verifiably listable (the same read the deployment probe makes) and exits loudly after 30 attempts — verified over repeated fresh-volume compose cycles.

Added

  • Directory folder references that claim this system must now resolve. A FOLDER.items OBJECT_REF whose namespace is local (or the server’s configured system id) and whose target is not a versioned object of that EHR refuses the commit with 422, naming each unresolvable reference at its tree path — on the /directory routes and on folder hierarchies committed through a CONTRIBUTION alike. References into foreign namespaces (and unknown) are stored verbatim, unchecked: openEHR object references are explicitly distributed, so only a reference that claims this system and cannot resolve here is a client defect. No released openEHR text assigns this behaviour in either direction (register AMB-211, reported upstream); reported by an external user as a missing safety net for mistyped composition references.

Changed

  • PostgreSQL 18.6 everywhere — the ferroehr-postgres image base, the CI service containers, and every documented pin move from 18.4 to 18.6, the latest PG 18 patch (2026-08-13; 18.5 was never released). 18.6 is itself a security release fixing 28 CVEs, several in surface this CDR uses (pgcrypto, pg_trgm, to_char()/regexp buffer overruns, plan-cache invalidation). Operators upgrading an existing cluster in place should note the upstream release notes advise checking (and possibly reindexing) btree_gist indexes — the temporal vo_version keys use btree_gist.
  • Helm chart 6.0.9 — no template changes: appVersion moves to 3.17.8 and the chart README states the PostgreSQL floor as 18.6 (6.0.8 was cut mid-cycle and never published; the publish lane ships 6.0.9).

Security

  • h2 0.4.16 — RUSTSEC-2026-0258: the HTTP/2 implementation under hyper accepted and queued empty DATA frames without limit (low severity, unbounded-memory class). In-range lock upgrade; no code change.

  • The published ferroehr-postgres image is rebuilt clean of its 15 fixable HIGH findings. Nine were Debian util-linux packages (CVE-2026-53615) whose fix sits in trixie-security while the pinned upstream base rebuilds on its own cadence — the image now applies Debian security updates at build time, so a published fix reaches the image at the next rebuild instead of waiting for upstream. The other six are new Go-standard-library advisories in gosu, the privilege-dropping helper the upstream postgres image bundles (CVE-2026-33818, CVE-2026-56853, CVE-2026-56858, CVE-2026-56859, CVE-2026-56860, CVE-2026-56862 — encoding/asn1, net/http, html/template, encoding/xml, net/url, crypto/tls): gosu sets uid/gid and execs, opening no socket and parsing no untrusted input, so none of that code is reachable; they join the per-CVE, path-scoped adjudication with published OpenVEX statements in security/vex/postgres-gosu.openvex.json, which upstream’s next gosu rebuild deletes.

  • The vulnerable quick-xml 0.26 is eliminated from the build, not ignored. pprof’s flamegraph feature pinned inferno ^0.11, whose quick-xml 0.26 carried two DoS advisories (RUSTSEC-2026-0194/0195, fixed upstream in 0.41). The feature is now off and the flamegraph SVG (the /management/flamegraph endpoint and the bench profiler) renders through a direct inferno 0.12 dependency on quick-xml 0.41 — the two advisory ignores and their VEX statements are deleted, and the only quick-xml in the dependency graph is the fixed line. Every remaining advisory exception now names the upstream event that deletes it; none of the remaining four has a fixed release anywhere to upgrade to (rsa’s Marvin advisory has no patched release, two are unmaintained-crate notices on compile-time proc-macros via latest-release carriers, and the rkyv pin exists only in the lock file with zero resolved nodes).

3.17.7 - 2026-08-15

Added

  • Releases archive to Zenodo with a citable DOI. Every release tag now deposits a snapshot under the concept DOI 10.5281/zenodo.21940279 (the README badge; it always resolves to the latest archived version — v3.17.6 is 10.5281/zenodo.21940280). The deposit metadata is generated from CITATION.cff in the flat legacy shape Zenodo’s GitHub integration documents, after the first deposit demonstrably ignored the InvenioRDM-record-shape file and archived raw repository metadata instead; the concept DOI is recorded in CITATION.cff for citation managers.

Changed

  • Helm chart 6.0.7 — no template changes: the chart re-releases because appVersion moves to 3.17.7 and the published 6.0.6 is immutable by the publish lane’s own refusal-to-overwrite.

Security

  • The postgres image’s scan gate is green again. The new Go-stdlib advisory CVE-2026-39821 (golang.org/x/net/idna Punycode via net/http) matches gosu, the privilege-dropping helper the upstream docker-library/postgres base ships unchanged; it joins the fifteen already-adjudicated gosu entries as a per-CVE, path-scoped exception with an OpenVEX not_affected statement (gosu resolves no hostnames and issues no HTTP requests, so the IDNA path is never entered). The entries go when upstream rebuilds gosu on a fixed Go line.

3.17.6 - 2026-08-15

Security

  • The rsa advisory’s published VEX statement was false, and is re-grounded. It claimed RSA is “reached only through openidconnect” with “no RSA private-key operation” — the OpenPGP signing path (pgp → rsa) was missing. The corrected statement rests on the true ground (the default signing.mode = "digest" performs no asymmetric operation; OIDC is public-key verification) and explicitly places an RSA-keyed pgp deployment outside its scope. The quick-xml statements now describe the current graph (the advisories fire on inferno’s 0.26, writer-only, not the upgraded object_store path), and a new CI gate fails any advisory exception whose advisory cargo-deny no longer raises — a resolved advisory cannot keep its exception silently.

Added

  • The documented development→stable read-time refusal now exists. Every accepted commit records whether the released generation’s own reader can express the body (vo_version.stable_compatible; the extra in-memory parse runs only under the development profile), and the one seam every stored version body leaves storage through refuses a stamped-incompatible version under spec_profile = "stable" with a 409 naming the profile, the version and the remedy — never a silent down-conversion. Rows from before the column existed are assessed on the fly at read. AQL takes the same stamp wherever it serves a version body: a whole-object projection is gated at result assembly, so a page reaching a version the released generations cannot express refuses the whole query with that same 409 — never a row silently elided from a RESULT_SET, which has no per-row diagnostic channel. Leaf projections over the identical rows stay ungated (they serve data values over paths the planning gate already bounded), and under the default development profile the assembly gate costs nothing. Our own extension; no openEHR spec governs runtime generation selection.
  • The conformance pipeline exercises both generation sets in the one record. The ferroehr lane composes a third server of the same image over the primary’s database under spec_profile = "stable" (host port 8082), declared as the ixit’s sut_stable instance, and the catalogue drives the profile boundary end to end: the development server commits and serves the development-only body, the stable server refuses it at the version read and at whole-object AQL (409, register AMB-210), and a released-surface body serves identically under both. The generation set a deployment runs is a new ixit/case declaration (spec_profile / requires.spec_profile), selected on at ISO/IEC 9646 selection time like the signing and terminology postures.
  • Archive restore is on the wire. POST /admin/archive/ehrs/restore and POST /admin/archive/parties/restore (admin-gated, our own extension — the SM declares no un-archive call) bring archived records back from the cold tier: all-or-nothing over the same existence checks as archiving, idempotent, 204 with no body. Archiving is no longer one-way for an operator.
  • The chart’s NetworkPolicies say what they admit, and can refuse the accident (chart 6.0.6). An empty ingressFrom renders an ingress rule with no from — every source admitted — while reading as default-deny. The open posture is now an explicit value (networkPolicy.ingressAllowAll and adminUi.networkPolicy.ingressAllowAll, both shipped true): the rendered object’s description, the install notes and the chart README all state it, and setting it false with no ingressFrom refuses at render — so “no open ingress” becomes a machine-checked fact where it matters.
  • A fhir_outbound readiness indicator. When the FHIR outbound emitter is enabled, /health/readiness now carries its broker-delivery liveness — non-required, so a broker outage reports DEGRADED (the outbox retains messages) and never fails readiness. Previously a PHI-bearing outbound stream could be down with nothing on the health surface saying so.
  • A boot advisory for RSA OpenPGP signing keys. With signing.mode = "pgp" and an RSA key, every signature is an RSA private-key operation — the operation the Marvin timing sidechannel (RUSTSEC-2023-0071) concerns, with no fixed rsa release. Boot now emits a prominent warning naming the advisory and the Ed25519/ECC rotation path (signing.retired_key_paths keeps history verifying); deliberately not a refusal, because signatures are immutable committed facts an operator must stay able to verify. The signing configuration page carries the guidance.

Changed

  • The ad-hoc-query scope cases are split per carrier, and every zero-row scope assertion now has something to exclude. The bundled bare-EHR case becomes three one-behaviour cases (grammar acceptance with the released #0 column-naming rule; scope via the ehr_id query parameter; scope via the openehr-ehr-id request header — the only released carriers), and the scoped zero-row cases each first commit a composition into a SECOND EHR, so a green row evidences the scope rather than an empty store. The conformance records regenerate on the new case ids; the EHRbase record is honestly redder where that server ignores the scope.
  • The openehr-* crates move to 0.0.30, and the published packages are self-consistent. openehr-term ships openEHR’s verbatim terminology XML (CC-BY-SA 3.0) but declared only MIT AND Apache-2.0 — the expression now names all three licenses and the text travels in the package. The published openehr-its package carries attribution for the openEHR JSON Schema it ships (pinned commit, upstream path, license). Packaged test code no longer embeds fixture paths the package does not carry, so cargo test inside a published .crate compiles again. Two new guards keep all of it true: a pinned attribution that stops travelling fails, and a packaged source embedding an unpackaged path fails. All eight bump in lockstep.
  • The documentation site was read end-to-end against the tree and rewritten. Every page’s substantive claims were verified against the code, the chart, the workflows and the committed conformance artifacts; drifted claims were corrected rather than catalogued (among them: the binary is self-contained but not statically linked, the FHIR surfaces are R4B, the TLS floor defaults to 1.3-only, rate limiting ships enabled, the query execution budget ships at 30 s, and a removed probes.exec chart option is gone from the docs too). Long chapters split into sub-pages: cluster hardening (five), the configuration reference (five), release verification, and the admin/messaging API walkthroughs.

Fixed

  • The FHIR read façade and outbound emitter respect the spec-profile read gate. Both read stored bodies around the gated versioning seam, so a development-only body was mapped to FHIR under the stable profile with no refusal; both now go through the gated read, mutation-proven.
  • The boot banner follows the resolved log format. With the default log.format = "auto" in a container (stdout not a terminal), logs render as JSON but the multi-line ASCII banner still printed first, handing every log collector unparseable lines on each boot. The banner now keys on the same TTY-aware resolution the log layers use: json, and auto off a terminal, are parseable JSON from the first byte.
  • Partial feature builds of the server compile again. The FHIR outbound emitter’s wiring was gated on the events cargo feature while its module exists only under fhir: a --features events build (no fhir) failed to compile, and a --features fhir build never started the emitter. Both gates now sit on fhir, which implies events at every level.
  • The Helm chart is 6.0.5. Reading the site against the tree corrected chart comments and operator-facing messages (the stale distroless base name, the secret-routing key list, the config-in-Secret cause in NOTES.txt, two book citations that moved in the hardening split, and the egress refusal’s pointer); the packaged bytes changed, so the version moves.
  • Wire and OpenAPI text corrected while verifying the site. The 406 answer for an unrecognized XML version parameter now names version=2 as the default (it said version=1); the expand_multimedia parameter description no longer claims the served body is byte-identical to the committed one (expansion restores data while keeping the offload-added uri/integrity fields); and DELETE /definition/artefact/adl2/{id} documents its real 409 still-referenced refusal.
  • The Helm chart is 6.0.4. The v3.17.5 cut moved the chart’s appVersion to 3.17.5 without moving the chart’s own version, and 6.0.3 was already published — a published chart version is immutable, so the publish lane refused rather than replacing it. Install with --version 6.0.4 to get the 3.17.5 application image by default.

3.17.5 - 2026-08-13

Security

  • The ferroehr-postgres image moves to a rebuilt PostgreSQL 18.4 base. The upstream postgres:18.4 image was rebuilt with current Debian packages, and this pulls that rebuild in. Nothing about FerroEHR changes: the schema, the extensions and the entrypoint are untouched.

3.17.4 - 2026-08-13

Added

  • A published threat model. A new Threat model chapter names the actors, the protected assets and eight trust boundaries, and for each boundary states the control that holds and the residual risk that survives it — plus an explicit list of what FerroEHR does not defend against, so silence is never read as coverage. It exists because the parts of this system that protect patient data are precisely the parts no openEHR specification governs: the five-stage authorization order, multi-tenancy isolation, the audit trail and the archive tier are all this project’s own design, and until now every deployer had to reconstruct them from configuration prose.

  • Governance, maintainer and support documents. GOVERNANCE.md (how decisions are made and how to become a maintainer), MAINTAINERS.md (the roster, every publishing identity, and what happens if the holder is unavailable), SUPPORT.md (question versus defect versus vulnerability), and a CODEOWNERS file routing review per area. All four state the current single-maintainer reality plainly rather than describing a process that does not exist.

  • A published support and end-of-life policy. SECURITY.md now states which versions receive security fixes (only the newest), when a version stops receiving them (the moment a newer release exists), and how a fix reaches you — covering the server, the Helm chart and the published crates, which follow separate version lines. The same policy is reachable from the Operations chapter’s upgrade guidance, where an operator planning change control actually looks.

  • Machine-readable licensing (REUSE 3.3). A LICENSES/ directory carries the full text of every licence any file in this tree is offered under, named by SPDX identifier, and a root REUSE.toml declares by glob which files are offered under which — so licensing now survives a file being copied out of the repository, which per-tree provenance files could not do. No vendored file was modified. Two positions are represented rather than flattened: the MPL 1.1 election under a tri-licensed corpus, and one upstream file whose own header contradicts its repository’s licence. reuse lint and a declarations-versus-documentation check both gate the merge.

  • A published VEX document for the Rust dependency advisories. Every advisory the cargo deny gate accepts now carries an OpenVEX statement with a controlled-vocabulary justification and a checkable impact statement, at security/vex/rust-advisories.openvex.json, so a downstream scanner ingests the argument instead of reporting unexplained findings. It also covers the advisory that only a Cargo.lock-reading scanner reports for a crate this project’s feature set never compiles. The document is generated from deny.toml joined with the published reasoning, and CI fails if the two disagree in either direction — an advisory cannot be accepted without its justification reaching you.

  • About 190 openEHR spec functions are now callable on the generated types. The Reference Model, Archetype Model, BASE and LANG classes declare functions the specifications define — ITEM_LIST.ith_item, DV_QUANTIFIED.magnitude, ARCHETYPE_HRID decomposition, Iso8601_timezone’s eight accessors, the AOM2 primitive-constraint validity checks, and many more. They existed only as declarations; each one now has an implementation written from the governing spec section, with that citation in its documentation and a test asserting the published post-condition. Functions the released text leaves undefined, or that need state the class does not carry, are deliberately still absent and recorded as such rather than guessed at.

  • Time_Definitions’s eleven validity functions are now public on openehr-base — valid_year, valid_month, valid_day, valid_hour, valid_minute, valid_second, valid_fractional_second, days_in_month, and the four valid_iso8601_* string predicates. They had no realization anywhere despite being declared by the spec.

Changed

  • Every first-party Rust source file now states its licensing inside itself. All 2469 tracked .rs files carry an SPDX-FileCopyrightText and an SPDX-License-Identifier header: the six published spec crates state MIT AND Apache-2.0 — the position their manifests already declared — and everything else states MIT. Licensing was previously declared only by glob in REUSE.toml, which is complete but does not travel with a file that is copied out, and file-level redistribution is the expected case for a project whose premise is that people build on it and ship it. The 1471 generated files receive their header from the code generator, not from a hand edit, and a new CI check fails the build if any first-party Rust file loses its header or states a licence other than the one declared for its path. Vendored trees are untouched and stay glob-declared, as they must be. The eight published openehr-* crates are bumped to 0.0.26 in lockstep, since their packaged sources changed.

  • BREAKING — enumeration-constrained OPT and AOM2 fields are now typed enums, not strings. EXPR_OPERATOR.operator becomes OperatorKind, and C_DATE/C_DATE_TIME/C_TIME.timezone_validity become Option<ValidityKind>. The XML schemas constrain these fields to a fixed set of values, and carrying them as free text meant an out-of-range value was indistinguishable from a valid one until something downstream misread it; they are now refused when read. The wire form is unchanged — the same characters go out, and every operational template in the corpus still round-trips byte-for-byte — so only code that reads these fields as strings is affected.

  • A Web Template’s tz_validity no longer silently disappears when the template is malformed. It was parsed with a fallback that turned any unreadable value into “no timezone constraint at all”, which is a different and weaker statement than the template made.

  • An ehr: URI must name the EHR attribute it addresses. The server accepted a short form that put the versioned-object id straight after the EHR id (ehr:/<ehr_id>/<uid>::SYSTEM::1); no released spec defines it. BASE master11-paths enumerates the locator’s values — they “come from attribute names of the class EHR … namely compositions, directory etc.” — and every example carries one. The short form is now refused; write ehr:/<ehr_id>/compositions/<uid>::SYSTEM::1 instead. This affects any DV_EHR_URI value stored in a LINK or elsewhere that used the short form.

  • A refused log-filter swap now answers 400 or 503, not always 400. POST/DELETE /management/loggers carried one flattened error string, so a reload layer that had gone away was reported as if the caller had written bad directives. The two faults are now distinct types: unparseable directives stay 400, a subscriber whose reload handle is gone is 503.

  • Errors carry their cause across the remaining request-path seams. Six sites in the platform library — Web Template building, FHIR terminology decoding, the log-filter seam, and admin dump/load payload reading — flattened their cause into a message, so nothing downstream could match on it. Each now carries a typed source (RFC 0201). The sites that remain are structural and say so at the site: prometheus::Error has no source-bearing variant, and the console’s error types must be serializable (server-fn boundary) or Clone + Eq (reactive-signal storage), which no underlying error is.

  • The Rust toolchain moves to 1.97.1. It carries an LLVM miscompilation fix whose underlying bug has been present since at least 1.87 — the pinned 1.96.1 sat inside that window. The MSRV stays 1.96: nothing here uses a 1.97 feature, and the published crates are the only artifacts whose consumers would feel a bump.

  • One ISO-8601 grammar, not two. openehr-rm validated dates, times, date-times and durations with its own hand-written reader while depending on the openehr-base one that owns them. The two drifted twice and both drifts shipped, each letting a value pass validation and then behave as invalid. The RM side now calls the spec functions; 341 lines of duplicate grammar are gone.

  • ITEM_TABLE’s twelve accessors on the published spec crates — row_count, column_count, row_names, column_names, ith_row, named_row, has_row_with_name, has_column_with_name, has_row_with_key, row_with_key, element_at_cell_ij and as_hierarchy, which transposes the stored row-per-CLUSTER form into the column CLUSTERs the spec asks for.

  • Every DV_* arithmetic and accessor function the openEHR RM defines is now implemented on the published spec crates — the whole quantity and date/time family, not a subset: DV_QUANTITY, DV_COUNT, DV_PROPORTION and DV_DURATION arithmetic; DV_DATE, DV_TIME and DV_DATE_TIME displacement and difference; the DV_AMOUNT and DV_ABSOLUTE_QUANTITY dispatchers over their subtypes; and DV_PERIODIC_TIME_SPECIFICATION’s period, calendar_alignment, event_alignment and institution_specified, parsed from the HL7 PIVL/EIVL syntax the spec publishes.

  • Accuracy now propagates through arithmetic as the RM specifies it. DV_AMOUNT defines the rule — accuracies sum for both addition and subtraction, an unrecorded accuracy on either side makes the result’s unknown, and a mixed percent/absolute pair is expressed in the form of the larger operand — and every descendant follows it. DV_ABSOLUTE_QUANTITY’s duration-valued variant is applied to the date/time types. A combination no valid value can carry (a percentage past 100) refuses the operation rather than returning a value that fails its own class invariant.

  • A decimal factor now means the decimal its author wrote. DV_COUNT and DV_PROPORTION arithmetic reads a Real factor as the number it denotes rather than as the binary approximation carrying it, so count * 0.1 is a whole count and 1/3 equals 0.1/0.3. The previous reading refused both.

  • DV_COUNT arithmetic and DV_MULTIMEDIA/COMPOSITION predicates on the published spec crates: add, subtract, multiply, is_inline, is_external, is_compressed, has_integrity_check, is_persistent. Overflow and non-whole products are refused rather than wrapped or rounded — openEHR defines neither, and a wrapped count is a silently wrong clinical value.

  • Releases are archived on Zenodo with a citable DOI. .zenodo.json is generated from CITATION.cff so the two cannot disagree — necessary because Zenodo ignores CITATION.cff entirely whenever a .zenodo.json is present, and a deposit’s metadata is immutable once its DOI exists. It adds what the citation file cannot express: which openEHR specifications a release is derived from, and where the published crates live.

  • A fuzz target for the identifier readers. OBJECT_VERSION_ID and its siblings are parsed straight from the {version_uid} URL path parameter — before any body is read, any content type negotiated, or any authorization runs — and were previously unfuzzed. The harness also checks the composite’s own contract: an identifier recomposed from the three parts its reader just produced must equal itself, which catches a mis-sliced separator that still returns Ok.

  • Every optional integration now documents how to enable it on Kubernetes. Nine chapters — change events, FHIR connectors, terminology servers, Subject Proxy, S3 multimedia, the admin console, observability, version signing and multi-tenancy — described only the environment-variable and Compose paths, so an operator deploying with the chart had no documented way to switch any of them on, even though the chart’s config passthrough has always reached every key. Each now carries a values snippet, the prerequisites stated before the reader hits a boot failure (a bucket that must exist, a claim the identity provider must issue, an egress rule an exporter needs), and how to turn it off again.

  • authz.rbac.ehr_access_default — object-level default-deny, as one setting. An EHR that carries no ACCESS_CONTROL_SETTINGS was reachable by any caller the coarse layers admitted, and the only way to change that was to author a settings object per EHR, through a CONTRIBUTION commit per record — so the safer posture was reachable only by the most laborious path. Setting this to restricted makes a setting-less EHR reachable only by authz.rbac.admin_role. The default stays open, because changing it changes who can read existing records; explicit per-EHR settings win over it in both directions. The admin carve-out is deliberate — a plain deny would leave such a record unreachable by the operator who would author the settings that fix it.

  • Every pod the Helm chart deploys now runs in its own user namespace (hostUsers: false, chart 6.0.0). Container UIDs are mapped onto an unprivileged host range, so a container escape arrives as a host user that owns nothing — verified on a live v1.36.1 node, where the pod’s /proc/self/uid_map reads 0 838860800 65536 while the process still sees uid 65532. This is the reason the chart’s Kubernetes floor moves to 1.36 (see Changed). Set hostUsers: true to opt out on nodes whose runtime cannot map UIDs; there, the pod refuses to start rather than downgrading silently.

  • The chart spreads replicas across nodes by default. A soft maxSkew: 1 constraint over kubernetes.io/hostname means two replicas prefer two nodes, so one node failure is no longer a total outage. It is ScheduleAnyway, so a single-node cluster still schedules. Supplying topologySpreadConstraints replaces it wholesale.

  • supplementalGroupsPolicy: Strict on every pod: the process gets only the groups the manifest names, so a group baked into an image cannot widen file access.

  • The migration Job no longer counts a drain as a migration failure. A podFailurePolicy ignores pod failures caused by disruption, so ordinary cluster maintenance during a release cannot exhaust backoffLimit and fail the upgrade with no migration error in any log.

  • New chart values: service.trafficDistribution (zone-local routing, off by default), autoscaling.behavior (scaling policies passed through), adminUi.terminationGracePeriodSeconds and adminUi.preStopSleepSeconds. The console also gains a startup probe, a preStop pause and the server’s full security posture.

  • The Helm chart can deploy the admin console. adminUi.enabled renders the console as its own Deployment/Service/ServiceAccount, with an optional Ingress and a NetworkPolicy that confines its egress to the CDR and DNS — the console is a REST client of the CDR by mandate, so that is now enforced rather than documented. Off by default, and off renders nothing.

  • The Helm chart validates your values file (values.schema.json, chart 5.0.0, unchanged in 5.0.1). helm install, upgrade, lint and template now refuse a values file that misspells one of the chart’s own keys, gets a type wrong, or names a value outside the permitted set, instead of rendering and ignoring it — a breaking chart change, which is why the chart’s major version moves. The config: tree stays deliberately open: that vocabulary belongs to the server and is validated by the binary at boot, so copying it into the chart would fork it. Artifact Hub renders the schema on the chart’s listing. (#2184)

  • The published Helm chart is signed with a keyless cosign signature, from chart 5.0.1 onward, alongside the SLSA build provenance attestation it already carried — the attestation says what the chart was built from, the signature says who signed the artifact you pulled. No key material exists anywhere; verification requires an identity and an issuer, and both are in the installation guide. helm install --verify still does not apply: the chart ships no PGP .prov. Chart 5.0.0 is published but carries neither signature nor attestation — its publishing run pushed the chart and then failed at the signing step, and a published chart version is never replaced. It installs and runs normally; pin 5.0.1 or newer if you verify what you deploy. (#2183, #2184)

  • The PGP signing key can be rotated without losing history. Signing now uses a signing-capable subkey selected by its OpenPGP key flags, which is the rotation mechanism the standard provides: issue a new subkey, the certificate retains the old one, and every previously-signed version keeps verifying with no configuration change. For a certificate that is genuinely replaced — a compromised key, an organisational change — signing.retired_key_paths holds the retired public keys, verify-only, so a retired key can never sign again. Verification also no longer ignores subkeys, which it did before.

  • deploy/helm/ci/boot-check.sh — runs a real ferroehr image against every values overlay the chart ships, replaying the complete delivery surface: the rendered ferroehr.toml, every config.files entry, every file-borne secret at its real value, and the environment the Deployment declares (including valueFrom.secretKeyRef). Point FERROEHR_IMAGE at the tag you intend to deploy and pass your own values file to check it before installing. (#2159)

  • CI lane chart-boot — boots every committed values overlay against an image packaged from the branch under test, on every pull request that touches the chart or the configuration vocabulary the overlays are written against. Previously the equivalent check ran only on a release tag, which is after the cut has already failed. (#2159)

  • The server can run with no DDL rights at all. db.migrate = "verify" makes it issue no schema statements: at boot it checks that the database already carries exactly this build’s migrations and refuses to start otherwise, naming the schema and what is wrong with it (never migrated, behind this build, ahead of it, a failed migration, or a migration applied from different source text). The runtime DSN can then authenticate as ferroehr_app alone, so an application-level SQL flaw reaches rows and never the schema. apply remains the default, so an empty configuration still boots against an empty database. (#2049)

  • ferroehr db migrate and ferroehr db verify — the out-of-band schema step and its read-only check, for running migrations under a ferroehr_migrator DSN separately from the serving process. (#2049)

  • A migration Job in the Helm chart (migrations.job.enabled): a pre-install,pre-upgrade hook Job that Helm waits on, so a failed migration fails the release instead of rolling pods against a schema that was never applied. It authenticates from its own Secret — deliberately a different credential from database.existingSecret — and rendering is refused if that Secret is not named. The install NOTES report the resulting posture, including the case where the Job is enabled but config.db.migrate is still apply. (#2049)

  • Tamper evidence on the audit trail. Every record in the local Audit Record Repository is linked into a SHA-256 hash chain maintained inside PostgreSQL, committing to its predecessor and to its own content, so the protection covers every writer rather than only the server’s own code. The table now refuses every rewrite path but the per-sink forwarding stamp — a content UPDATE, a DELETE and a TRUNCATE are all rejected — and retention pruning goes through the one sanctioned deletion path, which records which positions it removed so reaping is distinguishable from tampering. SELECT * FROM audit.verify_audit_chain() reports one row per damaged record naming what is wrong with it, and nothing at all when the trail is intact; it is also reachable in-process as AuditStore::verify_chain. Detection, not prevention: the chain is unkeyed, so the controls for a party with unrestricted write access remain the least-privilege role and the off-box syslog/ATX:FHIR sinks. (#2059)

  • openehr-query: a spanned lexing entry point, lexer::lex_spanned, returning each token together with the byte range of the source it was lexed from (lexer::SpannedTokens, with a byte_span mapping from token indices to a source range), plus parser::parse_spanned over that stream. lexer::lex and parser::parse are unchanged (#2145).

  • Six more OCI annotation keys on every image: authors, vendor, documentation (the docs site, which no image pointed at), url, and base.name/base.digest naming the exact pinned base image each one is built on — so “what is this built from” is answerable from the image rather than from the Dockerfile. Every image also reports its own title and description now, where all three previously inherited the repository’s.

  • The Helm chart can deploy by digest. image.digest renders the pod’s image as repository@sha256:… and takes precedence over image.tag. The hardening chapter had been telling operators to deploy by digest “so what you verified is what runs” — and the chart had no such key and no template that read one, so the instruction silently did nothing. A digest is what a build provenance attestation is made over, so this is what makes verification bind to the image actually running. The value is accepted with or without the sha256: prefix.

  • The OpenSSF Best Practices badge at the passing level, alongside the Scorecard badge in the README. Every criterion is answered from something a reader can check — the vulnerability process in SECURITY.md, the release signing and provenance in the security chapter, the test and analysis gates in CI — rather than self-asserted.

  • Fuzzing for every parser that reads attacker-controlled bytes off the wire: cargo-fuzz (libFuzzer) harnesses for the canonical-JSON reader, the canonical-XML reader, the AQL lexer/parser, the FLAT/STRUCTURED simplified formats, the ADL source parser and the OPT 1.4 template reader, each a pure parse of a byte slice, seeded from the openEHR corpora already committed in the repository. A bounded campaign per target runs on a nightly schedule with its corpus persisted between runs; long campaigns are a documented local command (fuzz/README.md). The harnesses live in their own workspace, so no ordinary build, clippy or test run is affected.

  • The cluster-hardening chapter now covers breach containment, logging, managed control planes and the supply chain. What scaling to zero does and does not stop (it halts clinical access immediately — a clinical-safety decision to make before an incident, not during one — but does not undo an append-only commit, does not stop an attacker holding the DSN, and does not truncate the audit trail). A rotation procedure for every credential, with the restart requirement stated, because configuration is read at boot and Kubernetes propagating a Secret into a volume does not make a running process re-read it. A checkable supply-chain map: each control against the artifact that satisfies it and the command you can run to verify it yourself, including the two gaps that remain. And an honest account of what becomes your cloud provider’s on a managed control plane, with the instruction to test NetworkPolicy enforcement rather than read the provider’s documentation and conclude.

  • The application log and the ATNA audit trail are documented as two streams that must not be treated alike. The application log is diagnostics on stdout and may be sampled or dropped; the audit trail is the accountability record — who accessed which patient’s data — travelling by a different path, with its own store, its own ITI-81 retrieval endpoint and its own retention. A collector configured to drop a noisy stream under volume is a reasonable policy for the first and a compliance failure for the second, and the two are separated precisely so that one can be lossy.

  • Rotating a PGP version-signing key is documented as the operation it actually is. In the default digest mode there is no key and nothing to rotate. In pgp mode, replacing the key makes every previously-signed version fail verification: the stored signature carries no key identifier, verification checks against the single currently-configured key, and with the default verify_on_read: strict that is served as a 5xx on reading historical data. The signatures cannot be re-issued, because a version’s signature is an immutable committed fact. The three real options — treat the key as long-lived, move to verify_on_read: warn across a rotation as a recorded reduction in guarantee, or use digest mode — are documented with their trade-offs, along with the fact that no keyring or multi-key verification exists today.

  • A default-deny egress policy you can actually write correctly. networkPolicy.egress.enabled refuses all outbound traffic except what you list; DNS is always included, and the database is now a first-class networkPolicy.egress.database key rather than something to remember. Enabling egress with no database destination is refused at render time, because that mistake presents as a database fault — readiness reports the DB down, the log shows a connect timeout, and nothing mentions the network policy. The cluster-hardening chapter carries the full destination table derived from the configuration tree: every outbound destination the server can make, which config key turns it on, and the port it needs, so a policy can be assembled from what you have enabled rather than guessed from what a default install happens to use.

    Two failure modes are documented beside it because neither announces itself. A blocked OTLP collector drops spans without failing any request, so an over-tight policy produces a server that is healthy by every check and has silently stopped being observable. And tightening egress under a running pod appears to work when it has not: a NetworkPolicy is enforced on new connections, so a pod whose connection pool is already established keeps serving after you remove its database rule and fails at the next restart — measured, along with the recovery, on a live cluster.

  • The published container images are re-scanned every week, not only when they are built. CI scanned an image at the moment it was built, which catches what was known then and nothing after — so a CVE published the week after a release applied to the image you were running and nothing looked again. A scheduled lane now scans all three published images at the tag you pull, with the same severity floor and the same adjudicated exceptions, and with the project’s OpenVEX documents applied so an accepted finding keeps its argument attached. A finding both opens a tracking issue and fails the run, because a red scheduled run nobody watches is not a control. At the time of writing all three published images carry zero fixable HIGH/CRITICAL findings.

  • A copyable admission policy for verifying image provenance, for Kyverno and for sigstore-policy-controller, with the exact certificate identity and OIDC issuer our publishing lanes use — the deployment-time half of the signing work. It is documented rather than shipped in the chart, because an admission policy is cluster-scoped and governs workloads the chart knows nothing about, so helm uninstall must not be able to remove it. The chapter is explicit that the policies are written from the workflow definitions and not yet verified against a live attestation, and gives the command to reconcile them once the first attested release exists.

  • A cluster-hardening chapter, split by who can actually apply each control. The OWASP Kubernetes Security Cheat Sheet mixes controls a workload chart owns (pod security context, resource bounds, NetworkPolicy, ServiceAccount, secret consumption) with controls only a cluster operator can reach (node patching, etcd encryption, kubelet authentication, admission control, the API server’s own access model). The new chapter carries an ownership map naming, per control, who applies it and what happens if nobody does — including the ones no application-level hardening can compensate for, since an anonymous-auth kubelet or a readable etcd bypasses every control this server ships. Claims about the running deployment come from kubectl, crictl and /proc on a live cluster, not from reading values.yaml.

    Two facts the chart had been conflating are now stated separately: kubeVersion: ">=1.25.0-0" is a compatibility floor (the newest API the chart uses is HPA autoscaling/v2, GA in 1.23, so it genuinely runs there), while the supported platform is upstream’s three most recent minor releases — a window that moves three times a year, which is why it lives in the book rather than in a constraint that would refuse to install on a cluster the chart works on. The chapter also records a decision: this project does not watch Kubernetes platform advisories, because we run no cluster and could only ever close such an issue as the operator’s — kubernetes-announce and the official CVE feed are named as the operator’s duty instead, beside what we do watch.

  • The chart’s rollout strategy is explicit (strategy in values), rather than inherited from the API server’s 25%/25% defaults. maxUnavailable: 0 guarantees served capacity never drops during an upgrade: a new pod must pass readiness before an old one is removed. At the default two replicas this is exactly what Kubernetes already computed, so nothing changes for a default install — it only bites above two replicas, in the safer direction. The trade-off is documented rather than hidden: with no spare scheduling capacity the rollout stalls visibly instead of proceeding at reduced capacity.

  • The Helm chart is published, to GHCR as an OCI artifact. Installing no longer means cloning the repository:

    helm install ferroehr oci://ghcr.io/rubentalstra/charts/ferroehr --version 4.0.0 \
      --set database.existingSecret=ferroehr-db --set image.tag=3.17.3
    

    OCI only — there is no HTTP chart repository, no index.yaml and no chart branch, so a chart version exists in exactly one place and the chart inherits the authentication, retention and access control of the images it deploys. The cost is stated rather than hidden: helm repo add does not apply to this chart, so a client older than Helm 3.8 cannot install it. The chart version and the image tag are separate SemVer lines — --version pins the templates and values schema, --set image.tag= pins the server binary — and the documentation now explains pinning both, which is the thing that is easy to get wrong.

    The chart carries signed keyless Sigstore provenance, verified the same way as the images: gh attestation verify oci://ghcr.io/rubentalstra/charts/ferroehr:4.0.0 -R rubentalstra/FerroEHR. It deliberately ships no PGP .prov, so helm install --verify does not apply: a .prov needs a long-lived private key in CI, which is exactly what this project’s publishing lanes are built to avoid.

    Two properties of the lane are worth knowing because a registry does not provide them. A published chart version is never overwritten — an OCI tag is mutable and helm push replaces one silently, so the lane refuses to push a version that already exists. And a chart that cannot deploy its own default image is not published: the lane renders the chart’s default configuration and runs the appVersion image’s own config check against it, because helm install with no image.tag pins appVersion and a configuration key that image does not know is a crash-loop, not a warning.

    The chart is listed on Artifact Hub with category, links, maintainers, screenshots, and all three published images with their platforms so the hub’s security report covers what the project ships. Its per-release changelog is derived from CHANGELOG.md at package time rather than maintained a second time, since Artifact Hub’s change kinds are Keep a Changelog’s subsections.

  • Bounds on the resources a single request can consume. query.max_result_rows caps an AQL query that neither the query text nor the request bounds (previously such a query returned every matching row, making one request an unbounded allocation); query.timeout_ms is now on by default and db.statement_timeout_ms backs it at the database, because the HTTP request timeout answers the client without cancelling the statement; and [server.connection] bounds the connection itself — an HTTP/1 header-read timeout for slow-header attacks, plus HTTP/2 stream-concurrency and keep-alive bounds, since every other limit engages only after a request head is parsed.

  • [server.tls].min_version — the TLS protocol floor, defaulting to 1.3 only, with "1.2" available as a named compatibility widening. Previously both were enabled unconditionally. TLS 1.1 and 1.0 are not selectable at all.

  • Per-caller request-rate limiting ([server.rate_limit], on by default): a caller past its rate is refused 429 Too Many Requests with Retry-After. Two tiers — an address tier outside authentication, so a flood is refused before the server verifies a signature per request, and a principal tier inside it keyed on the authenticated subject, because a hospital behind one NAT is a single address and address-keying a clinical API would throttle a whole site for one busy client. The defaults sit above this implementation’s own measured whole-server ceiling, so the limiter cannot refuse a caller the server could still have served — that boundary belongs to the existing max_in_flight shed, which answers 503. Three statuses, three meanings: 503 the server is full, 429 you are asking too fast, 413 your payload is too big. Turn it off before benchmarking, or the benchmark measures the limiter.

  • Configurable request-body limits ([server.limits]), replacing a fixed internal 16 MiB ceiling: a clinical tier plus a bulk tier for the routes that accept bulk by design (operational-template upload, /message/import, /message/tdd). Over-limit is 413. The defaults are sized against measured payloads — the largest operational template in the vendored CKM corpus is 5.4 MB — and a deployment whose compositions embed large DV_MULTIMEDIA data raises body_bytes deliberately.

  • Response security headers on every response, transport-layer ones included (Cache-Control: no-store, X-Content-Type-Options, Referrer-Policy, Cross-Origin-Resource-Policy, X-Frame-Options, and a minimal Content-Security-Policy). Strict-Transport-Security is deliberately not sent: RFC 6797 §7.2 requires a browser to ignore it over plain HTTP, which is how this server is commonly reached behind a terminating proxy, so the TLS edge owns it. The documentation site carries its policy as a <meta> element, since GitHub Pages sets no headers.

  • Signed build provenance and SBOM attestations on every artifact published from this release onward. The images already carried BuildKit’s SLSA provenance and an SPDX SBOM but unsigned — readable, not verifiable; they are now signed through Sigstore, and the release binaries gained both from nothing. Verify with gh attestation verify oci://ghcr.io/rubentalstra/ferroehr:TAG -R rubentalstra/FerroEHR. Signing landed in this cycle, so 3.17.3 and earlier tags carry no attestation and that command reports none for them — nothing to verify rather than a verification failure. Each release also attaches a CycloneDX SBOM of the Rust dependency graph with pkg:cargo/… purls, licences, checksums and the direct-versus-transitive dependency edges. The isolation-based SLSA build levels are deliberately not claimed, since the provenance is generated in the job that builds.

  • Container scanning in CI: image vulnerability scanning, Dockerfile lint, and secret plus misconfiguration scanning over the tree. Adjudicated exceptions carry their reasoning — for an unreachable finding in an inherited upstream layer, as a published OpenVEX document under security/vex/.

  • Boot warnings for two deliberate weakenings that are easy to leave switched on: cors_permissive, and authentication enabled on a plaintext listener bound to a routable address.

  • Every import from openehr-its now names the module that defines it. The crate’s nine convenience re-exports are gone, so consumers write opt14::types::CObject and xml::runtime::ToXml rather than borrowing those names from a parent module. Longer at the use site, and it says where a type comes from. Affects anyone depending on the published openehr-its crate; no behaviour, wire format or serialization changes with it.

  • The Keycloak quickstart now demonstrates the authorization it documents. It shipped one user holding every role with RBAC switched off, so every authenticated caller could use every surface and no refusal was possible. The demo realm now carries four identities — an admin, a clinician, a read-only auditor and a user with no roles at all — and the overlay enables RBAC, so a reader can watch a valid credential be refused 403 instead of taking it on trust. The existing ferroehr/ferroehr login is unchanged.

  • probes.exec is removed (breaking, for anyone who set it). It ran ferroehr healthcheck for liveness, readiness and startup, passed no --url, and that flag defaults to the openEHR status document rather than a health endpoint — so readiness never touched the database and a pod with a dead database reported Ready and took clinical traffic. The httpGet probes that were already the default are correct and are now the only path: /health/liveness for liveness and startup, /health/readiness for readiness.

  • The chart now requires Kubernetes 1.36 or newer (kubeVersion: ">=1.36.0-0", breaking; chart 6.0.0). It is a compatibility floor: 1.36 is the release where the newest field the chart renders (hostUsers) became stable, which is what lets every field render unconditionally with no version gates left. Chart.yaml carries the field-to-KEP table that is its evidence.

  • The PodDisruptionBudget sets unhealthyPodEvictionPolicy: AlwaysAllow, the documented recommendation. With the previous default a node drain waited for pods to become healthy that were unhealthy because of the drain, so it never completed. It and the preStop sleep action are now rendered unconditionally rather than version-gated, since both are stable below the new floor.

  • The chart’s render gate covers the admin console, as a fourth overlay with its own golden render — the per-container restricted-profile check exists precisely because a second workload is where a posture gets lost, so it has to see one. Two assertions ride along: pod isolation must be identical across every workload of a release, and a multi-replica Deployment must carry a spread or affinity rule.

  • The chart’s pinned Helm version moves to 4.2.3 (current release), with the golden renders regenerated on it.

  • Metric names change: /management/prometheus now derives its suffixes. The server had two metrics systems and now has one — OpenTelemetry, feeding both the Prometheus pull surface and the OTLP push from a single meter provider, so a metric can no longer exist on one surface and not the other. Counters are unaffected on the wire (compositions_committed_total and friends render exactly as before). Histograms and gauges lose their hand-written _seconds/_bytes suffix from the instrument name and gain it from the declared unit instead — the rendered name is the same in almost every case, but check any dashboard that pinned a name literally. Bucket boundaries are unchanged, deliberately and test-pinned: a re-bucketed latency histogram would silently invalidate every alert built on it.

  • telemetry.metrics_push now exports every family the Prometheus surface exposes. It previously carried four of ten, missing the build identity, the request-duration histogram and the ATNA audit counters, and the boot warning added earlier in this cycle to disclose that is removed as no longer true.

  • deploy/helm/validate.sh now ends every run by stating the properties it does not check — that the server accepts the render, that the selected image understands its keys, that a pod starts and serves — each with the command that does check it. A green render was being read as a working deployment. (#2159)

  • The Kubernetes, Compose, Operations, Audit, Security, From-source and configuration-reference pages now state both migration postures plainly — what the quickstart does and why, and what a least-privilege deployment does instead — and document the audit trail’s tamper evidence and its verification query. The [db] reference table gained migrate and the previously undocumented statement_timeout_ms.

  • Bearer-token refusals are now classified: the authentication layer records why a token was rejected — expired, not yet valid, bad signature, wrong issuer, wrong audience, missing claim, malformed, unusable key material and twelve more — instead of collapsing every one into a single opaque string. Each refusal reaches the log as a stable reason field alongside the authentication mechanism, so a burst of expired tokens (clock skew, or a token lifetime that is too short) is countable and no longer hides a burst of bad signatures (someone presenting tokens this server was never meant to accept). The underlying jsonwebtoken error is carried as the error’s source, so the cause chain can be walked instead of grepped.

    No wire change: RFC 6750 §3.1 assigns one invalid_token code to a token that “is expired, revoked, malformed, or invalid for other reasons”, so every refusal keeps answering 401 with the identical WWW-Authenticate challenge it did before — now asserted for every refusal kind.

  • openehr-query: a syntax failure now locates itself in the source. parser::SyntaxFault carries bytes: Option<Range<usize>> — the byte range of the offending text — alongside the existing token indices, so a caller can underline the offending characters or derive a line and column instead of mapping a token index back through the stream itself. parser::parse_str fills it in on every failure; parser::parse over a bare token slice reports None, because a bare slice carries no source positions. The rendered Display — and therefore the body of an AQL 400 — is unchanged (#2145).

  • The AQL parser reports a typed error instead of a string. openehr_query::parser::parse_str and parse now return ParseError, which separates a lexing refusal — carrying the lexer’s own error as a real source(), so the cause chain is walkable — from a grammar refusal, which carries every position the parser reported along with the token found there. A caller can finally branch on which pass refused without matching a substring. The message text is unchanged, so the 400 body for an invalid query reads exactly as before and no wire behaviour moves. The published crate also no longer exposes chumsky’s error type in its public API.

  • The Kubernetes hardening chapter’s image-provenance admission policies now carry the certificate identity the publishing lanes actually issue — read off a published image rather than inferred from the workflow files — and both are corrected to the form that can read a GitHub Artifact Attestation at all. Kyverno needs type: SigstoreBundle with an attestations: block (the field defaults to Cosign, which looks for a detached signature these images do not carry and so refuses every one of them), and sigstore-policy-controller needs signatureFormat: bundle with an attestations: entry. Minimum versions are stated, failureAction moves off the deprecated spec-level field, and each policy names the ref set it admits: as published they accept released images only and deliberately refuse the develop tag, with the one-word change a staging cluster needs spelled out.

  • Provenance verification commands across the security and Kubernetes chapters now point at artifacts that actually answer. gh attestation verify on the three :develop images succeeds today, and --signer-workflow is shown as the way to require the specific publishing lane rather than any workflow in the repository. The release-tag, release-binary and chart forms are marked as starting at the 3.17.4 cut instead of being demonstrated against a tag that returns HTTP 404.

  • Release binaries reach SLSA v1.0 Build Level 3, and carry provenance a scanner and an offline verifier can both read. Build L3’s distinguishing requirement is that signing material must be out of reach of the build steps — and every step of a GitHub Actions job shares one runner VM, so attesting inside the building job cannot satisfy it. The release build now happens inside a reusable workflow: it runs on its own VM, a caller passes declared inputs and cannot add steps, and the caller job has no steps at all. You can now require that signer rather than trusting any workflow in the repository: gh attestation verify <tarball> -R rubentalstra/FerroEHR --signer-workflow rubentalstra/FerroEHR/.github/workflows/release-build.yml. Each release also carries its provenance as <tarball>.intoto.jsonl beside the Sigstore bundles. The container images and the Helm chart remain Build L2 and now say so where they are built, rather than leaving the level to be assumed.

  • A direct push to develop is refused. The branch ruleset already blocked deletion and force-pushes and required signed commits, but not a pull request — so the discipline every change has followed was convention rather than enforcement, and the main ruleset had the rule all along. Requiring zero approvals, so it changes nothing about how a maintainer merges their own work.

  • The conformance baseline moves to 1014 of 1014, and three new cases pin behaviour that had none. The run that produced it went red first, with five cases failing because the two shared-definition deletes now require the admin role. Every red row was attributed before anything was touched, and the attribution landed on the catalogue: those five pin the semantics of a delete — none of them declares a forbidden outcome — so they were calling an operation their principal is not authorized for and never reaching it. They now drive the delete as admin, with their 204/404 expectations untouched. The refusal the authorization change introduced gets its own two cases, so the behaviour is tested rather than merely described, and a read-only case that had quietly stopped isolating anything — its subject route became admin-gated, and it was the only write on that surface — is joined by one on a route where the restriction is still observable. The comparison record for the upstream EHRbase server is regenerated from the same run of the same catalogue.

  • Server-fault log records now carry the underlying failure that caused them. A 500-class fault writes a cause field holding the full error chain — the PostgreSQL driver error, codec refusal or HTTP transport failure, and whatever caused that — instead of a single flattened sentence. Response bodies are unchanged: every status code and every message a client receives is byte-identical, and the cause never appears in one. ABAC attribute resolution and policy-engine failures additionally name which step failed (resolve the EHR subject attribute, resolve the template attribute, reach an authorization decision) rather than one generic authorization label.

  • The published conformance badges are derived by the conformance runner, from the same rule as the verdicts they sit beside. They were re-derived downstream by a separate ~90-line implementation of the tier semantics, which is how a badge once read FAIL 5/5 capabilities: the count was tier-local while the verdict was cumulative, and both were individually right. A count now quantifies over the very capability set its verdict was judged on, so that contradiction is no longer expressible and the self-check that used to catch it is gone rather than ported. Every committed badge regenerates byte-identically, apart from badge.json, where a middot is now a literal UTF-8 character instead of a · escape — the same string, differently spelled.

  • Deleting a shared definition artefact now requires the admin role, whichever route you use. DELETE /definition/archetype/adl1.4/{archetype_id} and DELETE /definition/artefact/adl2/{artefact_id} were clinical-class, so any authenticated principal could remove an archetype every EHR in the deployment validates against — while the neighbouring DELETE /admin/template/{template_id}, with exactly the same blast radius, required admin. The difference came from which path prefix a route happened to be given, not from a decision. This tightens the wire: a client deleting archetypes with a clinical role now gets 403. Uploads are unchanged and stay clinical, as they always were. The specification decides that the three must match rather than which class they take — the SM puts removal of archetypes and templates in one clause and one pair of interfaces, and its Admin component names no definition artefact at all — so the choice of admin is ours, following the one privilege recommendation in the conformance text for irreversible deletion.

  • The database DSN, the two AMQP broker URLs and a Basic user’s password hash are now delivered as mounted files instead of environment values. An environment variable is readable through /proc/<pid>/environ and is inherited by every child process, which is why the OWASP Kubernetes Security Cheat Sheet asks for a read-only volume instead — and the DSN is the credential that reaches patient data. All four now arrive at /etc/ferroehr-secrets/… (mode 0440, owned root:65532, not a subPath mount so a rotated Secret still propagates) with only the path passed as environment. A DSN supplied through database.existingSecret is projected from your own Secret, so nothing about how you create it changes. audit.fhir_feed.url and multimedia.access_key_id are the only credential-bearing keys still passed as values, because they are the only ones without a *_file sibling.

    What an upgrader must change: a Basic user’s hash moves from config.auth.basic.users[].password_hash to secrets.basicUserPasswordHashes.<username>, leaving only username and roles under config:. Setting the hash under config: is now refused, and the error names the key that carries it. Nothing else in your values file changes.

    A consequence worth knowing: configuring a Basic user no longer moves the whole rendered configuration into a Secret. That behaviour existed only because the hash had nowhere else to go; the ConfigMap is now used in every supported configuration, and the “no route” fallback has no reachable input — it is kept for the next secret key that arrives without a file sibling.

  • The chart requires a server build that understands the *_file keys above. A chart from this cycle deployed against 3.17.3 or earlier refuses to boot with unknown configuration key url_file. appVersion moves with the release, so a published chart and its default image always agree; a chart taken from a development checkout and pinned to an older image does not.

  • The Helm chart is version 4.0.0, and the major bump is a real break in its values contract. The chart version is SemVer over the chart’s own contract and is independent of the application version (appVersion still names the release). .Values.config is no longer rendered verbatim: a secret-bearing key under config: either fails the render naming the secrets: key that carries it, or moves the whole configuration into the Secret. What an upgrader must change: move config.audit.fhir_feed.url to secrets.auditFhirFeedUrl, move any config.terminology.external.oauth2_clients.<name>.client_secret to secrets.terminologyOauth2ClientSecrets.<name> (declaring the client’s token_url/client_id under config: as before), and move a DSN, AMQP URL, OIDC HMAC secret, signing passphrase or S3 secret key set under config: to its existing secrets: key. helm template names every offending key at once, so a dry run is the migration checklist. Values files that set none of these render byte-identically apart from the chart version label. A config.auth.basic.users entry keeps working and needs no change, but the release now carries its configuration in a Secret rather than a ConfigMap — automation that reads kubectl get configmap <release> must read the Secret instead.

  • The chart’s golden-render gate runs in CI. deploy/helm/validate.sh — the only check in the repository that pins the pod security context, the resource limits, the probes and the NetworkPolicy — ran in no workflow, so the golden renders went stale on develop and nothing failed. A helm-golden job now runs it (plus helm lint) on every pull request touching deploy/helm/**, inside the aggregate gate rather than beside it. The helm version is pinned in deploy/helm/.tool-versions and asserted by the validator, because helm template output is not byte-stable across helm releases.

  • A missing mandatory key in a terminology-provider, terminology-OAuth2-client or subject-proxy-system table is now reported by the boot-validation pass instead of by the file parser. Those sections read every default from one place, so a [terminology.external.providers.<name>] table with no url, an [terminology.external.oauth2_clients.<name>] with no token_url/client_id, or a [subject_proxy.systems.<name>] with no base_url still refuses to boot and still names the key — but it arrives with the rest of the configuration errors at once, with did-you-mean, rather than as a bare “missing field” that stops at the first one.

  • Chart secrets are mounted as read-only files instead of environment values wherever the configuration tree allows it. An environment variable is readable through /proc/<pid>/environ and is inherited by every child process, which the OWASP Kubernetes Security Cheat Sheet asks deployments to avoid. The OIDC HMAC secret, the signing passphrase and the S3 secret access key are now mounted at /etc/ferroehr-secrets/… (mode 0440, owned root:65532, not a subPath mount, so a rotated Secret still propagates) and only the path is passed as environment through the existing *_file configuration keys. The database DSN, the two AMQP URLs and the S3 access key id continue to travel as environment values, because those keys have no *_file sibling to point at a file; the chart says which is which at each value.

  • The chart renders a Prometheus Operator ServiceMonitor under metrics.serviceMonitor.enabled. The existing prometheus.io/* pod annotations are honoured only by a Prometheus that does its own Kubernetes service discovery; an operator-managed Prometheus ignores them entirely, so on a kube-prometheus-stack cluster the chart previously published no discoverable target at all.

  • The Compose deployment path is hardened to match the Helm chart. The Kubernetes path already ran non-root with a read-only root filesystem, no privilege escalation, a RuntimeDefault seccomp profile and every capability dropped; the quickstart Compose file — the one an evaluator downloads and runs — set none of the equivalents. It now does: every service drops all capabilities and adds back only what its entrypoint provably needs (five for PostgreSQL, two for the S3 gateway, none at all for the two distroless images), refuses privilege escalation, bounds its file descriptors, restarts unless stopped, and PostgreSQL and the admin console run read-only roots.

  • Every published port binds 127.0.0.1 by default. A published container port is DNAT’d ahead of the host firewall’s own chains, so ufw deny 5432 does not stop it — the quickstart previously exposed PostgreSQL, with the dev credentials that ship in the same file, on every interface. Set FERROEHR_BIND_HOST=0.0.0.0 (or a specific address) when a deployment genuinely needs to be reachable from another machine.

  • The SMART App Launch discovery document is now boot-validated, not relayed. /.well-known/smart-configuration is published to third-party applications, so an empty or plaintext value is a claim they act on. With smart.enabled = true: every advertised endpoint must be an absolute https URL (new smart.endpoints.allow_insecure_endpoints opts out for development, mirroring auth.oidc.allow_insecure_issuer); the advertised issuer must carry no query or fragment (RFC 8414 §2 — the same rule auth.oidc.issuer follows, since it is the same identity); response_types_supported must be non-empty (RFC 8414 §2 marks it REQUIRED); token_endpoint_auth_methods_supported must name at least one method; and code_challenge_methods_supported must include S256, because SMART App Launch requires PKCE (RFC 7636) and publishing a list without it tells every app the authorization server cannot do it. A deployment relying on any of those being empty now fails to boot with the key named.

  • smart.endpoints.issuer and auth.oidc.issuer must agree. One tells applications where to obtain a token; the other is what this server accepts. Configured independently and never compared, a mismatch was silently broken in the most confusing way available — every app obtained a valid token and every request was refused as invalid. Both are known at boot, so a mismatch (and smart.enabled with no [auth.oidc] at all) is now a boot error naming both keys. A trailing slash is not a mismatch.

  • Authentication and authorization rewritten against the OAuth2/JOSE RFCs (the wire-visible half; the boot-validation half is under Added/Fixed). What an integrator sees change:

    • A malformed Authorization header — unparsable, an unknown scheme, or a bearer credential outside the RFC 6750 §2.1 b64token grammar — now answers 400 instead of 401. The old answer said “your credential was rejected” about a request the server never read a credential from.
    • The WWW-Authenticate challenge now names the outcome: error="invalid_token" for a rejected credential, error="insufficient_scope" on a bearer 403, error="invalid_request" for a malformed header, and no error code at all when the request carried no credentials (RFC 6750 §3.1 asks for exactly that). The Basic challenge advertises charset="UTF-8", and a bearer 403 now carries a challenge too.
    • An unreachable token issuer (the JWKS cannot be fetched) answers 503 with Retry-After instead of 401. A client holding a valid token was previously told its credential was bad, which a well-behaved client responds to by not retrying.
    • A Basic credential must be padded base64 (RFC 7617 §2 → RFC 4648 §3.2). The unpadded form was accepted by a decoder setting no specification requires.
    • An external policy server answering 5xx is now a fail-closed 500, not a deny. A broken PDP could previously refuse clinical access indefinitely while looking like policy.
    • Bearer tokens are validated more strictly: nbf is honoured, iss/aud/sub are required alongside exp, an unexpected typ is refused, and a token claiming the RFC 9068 at+jwt profile must carry iat, jti and client_id. An audience-less configuration is no longer possible, so a token minted for another resource server — or an OIDC ID token — can no longer authenticate.
  • scope no longer grants roles — read this if you use OIDC. Roles were mined from realm_access.roles AND the OAuth2 scope claim. A scope is a grant of delegated authority to a client (RFC 6749 §3.3), not an assertion about the subject, and treating it as one made the “at least one role” check pass for every OIDC token, because openid counted as a role. Roles now come from the RFC 9068 §2.2.3.1 carriers — roles, groups, entitlements, then realm_access.roles — configurable as before via authz.rbac.role_claims. A deployment whose callers relied on a scope naming a role must move that role into a role claim; scopes remain on the principal for SMART enforcement.

  • New auth.oidc.require_at_jwt (default false): refuse a token that does not claim the RFC 9068 access-token profile. Off by default because §2.1 makes the type a SHOULD for the authorization server, so requiring it would reject conforming issuers.

  • Authentication and authorization configuration is now validated at startup, and six previously-accepted configurations refuse to boot. Each one could only have produced a silently weaker deployment than it appeared to describe, so the server names the offending key and stops instead. What fails, and the fix:

    • auth.enabled = true with no mechanism configured (no [[auth.basic.users]], no [auth.oidc]). Such a server can only refuse every request while advertising an authentication scheme it does not implement, which RFC 9110 §11.6.1 forbids. Fix: configure a mechanism, or set auth.enabled = false for a development server. (This was a boot warning during the current development cycle and never shipped as one.)
    • [auth.oidc] with empty audiences. This disabled audience checking altogether, so the server accepted access tokens minted for a different resource server, and could not tell an OpenID Connect ID token (whose aud is a client id) from an access token — RFC 7519 §4.1.3, RFC 9068 §4 step 4, RFC 8725 §3.9/§3.12. Fix: set audiences to the aud value your identity provider issues for this CDR. There is no opt-out.
    • [auth.oidc] issuer that is blank, not a URL, not https, or carries a query or fragment. RFC 8414 §2 defines the issuer identifier as an https URL with no query or fragment, and §6.2 requires TLS — over plain HTTP an attacker can serve their own signing keys. Fix: correct the issuer; for a development identity provider on plain HTTP, set allow_insecure_issuer = true (the no-query/no-fragment rules still apply).
    • [auth.oidc] hmac_secret shorter than 32 bytes. RFC 8725 §3.5: a human-memorizable password must not be used directly as the key to a keyed-MAC algorithm such as HS256. Fix: supply at least 32 bytes of high-entropy material, or move to a JWKS/discovery key source. Configuring any symmetric secret now also logs a boot warning that it is a development posture: the key is shared with the authorization server, so the CDR can mint the tokens it accepts.
    • [auth.oidc] clock_skew_leeway_seconds above 300.
    • any [[auth.basic.users]] password_hash below the OWASP Argon2id floor (m>=19456, t>=2, p>=1, algorithm argon2id), or one that is unparsable, or an entry with a blank username. The password verifier takes its cost parameters from the stored hash, so a deliberately cheap hash verified happily and silently weakened that account. Fix: re-hash the affected passwords at the floor or above. The bundled development and quickstart hashes were regenerated accordingly (same passwords).
    • With abac.engine = remote, a missing policy entry for any resource kind the enforcement point consults (ehr, ehr_status, composition, contribution, query, plus directory when check_directory = true). Previously an unconfigured kind was silently permitted, i.e. a whole class of requests bypassed the decision point. Fix: add the missing [authz.abac.policy.<kind>] entries. At runtime an unconfigured kind is now denied and logged, never permitted.
  • The conformance comparison lane is named for the system it composes: CONF_SUT=ehrbase (was ehrbase-java), with its Compose file, party set and committed artifacts under the same name, and the two environment knobs FERROEHR_EHRBASE_IMAGE / FERROEHR_EHRBASE_PORT (was FERROEHR_JAVA_*). Committed measurement records keep the container names and topology they observed — a rename never restamps a measured run.

  • The openehr-its Web Template builder for ADL2 sources is renamed to match the generation modules: build_web_template_am24 becomes build_web_template_v2_4 (the am14/am24 module names became v1_4/v2_4 in the multi-generation refactor, and the retired spelling survived in this API, two file names and their prose). A crate-API change with no behaviour change.

  • Snapshot tests now resolve their workspace root from a pinned INSTA_WORKSPACE_ROOT rather than a runtime lookup, so cargo-lock contention can no longer make stored snapshots appear missing or write pending snapshots to a doubled path.

  • Dependency updates: base64 0.23, fancy-regex 0.19, jsonschema 0.49 and jsonwebtoken 11. No behaviour change — each was checked against its upstream changelog and then against this workspace’s own call sites, including that jsonschema’s new single-error validate signature does not reduce our error reporting (both call sites already collect every error).

  • The documentation site’s search metadata is produced by a small Rust tool (tools/docs-meta) rather than a Python script — the docs workflow is Rust-only by design, and this repository ships no Python.

  • The bearer-token command in the documentation uses jq instead of python3. The quickstart told you to pipe a Keycloak token response through a Python one-liner, which is an interpreter you may not have on a machine running a pure-Rust server. The site-build and corpus-vendoring scripts lost their embedded Python for the same reason — a vendoring script must still run years from now, and a hidden python3 requirement is exactly what stops it. Their output is unchanged: both regenerated corpus provenance files were checked byte-for-byte against the committed ones, including two archetype display names that genuinely end in a space.

  • Search-result presentation on the documentation site. Every documentation page now carries a canonical URL (so the dev, latest and frozen-release copies of a page no longer compete in the index), its own meta description taken from the page’s opening prose (they previously all shared one generic sentence, which is why search snippets repeated), Open Graph and Twitter card tags with the brand social card, noindex on non-canonical versions, and BreadcrumbList structured data mirroring the URL hierarchy. The landing page gains WebSite, Organization and SoftwareSourceCode structured data naming the site’s top-level sections.

  • The documentation version picker works again. It hardcoded the site base path /ferroehr, left over from before the move to the ferroehr.eu apex domain, so it requested /ferroehr/versions.json, got a 404, and could never populate the selector — the docs appeared to have no other versions at all. The base is now derived from the page’s own location, so it is correct at a domain root and under a project path alike, and cannot go stale on the next move. The published version set is also pruned to the current release: the manifest had 28 entries, 22 of them unreachable because their paths carried the base in use when they were cut (and their frozen books still carry the pre-rename product name).

  • The configuration guide now documents spec_profile in full — what it selects, why it is one coupled key rather than three, where to see the active profile, what changes on the wire in both directions, and the asymmetric change contract — and the “Migrating from 3.x environment variables” section is gone: it mapped spellings from a layout that was never released, so it documented a migration nobody can need. The variables that genuinely are not configuration keys (the config-file pointer, the healthcheck URL, build stamps, Compose parameterization) are documented as such instead.

  • The eight published openehr-* crates’ crates.io front pages (their README.md and one-line descriptions) now describe the crates as they actually are: the generation modules they expose and which is current, the Generation enum as the way to ask which openEHR version a generation implements, and — for the five generated crates — that no SPEC_VERSION constant exists (every one they advertised was removed by the multi-generation refactor). Also corrected: the retired am14/am24 module names, openehr-lang’s conflation of specification units with generations, and three capability claims that were simply untrue (a feature that gates nothing, a parser library the crate does not depend on, and sibling behaviour files that do not exist).

  • The project README now documents the two selectable openEHR specification generations (the spec_profile key, with the change contract in both directions) and the eight openehr-* specification crates published on crates.io — neither was mentioned before, and the architecture diagram, CI-gate list and roadmap link are corrected to current reality.

  • The vendored openEHR specification tree no longer carries upstream’s own dependency manifests (the CNF Robot harness’s requirements.txt, the ITS-REST and TERM PHP tooling’s composer.json). They are upstream build tooling, not specification text, and their presence made this repository’s dependency graph claim Python and PHP ecosystems it does not use — raising advisories against pinned third-party versions nothing here installs. The vendoring script now excludes dependency manifests across the common ecosystems, so a future re-vendor cannot reintroduce them; the vendored spec text itself is unchanged.

  • Hand-written spec-behaviour files that were byte-identical across the openehr-rm/openehr-base generation modules (89 twin families, ~29k lines of duplicated source) are now generation-twin TEMPLATES: one hand-written source under the code generator’s templates/ tree, with the per-generation copies stamped by emit under an @generated-from-template header — generation divergence becomes impossible instead of policed, and an emitter invariant refuses any new unconverted twin. The one genuinely generation-specific file (ITEM_TAG’s construction under RM 1.1.0’s field order) is an explicit per-generation override carrying its adjudication. Crate behaviour is unchanged; the packaged sources now say which file is stamped from which template.

  • The Docker Compose quickstart is now a standalone, zero-configuration file. docker-compose.yml pulls the published images (pinned to the release it shipped with) instead of building, and carries the server configuration inline, so docker compose up works in any empty directory — no repository checkout, no bind mounts, no environment variables. It needs Docker Compose 2.23.1 or newer.

  • The admin console moved behind the admin-ui Compose profile: it is opt-in (docker compose --profile admin-ui up, then http://localhost:3000) rather than started by every docker compose up. As a side effect, the observability overlay’s Grafana no longer collides with it on port 3000.

  • The quickstart’s authentication posture is now Basic auth with a single user (ferroehr / ferroehr, holding the ADMIN and USER roles) and role-based access control disabled, so every enabled surface works out of the box; the admin API, the management endpoints, and permissive CORS stay on. The previous three-user, RBAC-enabled development posture moved to the repository development override. Both are development defaults — see the security chapter for turning RBAC on.

  • The conformance stack composes its own self-contained docker/sut-ferroehr.yml instead of overlaying the root Compose file, so the end-user quickstart and the conformance instrument evolve independently. Starting the development FHIR terminology server is now docker compose -p ferroehr-cnf --project-directory . --profile terminology -f docker/sut-ferroehr.yml -f docker/sut-terminology.yml up -d --wait ferroehr from a repository checkout.

  • openehr-lang now models generations by COMPONENT VERSION like every other spec crate: one openehr_lang::v1_1 generation holding the version’s published specifications side by side — the STABLE, tool-implemented BMM v2.x model (bmm/bmm_persistence/beom, on the prelude) beside the PAUSED BMM3 model (bmm3, full-path only) — together with the hand-written ODIN/BEL/EL readers and the shared lexer for that version’s notations (all previously crate-root modules). The released LANG 1.0.0 machine-readable BMM is recorded as unusable for code generation (no includes, unnamed BMM_CLASS/BMM_PACKAGE, an explicitly obsolete package; BMM is TRIAL in that release), so no 1.0.0 generation exists until upstream republishes usable artifacts.

  • The published openehr-* spec crates now expose every BMM generation under a version-named top module — openehr_base::v1_3, openehr_rm::v1_2, openehr_lang::v2/v3 (replacing bmm/bmm_persistence/beom and bmm3 at the crate root), openehr_am::v1_4/v2_4 (replacing am14/am24), openehr_term::v3_1 — with a new per-crate Generation enum (derived Default marking the current generation, spec_version()/as_str(), FromStr/Display) and the crate prelude re-exporting the current generation only. Import paths into these crates change accordingly; the served wire formats are unchanged.

Removed

  • authz.rbac.management_access and management.access_default are gone (breaking, for a configuration that set either). Both were inert: the management surface is gated per endpoint, and neither key was ever consulted by the code that gates it. That was the dangerous kind of dead setting — someone setting management_access = "admin_only" to lock the surface down had not locked anything down, and the security chapter presented it as the gate. [management.endpoints] is now the single authority, one level per endpoint, with no global default beside it: an endpoint you do not name is off and is not mounted. A test pins that the levels are independent, so no endpoint’s level can leak onto its neighbour.

    If you set either key, delete it — the server refuses unknown configuration keys at boot, so this fails loudly rather than silently. The behaviour you had is unchanged, because neither key did anything.

  • openehr-rm’s ehr-extract cargo feature, which gated nothing (the EHR Extract modules were always compiled), and openehr-its’s empty bmm placeholder module — a published module that promised future surface nobody could use.

  • The generated openehr-* spec crates no longer carry a crate-level SPEC_VERSION constant: a multi-generation crate has no single implemented spec version, and a fixed crate-root pin would contradict a configured non-current generation. The ONLY pin authority is the emitted Generation enum (per-variant const fn spec_version(); the derived Default variant is the current generation) — the generation modules carry no version constant either; the hand-written single-spec crates (openehr-its, openehr-query, openehr-adl) keep their literal constant.

Fixed

  • A TERMINOLOGY_ID is no longer refused for its lexical form. The CDR rejected composition values such as snomed_ct(3.1) and SNOMED CT with Invariant Value_valid failed on type TERMINOLOGY_ID — an invariant the released BASE model does not declare anywhere: the class table carries no Invariants row and the machine-readable model no invariants key. Released QUERY 1.1.0 settles it from the other side by publishing terminology_id/value='snomed_ct(3.1)' in its own normative example, which the enforced production forbade. Values that were being rejected at commit are now accepted.

  • Five CI gates ran without gating anything. Branch protection routes through a single aggregate check, which is what lets jobs be added or renamed without editing repository settings — but a job missing from that check’s dependency list still runs, still goes red, and still merges. The compose hardening guard, the error cause-chain ratchet, the no-Python rule, the spec-citation check and the chart boot check were all in that state. All are wired in, and a new guard fails the build when a job is missing from the list in either direction, so the next one cannot slip through silently. The compose guard is the one worth calling out: docker-compose.yml and the operations chapter both told readers it enforced on every compose artifact, and it had never run.

  • The forbidden-licence check now actually runs. It was configured as a FOSSA customLicenseSearch, which could not upload on this organization’s plan and did not honour the configured path exclusions — every one of its findings came from vendored third-party trees rather than from this project’s own code. It is now a local check that shares the same exclusion list and fails the build.

  • An uploaded operational template no longer silently loses its slot constraints. A slot assertion was rendered into ADL by writing the raw numeric operator code the OPT XML encodes — text the assertion parser cannot read — so any slot whose template did not also carry a pre-rendered string expression ended up with no constraint at all, and would admit archetypes it was written to exclude. The operator is now rendered as its ADL symbol, and a test reads the result back through the parser.

  • Temporal constraint conformance answers where it used to give up. A child archetype narrowing a date, time or duration to a range under a parent that constrains it by pattern reported “cannot tell” instead of judging it, on the stated grounds that the generated date/time types offered no ordering. They have offered ordering for some time, so the check now reaches a verdict.

  • An operational template whose archetype id carries an ADL 2 version is accepted again. Uploading an OPT whose root id is openEHR-EHR-COMPOSITION.minimal.v1.0.0 was refused as a malformed archetype identifier (VARID), even though deployed OPT 1.4 exports carry exactly that form and the validator documented the tolerance. The check had been routed through a grammar whose version production is single-part by definition, so the tolerance it described could never fire.

  • AQL archetype subsumption works again for ADL 2 identifiers. A query naming a parent archetype is supposed to match data created with its specialisations. For an AOM2-era identifier — openEHR-EHR-OBSERVATION .lipid_panel.v1.0.0, the form an ADL 2 archetype actually carries — it did not: the predicate read the identifier through a decomposition that accepts only the single-part .v1 version, silently fell back to matching the whole identifier string, and left the index built for the subsumption query unused. The answer was narrower than the spec requires, with no error. Both eras’ forms are now read through the same decomposition the lineage index is built from.

  • TERMINOLOGY_ID now enforces its grammar. A value was accepted if it was merely non-empty and printable, so "SNOMED CT" and "http://snomed.info/sct" passed. The identifier is now the production BASE base_types master05 §Syntaxes states. The version part in parentheses admits a leading digit, because every example the same chapter gives (ICD9(1999), ICD10AM(3rd_ed)) starts with one — reported upstream, since the grammar as written refuses openEHR’s own identifiers.

  • A Web Template no longer reports a query string as a terminology. A C_CODE_REFERENCE whose reference-set URI carries parameters — terminology:NSI?subset=doctor_category&language=en-GB — had the entire remainder copied into the input’s terminology field, which is a terminology identifier. Only the identifying part is taken now, for both the bare form the conformance templates use and the addressed form AQL documents.

  • The release lane would have shipped a release with no binaries. GitHub release immutability is now enabled, which freezes a release’s assets the moment it is published — and the binaries, SBOMs and Sigstore bundles are attached by a job that runs after the release is created. The release is now created as a draft, its assets are attached, the expected set is verified, and only then is it published, so nothing can be locked out.

  • Three list-typed configuration settings could not be set from the environment at all. signing.retired_key_paths, smart.endpoints.capabilities and every authz.abac.policy.<kind>.parameters are list-typed but were missing from the loader’s list registry, so an env value was not merely mis-split — it was refused at boot: invalid type: string "…", expected a sequence. All are now registered, and tests assert that every env-settable list key parses from one value, splits on several, and that a list nested under a map key works too.

    signing.retired_key_paths is the key that carries PGP key rotation: it holds the retired PUBLIC keys that keep versions signed before a rotation verifiable. Until now it was reachable only from a TOML file, so any environment-driven deployment — Compose, plain Docker, a chart’s extraEnv — could not configure key rotation at all, and found out only when the server refused to start.

  • The fuzzing lane had not run at all. Every one of the six libFuzzer harnesses failed to build on every scheduled run: cargo-fuzz defaults the build target to the triple it was itself built for, CI installs its musl asset, and musl’s static libc cannot carry a sanitizer — so the campaign died before executing a single input. The target triple is now named explicitly, and the harnesses are compiled on the pull-request path whenever they or a fuzzed crate change, so a scheduled-only lane can no longer rot unnoticed.

  • Multimedia committed inside an EHR_STATUS or a FOLDER can now be read back. Externalization is applied by the versioning path every versioned object commits through, but re-inlining was wired to the COMPOSITION read alone — so a large DV_MULTIMEDIA committed in an EHR status left the database, sat in the object store, and no API call returned it; ?expand_multimedia=true on that read was an undeclared parameter and was ignored. All nine reads that can serve externalized content now honour it (the bare COMPOSITION / EHR_STATUS / FOLDER reads and the VERSION envelopes that wrap them), and the OpenAPI document declares it on exactly those operations — pinned by a test, so the declarations and the handlers cannot drift apart again.

  • Turning multimedia externalization off no longer strands the blobs already offloaded. ?expand_multimedia=true against an already-externalized record was silently ignored once multimedia.enabled went back to false: the read answered 200 with the compact s3:// reference and no indication that the expansion had not happened. enabled now governs new offloads only — the fetch-and-verify path stays available as long as an endpoint is configured, so content this server externalized stays readable. With no store reachable at all, an expansion request now fails instead of quietly answering with the reference.

  • FERROEHR__MULTIMEDIA__* set in your shell now reaches the compose stack. Every other tunable the quickstart documents is a pass-through; the multimedia keys were not, so the documented S3 recipe brought the stack up healthy with multimedia silently off. The whole FERROEHR__MULTIMEDIA__* set is now declared on the ferroehr service with no value, which is the Compose form that forwards a variable when it is set and removes it entirely when it is not — so an unset key stays absent rather than arriving as an empty string, which for endpoint is the difference between “use the default” and a boot refusal.

  • The development Keycloak realm now mints tokens the server accepts. The realm declared no audience mapper, so Keycloak emitted no ferroehr audience, while the same development configuration requires that audience — every bearer token it issued was refused 401 InvalidAudience, and no user of that realm could ever produce an accepted one. The server was right in both halves (RFC 7519 §4.1.3); the realm was wrong. An oidc-audience-mapper on the ferroehr client fixes it, verified live: the access token now carries aud: ferroehr. The realm also carried identifiers from the deployment it was exported from — a hardcoded tenant claim naming a foreign UUID, a foreign client scope, and a default role named after another product — all removed.

  • migrations.runByMigratorRole now does what its documentation says. The key was described as “rendered into NOTES for the operator” and no template read it, so the chart, its generated README and the book all described a marker that did nothing. It now surfaces at install time when set to false, stating plainly that the chart cannot verify the claim.

  • A stock helm install no longer appears to succeed and then crash-loops. config.auth.enabled defaults to true and the server requires a mechanism with it (RFC 9110 §11.6.1 — a 401 challenge must name a scheme the server implements), so the default values produced a pod that exited at boot with the reason visible only in its log. The chart now refuses to render that combination, naming both ways to configure authentication and the explicit development-only opt-out. The default was deliberately not changed to disable authentication: that would hand a fresh install an unauthenticated CDR serving patient data, which is a worse outcome than a loud failure.

  • The CDR Service no longer load-balances openEHR traffic into the admin console. A Service selector is a subset match, and the console’s pods carried the server’s app.kubernetes.io/name plus a component label — so they matched the server’s own Service, whose targetPort: http then resolved to the console’s port 3000. With the console enabled, a share of API requests were answered by a web UI. The same overlap inflated the PodDisruptionBudget’s disruptionsAllowed from 1 to 2, enough for a node drain to evict both server replicas at once. The console now carries its own app.kubernetes.io/name. Verified on a live cluster: the CDR Service’s endpoints are exactly the two server pods and the PDB reports expectedPods=2.

  • An Ingress with a path but no pathType is no longer rejected by the API server. pathType is required on networking.k8s.io/v1 but optional in the chart’s schema; it now defaults to Prefix.

  • A Service with custom annotations no longer loses its description. The template emitted two annotations mappings under one metadata, so the later one won and kubernetes.io/description silently disappeared.

  • An autoscaler with every metric disabled is refused instead of installed inert. Setting both HPA targets to 0 rendered an HPA with no metrics, which the API accepts and the controller can never act on — while the Deployment omits replicas under autoscaling, so a fresh install silently ran a single replica.

  • The install notes no longer advertise a Prometheus endpoint that is off. The endpoint was reported as public whenever metrics were enabled, ignoring config.management.endpoints.prometheus, which defaults to off.

  • The chart no longer promises drain-safety it cannot deliver on older clusters. PodDisruptionBudget.unhealthyPodEvictionPolicy was rendered unconditionally under a kubeVersion floor of >=1.25.0-0, but the field is alpha in 1.26, beta in 1.27 and stable only in 1.31 (KEP-3017). Below 1.27 the API server prunes it or leaves it behind a disabled gate, so the install succeeded while the policy silently did not apply. It is now version-gated and rendered only where it takes effect; clusters below 1.27 get the documented API default (IfHealthyBudget) visibly rather than an ignored setting.

  • The chart-publish gate now judges every values overlay, at real secret values. The tag lane carried its own copy of the boot replay, which wrote a placeholder into every file-borne secret (so the 32-byte HMAC floor and the Argon2id password-hash parse were checked against the placeholder, not the configured secret), rendered only default-values.yaml, and skipped valueFrom.secretKeyRef environment entirely. It now calls deploy/helm/ci/boot-check.sh, which mounts each secret at its rendered value, resolves secretKeyRef, and boots every committed overlay — and the replay logic exists in exactly one place.

  • boot-check.sh no longer hands one workload’s environment to another image. With the admin console added, the script collected env: from every rendered Deployment, so the console’s FERROEHR_ADMIN__* variables (a separate binary with its own config root) reached the CDR image, whose strict sweep correctly refused them and reported a crash-loop for a deployment that runs fine. Environment collection is scoped to the CDR Deployment.

  • The Helm chart’s shipped values overlays now produce configurations the server accepts. All three rendered, linted and matched their goldens while none of them would boot: default-values.yaml enabled authentication with no mechanism, basic-auth-values.yaml carried a placeholder that is not an Argon2id PHC string, and all-features-values.yaml carried a 15-byte HMAC secret under the RFC 8725 §3.5 32-byte floor and enabled SMART without its required public base URL and endpoint metadata. The tag lane boots the appVersion image against the rendered default, so the next release cut would have failed. (#2159)

  • The audit schema was unreachable under the least-privilege role model. It carried no grants at all, so only the database owner could write audit records — a deployment connecting as ferroehr_app could not record one. The runtime role now holds exactly SELECT, INSERT, a column-scoped UPDATE on the two delivery stamps, and EXECUTE on the reaping and verification functions. The compose PostgreSQL image also pre-creates the audit schema alongside ehr and ext. (#2049, #2059)

  • Publishing a chart-only fix between releases works again. The chart lane’s pre-flight check — that a plain helm install of the chart will not crash-loop — now compares the chart against the image it is actually judged with: the released image on a vX.Y.Z tag, where appVersion is that release, and an image built from the current tree on a manual chart-only run, where appVersion still names the previous release and therefore rejects every configuration key added since it. The check itself is unchanged in strength; only the image it compares against is now the right one, and the run log and job summary name that image and why it was chosen.

  • A vX.Y.Z tag starts the image build and the chart publish at the same moment, and the chart lane’s pre-flight pulls the image the other lane is still pushing. The chart lane now waits for that specific build rather than retrying a registry pull in the dark: if the image build fails, the chart lane fails immediately with a link to it instead of presenting the race as a broken chart.

  • The published images carry OCI annotations, so registries can read their description. GHCR showed “No description provided” for all three packages even though each carried a description label: a label lives in the image config blob, while a registry reads the description from the image index annotation, and no annotation was being written at all. All three lanes now emit annotations at both index and manifest level.

  • The container images no longer claim the wrong licence. All three Dockerfiles declared org.opencontainers.image.licenses="Apache-2.0" for an MIT-licensed project. Images published by CI were unaffected — the workflow’s label overrode the Dockerfile’s — so this was only ever visible to someone building the Dockerfiles directly, which is the documented compose path. The app and console images now declare MIT, and the postgres image the SPDX expression MIT AND PostgreSQL, which is what it actually contains.

  • Service-layer refusals now report the precise openEHR Service Model call status they were raised with, instead of a generic stand-in. An unusable archetype-id regex pattern answers invalid_id_pattern (not precondition_violation) and an invalid stored query invalid_query (not content_invalid), and every conflict, bad request and semantic refusal keeps its own status through the service boundary. The HTTP status and response body of every affected request are unchanged — only the status a Service Model caller reads back becomes accurate.

  • An unusable id pattern, an unparseable stored query and an unparseable ad-hoc AQL query now carry the underlying parser failure as an error cause, so an operator can read the full diagnosis from the server log.

  • auth.oidc.hmac_secret_file and auth.oidc.jwks_json_file now appear as their own reference lines in the shipped ferroehr.toml template, so ferroehr config default teaches them. Both were real configuration keys mentioned only inside a trailing comment on another key’s line, in a file whose header calls itself the complete server configuration.

  • Documented gh attestation verify invocations no longer promise a result on the Helm chart or on 3.17.3: no chart version has been published, and signing landed after that tag was built. Both chapters now say so, and the stale “no published artifact carries an attestation” warning on the admission policies is replaced by a narrower note about the one thing still unproven — that no admission controller has yet run these policies against a live image.

  • The AQL printer emitted queries that meant something different when read back. Found by the new fuzzer, and it is a correctness defect rather than a formatting one: to_aql dropped parentheses that the grammar needs, so re-parsing the printed text produced a different query. Two causes. The AND/OR operators are stated in the grammar as binary alternatives of one recursive rule, which resolves left-associatively — so a same-precedence operand survives a re-parse on the left and silently re-associates on the right, and the printer treated both sides alike. And a CONTAINS chain used as a boolean operand absorbs whatever operator follows it, moving that operator inside the CONTAINS scope. The printer’s own documented invariant — parse(print(q)) == q — now holds across every boolean shape, asserted rather than assumed.

  • The container images declare their user numerically, so Kubernetes can actually verify it. Both images stated USER nonroot:nonroot, which reads well and defeats the check it was meant to support: the kubelet cannot resolve a username against an image it does not read, so runAsNonRoot: true without an explicit runAsUser refuses the pod with “cannot verify user is non-root”. The distroless base already declared the numeric 65532, and that line was overriding it with a name. Both now declare USER 65532:65532 — the same identity, stated the way every consumer can check. The Helm chart was unaffected because it pins runAsUser: 65532 itself; a plain kubectl run of these images was not.

  • Release artifacts carry their signature with them, so verification no longer needs GitHub. Each release now attaches the Sigstore bundles as assets beside the binary and its SBOM, which means a consumer can verify an artifact on an air-gapped host with nothing but the download in hand: gh attestation verify <tarball> --bundle <tarball>.sigstore.json --repo rubentalstra/FerroEHR. The attestations were already being produced and stored in GitHub’s attestations API; what was missing was the form in which a signature travels with the file.

  • The Helm golden-render gate was inert in CI, and the chart publish lane could not run at all. Both jobs passed version-file: to azure/setup-helm, which has no such input — its inputs are version, token and downloadBaseURL. An unrecognised input is only a warning, so the action quietly installed the newest Helm instead of the pinned one; the golden comparison then skipped itself (“running helm 4.2.3, goldens are pinned to 4.1.3”) and the job went red without ever comparing a render. The workflows now read the version out of deploy/helm/.tool-versions and pass it as the input that exists, so the pin means what the file says it means.

  • The conformance lane could not boot its own server. Two configuration rules added by the SMART hardening this cycle — a plaintext smart.public_base_url is refused without the named development flag (RFC 6749 §3.1.2.1, RFC 8414 §6.2), and an empty token_endpoint_auth_methods_supported is refused because a document advertising a token endpoint and no client-authentication method describes an authorization server that authenticates no client at all — are both correct, and the lane’s compose overlay satisfied neither. The system under test exited during compose, so not one conformance case ran. Fixed in the overlay, with the reason for each key stated where it is set.

  • The chart’s publish lane checked only half of what it hands the server. It validated the rendered ferroehr.toml against the appVersion image but not the FERROEHR__* environment the Deployment declares — and the strict boot-time sweep refuses an unknown variable just as firmly as an unknown TOML key. The lane now replays the declared environment, with a real file behind every *_FILE path, so a chart whose environment grammar outruns its image is refused before publication rather than discovered as a crash-loop.

  • Two scripts on the chart’s build and publish paths required Python. The Artifact Hub changelog derivation and the chart validator’s manifest check are now awk, matching the repository-wide rule that there is no Python here. The rewritten changelog derivation was verified byte-identical to the previous implementation across both a released section and the unreleased one (21 and 88 entries). The validator’s structural check kept the coverage it uniquely provided — every rendered document must carry a top-level apiVersion and kind — after confirming that YAML well-formedness is already enforced twice, by helm template and helm lint.

  • The Swagger UI showed a blank page at the URL the documentation gives you. /ferroehr/rest/swagger-ui answered 200 with the right HTML, and nothing on it loaded: the bundled index.html references its stylesheet and scripts relatively (./swagger-ui.css), so without a trailing slash the browser resolved every one of them against the parent directory and got a 404. You had to know to type /ferroehr/rest/swagger-ui/index.html. The mount path now redirects there, so the documented URL works. The target is index.html rather than the trailing-slash form deliberately — the path-normalizing middleware strips a trailing slash, which is what made the previous attempt at this an infinite redirect loop.

  • The four credential keys that had no file route now have one, the database DSN among them. The configuration rule is that every secret gets a *_file sibling so it can arrive as a mounted read-only file instead of an environment value — an environment value is readable through /proc/<pid>/environ and is inherited by every child process. Four keys were missing theirs, and they included the most valuable secret in the deployment: db.url (the DSN), events.url and fhir.outbound.url (AMQP URLs carrying credentials), and auth.basic.users[].password_hash, whose only route was inline in a configuration file. So the three secrets that already had a file route arrived as mounted files while the DSN travelled as environment. Now db.url_file, events.url_file, fhir.outbound.url_file and password_hash_file all work, through the config file or the environment grammar (FERROEHR__DB__URL_FILE, …), with redaction and boot validation identical on both paths — a file-delivered DSN redacts its credentials exactly like an inline one, and a file-delivered password hash faces the same Argon2id parameter floor. The inline and environment forms are unchanged: this adds a route rather than removing one, so a Compose deployment needs no changes. Setting a key both ways is a boot error, and the built-in development default does not count as “set”, so a *_file alone works without blanking anything first.

  • The image scanners were ignoring a setting their own configuration claimed to carry. trivy.yaml documented ignore-unfixed as one of its two deliberate choices, but the key lived only as an input in the container workflow — so any other lane reading that config got a different gate. Measured on the published PostgreSQL image, the same scan reports 55 HIGH/CRITICAL findings without the setting and 0 with it, all 55 being CVEs with no fix available. The setting now lives in trivy.yaml under vulnerability:, where every lane inherits it. (The nesting matters: a top-level ignore-unfixed key is silently ignored by Trivy.)

  • A partial database wipe bricked the deployment permanently, and now it explains itself instead. Almost every object the server creates lives in the ehr schema, but the cold archival tier lives in its own cold schema — so a DROP SCHEMA ehr CASCADE (a restore gone wrong, a recreated volume, a wiped test database) took the primary tier and the migration bookkeeping and left the archived clinical rows standing. The next boot then failed with a bare relation "vo_version" already exists and every restart retried the same failure: permanent CrashLoopBackOff, with no message naming the cause and no path out short of hand-written SQL. Observed on a live cluster, not reasoned about.

    The server now detects the state before migrating and refuses with the cause and both remedies in the message — restore the whole database from backup, or, if the wipe was intended, DROP SCHEMA cold CASCADE. Making the migration silently re-runnable was considered and rejected: those mirror tables were created from the primary tables as they stood at the time, so adopting a survivor could leave the archive tier a different shape from the tier it mirrors, and its rows are clinical content belonging to a repository that no longer exists. A crash with a remedy beats a database that quietly re-adopts another repository’s data. The operations page documents both directions of a partial wipe, including the reverse one the boot check cannot see, and says what to do instead: drop the database, which has no partial-state failure mode.

  • A Helm-configured Basic user’s Argon2id hash was written into a ConfigMap. The chart rendered .Values.config verbatim, so any secret in the configuration tree landed in an object nothing treats as sensitive: readable with namespace read, pasted wholesale into support tickets, collected by backup tooling that skips Secrets, and not covered by Secret encryption at rest. An Argon2id hash is not a plaintext password, but it is an offline cracking target — which is the whole reason this server enforces the OWASP parameter floor at boot. Two more live instances of the same defect came out with it: audit.fhir_feed.url and a terminology OAuth2 client secret, both of which the chart had no secure route for at all.

    The chart now classifies every key it renders, by name shape rather than against a list of today’s keys, so a secret key added to the configuration tree in a future release cannot leak by default. A secret with a secrets: route (authOidcHmacSecret, signingKeyPassphrase, multimediaSecretAccessKey, the new auditFhirFeedUrl and terminologyOauth2ClientSecrets, the DSN and the two AMQP URLs) is refused under config: with the key that carries it named in the error. A secret with no route — today only auth.basic.users[].password_hash, whose configuration key has no password_hash_file sibling to point at a mounted file — moves the whole rendered ferroehr.toml into the chart’s Secret, and no ConfigMap is created at all; the install notes say which object your release used. Verified on a live cluster: the Deployment serves authenticated openEHR requests with its configuration mounted from the Secret, and no ConfigMap in the namespace contains the hash.

  • A rotated secret or an edited ABAC policy did not roll the pods. The checksum/config annotation hashed the ConfigMap template only, so a change under secrets: or config.files reached the mounted volume while every running pod kept serving with the value it read at boot. The annotation now covers the rendered configuration, the mounted files and the secret values together.

  • The Helm chart could not start a pod at all, and only a real cluster could show it. Everything about the chart had been verified by rendering its YAML; deployed for the first time, every pod crash-looped. Kubernetes injects a set of Service link environment variables for each Service in the namespace, and for a Service named ferroehr* those (FERROEHR_SERVICE_HOST, FERROEHR_PORT_8080_TCP_ADDR, …) land inside the server’s reserved FERROEHR_ configuration namespace, where the strict boot-time sweep rejected them and refused to start — the install command from the documentation could not work. The chart now pins enableServiceLinks: false, which this workload never needed, and the chart’s own validation gate asserts the line so it cannot be dropped again.

  • The chart’s default-deny ingress NetworkPolicy was documented as narrower than it is. With networkPolicy.ingressFrom empty the rendered rule carries no from selector, and a rule without from admits every source — other namespaces included, not “any pod in this namespace” as the values file claimed. Only the port list is narrowed in that state. The values file, the template and the documentation now say so, and point at setting ingressFrom for a PHI workload.

  • The XML reader had no nesting bound, so a deeply nested document recursed the generated readers off the stack — and a Rust stack overflow aborts the process rather than unwinding, which means the catch-panic layer that renders this server’s clean 500 could not intercept it: one request would have taken the process down for every caller. Nesting is now bounded (256 levels, released as elements close, so wide documents are unaffected) and refused with a typed error. A DOCTYPE declaration is also refused outright now: entity attacks were already impossible because the XML library parses no DTDs, but that is a property of a dependency’s current behaviour, and canonical openEHR XML has no use for a DOCTYPE.

  • The API’s Content-Security-Policy was applied as an overriding outermost layer over a tree that includes Swagger UI, so default-src 'none' would have left the documentation page blank in a CSP-enforcing browser — in the default configuration, since Swagger UI is on by default. Swagger UI now carries its own policy, which needs no inline allowance at all (measured, not assumed, and pinned by a test).

  • An over-limit chunked request body (no Content-Length) was answered 400 as a malformed body instead of 413: the body-collection step turned the limit error into an empty body. Chunked and declared-length bodies now both refuse 413.

  • The unauthenticated /health family could disclose database connection detail: two indicators interpolated the raw driver error into the component detail, whose text can name the DSN host, the database and the connecting role, and a third served a failed check’s panic payload. All three now serve the fact and log the cause.

  • Request paths reached the log with their query string at debug level, and subject_id — an external patient identifier — is a query parameter by the specification’s own design. The request span now records the path only.

  • New top-level configuration key spec_profile (development | stable, default development; env FERROEHR__SPEC_PROFILE): selects the openEHR specification generation set the server runs as ONE coupled choice — development = RM 1.2.0 + BASE 1.3.0 + LANG 1.1.0 (today’s behaviour), stable = the latest RELEASED generations, RM 1.1.0 + BASE 1.2.0 + LANG 1.0.0. The active profile and its generation versions appear on the boot banner and GET /management/info. Under stable, AQL that addresses specification surface the released generation does not define is refused with an error naming the active profile. Profile-change contract: stable → development is always safe (openEHR minor releases are additive); development → stable is supported only for data that never used development-only constructs — such stored objects are refused loudly at read, never silently down-converted.

  • docker-compose.keycloak.yml — a standalone OIDC quickstart overlay. Download it beside the quickstart file and run docker compose -f docker-compose.yml -f docker-compose.keycloak.yml up to get a Keycloak identity provider on port 8081 with a ready-made demo realm (a confidential ferroehr client with the password grant enabled, and the user ferroehr / ferroehr carrying the ADMIN and USER realm roles) and the server’s bearer-token validation pointed at it. Basic auth keeps working, so the server advertises Basic, Bearer. The realm travels inside the file — no repository checkout needed.

  • docker-compose.override.yml — the from-source development stack for anyone working on FerroEHR itself. In a checkout it is merged automatically onto a bare docker compose up, so docker compose up --build builds the server, database, and admin-console images from the current sources and runs them with the fuller development configuration (docker/ferroehr.dev.toml: the three users ferroehr / ferroehr-admin / ferroehr-readonly, role-based access control enabled, and a keycloak profile with the full development realm).

  • The quickstart Compose files are attached to every GitHub release, so a fresh evaluation is downloading docker-compose.yml from the release page and running docker compose up.

  • Two new [auth.oidc] keys. clock_skew_leeway_seconds (default 60, env FERROEHR__AUTH__OIDC__CLOCK_SKEW_LEEWAY_SECONDS) sets the leeway applied to the time-based token claims; values above 300 are refused at boot, because expiry leeway may be “no more than a few minutes” (RFC 9068 §4 step 6) and a large one silently extends every token’s life. allow_insecure_issuer (default false) accepts a non-https issuer for development and test deployments only — it exposes token verification material to anyone on the network (RFC 8414 §6.2).

  • A new [authz.abac] key check_directory (default false): submits DIRECTORY (FOLDER) operations to the policy decision point. It replaces the old rule that inferred the opt-in from a directory entry in the remote-PDP policy map, so the check now works under the embedded Cedar engine too — previously a Cedar deployment could only enable it by inventing a policy name Cedar never reads. If you enabled DIRECTORY checks that way, set check_directory = true.

  • A CI guard that keeps the Compose files’ default image tags pinned to the workspace version, so a downloaded quickstart file can never reference images from a different release.

  • The openehr-lang crate’s v1_0 generation is now the TRUE LANG Release-1.0.0 surface, not a copy of the development line: it is generated faithfully from the release’s own machine-readable BMM, its vendored grammar set is the release’s actual syntax-appendix files (vendor/grammar/v1_0/, incl. that era’s base_lexer.g4), its lexer carries the ODIN reading alone (1.0.0 publishes no EL grammar and BEL first appears in 1.1.0), and its ODIN reader enforces the release’s own syntax: lowercase-only attribute keys, ANY primitive comparable value as a container key (real/boolean/character/term-code/duration keys, signed numerics), whole-document plug-in fragments, comma-only fractional seconds on times and dot-only on durations. Under the stable spec profile these are the rules ODIN text is read by. The ADL/cADL grammar set is likewise version-scoped by AM generation (crates/openehr-adl/vendor/grammar/{v1_4,v2_4}/, vendored by the new scripts/vendor/adl-grammars.sh).

  • The openehr-rm and openehr-base crates now ALSO emit the latest RELEASED specification generations beside the development pins — openehr_rm::v1_1 (RM 1.1.0, resolving against BASE 1.2.0 per its own BMM includes) and openehr_base::v1_2 (BASE 1.2.0) — each a complete peer: full type model, canonical-JSON codecs, RM attribute model, invariant cores, and validation behaviour. The served wire is unchanged (the current generations stay v1_2/v1_3); runtime selection between generations arrives with the spec_profile configuration key.

  • The s3 profile’s SeaweedFS healthcheck could never pass: it probed http://localhost:8333/, and localhost resolves to ::1 first while the S3 gateway listens on IPv4 only, so docker compose --profile s3 up --wait waited forever. It now probes 127.0.0.1.

  • An unknown Basic username now performs the same key-derivation work as a known one, against a fixed decoy hash, so response time no longer reveals whether an account exists.

  • The embedded Cedar policy engine now receives the caller: the principal was a constant with an empty role and scope set, so no role-aware or scope-aware policy could ever match, and every decision log named the same anonymous entity. The authenticated subject, its roles, its scopes and the operation id are now all available to a policy.

  • A Cedar policy that errors during evaluation no longer changes the outcome silently. Cedar skips an erroring policy and reports it in its diagnostics, which were discarded — so an erroring forbid stopped forbidding. A policy set that cannot be evaluated is now a fail-closed 500.

  • An external policy server with no configured rule for a resource kind now denies (and logs) instead of permitting, and the missing rule is refused at boot.

  • Three authorization gates reached for the authenticated caller and, finding none, invented one named UNKNOWN — writing a false subject into the audit trail. They now refuse, as the fourth gate already did.

  • A CONTRIBUTION’s template ids are read from the typed COMPOSITION value rather than a JSON pointer that a shape change could silently stop matching.

  • auth.oidc.algorithms is now bound to the configured key source: HS* with a JWKS (or RS* with only a symmetric secret) is a boot error, closing the algorithm-confusion setup where a public key is accepted as an HMAC secret (RFC 8725 §3.1). none is refused outright.

  • JWK selection honours the use, key_ops and alg facets and refuses an ambiguous key set when the token carries no kid, instead of taking the first key in the document.

  • The upstream-EHRbase comparison lane could not boot. A product-wide rename had swept the upstream image’s OWN contract along with our own names: the database container’s EHRBASE_USER*/EHRBASE_PASSWORD* init variables, the database name baked into that image’s init script, and EHRbase’s /ehrbase/rest/... base path (in both the party ixit and the readiness probe). The lane composes, becomes ready and drives cases again; only the strings upstream owns were restored, and the credentials we choose are unchanged.

  • Under spec_profile = stable, a demographic party body carrying the RM 1.1.0 PARTY.reverse_relationships attribute is now accepted: the payload is validated by the released generation’s own strict reader and the attribute — derived data the development line removed as redundant (upstream SPECRM-124) — is dropped on ingress rather than stored. Previously the stable server refused a valid instance of its own advertised generation with 400. The development profile still refuses the attribute as undeclared, and the generation delta ledger now pins the REMOVED direction too, so a future spec re-vendor changing either direction fails the build until adjudicated.

  • An ODIN integer container key whose lexeme cannot be evaluated (an out-of-i64-range magnitude, an inexact negative exponent) is now a parse refusal; previously it silently became the key 0, so two such keys collided as duplicates and stored data could be mis-keyed. Signed integer keys ([-1], [+2]) and exponent forms ([29e2]) now parse per the grammar’s integer_value : ('+'|'-')? INTEGER. ODIN parse errors also now report the position of the real defect instead of the first backtracked branch failure.

  • A bearer token whose sub claim is missing or blank is now refused with 401. Previously a validated token without a subject authenticated as a principal literally named unknown, and that fabricated identity was stamped into the audit trail; an unattributable caller is now rejected instead of silently mis-attributed.

  • Configuring both a symmetric secret (auth.oidc.hmac_secret/_file) and a static JWKS (auth.oidc.jwks_json/_file) is now a boot-time configuration error naming both keys. Previously the server silently used the symmetric secret and ignored the JWKS.

  • An unreachable or unresponsive OAuth2/OIDC issuer no longer stalls bearer-token requests or hammers the issuer with outbound connections. The client that fetches the issuer’s discovery document and JWKS now carries explicit timeouts, and a failed fetch is remembered briefly, so bearer requests during an issuer outage are refused fast instead of each one opening a fresh connection and waiting for the operating system’s TCP timeout. Three new [auth.oidc] keys tune this (they apply only when signing keys come from discovery — that is, when neither hmac_secret nor jwks_json is set): connect_timeout_ms (default 3000), request_timeout_ms (default 5000), and negative_cache_ttl_seconds (default 10, 0 disables). Successfully fetched keys keep their existing five-minute lifetime, and recovery is automatic once the negative entry expires.

Security

  • Release tags are now protected by a ruleset. Three lanes publish off a raw tag push — the release, the Helm chart, and the documentation version cut — and until now the tags driving them could be created, moved or deleted freely. A release-tags ruleset on refs/tags/v* blocks tag deletion and non-fast-forward tag updates and requires signatures, codifying the signed-tag practice the project already followed. Release immutability protects the window after a release is published; this protects the window in which a tag drives a build, an image push and a chart publish. SECURITY.md now records the full expected repository security posture, with the two read-back commands that detect a settings reset.

  • Every release tarball now ships a .sha256sum. It is the only verification available to an operator with neither gh nor cosign installed, which is what a locked-down clinical environment tends to be. The documentation is explicit that a checksum detects a corrupt download and not a substituted release — only the Sigstore bundle answers “who built this”.

  • A release now carries an SPDX SBOM of the repository, generated from the release commit, alongside the per-binary CycloneDX SBOM and the per-image SPDX SBOM. The three answer different questions — what is inside the binary I am about to run, what am I redistributing and under what terms, and what is in the image’s OS layer — and the security chapter now says which is which, so a reader picks the right one instead of assuming they are redundant. The CycloneDX specification version (1.5, the highest the generator emits) is recorded with its reason.

  • The release asset-completeness guard now matches asset names exactly. It compared by substring, so the check for the .tar.gz was satisfied by .tar.gz.sigstore.json alone and a release missing its actual tarball would have been published — unfixably, since publishing freezes a release. It now requires a whole-line match, and additionally checks the checksum, the repository SBOM and the quickstart compose files.

  • Dependabot now covers the fuzz workspace and the container base images. The fuzz harnesses are a separate Cargo workspace with their own lock file that a cargo entry at the repository root could never reach, so they were receiving no security bumps. The three base images are digest-pinned, which is correct for reproducibility and means they never receive a patch unless something proposes one — so the weekly image scan kept finding CVEs in the clinical images while nothing offered the fix.

  • The admin console’s Content-Security-Policy no longer allows inline scripts. script-src is now 'self' 'wasm-unsafe-eval' 'nonce-…', where the nonce is freshly generated for every response and stamped on the only inline script the console emits — Leptos’s hydration bootstrap and its resource-serialization chunks. An injected inline script no longer runs. style-src keeps 'unsafe-inline', because the console’s component library creates its stylesheets in the browser through the DOM without a nonce attribute; adding a nonce there would suppress the inline allowance under CSP Level 3 and leave the console unstyled, so the allowance stays with its reason recorded rather than being traded for a policy that looks stricter and works worse.

  • The error-body hygiene check now also covers Service Model faults (SmError::exception and exception-coded call statuses), which are a second route to a 500 response body it previously did not inspect. Three boot-time terminology and subject-proxy failures were rewritten to keep the underlying diagnostic out of the message and carry it as an error cause instead, so it reaches the server log and never a client.

  • Every one of the 20 GitHub Actions referenced by the build, test, release and publish pipelines is now pinned to a full commit SHA instead of a mutable tag, so a retagged or compromised upstream release can no longer change what runs against the tokens that publish this project’s releases, container images and crates. Each pin carries its human-readable version in a trailing comment and was verified to belong to the named repository.

  • 38 of the 40 repository checkouts in CI no longer leave the job’s API token in .git/config for the rest of the job (persist-credentials: false), so a later step — a third-party action, a build script, an uploaded artifact — can no longer pick it up and push with it. The two exceptions are the documentation jobs that genuinely use git against the remote, and both are now annotated with the reason.

  • All nine workflows now start from permissions: {} and grant each job only what that job actually uses, so write access to releases, packages, issues or Pages exists in the four jobs that publish and nowhere else. Previously a single workflow-level grant applied to every job in the file — contents: write to all of release.yml, packages: write to all of containers.yml — and codeql.yml declared nothing at all and inherited the repository default.

  • The release tarballs and the crates.io uploads are now built with no compile cache at all, so a published artifact can only contain bytes produced from the tag being released. Both lanes were silently restoring one: the toolchain action they use enables caching by default, which the release job’s own comment already assumed it did not.

  • The per-architecture release tarballs are now built from the exact commit the release notes were read from, resolved once and passed on as a SHA, instead of each job resolving the release tag separately. A tag that moved between the two jobs could previously publish assets that did not match the release they were attached to.

  • The build pipeline is now itself statically analysed on every pull request: a new zizmor gate audits the workflow definitions, and CodeQL analyses them as source alongside the Rust code. The properties above therefore cannot silently regress — an unpinned action, a checkout keeping its credential without a recorded reason, or a context value spliced into a shell command each fail the build.

  • The PGP signing private key is no longer world-readable inside the pod. The secrets volume was projected at 0440, but the sibling config volume carried no mode and so defaulted to 0644 — and that volume holds every config.files entry, whose documented use includes the PGP signing private key (signing.key_path), mutual-TLS PEMs and a JWKS blob, plus ferroehr.toml itself when the configuration holds a secret the chart cannot route out. Now 0440, matching the secrets volume. Verified on a live cluster: the applied defaultMode is 288 (0440) and both server pods read their configuration and became Ready.

3.17.3 - 2026-08-05

Added

  • Archiving an EHR or a party (POST /admin/archive/ehrs, POST /admin/archive/parties) now physically moves the archived objects’ rows out of the primary storage tables into a cold storage tier held in the same database (a new cold schema, added by migration 0007_cold_archive_tier), instead of only flagging them. The primary tables — and their indexes — shrink by exactly what was archived, while the wire is unchanged: an archived EHR, composition, folder or party is still retrievable, still carries its full revision history, and is served from the cold tier; unarchived reads are untouched and never consult it. Writing to an archived object brings it back automatically, a physical delete clears both tiers, and an admin export still dumps archived content. Multi-tenant isolation is enforced on the new tier by the same row-level-security policy as the primary tables.
  • ferroehr-ext — the optional-integration crate. The FHIR conversion core (mapping model + FLAT builder, outbound reverse-map, feeder-audit probe), the events transport (the EventPublisher seam, the AMQP publisher, the routing-key grammar), and the multimedia engine (content-addressed S3-compatible blob store + offload/expand transforms) now live in their own crate behind one additive cargo feature each (fhir, events, multimedia). Default builds are unchanged (all features on, wire-identical); a --no-default-features build produces a slim CDR that compiles the integrations out entirely and refuses an enabled-but-unbuilt integration loudly at boot. Configuration sections are unchanged and stay in the one config tree.
  • The typed FHIR R4B surface. The ATNA AuditEvent the audit trail stores and forwards, and the Parameters/ValueSet responses an external terminology server answers with, are now built and read through a typed FHIR R4B resource model (fhir-model, contained entirely in the optional-integration crate behind its fhir feature) instead of hand-written partial structs. The audit wire bytes are unchanged. The terminology client is correspondingly stricter: a server response that is not a valid R4B resource — for example a $expand result missing the required ValueSet.status or expansion.timestamp — is now reported as an upstream fault instead of being partially read, and its response cache holds decoded results rather than raw JSON. A binary built with --no-default-features refuses at startup when audit.store, audit.fhir_feed, or an external terminology provider is configured; the DICOM/syslog audit feed and the in-process terminology bundle stay available.
  • The OPT 1.4 constraints that target computed RM functions (EVENT.offset, DV_PROPORTION.is_integral, the US-spelled null_flavor) are now visible as a typed per-template report of unenforceable constraints instead of a silent skip; nothing new is rejected.
  • The BMM v3 model gains MODEL-level navigation (type_conforms_to, ancestor walks, flattened property lookup) and generic-substituted property synthesis per the LANG generic-inheritance semantics; a new P_BMM schema-validity pass reports duplicate package listings, case-folded class-name collisions, and non-conformant property redefinitions with spec citations.
  • The Expression Language has a native parser (hand-written over the vendored normative EL grammars): BMM_ASSERTION class invariants and routine pre/post-conditions in the BMM v3 model now parse into the published EL expression classes, with unparseable published-schema invariant strings collected as typed findings (319 of 400 pinned-schema strings parse; the remainder are Eiffel-flavoured forms the normative grammar does not admit, reported upstream).
  • ADL rules and slot assertions are now modeled as full expression trees (the BEL expression object model), with the string form derived from the tree. Printed ADL 2 output changes minimally where the old form was wrong: each assertion in a multi-assertion block carries its own string form (previously the whole block repeated), include/exclude emit their keyword once per list per the grammar, a symbolic ∈ prints as matches (one operator in the model), and an archetype-id constraint’s ; "assumed" value is no longer dropped.

Changed

  • The openehr-adl serializer seam is fallible. openehr_adl::print::print and openehr_adl::print::assertion_text now return Result<String, openehr_adl::print::PrintError>. An in-memory archetype whose rules assign an EXTERNAL_QUERY is refused instead of serialized with the assignment’s right-hand side silently empty: no released grammar spells EXTERNAL_QUERY, so no rendering of it could be valid ADL. The same refusal now covers a function-call expression node carrying no string name and a value-reference node carrying no string path — the two remaining shapes the printer used to render as empty text. Printed text for every other artefact is byte-identical.
  • Served OpenAPI descriptions no longer reference the internal conformance register (“register-documented …”); each affected description states its adjudicated handling with the released citation it already carried. Wire behaviour is unchanged.
  • The documentation site’s comparison chapter (and every page that echoed it) no longer frames EHRbase as “upstream”: FerroEHR and EHRbase are presented as two independent open-source openEHR CDRs measured by the same neutral instrument. The rendered comparison charts and generated tables carry the new labeling; provenance and licensing statements are unchanged.

Removed

  • BREAKING: the deprecated auth.admin_scope configuration key is retired. The management surface’s AdminOnly access level now gates on the RBAC admin role (authz.rbac.admin_role, default ADMIN) — the same gate every Admin-class API operation already uses. Deployments that disabled RBAC keep the previous behaviour (any authenticated caller passes AdminOnly); deployments that set admin_scope should grant the admin role instead (a JWT scope entry naming the role continues to surface as that role via scope→role extraction).

Fixed

  • A CONTRIBUTION commit whose version data cannot be converted to its RM resource — an empty mandatory 1..* container, a missing mandatory attribute — is now refused as 400 Bad Request, exactly as the same bytes are refused on every direct commit route, instead of 422. The commit seam runs the same strict canonical-JSON door the direct routes run (the released responses/422.yaml scopes 422 to content that “could be converted to a resource”); demographic party and party-relationship bodies take the same correction. Incomplete (553) commits keep their master06 relaxation.
  • EHR_ACCESS.settings now constructs typed: the spec leaves ACCESS_CONTROL_SETTINGS open for scheme-defined subtypes, and the generated model carries such an instance verbatim through a validated open-subtype carrier instead of refusing every legal scheme instance. Canonical JSON round-trips a scheme instance byte-identically; canonical XML — which defines no mapping for scheme members — refuses one honestly instead of dropping content.
  • Commits into an OPT 1.4 archetype slot are now checked against the slot’s allowed archetypes. The slot’s include/exclude archetype-id patterns were read only from each assertion’s optional string_expression string, which most operational templates do not emit — so for those templates the slot admitted any archetype of the right RM type, and a composition carrying a filler the template never allowed was accepted. The patterns are now read from the assertion’s expression tree (the constraint itself), with the string form used only as a fallback, so slot fillers are validated against what the template actually constrains. Templates whose assertions did carry the string form behave exactly as before, and a slot whose assertion constrains something other than the archetype id stays open as before rather than being narrowed by a pattern that does not apply to it.
  • Uploaded OPT 1.4 templates no longer lose their archetype-slot constraints. The XML codec discarded the content of every element the schemas declare as xs:anyType, which on an operational template is the EXPR_LEAF.item carrying each slot assertion’s archetype-id regex and its left-hand attribute path — so every OPT 1.4 slot constraint read back empty. Such an element is now kept verbatim (attributes, text and child elements) and re-serialized unchanged, so slot patterns survive a template round-trip and the OPT-to-ADL2 conversion and template-validation paths that read them now see the real payload. Canonical RM XML is unaffected.
  • A served Web Template (application/openehr.wt+json) no longer reports a field as mandatory when the template leaves it optional. A node’s min was taken from the constraint’s occurrences alone, ignoring the owning single-valued attribute’s existence; the two are orthogonal, and for a single-valued attribute existence is what governs presence. So an optional attribute carrying a 1..1-occurrences constraint — for example ISM_TRANSITION/careflow_step, which openEHR declares optional — was published as min: 1. min is now the lower of the two. Container attributes, node max, and commit-time validation are unchanged.
  • ADL2 slot-narrowing validation (VDSSM) no longer stops at the first include it cannot read. A specialised ARCHETYPE_SLOT whose include list mixes archetype-id regexes with constraint-based assertions was skipped entirely, so a genuinely widening literal after such an assertion went unreported. Each include is now judged on its own — an unreadable one is skipped, the rest are still checked. Symmetrically, no widening is claimed when the PARENT slot’s admitted set is itself unreadable, and a restatement is now judged over all assertions rather than the regex ones alone, so neither case invents a prohibition.
  • Flattening a specialised archetype no longer drops an inherited tuple constraint. A child node’s [a, b] attribute-tuple wholly replaced the flat parent’s tuple set, so a parent tuple over a disjoint attribute group silently vanished from the flat form (and from every operational template and Web Template built from it). Tuple overlay now merges by member-attribute group: a child tuple redefines the parent tuple over the same group, tuples over other groups are inherited, and a group the parent does not carry is added.
  • A ROLE carrying an empty capabilities list (and a party carrying an empty relationships list) is now refused at parse (400) — the RM invariants Capabilities_valid/Relationships_validity forbid present-but-empty; the generated model now emits these optional lists as Option<NonEmptyVec<…>> (the emitter’s invariant matcher learned the BMM’s .empty spelling and conjunction form), so a lenient acceptance is unrepresentable.
  • Template-mediated Simplified-Format commit failures now answer 422 instead of 400: a missing mandatory ctx field (language/territory), a |code outside a closed value set with no |value, a datatype mismatch, or any other post-conversion validation failure of a body that was readable as FLAT/STRUCTURED (register entries AMB-207/AMB-208). Refusal messages now name the actual defect. Template-independent FLAT syntax violations (|other conflicts, malformed keys) keep answering 400.
  • The six missing-mandatory conformance expectations, the create_ehr and update_directory binding outcomes, the invalid-other-details fixture, and the shared EHR-Extract import identity were corrected on the catalogue side after spec-adjudicated triage; the CNF runner’s ETag matcher gained released-grammar structural tokens for the object-id and template-id segments plus a validate gate that refuses un-resolvable matcher placeholders.

3.17.2 - 2026-08-04

Added

  • The eight openehr-* spec crates are published on crates.io — openehr-base, openehr-rm, openehr-am, openehr-adl, openehr-term, openehr-lang, openehr-query, openehr-its — each with its own README, packaged license texts (MIT, plus Apache-2.0 where openEHR-derived material is embedded), and docs.rs documentation. Packages version on their own independent SemVer line (starting at 0.0.x, permanently decoupled from the openEHR spec versions); the implemented spec version is carried by each crate’s SPEC_VERSION constant. Releases after the first go through an OIDC trusted-publishing workflow (publish-crates.yml) — no long-lived registry token (#1886).

  • RM validation now realizes every remaining register-visible class invariant: RESOURCE_DESCRIPTION(_ITEM), AUTHORED_RESOURCE, EXTRACT and EXTRACT_UPDATE_SPEC gained generated invariant cores wired into the typed dispatch — the machine-classified Unrealized register is at ZERO rows (#1623); EXTRACT_SPEC criteria and OPT-carried REVISION_HISTORY refusals are pinned by twins (#1648, #1737).

  • EHR-Extract export now evaluates EXTRACT_SPEC.criteria: AQL criteria queries select each entity’s primary set $ehr-bound (the entity’s EHR scopes the query and a literal $ehr parameter binds to its id); a non-AQL formalism or an unparseable criterion is refused with 400. The former blanket criteria refusal is gone (#1736).

  • On-demand CPU flamegraph of the running server: GET /management/flamegraph — a new opt-in management endpoint (default off, like the whole surface) that samples the process with an in-process pprof profiler for a bounded window and answers with the rendered flamegraph SVG. seconds/frequency query parameters are capped by the new [management.profiling] config keys (max_seconds, max_frequency); a request beyond a cap is refused with 400, a concurrent sample window with 409. No openEHR spec governs the management surface — our own operational extension (#1861).

  • Span-timing flamegraph capture: the new telemetry.flame_file config key installs a tracing-flame layer that writes folded stack samples of every span to the given file for offline rendering with inferno (inferno-flamegraph < file > flame.svg) — the async-attribution complement to the sampled-stack endpoint. Unset (the default) the layer is not installed at all. Our own telemetry extension (#1862).

  • EHR Extract import now enforces the copy closure (RM common master06 §Copying): a received branch version is refused with 400 unless its fork-point trunk version and same-branch predecessor travel in the same extract or are already stored (#1770).

  • FOLDER (DIRECTORY) resources can now carry ITEM_TAGs. ITS-REST overview Requests_and_responses.md §openehr-item-tag and openehr-version-item-tag names FOLDER among the change-controlled resources the wrapper headers associate tags with, and the DIRECTORY write routes have always carried those headers — but the tag store rejected the FOLDER target type, so a tagged directory commit answered 409 with a raw PostgreSQL constraint string AFTER the directory version had already been created. FOLDER tags now store, echo and appear in the EHR-wide tag listing (GET /ehr/{ehr_id}/tags). No dedicated /directory/…/tags routes appear: the release defines none, and a FOLDER tag is reached through the wrapper headers and that listing.

  • Tag mutations emit IHE ATNA audit records. Creating, replacing or deleting ITEM_TAGs now leaves an audit trail under the tagged resource’s own DICOM class. openEHR is silent here — tags are outside change control, so they correctly produce no CONTRIBUTION and no AUDIT_DETAILS, and no released text substitutes anything — so this is our own design for a clinical repository.

  • A 553|incomplete| commit may now omit mandatory data for every committable kind. RM common master06 §Incomplete Content states that in the incomplete state “mandatory attributes may be absent … single-valued attributes may have null values and container attributes may be empty, even though they may have minimum existence and cardinality respectively of one”, and that such data “respects the same template and archetype(s), but with all existence and cardinality lower limits set to zero”. Until now only the template/archetype layer relaxed, and only for COMPOSITION: a 553 commit whose content left a mandatory attribute absent or a 1..* container empty was still refused 422 by the reference-model layer. It is now accepted — for COMPOSITION, FOLDER, EHR_ACCESS and the demographic party / relationship kinds, on the CONTRIBUTION route and on the direct COMPOSITION/DIRECTORY routes (via openehr-version: lifecycle_state.code_string="553"). EHR_STATUS is unchanged: the CNF schedule holds that the incomplete state does not apply to it, pending SPECPR-368. Only the presence and lower-bound checks relax — types, _type slots, terminology bindings, patterns, coded values and every other class invariant are enforced exactly as before, so an incomplete commit carrying data that is WRONG rather than merely missing is still refused (“data may be missing, but it may not be wrong”).

  • The conformance catalogue can now state a CONTRIBUTION’s commit audit. A case’s audit: block states only its delta against the derived envelope audit, and the reserved absent sentinel omits a member outright — the seam the mandatory-member refusals need (RM common UML/classes/org.openehr.rm.common.audit_details.adoc §Attributes makes change_type and committer 1..1; the released OAS specifications/schemas/common/UpdateAudit.yaml §required lists both on the commit DTO). Three cases land on it: an omitted change_type, an out-of-group change_type code, and the conformant twin that states both members and proves the client-supplied time_committed is not the recorded one. Cases that already authored an audit: block now actually put it on the wire.

  • A case pins which lineage head a default EHR-Extract exports. With a trunk head and a branch head both open in one container, RM ehr_extract master04-common_package.adoc §Version Specification never says which is the “latest available version”. The choice — the trunk head — is now adjudicated in the ambiguity register (AMB-206) and pinned by I_EHR_EXTRACT_SERVICE.export_ehr_extracts-latest_across_lineages.

  • Canonical-XML responses are now checked against the published openEHR XSDs. A new gate serializes documents through the shipped codec and validates them with an XSD processor against both vendored ITS-XML bundles, and it records exactly where a served document and the schema its namespace declares disagree. The finding it makes visible: the v1 bundle openEHR still publishes as the STABLE one is frozen at an older Reference Model generation, so a document that is a correct RM 1.2.0 instance can carry attributes that bundle never declared — FOLDER.details on a directory is the case you are most likely to meet, and 50 RM classes (EHR, EHR_STATUS, CONTRIBUTION, the demographic party types, …) have no v1 schema at all. Nothing served changes: the default namespace is still v1, Accept: application/xml; version=2 still selects v2, and the codec still writes the complete Reference Model rather than dropping clinical content to fit an older schema. Deployments that validate responses against the published XSDs should be aware that the v2 lineage is the one that models RM 1.2.0 — and that it currently cannot be compiled by a standards-conformant XSD processor because of an invalid pattern in the upstream schemas. The full per-attribute breakdown ships in the conformance ambiguity register as AMB-185.

  • A conformance case for a party’s inline, by-value relationships. A PARTY_RELATIONSHIP is modelled twice and openEHR reconciles the two nowhere: the Reference Model stores it inside its source party (“the relationships attribute is by value”), versioned with that party and with no version container of its own, while the Service Model gives every relationship its own independently-addressed container. This server serves both, and they are disjoint — committing a party that carries an inline relationships list does not create a relationship container, and creating a relationship container does not append to any party’s list. The catalogue now pins the inline half: a PERSON committed with a by-value PARTY_RELATIONSHIP is accepted and reads back with that relationship unchanged, unexpanded and un-repointed. Behaviour is unchanged; the adjudication ships in the conformance ambiguity register as AMB-187.

  • New conformance cases for the LOCATABLE root rules and the feeder-system audit. The catalogue now pins the two refusals above from the wire side (a COMPOSITION root whose archetype_node_id contradicts its ARCHETYPED block; a COMPOSITION carrying an empty links list; a COMPOSITION whose inner ITEM_TREE carries an empty archetype_node_id), and four cases cover FEEDER_AUDIT end to end: a commit carrying audits at the COMPOSITION root and on an interior data node round-trips every modelled attribute (identifiers, inline original_content, both system audits, and the originating audit’s other_details), an update that retains the feeder audit keeps it on the new version, an update that drops it is accepted and does not carry it forward, and an update whose content is identical to the preceding version still creates version 2.

  • Conformance cases for populated LINKs on a COMPOSITION. A commit carrying a complete LINK at the COMPOSITION root and on an interior ENTRY now round-trips with both links intact (the accepting twin of the empty-links refusal), and a LINK whose target is not an ehr:// URI is refused (422) — placed on the interior ENTRY so the case also proves the rule is applied below the resource root.

  • Conformance cases for FOLDER.items — the directory’s reference slot. The catalogue exercised directory folders, names, links and details, but never the attribute the whole abstraction rests on: a folder’s items list holds references to other objects, never the objects themselves, and the same object may be referenced from more than one folder (that is what lets one directory classify a composition as both an episode and a problem). A directory whose two sibling folders both reference the same target — beside a second, distinct target — is now committed and read back with both references intact, so a server that collapsed the duplicate would fail; and a folder that carries a composition by value in items instead of a reference to it is refused with 422, leaving the EHR without a directory. Two further cases pin how wide an identifier that reference slot accepts. A folder reference may be version-pinned — its id a three-part OBJECT_VERSION_ID naming one particular version of a composition — and it now round-trips with all three parts intact, so a server that truncated it to the leading UUID would fail. And a reference identified in a foreign scheme (a GENERIC_ID) is accepted and served back unchanged: the Reference Model types the slot at OBJECT_ID, whose family has six concrete members, while the published OpenAPI schema for the same slot enumerates only two — the adjudication ships in the conformance ambiguity register as AMB-186.

  • Conformance cases for the ATTESTATION wire family. The catalogue now drives the 666|attestation| CONTRIBUTION member end to end: attesting an existing COMPOSITION version is accepted, adds no new version, reports the attestation-only aggregate change type, and surfaces the completed ATTESTATION on both the version envelope and the revision history; the pending-then-signed pattern leaves both attestation objects on the one version; and three refusal twins pin the ATTESTATION invariants on the wire (a missing reason, a coded reason whose code sits outside the openEHR attestation reason group, and a present-but-empty items list). An ORIGINAL_VERSION carrying attestations is also pinned as a canonical-JSON/XML serialization vector. An attestation carrying an inline DV_MULTIMEDIA attested_view — the screen image of what was signed — now round-trips with its media type, size, inline data and alternate text intact on both the version envelope and the revision history.

  • Conformance cases for the generic-package party and participation rules. A COMPOSITION whose context participation carries a bounded DV_INTERVAL<DV_DATE_TIME> time and whose ENTRY-level other participation carries an open-ended one now round-trips with every interval boundary intact, and four refusals pin the terminology and identity rules those classes carry: a coded PARTICIPATION.function outside the openEHR participation function group, a PARTY_RELATED participation performer whose relationship code is outside the subject relationship group, and — on the commit audit’s committer, which sits beside the committed content and is therefore missed by a content-only validation walk — the same out-of-group relationship, a PARTY_IDENTIFIED carrying none of name, identifiers or external reference, and one whose name is the empty string. A PARTY_RELATED committer with an in-group relationship is pinned as the accepting twin, so refusing that party type wholesale no longer passes.

  • A conformance case for the third PARTY_SELF referral scheme. The RM names three ways to refer to the record subject from inside an EHR, and the catalogue only exercised two of them. A COMPOSITION whose interior ENTRY carries a PARTY_SELF subject with a complete external_ref PARTY_REF (id, namespace and type) now round-trips with that reference intact, so a server that dropped or refused a per-instance subject reference — a spec-supported deployment style — no longer passes the catalogue.

  • A canonical-JSON output mode for the corpus fixture generator. The openehr-its canonical_convert example now emits canonical JSON when the output path ends in .json (and handles ORIGINAL_VERSION documents under the published <version> root), so a committed JSON fixture can be the codec’s own output rather than a hand-typed approximation of it.

  • Every figure the vendored specs reference is now vendored too. scripts/vendor/spec-docs.sh additionally fetches, from the same pinned commits, exactly the figures the vendored chapters reference: the 129 UML class-diagram SVGs ({uml_diagrams_uri}, under docs/specs/openehr/<COMPONENT>/docs/UML/diagrams/) plus the 200 per-document diagrams and images ({diagrams_uri} / {images_uri}, under <COMPONENT>/docs/<doc_name>/diagrams/ and .../images/). Spec chapters are now readable offline with their figures intact instead of carrying dangling links. Only referenced files are taken, byte-for-byte; a referenced figure missing at the pin fails the vendoring run.

Changed

  • The declared FHIR integration surface is now stated as R4B (4.3.0) everywhere — the served OpenAPI descriptions, configuration docs, and website pages previously said “R4”. Wire behaviour is unchanged: the resources FerroEHR touches (AuditEvent/BALP, terminology Parameters/ValueSet, the inbound connector starter set) are byte-identical between R4 (4.0.1) and R4B (4.3.0); the /fhir/r4/… connector paths are unchanged (#1885).

  • The openehr-* spec crates are republished as 0.0.3 (lockstep): the generated sources are re-emitted in rustfmt-normalized form after the emitter changes that shipped with the typed-DTO campaign — no semantic change to any generated type or impl.

  • The tenant admin API (/admin/tenant) and the event-subscription admin API decode their request bodies into typed definitions: a field of the wrong JSON type ("enabled" as a string, a non-string predicate or name) is now refused with 400 naming the offending member, where it was previously coerced to a default or treated as absent. Well-typed requests are unaffected; the response record shapes are byte-identical (#1694).

  • BREAKING (error class): the spec model is now complete-by-construction across every crate: the cross-schema re-emission closure applies uniformly (openehr-rm re-emits BASE’s Interval/Iso8601 family, am14 re-emits AUTHORED_RESOURCE/RESOURCE_DESCRIPTION), and every 0..1 list carrying a present-implies-non-empty invariant is emitted Option<NonEmptyVec<T>> — a present-but-empty list ("links": [], "contacts": [], "mappings": [], …) now refuses with 400 at parse instead of 422, on every strict write surface; the 553|incomplete| relaxation is unchanged (#1699, #1730).

  • BREAKING (v1-pinned XML consumers): the default canonical-XML lineage served for application/xml is now the ITS-XML v2 namespace (http://schemas.openehr.org/v2) — the only published schema bundle that models the RM 1.2.0 this server emits. The v1 lineage stays selectable per request with Accept: application/xml; version=1 (a non-default v1 response is labelled Content-Type: application/xml; version=1). Request payloads are unaffected — both namespaces are read regardless (#1666).

  • An ITEM_TAG whose key or value violates its own RM invariants is now refused when the payload is read, not after it is built. RM UML/classes/org.openehr.rm.common.item_tag.adoc §Invariants states Inv_key_valid (not key.is_empty and key.is_justified — no leading or trailing whitespace) and Inv_value_valid (a present value must be non-empty) over ITEM_TAG’s own fields. Both now run at the type’s construction door, which the canonical-JSON and canonical-XML readers build through, so an empty or whitespace-padded tag key is rejected at parse — in any document position, named by its path — instead of producing a value that existed in violation of its own class definition until something validated it. Tag requests that already conformed are unaffected.

  • The canonical-JSON reader honours the default values the vendored meta-model states. A Point_interval payload that omits lower_included/upper_included/lower_unbounded/upper_unbounded now reads back with the schema’s own declared defaults (included, bounded), matching what Proper_interval and Multiplicity_interval already did — the meta-model states those four values on Point_interval and only there, and the reader had been ignoring them.

  • BREAKING: a structurally invalid RM request body now answers 400, not 422, and the commit audit’s coded members carry their released wire shape. The commit routes (COMPOSITION, EHR_STATUS, DIRECTORY, EHR creation, the demographic party and relationship routes, and the EHR-Extract import) decode the request body into the concrete openEHR type before anything else runs, so the strict canonical reader is what judges its shape. A body that is not an instance of the addressed class at all — a wrong or missing _type, an undeclared member, a repeated member, an absent mandatory attribute, an empty 1..* container, a malformed identifier, a PERSON posted to /demographic/agent — is refused there, and the ITS-REST overview (Requests_and_responses.md §HTTP status codes) assigns that class 400 (“could not be parsed or is invalid”), reserving 422 for a body that is “well-formed but was unable to be followed due to semantic errors”. Bodies that ARE valid instances but break a semantic rule — a terminology binding, a template mismatch, a body uid naming a different versioned object — keep their 422 exactly as before. In the same change the commit envelope adopts the released UpdateVersion.yaml / UpdateAudit.yaml shapes end to end: lifecycle_state and commit_audit.change_type are DV_CODED_TEXT ({"value": …, "defining_code": {"terminology_id": {"value": "openehr"}, "code_string": "532"}}) rather than the flat TERMINOLOGY_CODE spelling, commit_audit.description is a DV_TEXT object rather than a bare string (so a description may now be CODED when it travels in the body — the openehr-audit-details header still carries only the plain description.value), and commit_audit accepts its UPDATE_ATTESTATION subtype, which the released schema’s discriminator has always allowed and which RM common master06 §Committal and Audits names (“AUDIT_DETAILS … or its subtype ATTESTATION”). The committal request headers, the CONTRIBUTION route and every response body are unchanged. Simplified-Format (FLAT/STRUCTURED) input failures split along the same line: a TEMPLATE-INDEPENDENT format violation — a key breaking the FLAT field-identifier grammar, a ctx/ key outside the master06 vocabulary, the forbidden |other + |code/|value/|terminology combination — is 400 (the document is not readable as FLAT), while template- or RM-mediated conversion failures (an unknown path, an undefined suffix, a missing mandatory ctx field, a closed value set) keep the composition endpoints’ own 422.

  • BREAKING (canonical XML): a version read now serves the <version> document element the published XSDs declare, instead of <original_version> / <imported_version>. ITS-REST overview Resources.md §“XML Format” requires that “both request payloads and responses MUST conform to the [published XSDs]”, and the ITS-XML schemas declare exactly one document element for a VERSION — <xs:element name="version" type="VERSION"/> — over an ABSTRACT VERSION type; neither published lineage declares an element named after a concrete subtype. The concrete class therefore rides on the root’s xsi:type (ORIGINAL_VERSION or IMPORTED_VERSION), as XML Schema requires of an instance of an abstract type. Every GET .../versioned_composition/{uid}/ version[/{version_uid}] and .../versioned_ehr_status/... read requested with Accept: application/xml is affected, in both the v1 and v2 lineages. A schema-validating client rejected the old roots; a client that pattern-matched on them must now read <version> plus xsi:type. Canonical JSON is unchanged (the envelope keeps its _type self-tag), and no other resource’s root changes.

  • lifecycle_state is now required on every CONTRIBUTION version (400). SM master03 §Version Update Semantics says “The lifecycle_state must be supplied in all cases”, and the released UpdateVersion schema lists it under required; a member that omitted it was previously accepted and silently defaulted to 532|complete|. A 666|attestation| member is exempt — it commits no new version, so it has no version lifecycle state to supply.

  • A DELETE on a change-controlled resource now refuses a committal header that names a lifecycle other than 523|deleted| (400). The header was previously parsed and discarded, leaving a client believing an instruction had been merged that was not; a DELETE commits the logical-deletion procedure, which fixes the state. A DELETE with no lifecycle attribute is unaffected.

  • other_input_version_uids is refused on the CONTRIBUTION write wire (400). The released UPDATE_VERSION schema declares no such property (and NewContribution.versions items are UpdateVersion), so the merge commit has no released shape — the same absence the import commit has. Merge provenance stays produce-only: it is still served on ORIGINAL_VERSION reads, and it is still preserved verbatim by the EHR-Extract import and the archive load, which reproduce a foreign version unchanged.

  • 422 Unprocessable Content messages now follow one uniform shape — <RM attribute path> <what is wrong> (<invariant name>), for example ATTESTATION.items must be a non-empty list when present (ATTESTATION.Items_valid). Internally the service layer carries every such refusal as structured data (the attribute path, the named openEHR invariant, and the nested class-invariant violations a validation pass produced) instead of a pre-formatted sentence, and renders it into the response body exactly once, at the REST edge. Most messages are byte-identical to before; the remainder were reworded into the uniform shape (chiefly the CONTRIBUTION version rules, item-tag rules, operational-template rules, PARTY_REF refusals, and the VERSIONED_COMPOSITION cross-version invariants, where an attribute is now written COMPOSITION.category rather than COMPOSITION category). A PARTY_RELATIONSHIP with an absent source/target, and a party with an empty identities list, are now reported by the RM decoder that already refuses them rather than by a second hand-written check — same 422, different wording. Response body SHAPE, status codes and all validationErrors[] contents are unchanged, and the 422 body is spec-silent (responses/422_COMPOSITION.yaml declares no schema), so no declared contract changes.

  • A canonical-JSON object that repeats a member name is now REFUSED with 400, naming the repeated member. The reader previously let the last occurrence win. RFC 8259 §4 says object member names “SHOULD be unique” and that “when the names within an object are not unique, the behavior of software that receives such an object is unpredictable”, and no conformant openEHR writer emits a repeated member — every attribute is written once, in model declaration order — so a repeated member is refused rather than silently resolved to one of its values. This applies to the _type discriminator as well as to modelled attributes.

  • Canonical-JSON refusal messages now carry the full JSON path to the offending node (for example (at $.content[0].data.items[2].value)), so a client can locate the defect in a large document without bisecting it. The refusal wording itself is unchanged in kind: the offending member, the class that does not declare it, and the members that class does declare.

  • The canonical-JSON lexeme for a REAL in exponent form now writes a signed exponent (1e+21 rather than 1e21). No openEHR specification governs the rendering of a REAL — both forms denote the same value and RFC 8259 §6 admits both — and the new form is what the reference JSON encoder produces. Only magnitudes outside the plain-decimal window (roughly 1e-5 to 1e16) are affected; no clinical quantity in the conformance corpus changes by a single byte.

  • A canonical-JSON payload carrying an attribute the openEHR RM does not declare is now REFUSED with 400, naming the path and the offending member. The reader previously ignored an undeclared key. It is refused because that is the only reading under which the JSON and XML encodings of a resource share one data model: ITS-REST/specifications/docs/overview/ Resources.md requires an XML payload to validate against the ITS-XML schemas, which declare no wildcards — an undeclared element cannot validate — and states the same for the JSON encoding at SHOULD strength, while openEHR’s own published ITS-JSON schemas close 128 of their 134 object definitions with additionalProperties: false. The status is 400 rather than 422 because a document the reader cannot read never converts (the released status table: 400 is content that “could not be parsed or is invalid”; 422 is content that is “well-formed but was unable to be followed due to semantic errors”). The set of accepted attributes is the RM at this server’s pinned version, so a payload that is valid openEHR is unaffected; a client sending a private extension member must move it into a modelled slot (for example ITEM_TREE other_details, or a FEEDER_AUDIT).

  • A malformed identifier is now refused wherever it appears in a document, not only in a request path. HIER_OBJECT_ID, OBJECT_VERSION_ID, VERSION_TREE_ID and the UID family are built through a constructor that runs the released identifier grammar (BASE/docs/base_types/master05-identification_package.adoc §Syntaxes), and the canonical JSON and XML readers construct through it — so an identifier such as PractitionerRole/12345-mock in a PARTY_REF, or a one-part value tagged OBJECT_VERSION_ID, is rejected at parse with 400 instead of being stored and served back.

  • UpdateItemTag request bodies reject undeclared members. The released OAS declares the schema additionalProperties: false, so an unexpected member is now a 400 rather than being silently dropped.

  • A committer external_ref.id supplied through the openehr-audit-details header must be a well-formed HIER_OBJECT_ID. A malformed value is refused with 400 instead of being written into the commit audit.

  • OBJECT_VERSION_ID values on the wire are now checked against the full openEHR identifier grammar. A version identifier in a request path, an If-Match header or a VERSION.uid previously only had to have three ::-delimited parts with a well-formed version-tree id; its object_id and creating_system_id parts are now also required to be legal uid values — an ISO OID, a UUID, or an internet id (BASE/docs/base_types/master05-identification_package.adoc §Syntaxes). Values such as bad id::sys::1 or 1234-5678::sys::1, which no conformant client sends, are refused with 400 instead of being accepted; every identifier the spec admits is unaffected. The same grammar now backs the validating constructors of HIER_OBJECT_ID, OBJECT_VERSION_ID and the UID family, so a malformed identifier cannot be built through them.

  • AQL now rejects an [archetype_node_id='…'] predicate whose value is neither an archetype identifier nor a node code. CONTAINS COMPOSITION c[archetype_node_id='openEHR-garbage'] used to be planned as an archetype constraint that could never match (a 200 with an empty result set); such a value is now a 400 naming it. The QUERY spec defines that bracket predicate as equivalent to the archetype ([openEHR-EHR-…]) and node ([at0002]) shortcut predicates (QUERY/docs/AQL/master03-syntax.adoc §“Archetype predicate” / §“Node predicate”), so those two forms are its whole admissible operand set — matching what the RM lets LOCATABLE.archetype_node_id hold. Well-formed archetype ids and at/id codes are unaffected.

  • Operational-template upload now enforces AOM2 VCACA’s numeric arm. A template that states a container cardinality wider than the reference model’s — for example CLUSTER.items cardinality {0..3} against the RM’s List<ITEM> [1..*] — is refused with 422 naming the rule (AM/docs/AOM2/master04.5-constraint_model-class_definitions.adoc §VCACA; master08-validation.adoc §Validate Definition). The fully-open {0..*} that published templates commonly carry is not affected: ADL 1.4 makes C_MULTIPLE_ATTRIBUTE.cardinality mandatory where AOM2 makes it optional and set “only if it overrides the underlying reference model”, and cADL’s open constraint means “any value permitted by the underlying information model” (AM/docs/ADL1.4/master05-cadl.adoc §“‘Any’ Constraints”), so an open interval states no override and defers to the RM. Every template that uploaded before still uploads.

  • Commit-audit and EHR-resource validation failures now report through the structured error body. The AUDIT_DETAILS.committer and the EHR_STATUS / EHR_ACCESS / FOLDER / demographic-party commit checks are now produced by the shared Reference Model validator instead of duplicated by hand, so the same defects are refused with the same 422 status but render as the openEHR Error object ({ message, validationErrors[] }) rather than a flat message — the shape every other validation failure already used. The one message-detail change: a PARTY_RELATED committer whose relationship code is outside the openEHR subject_relationship group is still refused naming Relationship_valid, but no longer echoes the rejected code.

  • System-generated commits are now attributed to the product’s own identity. A write with no authenticated principal (auth disabled, or an internal write such as an import or a synthesized composition) records an AUDIT_DETAILS.committer of PARTY_IDENTIFIED name FerroEHR. It previously read EHRbase on the service paths and ferroehr.local on the REST adapter path; all layers now emit the one value. The deployment’s configured system_id ([server] system_id, the value stamped into AUDIT_DETAILS.system_id, EHR.system_id and every OBJECT_VERSION_ID.creating_system_id) is unchanged and stays separately configurable. Audits already committed keep the committer they were written with.

  • Audit, attestation and revision-history wire bodies now follow canonical BMM field order. AUDIT_DETAILS, ATTESTATION, REVISION_HISTORY and REVISION_HISTORY_ITEM are built as their RM types and serialized through the canonical codec instead of being assembled by hand, so every such body — the version commit_audit, the CONTRIBUTION audit, and both the EHR and demographic revision histories — now carries _type first followed by the Reference Model’s own attribute order. Only key order changes; the same attributes with the same values are emitted, and JSON key order is not semantic, so parsing clients are unaffected.

  • Malformed attestation fields are now refused instead of stored. ATTESTATION.attested_view, proof and items used to be copied into the stored attestation without being read; a submission whose attested_view is not a DV_MULTIMEDIA, whose proof is not a string, or whose items carry a member that is not a DV_EHR_URI is now rejected with 422 naming the offending attribute (and, for items, its index).

  • Outbound openEHR spec-defect reports moved to the issue tracker. The docs/conformance/upstream-reports.md ledger is deleted; every report is now a GitHub issue labeled upstream-report (what the released spec says, what this implementation does, the resolution sought). The ambiguity register’s upstream_ref field is renamed to upstream_issue and carries the GitHub issue number — the published ambiguity-register.schema.json changed accordingly.

Removed

  • The AQL RESULT_SET no longer carries a top-level id. Responses from POST/GET /query/aql and the stored-query execute routes previously added an id field holding a freshly minted UUID. The released ITS-REST ResultSet schema declares exactly meta, name, q, columns and rows with no additionalProperties, so the wire has no slot for it and the field was an undeclared property on a closed object schema. Clients that read id should use the response’s ETag header instead — the released ITS-REST text names it “a unique identifier of the resultSet” (query/Request.md §“Common Headers and Query Parameters”), and it is unchanged by this removal.

Fixed

  • Licensing coverage corrected (#1883): the vendored openEHR specification text and the CKM-derived clinical models are CC-BY-SA 3.0 (per-file licence metadata for clinical models), not Apache-2.0 as the README previously implied. The repo now ships the CC-BY-SA 3.0 text (LICENSE-CC-BY-SA-3.0), vendors each upstream LICENSE alongside its tree, records the license in every vendored tree’s PROVENANCE.md, and the documentation site gained a Licensing & legal page with the full reckoning. FerroEHR’s own code stays MIT; the openEHR machine-readable artifacts and test corpora stay Apache-2.0.

  • EHR-Extract import: the copy-closure check now matches the fork-point trunk version with ANY creating system (RM common master06 §Distributed Versioning — a branch legitimately forks off a foreign trunk); /management/flamegraph answers a well-formed SVG instead of a zero-byte body when the sample window catches an idle process.

  • CNF runner: an unresolvable <name> placeholder in an outcome header matcher is now a loud case failure instead of silently wildcarding to .*; the structural tokens <n>/<system_id> resolve to their real grammars, and outcome matchers see the same merged variable scope as request building (#1852).

  • AQL LIKE and matches predicates on multi-valued paths now use the existential (any-match) lowering the comparison operators already use — a row is matched when ANY node on the path satisfies the predicate, instead of an order-undefined single-node pick (#1448).

  • The admin EHR dump/load archive now carries every vo_attestation row (with its at_committal flag), and load re-persists them verbatim — a restored version keeps its attestations and its stored signature verifies under verify_on_read = strict (#1685).

  • A node id carrying the at/id code leader but failing the code grammar is refused (at0abc). AOM2’s own code predicate is leader-based, so such a string claims code-hood and must satisfy the code syntax — previously it fell between the code family (whose grammar it fails) and the free-text family (whose leader-freedom it lacks) and no rule caught it.

  • The OPTIONS / conformance manifest serves the generated contract DTO, and the System API group joins the authorization and audit classifiers. The manifest body is now the emitted Options DTO (byte-identical wire; a lockstep test pins the served OpenAPI’s documentation shape to it), the System route table joins the RBAC route map explicitly, and the operation carries explicit authorization (any authenticated principal) and audit (application-activity) classifications instead of the fail-closed defaults.

  • A malformed ITEM_TAG refuses at construction, not after the fact. The generated ItemTag type gains a validated constructor running its RM invariants (Inv_key_valid/Inv_value_valid), so a violating tag cannot exist as a typed value anywhere in the application — the JSON and XML readers refuse it at parse, path-named. Wire statuses are unchanged (the tag routes’ 422 mapping stays); a stored tag row that no longer constructs is reported as the server fault it is instead of being served.

  • An undeclared key on a CONTRIBUTION version member is refused (400, named at its member path) instead of silently ignored. The released commit wire declares exactly six member properties (UpdateVersion.yaml) plus the adjudicated _type self-tag; the member seam was the last non-strict reader, accepting arbitrary extra keys without a diagnostic while every other read surface refuses them.

  • A CONTRIBUTION whose audit omits committer is now refused (422) instead of being attributed to the server’s default identity. The same released commit schema that requires change_type requires committer (NewContribution.yaml over UpdateAudit.yaml), and a server-invented committer would put an identity the client never named into the audit trail. The direct COMPOSITION/DIRECTORY routes are unchanged — there the committal headers stay optional and the authenticated default applies, exactly as the ITS-REST overview requires.

  • An emptied ITEM_TAG collection is echoed as no header, never an empty one. The EHR-side write routes echoed an EMPTY openehr-item-tag header when the stored collection was empty — but the empty header value is the release’s “remove all ITEM_TAGs” request instruction, so a mirroring client would read the confirmation of its own wipe as an instruction to wipe again (harmless) or, worse, treat state responses as carrying the destructive form. Both echo paths now share one rule: an empty collection emits no wrapper header.

  • Terminology failure bodies no longer disclose deployment configuration on the remaining two surfaces. A terminology 404 named WHICH configured provider answered, and a commit whose archetype constraint binding had no configured terminology route answered a 500 revealing that routing gap; both bodies now carry only what the client can act on, with the operator detail on the trace record — completing the operator-detail adjudication for the terminology surface.

  • The generated ITS-REST contract is typed end to end. The emit-rest generator resolved $refs before emitting, so every request/response body and parameter lost its schema name and the trait/DTO surface degraded to untyped JSON values; RM/BASE payload references likewise degraded on a rationale that expired with the foundation rewrite (the spec types carry emitted strict serde impls). Body and parameter references now keep their names, RM/BASE references resolve to the typed spec structs — making every DTO field strict by construction — and a discriminator.mapping schema (Versionable) emits a real _type-dispatched enum instead of an untyped alias. Remaining untyped spots are honest: anonymous oneOf responses, query result rows, and schema-less OPT objects.

  • A node claiming a _type foreign to its slot is now refused everywhere, from the RM model. The whole-instance validation pass dispatched each node on its own wire _type, so a tagged object sitting in a slot declared as something else was validated as the type it claimed to be — a DV_TEXT inside COMPOSITION.content (declared List<CONTENT_ITEM>) validated cleanly as a DV_TEXT. One model-driven rule now asserts every tagged node (root, single slots, and list members alike) conforms to its slot’s declared RM type, read from the generated BMM attribute model; a scalar member of a class-typed list slot is refused the same way. The rule never relaxes on a 553|incomplete| commit (“data may be missing, but it may not be wrong”). The hand-written FOLDER member checks this replaces are removed; their refusals now come from the general rule.

  • The conformance suite’s spec-citation gate now resolves the cited document and section. It previously took only the second whitespace token of a citation and asked whether ANY path under the component directory contained it as a substring, so a citation naming a real component plus any common word passed even when the document and §section did not exist. The gate now resolves the whole path hint to a real vendored document (or chapter directory) and verifies every §section names a real section of it — following the include:: directives that pull the UML class and interface tables into a chapter, and reading the class tables’ own labels, the markdown chapters’ titles, the AM validity-rule anchors and the OAS files’ keys. It also covers the citations of fixture-set rows and of the corpus manifest, not just case cores. The 104 citations the strengthened gate found unresolvable — phantom chapters, sections that never existed, class tables cited under the wrong directory, and one citation of an internal proposal instead of a released spec — were re-derived first-hand and corrected.

  • A vendored corpus that had no vendor script has one. The real-world canonical-JSON corpus under crates/openehr-its/tests/vendor/ was hand-downloaded; it is now reproduced byte-identically from its pinned upstream commit by scripts/vendor/openehr-sdk-json.sh (with --check to report drift and write nothing). Its provenance record also named the wrong upstream repository — a product-rename sweep had rewritten ehrbase/ to ferroehr/ in the pin — which is corrected.

  • authz.abac.enabled = true now actually enables ABAC. The server binary built an RBAC-only authorization handle unconditionally — the ABAC policy engine and its attribute resolvers were never constructed on the shipped run path, so a deployment that configured attribute-based rules ran without them, silently. The binary now boots the configured engine (Cedar or remote PDP) with database-backed attribute resolvers, logs the active authorization layers at startup, and refuses to start when an enabled ABAC block cannot be built (missing/invalid policy directory, unbuildable PDP client) — configuration that promises fine-grained authorization never degrades to authorization-off.

  • A one-group ISO OID now classifies as ISO_OID, not INTERNET_ID. BASE base_types master05-identification_package.adoc §Syntaxes gives iso_oid = number, { '.', number } — one or more groups — while the UID subtype dispatch required two, so a bare numeric root such as 12345 was tagged INTERNET_ID, whose own production it violates (a multi-character internet_id label must begin with a letter). This picks the _type on the wire, so the value now round-trips under the subtype the grammar assigns.

  • Non-finite Real values now serialize to canonical XML in the xs:double lexical form. The vendored XSDs type every Real element xs:double, whose lexical space spells the special values INF, -INF and NaN (XML Schema Part 2 §3.2.5); the serializer emitted Rust’s inf/-inf spellings, producing a schema-invalid document. Finite values are unchanged (a whole Real still writes 120.0).

  • The generated ITS-REST contract no longer drops single-$ref allOf composition. Seven DTOs the released OAS defines as a named alias of another schema — ItemTagOfComposition, ItemTagOfEhrStatus, the five demographic ItemTagOf* — degraded to an untyped string map instead of resolving to their referent.

  • The generated ITS-REST contract now includes the SYSTEM API group. The STABLE System API declares one operation, OPTIONS / (Options and Conformance), which the contract generator skipped: its group list omitted system and its HTTP-method table omitted OPTIONS entirely.

  • A composition committed against an ADL2-registered template is no longer refused with a 409, and an in-use ADL2 template now refuses physical deletion. The stored template identity on a version row was foreign-key-checked against the OPT 1.4 store only, so a commit whose template was provisioned through the ADL2 DEFINITION surface failed the constraint; the key now targets a registry spanning both template dialects. With that, the ADL2 template delete gains the same never-orphan guard the OPT 1.4 delete always had — deleting a template still referenced by committed versions answers 409 with the reference count (previously the row was deleted silently) — and it now also evicts the template’s compiled WebTemplate, so a re-uploaded template is never served from the deleted artefact’s cached form.

  • Template-scoped authorization rules now bind to compositions committed through the direct routes. A COMPOSITION committed with POST/PUT /ehr/{ehr_id}/composition stored no template identity alongside its version, while the same composition committed inside a CONTRIBUTION did — so an ABAC policy scoped to a template silently failed to match direct-route compositions (the attribute resolved to “no template” rather than to the template the composition declares). Both direct routes now record the template the version was committed against, exactly as the CONTRIBUTION route always has. Compositions committed before this release carry no template identity on their existing version rows; re-committing a new version records it.

  • A CONTRIBUTION whose audit omits change_type is now refused (422) instead of being given a server-invented one. The released commit schema makes the change set’s audit mandatory and its change_type a required member (NewContribution.yaml over UpdateAudit.yaml, for both the EHR and the demographic contribution routes), and RM common master06 §Contributions calls the contribution-level value approximate and “not expected to be used as a computable value” — it is the client’s account of its own change set. The server previously derived an aggregate from the member versions when the attribute was absent, putting an approximation into the audit trail under the client’s name; it now answers 422 naming CONTRIBUTION.audit.change_type. A conformant client is unaffected. (The direct COMPOSITION/DIRECTORY routes are unchanged: there the committal headers stay optional and the server default still applies, exactly as the ITS-REST overview requires.)

  • An undecodable request-header value is refused instead of silently ignored. A header whose bytes are not decodable as text — including the committal (openehr-version, openehr-audit-details) and item-tag wrapper headers — was dropped from the request as if it had never been sent, so a commit could carry different audit attributes or tags than the client supplied, with nothing on the wire saying so. Such a request now answers 400 naming the header.

  • An AQL query that cannot get a database connection now sheds with 503 + Retry-After instead of reporting 500. The query path’s database leg is classified like every other one, so a pool-acquire timeout is reported as a temporary overload (retryable) rather than as a server fault. A corrupt stored FHIR mapping definition is likewise reported as a server fault (500) rather than as a client error (422) — it is not something the caller supplied.

  • A round-tripped demographic party carrying inline relationships is no longer refused. PARTY.Relationships_validity requires every inline relationships[i].source to reference the party itself, and RM demographic master02 §Party Relationships requires that reference to be a HIER_OBJECT_ID denoting the party’s VERSION CONTAINER “rather than OBJECT_VERSION_IDs, which would denote particular versions” — but the check compared it against the body’s uid, which on a served party is the three-part OBJECT_VERSION_ID. The two could never be equal, so a client that read a party, added a relationship and wrote it back got 422. The comparison now uses the container id (the object_id of the version uid); a relationship sourced at another party is still refused.

  • The demographic create/update routes now honour the openehr-version lifecycle state. ITS-REST overview Requests_and_responses.md §“openehr-version and openehr-audit-details” requires that “whatever is provided it MUST be merged with the default VERSION and VERSION.audit_details attributes on commit runtime”; the direct party and party-relationship routes threaded only the audit half, so a 553|incomplete| demographic commit was reachable through a CONTRIBUTION alone. Both halves now merge on those routes (and on the SM-envelope entry points, which carry the same two attributes).

  • A read-only principal can export EXTRACTs again. POST /message/export realizes SM I_EHR_EXTRACT_SERVICE.export_ehr_extracts — a query over held versions whose selector is a whole EXTRACT_SPEC — but the read-only gate classified the extension route by its HTTP verb and refused it 403. It is now classified as the read it is (as the released ad-hoc AQL POST already was); the import routes stay writes.

  • A query result-set ETag is now stable across executions. The tag is documented as identifying the RESULT_SET (“it changes as soon as the resource changes”, overview Requests_and_responses.md §ETag and Last-Modified), but the digest covered meta._created, which is stamped per response — so every execution minted a fresh tag and conditional-request caching never hit. The digest now covers the result-determining content (q, the executed AQL, columns, rows) only.

  • A malformed server.system_id is now refused at boot. The value occupies the creating_system_id position of every OBJECT_VERSION_ID this CDR mints (BASE master05-identification_package.adoc §Syntaxes: creating_system_id = uid), but the boot check only rejected an empty value and one containing :: — so a configuration legal at startup could mint version identifiers this server’s own reader refuses. The configured value is now validated against the openEHR uid grammar itself (iso_oid | uuid | internet_id).

  • 500-class responses no longer echo internal diagnostics. A server-side fault previously rendered whatever produced it straight into the response body: serde’s parser message (naming Rust fields and byte offsets), the AQL executor’s PostgreSQL driver string (naming generated SQL and schema objects), the node codec’s RM attribute names and internal row shape, the authorization engine’s failure reason, and the XML/Simplified-Format serializers’ diagnostics. Every 500-class body now carries a curated, opaque message and the full detail goes to the server’s own log instead. 4xx refusals are unchanged and still name the client-caused defect — a malformed request payload is still refused 400 with the parse error that explains it, which is the only thing a caller can act on.

  • A tag PUT body is now validated against the released write schema. schemas/common/UpdateItemTag.yaml declares exactly key (required), value and target_path, with additionalProperties: false. Previously the body was read untyped: an undeclared member (target, owner_id, _type, anything) was silently dropped, and — worse — a value or target_path of the wrong JSON type was silently stored as ABSENT, losing a clinical annotation outright or changing the tag’s identity so a later delete addressed nothing. All three are now refused 400, naming the offending member by its JSON path, on the COMPOSITION, EHR_STATUS and all five demographic tag PUTs alike. An empty target_path still normalizes to absent, identically on both families.

  • A defective tag on a write no longer leaves the content committed. The openehr-item-tag / openehr-version-item-tag headers are now parsed and invariant-checked BEFORE the content commit, so a request carrying an invalid tag is refused with nothing created. Previously the refusal arrived after the COMPOSITION / EHR_STATUS / DIRECTORY / party version was already durable, on a response with no ETag and no Location — leaving the client no way to learn what it had just created, and no recovery but a re-POST that duplicated the content. The tag write itself still happens after the commit, so tagging continues to cause no re-versioning of content.

  • The tag response header can no longer instruct a client to wipe its tags. A valueless tag echoed as value="" (a shape the reference model forbids), and a tag list that could not be rendered as an HTTP header value — a control character in a tag key, which nothing in the reference model bars — fell back to an EMPTY header, which is exactly the byte sequence the spec defines as “remove all ITEM_TAGs”. A client mirroring that echo back on its next write would have cleared the collection. A valueless tag now echoes without a value, and an unrenderable list omits the header entirely.

  • The tag wrapper-header parser is quote-aware and no longer silently drops entries. A target_path containing a ; inside quotes (an AQL predicate, say) shattered into fragments that then parsed as garbage; quoted runs are now opaque at the entry separator. An entry carrying no key was skipped past, silently discarding a tag the client believed it had set; it is now refused 400.

  • Database integrity errors no longer leak schema names into client responses, and are no longer all reported as conflicts. Every SQLSTATE class-23 violation mapped to 409 Conflict carrying the raw PostgreSQL error text, so constraint, table and column names reached client bodies — and a CHECK or NOT NULL violation, which is a server-side invariant failure rather than anything a client can resolve, was presented as an optimistic-lock conflict to retry. Unique, foreign-key, restrict and exclusion violations keep their 409; CHECK and NOT NULL now answer 500. No branch returns a driver string: every client message is a fixed, actionable sentence, with the SQLSTATE, constraint and table recorded on the server’s own trace record.

  • A tag that survives a whole-list replace keeps its creation instant. The PUT is a full-collection replace, but re-asserting an existing tag identity is not the same as creating a new tag; previously every surviving tag’s stored creation time was reset on any edit to a sibling, which the admin export then reported. Visible through POST /rest/admin/… EHR export.

  • A CONTRIBUTION version that declares a foreign version identity is now refused (400, naming the offending key). The released commit wire declares six member properties (preceding_version_uid, signature, lifecycle_state, attestations, data, commit_audit) and no import shape at all — master06 §Copying puts the import behind commit_imported_version, whose “details of version id etc come from the ORIGINAL_VERSION”. Previously a member shaped like an IMPORTED_VERSION (_type: IMPORTED_VERSION, an item wrapping a foreign ORIGINAL_VERSION, or its own uid) was accepted and committed as a locally created ORIGINAL_VERSION under a freshly minted local identifier, silently discarding the identity and provenance the client had declared. All three keys are now refused. A member self-tagged _type: ORIGINAL_VERSION or _type: UPDATE_VERSION is unaffected — those name the class this wire commits — and importing versions that keep their foreign identity remains available through the EHR-Extract import route.

  • A version-container trunk position is now unique across creating systems. master06 §Copying has a second system BRANCH rather than extend the trunk of a copied container, and §Moving Version Containers continues the trunk increment under the new system’s id, so a trunk line is one global sequence however many systems contributed to it. The schema previously admitted two versions of one container both claiming trunk position 2, one per creating system; the archive-load path could write such a pair. It is now refused with a message naming the container, the position and the system already holding it. Branch identifiers still legitimately repeat across systems, which is what the three-part version identifier disambiguates.

  • A version that carries data can no longer claim the 523|deleted| lifecycle state (422). master06 §Logical Deletion states deletion as one procedure — create a new version, delete its data, set the state to deleted, commit — so a data-carrying deleted version is not producible by the spec’s own steps. Previously such a commit was accepted through a CONTRIBUTION member pairing a content change type with the deleted state, or through openehr-version: lifecycle_state.code_string="523" on a PUT; the resource then read back as deleted (204) while its content stayed stored and AQL-queryable. Both routes now refuse it. Deleting through DELETE, or through a data-less 523 CONTRIBUTION member, is unaffected.

  • The demographic endpoints accept PARTY_REF.type ANY. A PARTY_REF inside a party or PARTY_RELATIONSHIP body — ACTOR.roles, ROLE.performer, PARTY_RELATIONSHIP.source/target — was refused with 422 when its type was ANY, even though the composition endpoints accepted the same value. The demographic write boundary kept a second copy of the legal PARTY_REF.type set that had drifted from the single spec-cited one; it now judges every reference through that one definition, so the two surfaces give the same answer. Unknown type strings are still refused.

  • A PARTY_REF missing a mandatory attribute is refused. A reference in a demographic body without an id, namespace or type (all 1..1 on OBJECT_REF) passed the write boundary and was only caught, if at all, further in. It is now a 422 naming the missing attribute.

  • The authorization gate refuses an unaddressable resource id instead of guessing. A malformed {uid_based_id} — not a UUID, and not a well-formed three-part OBJECT_VERSION_ID — was previously read as if the whole string were the versioned-object id, so the template attribute the ABAC/SMART policy binds on silently came back empty and the request could pass a template-scoped rule. Such a request is now denied with 403, in line with the gate’s existing fail-closed handling of every other attribute-resolution failure. Well-formed ids (bare HIER_OBJECT_ID and full OBJECT_VERSION_ID, trunk or branch) are unaffected.

  • A version’s digital signature now covers the attestations it was committed with. openEHR signs “the entire Version object”, excluding only the signature attribute itself, so an attestation supplied on the commit (attestations on the committed VERSION) belongs inside the signed form. It previously did not: such an attestation could be altered in the database without any signature check noticing. It is now part of the signed canonical form at commit and at read, in both the local-commit and the EHR-Extract-import paths, so tampering with it is caught by strict read-time verification and the served version verifies for an external reader that recomputes the digest itself. Attestations added after committal (a 666|attestation| contribution) keep their existing behaviour: they post-date the signature by definition, and are served outside it. Versions committed before this change are unaffected — nothing is re-signed — unless they carried commit-time attestations, in which case their stored signature no longer matches and a strict verify_on_read deployment will report them; re-commit or set verify_on_read = warn while auditing.

  • An attestation whose optional items is sent as JSON null is no longer rejected. A null optional means absent, and the sibling optional attributes (proof, attested_view, description) were already read that way; items alone treated it as a malformed list and returned 422. That made commit-time attestations unusable through the typed API, which emits null for an omitted list. A present-but-empty list ([]) is still refused, which is what the RM invariant actually forbids.

  • A contribution mixing amendments and deletions now reports the amendment aggregate. When a client sends no contribution-level change type, the server derives one; openEHR names 250|amendment| for “a mixture of amendments and deletions that logically constitute a correction”, which previously fell through to the general 251|modification|. Uniform change sets and every other mixture are unchanged.

  • Versions received from another system are now served as IMPORTED_VERSIONs, and imported records report their local chronology. openEHR wraps a copied version: an EHR Extract import commits the received ORIGINAL_VERSION inside an IMPORTED_VERSION whose own contribution and commit audit record this server’s act of importing, while the wrapped original keeps the source system’s contribution reference, commit audit and signature. This server had never materialised the wrapper — it wrote the foreign commit audit as the version’s own and discarded the received contribution reference entirely. Four visible changes:

  • A VERSION read of an imported version (…/versioned_composition/{uid}/version/{version_uid}, …/versioned_ehr_status/version[/{version_uid}], and a resolve_refs contribution read) now returns "_type": "IMPORTED_VERSION" with the received original under item. An IMPORTED_VERSION carries no uid of its own — it shares the wrapped version’s identity — so read the version id from item.uid.value; the ETag is unchanged, so If-Match round trips are unaffected. Locally created versions are still ORIGINAL_VERSIONs.

  • An imported version container’s VERSIONED_OBJECT.time_created, its Last-Modified header, its revision history and every as-of-instant read now report the local import instant instead of the source system’s earlier clock, so a query for the record’s past state returns what this repository actually held at that time.

  • Re-exporting imported content now reproduces the received ORIGINAL_VERSION verbatim, including its source contribution reference; it previously carried this server’s local contribution id under the source version’s identity.

  • With version signing enabled, the import act signs the IMPORTED_VERSION wrapper it creates; the wrapped original’s own signature is stored and served untouched, and is never re-verified.

    An ORIGINAL_VERSION arriving in an extract without the mandatory contribution (or with a commit audit that is not a canonical AUDIT_DETAILS) is now refused with 400, instead of being imported with the provenance silently dropped. Content imported before this release keeps the provenance it was stored with.

  • A demographic version container’s owner_id now names the serving system, consistently, everywhere it is emitted. VERSIONED_OBJECT.owner_id is a mandatory reference to “the containing EHR or other relevant owning entity”, but a demographic party has no containing EHR — and this server had drifted into three different answers for it. GET /demographic/versioned_party/{uid} (and the party-relationship container read) served an OBJECT_REF in namespace demographic whose id was the container’s own uid, a self-reference that merely duplicated the sibling uid field; the X_VERSIONED_PARTY wrapper of an EHR-Extract export served namespace demographic over the system identifier; and the demographic ITEM_TAG surface already served the shape the published openEHR VersionedParty example shows. All three now emit that one shape — an OBJECT_REF with namespace: local, type: SYSTEM, and a HIER_OBJECT_ID carrying the deployment’s configured system_id. Clients that read owner_id off a demographic container will see the namespace, type and id all change; nothing else about those responses moves. The served OpenAPI example for the container read is corrected to match (it advertised a PARTY_REF, a type the published schema does not name there), the conformance catalogue now asserts the two tokens on the container read, and the adjudication is restated in the ambiguity register as AMB-69 — whose previous text incorrectly claimed this server already emitted the published shape.

  • A path predicate carrying a parenthesised uniqueness modifier is now refused instead of silently matching nothing. The Reference Model’s directory chapter shows folder paths written with a name and a bracketed uniqueness modifier (/folders[hospital episodes(car accident Aug 1998)]), a form the formal openEHR path grammar this server implements does not define. Such a predicate used to be accepted and bound whole as an archetype node id, so the path quietly resolved to nothing; it is now a loud unsupported-predicate error. Plain node-id and archetype-id predicates ([at0003], [openEHR-EHR-COMPOSITION.x.v1]) and bare name tokens are unaffected.

  • An OAuth2/OIDC committer’s identifier now names the token issuer. Every authenticated write stamps the committing principal into AUDIT_DETAILS.committer as a PARTY_IDENTIFIED carrying a DV_IDENTIFIER, whose issuer used to read ferroehr for every mechanism — including federated principals, whose subject the identity provider minted rather than this server. A Bearer principal’s identifier now carries the validated token issuer (iss) as its issuer; a Basic principal, whose credential this deployment holds, keeps ferroehr. Audits already committed keep the issuer they were written with.

  • ATTESTATION commit audits and rich audit descriptions now round-trip instead of being silently flattened. A CONTRIBUTION version whose commit_audit is an ATTESTATION — the openEHR way of committing content that is already signed, or that is marked as awaiting signature (is_pending: true) — used to be stored as a plain AUDIT_DETAILS: the concrete type and every attestation attribute (reason, is_pending, proof, items, attested_view) were dropped without an error, on the REST commit and on EHR-Extract import alike. They are now decoded, validated against the RM invariants, stored, and served back as an ATTESTATION on the version envelope, in the revision history, in the CONTRIBUTION rendering, and in exports/archives. A commit_audit whose _type names neither AUDIT_DETAILS (UPDATE_AUDIT) nor ATTESTATION (UPDATE_ATTESTATION) is now refused with 422 instead of being read as a plain audit. AUDIT_DETAILS.description is likewise kept whole: a DV_CODED_TEXT description keeps its defining_code instead of being reduced to its display string, and AQL can now address commit_audit/description/value, commit_audit/description/defining_code/code_string and .../defining_code/terminology_id/value as distinct fields (a coded description could previously never match). The audit table’s baseline schema changes with this: description becomes jsonb (the whole DV_TEXT) and a nullable attestation column is added.

  • A coded description on an attestation keeps its code. The 666|attestation| commit path completed a submitted UPDATE_ATTESTATION by reducing its description to the plain text of a DV_TEXT, so a DV_CODED_TEXT description lost its defining_code permanently at committal — while the same attribute on a version’s own commit_audit was already kept whole. An attestation’s description is now stored and served back exactly as submitted (_type, display value and defining_code), and a description that is neither a string nor a valid DV_TEXT is refused with 422 instead of being dropped.

  • A simplified-format ctx participation without a function is now refused. ctx/participation_* keys build EVENT_CONTEXT.participations, whose function the Reference Model requires; a FLAT/STRUCTURED commit that began a participation at some index (a name, an id, a mode or identifiers) but supplied no ctx/participation_function:<i> used to be completed with an empty function, committing a participation whose mandatory attribute carried no information. It is now rejected with an error naming the exact missing key (e.g. ctx/participation_function:0 is required).

  • The FLAT _link:i builder now reports a missing mandatory suffix instead of inventing an empty value. |meaning, |type and |target are all required on a simplified-format _link:i datum; a submission omitting one used to be silently completed with an empty DV_TEXT/DV_EHR_URI, storing data the client never sent. The conversion is now refused with an error naming the exact key (e.g. .../_link:0|meaning is required).

  • Compositions and directories that contradict the RM’s archetype-root rule are now refused (422) instead of stored. At an archetype root the archetype_node_id is the archetype identifier in string form, so a node carrying archetype_details whose archetype_id names a different archetype declares two conflicting identities and can no longer be committed. Payloads that were accepted before and are affected by this must correct the mismatched root before they will commit.

  • An empty archetype_node_id is now refused (422) on every node type. The RM requires a non-empty archetype_node_id on every archetypable node, but the check only ran on the node types with a hand-written invariant (ENTRY subtypes, CLUSTER, ELEMENT, SECTION, FOLDER, HISTORY and events). It is now applied to every RM type that inherits it — notably ITEM_TREE, ITEM_LIST, ITEM_SINGLE, EHR_STATUS, and the demographic and EHR-extract locatables — so a payload carrying "archetype_node_id": "" anywhere is rejected rather than stored. An absent archetype_node_id is unchanged: it is still reported as a missing mandatory attribute.

  • A present-but-empty links list is now refused (422). links is optional, but the RM forbids it from being present and empty, so "links": [] on any node of a committed COMPOSITION — or on any FOLDER of a committed directory — is now rejected rather than stored. Omit the attribute instead of sending an empty array.

  • RM class invariants are now enforced on every commit kind, not only on COMPOSITIONs. The whole-instance RM + terminology pass previously ran only for compositions, so anything below the root of an EHR_STATUS, EHR_ACCESS, directory FOLDER, party or party-relationship body went unchecked. Payloads that were accepted before and are now refused with a 422 include: an empty archetype_details.rm_version; a links member that is missing meaning, type or target, or whose target is not an ehr:// URI; a present-but-empty links list on a node nested inside EHR_STATUS.other_details or a party body; and an empty feeder_audit system_id. The same defects were already 422 inside a COMPOSITION. EHR_ACCESS.settings is deliberately unaffected — the RM leaves that slot’s type to the implementation, so it carries no RM rules to enforce.

Security

  • Admin dump/load failures no longer expose server filesystem paths. The file_not_writable and container-fault bodies of the EHR export/import operations carried the configured archive path (server deployment layout) and the raw archive-parse diagnostics. The bodies now carry a curated message — a defect in the CALLER’s archive still names the offending archive ENTRY, which is the actionable fact — and the path plus the underlying diagnostic go to the server’s trace record only.

  • Terminology-server failures no longer expose the deployment’s terminology configuration. A 500 raised by an upstream FHIR terminology server (or by its OAuth2 client-credentials grant) named the configured provider, its operation and the upstream error in the response body. The body is now the curated internal-error message; the operator detail (provider name, operation, upstream diagnostic) is emitted on the trace record. Boot-time configuration errors are unchanged — they never reach a response body.

3.17.1 - 2026-08-01

Added

  • Deprecated ADL 1.4 spellings are warned at ingest. The paren-less domain-block spelling (C_DV_QUANTITY <…>) is marked deprecated by the ADL 1.4 text while its normative grammar defines only that form; it stays accepted (refusing it would reject the whole published CKM library), and every occurrence now raises the advisory W14DEP validation warning naming the preferred parenthesised spelling — the deprecation is enforced at exactly the strength the spec gives it, never silently absorbed.

  • The largest OPT the official CKM publishes is now a conformance scale probe. The Congenital syphilis case investigation form (~5.2 MB of OPT 1.4 XML, the biggest template openEHR publishes) joins the curated CKM journey pack, and two new asserted CNF cases pin size-independence of the released contracts — the template upload and a COMPOSITION commit of its ~526 KB example must succeed and read back content-equal, because the released ITS-REST text defines no payload-size condition, so an implementation cap that refuses, truncates, or corrupts a large valid payload is a conformance defect. The template also rides the measured hospital simulation’s case-investigation journey, so every future performance record exercises the large-payload path.

  • ADL 2 (AOM2) archetypes can arrive as XML, in both serializations openEHR publishes. Only the AM 1.4 archetype XML was readable before; the generated canonical-XML codec now also covers the AOM2 persistent form (P_Archetype.xsd, root <archetype> = P_AUTHORED_ARCHETYPE — the shape the 8 example documents in the openEHR ITS-XML bundle carry) and the AOM2 model form (Archetype.xsd, root <archetype> = AUTHORED_ARCHETYPE). The two are separate codecs because the two schemas declare the same top-level element with different root types.

Fixed

  • openehr-its builds again for minimal consumers. The opt14 module (OPT 1.4 model + XML codec) was declared without the full feature gate its codec dependency carries, so any default-features = false consumer of the crate — whose documented minimal surface is only the dependency-free SMART scope grammar — failed to compile with 1,191 errors (first visible on the admin console’s WebAssembly lane). The module is now gated like its siblings; default builds are unchanged.

  • Empty inline dADL domain blocks parse (C_DV_QUANTITY <>). The ADL 1.4 reader refused an empty domain block outright; the dADL chapter’s own grammar admits the empty block and §Empty Sections allows it anywhere, so it now lowers to the open constraint (DV_QUANTITY matches {*}) in both the deprecated paren-less and the parenthesised spelling. 9 live CKM archetypes use the form. (The upstream regression fixture claiming the opposite is reported as a spec contradiction.)

  • The ADL 1.4 → 2 converter no longer emits unparseable tuples for heterogeneous C_DV_QUANTITY list rows. Rows constraining different member sets (a units+magnitude row beside units-only rows, as in the published range_of_motion archetype) used to convert into one tuple with EMPTY members ([{"mm"}, {}]) that no conformant ADL 2 reader accepts — 28 of the 1,142 published real-world archetypes were affected. Such rows now partition into one DV_QUANTITY alternative per member set, each tuple row constraining every member; co-constrained pairings and assumed values are preserved, and the whole real-world corpus now converts and re-parses.

  • ADL 1.4 archetypes that openEHR CKM actually publishes are no longer rejected over two spec-silent shapes. Uploading a real-world ADL 1.4 archetype failed to parse when it used either (a) a qualified term constraint with an empty code list — defining_code matches {[local::]}, media_type matches {[openEHR::]}, which names the terminology and constrains the code no further — or (b) the openEHR-profiled ordinal shorthand alongside another alternative in the same block, in either order (DV_TEXT matches {*} beside 0|[local::at0008], 1|[…]). Both forms are read by the reference grammar and are used throughout CKM’s own library, so the reader now accepts them; every other refusal is unchanged.

  • Mandatory 1..* container attributes are enforced. The canonical-JSON reader deliberately treats an absent list and an empty list as the same value (wire tolerance), so a committed CLUSTER with no items, a demographic PERSON with no identities, or a REVISION_HISTORY with no items was accepted although the RM declares those containers 1..*. The validation walk now checks every model-declared mandatory container — absent, or empty where the lower bound is 1 — uniformly from the generated RM model, and such commits are refused 422.

  • Structurally defective RM nodes are now refused for EVERY openEHR class, not only the ones carrying a class invariant. The per-node validation step deserialized a wire node into its concrete RM type — the check that surfaces a missing mandatory attribute or a wrong nested type — only for the classes with an RM invariant to run, so a defective node of any other type passed unnoticed: a PARTICIPATION with no performer, an ISM_TRANSITION with no current_state, a LINK with no target, a DV_BOOLEAN with no value, the demographic ADDRESS/CONTACT/PARTY_IDENTITY shapes, and more. The class-to-type dispatch is now generated from the openEHR meta-model and covers every emitted class, so such a commit is rejected with 422 naming the offending path and field. Wires that were already valid are unaffected.

  • A generated example COMPOSITION for a template that constrains other_participations is committable again. The example carried a PARTICIPATION without its RM-mandatory performer (plus a stray name, which PARTICIPATION does not have), so posting the generated example back was rejected. The participation is now completed with an identity-free PARTY_SELF performer — no fabricated person — and carries no name.

3.17.0 - 2026-08-01

Changed

  • The documentation website moved to its own domain, https://ferroehr.eu/. It was published at rubentalstra.github.io/ferroehr/, a GitHub Pages project sub-path; it is now served at the root of the ferroehr.eu apex over HTTPS, with www.ferroehr.eu redirecting to it. Every published URL loses the /ferroehr prefix — the user guide is at /docs/latest/, the OpenAPI endpoint reference at /api/. The old GitHub Pages URLs keep working through GitHub’s own redirect, so existing links and bookmarks do not break. This affects the website only; the CDR’s REST base path (/ferroehr/rest/openehr/v1) is unchanged.

Fixed

  • RM validation reaches legally untagged canonical-JSON nodes. Canonical JSON requires _type only on polymorphic slots, so a node under a concretely-declared attribute (COMPOSITION.context, EVENT_CONTEXT.participations, …) may omit its tag — and the validation walk, dispatching on the wire tag alone, silently skipped every RM class invariant and terminology binding on such nodes (the same content was refused over canonical XML but committed over JSON). The walk now resolves an untagged node’s effective RM type from the parent’s declared attribute type (the BMM-generated static RM model), so commits like an out-of-group EVENT_CONTEXT.setting are refused 422 regardless of tag presence or format.

  • The documentation version switcher pointed at dead URLs. Entries in the published version manifest kept whichever site base path was current when they were archived (/ehrbase-rs/… before the rename, /ferroehr/… before the domain move), so selecting an older release led to a 404. Each entry is now re-anchored to the live base at build time, which repairs every archived version without rebuilding the frozen documentation trees.

  • EXTRACT_SPEC.extract_type accepts the TERM extract_content_type vocabulary. The EHR-Extract export refused the codes TERM 3.1.0 binds to the attribute (803 “openEHR EHR” … 808 “other”, openEHR terminology — SupportTerminology §Vocabularies) and accepted only the RM-named string tokens (openehr-ehr, …). Both value spaces are now accepted; anything outside their union is still refused with a 400-family precondition error.

3.16.0 - 2026-08-01

Added

  • ODIN plug-in syntax blocks parse. attr = (syntax) <# … #> (master09-plug_in_syntaxes) is accepted, with the tag and the verbatim foreign-text body exposed as OdinValue::PlugIn; the body is handed to consumers uninterpreted, and a plug-in block is refused as an archetype default_value (it denotes no RM instance).

  • ODIN document artefacts parse in all three master04 forms. The ODIN reader now accepts Identified Object Documents (top-level ["id"] = <…> keyed objects) and the optional leading @schema = <uri> schema identifier, exposed via the new odin::parse_document; the anonymous and implicit forms were already supported.

  • ODIN cross-object references parse. Reference paths rooted at an object identifier (<["tourism_db_13"]/hotels["sofitel"]>, master06-references §Across Objects) are accepted alongside the existing within-object reference forms.

  • ODIN path extraction. OdinValue::paths() extracts the master05-content tree-path set (attribute paths, attr[key] container paths, nested bare-key segments) from any parsed ODIN structure.

  • The BMM v3 behavioural surface: type lattice, class/feature functions and a v3 materialisation. openehr-lang now answers the master06-core-types meta-type lattice (is_abstract, is_primitive, type_base_name, unitary_type, effective_type, effective_base_class, is_open/is_closed/is_partially_closed), the master07/master08 class and feature functions (BMM_SIMPLE_CLASS.type, BMM_GENERIC_CLASS.type + generic_parameter_conformance_type, has_ancestor_class, all_ancestors, flat_features, BMM_ENUMERATION.name_map, signature, arity, is_boolean), and materialises a v3 BMM_MODEL from a P_BMM schema (create_bmm3_model) — which lands three things the v2.x transform cannot: a generic ancestor’s parameter substitution, a class’s routines and constants, and a value_constraint on the type it constrains.

Fixed

  • A malformed enumeration is refused instead of loaded silently. A P_BMM enumeration with more than one ancestor, or with item_values that are not 1:1 with item_names, now fails model materialisation with a typed error (EnumerationAncestorCount, EnumerationItemListsNotOneToOne) — the two rules master07-core-classes §Range-Constrained Classes and the BMM_ENUMERATION class definition state. Stating names without values stays valid (the assumed values 0, 1, 2, … apply).

  • Container literal values are typed on the wire. BMM_CONTAINER_VALUE and BMM_INDEXED_CONTAINER_VALUE bound their inherited type slot to free-form JSON; they now carry BMM_CONTAINER_TYPE / BMM_INDEXED_CONTAINER_TYPE, so a container literal’s type member is a _type-tagged container type in canonical JSON instead of an arbitrary value. The two slots openEHR genuinely leaves untyped (BMM_INTERVAL_VALUE.type, EL_CASE.value_constraint) now carry their adjudication as a generated NOTE at the field.

  • The generated LANG model carries both extant BMM generations in full. openehr-lang was generated from the two vendored LANG schemas name-merged into one class map, so 18 class names both declare emitted the stable v2.x shape at the v3 (bmm3) module paths and the v3 attribute sets were discarded — a BMM_CONTAINER_TYPE had no is_ordered/is_unique, a BMM_CLASS none of its feature/invariant maps, and BMM_MODEL_TYPE + BMM_MODULE were never emitted at all (185 of 187 declared classes). Both generations are now emitted completely, each at its own source-package path (v2.x under bmm/, bmm_persistence/, beom/; v3 under bmm3/), and the canonical-JSON codec covers both.

  • ODIN leaf lists are type-homogeneous and admit interval lists. Mixed-kind lists (<1, "x">) are refused per the per-type list productions of the ODIN syntax specification, and interval lists (<|0..5|, |8..9|>, including the open , ... form) now parse.

  • Duplicate ODIN container keys are refused (rule VDOBU). Sibling keyed objects sharing a key ([1] = <…> [1] = <…>) were silently accepted; they now fail with a typed error per LANG/docs/odin/master05-content §Container Objects — archetype uploads whose terminology sections duplicate a term code are refused at the ODIN layer.

  • ODIN sections accept true/false/infinity as attribute names and semicolons between keyed objects. ODIN reserves no keywords (LANG/docs/odin/master03-basics §Keywords), so archetype/template uploads whose ODIN sections use those three words as attribute names — previously refused because they lex as ODIN value words — now parse; and the §Semi-colons separator is accepted between keyed-object entries, not only attribute pairs.

3.15.3 - 2026-07-31

Fixed

  • The quickstart dev credentials authenticate again. The v3.15.2 rename updated the dev usernames and documented passwords to ferroehr but left the committed Argon2id hashes (and the dev Keycloak realm’s stored password hashes) verifying the old secret, so every Basic-auth request against the compose stack returned 401. The dev user store now carries hashes of the documented password, and the Keycloak realm import sets the dev passwords at import time. Development stacks only; no production surface stores these credentials.

3.15.2 - 2026-07-31

Changed

  • The project’s own code is now MIT-licensed (owner decision 2026-07-31). Vendored third-party material keeps its upstream terms: the openEHR machine-readable specification artifacts and CKM-derived clinical models remain under Apache-2.0 (LICENSE-APACHE-2.0). The upstream NOTICE file was removed together with the relicense — no upstream code is present in this tree.
  • The product is named FerroEHR (owner decision 2026-07-31, tracked on #1353 — from ferrum, iron, the element Rust is named for). Every product-branded surface changes with it; deployments upgrading across this release must update:
    • Configuration: the environment prefix EHRBASE_* → FERROEHR_*; the config file search path ./ehrbase.toml / /etc/ehrbase/ehrbase.toml → ./ferroehr.toml / /etc/ferroehr/ferroehr.toml.
    • REST base path: /ehrbase/rest/openehr/v1 → /ferroehr/rest/openehr/v1 (likewise the admin/management extension routes under /ehrbase/…).
    • Binary and crates: the server binary ehrbase → ferroehr; the application crates ehrbase/ehrbase-rest/ehrbase-server/ ehrbase-admin-ui → ferroehr/ferroehr-rest/ferroehr-server/ ferroehr-admin-ui. The generated openehr-* specification crates are unaffected (they are versioned by the openEHR spec they implement).
    • Containers and Helm: the OCI images and the Helm chart are published under ferroehr/ferroehr-admin-ui; compose service names and the dev-stack database/realm names follow.
    • Repository: the GitHub repository is now github.com/rubentalstra/FerroEHR (old URLs redirect).
    • Conformance artifacts: the SUT identifier ehrbase-rs → ferroehr (docs/conformance/ferroehr/); measured numbers are unchanged.
    • The startup banner, served OpenAPI title, telemetry service.name, and the documentation website carry the new name and logo. The openEHR wire behaviour itself (canonical JSON/XML, AQL, status codes/headers) is unchanged.

3.15.1 - 2026-07-31

Added

  • An archetype or template may now give an interval as a node’s default value. ADL 2 lets a _default block hold any ODIN value, and ODIN counts intervals of the ordered types (|0..5|, |>=1939-02-01|, |<10.5|, |5.0 +/-0.5|, the single-value |5|) among those values. Such a default used to be rejected outright; it is now read, stored as a proper interval with its own bounds and open/closed flags, and written back out in the same interval syntax. A centre +/- delta interval over dates, times or durations is still refused — reducing it to bounds would need calendar arithmetic the source does not state — with a message that says so.

Fixed

  • Querying a parent archetype now also finds data recorded under its ADL 2 specialisations. An AQL archetype predicate naming a parent used to recognise a specialisation child only when the child’s concept extended the parent’s with a hyphen — the ADL 1.4 naming convention, which ADL 2 dropped. The server now reads the specialisation lineage from the ADL 2 archetypes and templates you have uploaded, so a query for the parent returns data committed under any of its stored specialisation children (and their children in turn), whatever their concept names are. Hyphenated ADL 1.4 identifiers keep their existing behaviour; a hyphenated ADL 2 identifier is no longer treated as a specialisation, because in ADL 2 the hyphen carries no such meaning. A parent with no uploaded family still matches only itself.
  • A CONTRIBUTION commit response now tells the client when the change set was committed. POST /ehr/{ehr_id}/contribution (and its demographic sibling) returned only the ETag/Location identity; it now also sends Last-Modified, carrying the commit audit’s recorded time. The header is present under both Prefer: return=representation and Prefer: return=minimal — on the minimal branch, where the response has no body, it is the only place the commit instant appears at all.
  • A backslash sequence the ADL/ODIN escape rules do not define is now refused with a clear message instead of read as literal text. The escape set is closed — \r, \n, \t, \\, \", \' and the two \u unicode forms — and anything else in a quoted string (a stray \q, a regex class such as \d written outside a regex, a string ending in a lone backslash, a \u with the wrong number of hex digits) is an authoring defect. Such a string used to be carried through as-is, so the defect reached the stored archetype as text; it is now reported at the offending literal. Regular expressions are unaffected: the backslash patterns inside a matches {/…/} constraint are still passed to the regex engine untouched.
  • A type cast written with its package path now parses. ODIN and ADL 1.4 dADL allow a type identifier to be qualified with dot-separated package names — (org.openehr.rm.ehr.content.ENTRY), (Core.Abstractions.Relationships.Relationship), and the same inside a generic such as (List<org.openehr.rm.ehr.content.ENTRY>) — which is how authors disambiguate same-named types from different models. Such a cast used to fail the parse and reject the whole archetype. The qualified name is kept exactly as authored; where the cast becomes a JSON _type tag, the class name is used.
  • Archetype text that carries a non-BMP character now reads correctly. ADL and ODIN allow a unicode character above the base multilingual plane to be written as an eight-hex-digit \uHHHHHHHH escape (emoji, historic scripts, rare CJK). Those escapes were accepted but decoded wrong — the cADL reader kept the escape as literal text and the ODIN reader produced nothing usable — so the character was silently lost. Both forms now decode, in archetype definitions, ODIN sections and rule expressions alike, and an escape that names no character (a value outside the range the eight-digit form covers, a broken surrogate pair) is reported as a syntax error at the offending literal instead of being silently substituted.
  • Uploading an ADL 2 archetype now runs the complete validation schedule. The parse-and-validate path skipped the final, flat-form phase, so two defects could pass: an internal use_node reference whose target path does not exist, and a container whose cardinality cannot hold the child nodes it declares as mandatory. Both are now reported. Archetypes that were valid remain valid — including specialised archetypes that redefine one parent node into several children, whose node identifiers are no longer mis-reported as duplicates.
  • A fixed numeric value written as a closed range now behaves like the point value it is. {5..5} and {5} are the same constraint, and an ADL 1.4 assumed value spelled either way is read identically; conversely, a “point” whose bound is open or unbounded is no longer treated as a fixed value.

3.15.0 - 2026-07-30

Changed

  • Release binaries are hardened for diagnosability: production panics now carry file:line backtraces (debug = "line-tables-only"), and the unwind panic strategy the clean-500 error contract depends on is pinned explicitly in the release profile so it can never silently regress to abort. Runtime behaviour is otherwise unchanged.

  • The workspace now enforces the official Rust best-practice baseline as compile-time policy (Rust API Guidelines, Clippy/rustdoc/Cargo books): every public item is documented (including the generated openEHR spec crates, whose docs now come from the BMM meta-model), panicking indexing and string slicing are compile errors outside tests, wall-clock/env/UUIDv4 API bans are machine-enforced, lint suppressions must carry a machine-read reason, and doc links are verified by a new rustdoc CI gate. No wire or storage behaviour changes.

  • An ADL 1.4 upload is now validated as ADL 1.4, not as a permissive superset of ADL 2. ADL 1.4’s cADL keyword set is closed (master05 §Keywords), so constructs that only ADL 2 defines are refused in a 1.4 text with a syntax error naming the construct: use_archetype, the archetype-slot closed marker, the _default pseudo-attribute, second-order attribute tuples, term-constraint strengths (required/extensible/preferred/ example), and @terminology operational bindings. before/after sibling order stays accepted — master05 lists both as ADL 1.4 cADL keywords. ADL 2 uploads are unaffected.

Fixed

  • A malformed deprecated concept section in an ADL 2 archetype is now rejected with the SACO syntax code (“must consist of the ‘concept’ keyword and a single local term”) instead of being ignored silently; well-formed deprecated concept sections stay accepted and ignored.

  • A use_node internal reference targeting its own ancestor, or another internal reference, is now rejected at validation (VUNP for ADL 2, VDFPT for ADL 1.4): an ancestor target defines an infinitely recursive expansion and previously validated clean; sibling and cross-branch targets are unaffected.

  • Illegal backslash escapes in ADL character values and rules-section strings are now rejected at parse (only \r, \n, \t, \\, \", \' are legal quoted forms, plus \uHHHH unicode escapes in strings): previously a form like '\q' was accepted silently across the ODIN, assertion and ADL lexers. Unicode content and the legal forms parse as before.

  • Converting an ADL 1.4 archetype to ADL 2 now carries its extended meta-data across. The standardised description/other_details items of ADL 1.4 App.B (Extended Meta-data Guide) — items “intended to be implemented by any ADL 1.4 => ADL 2 conversion tool” — previously survived conversion only as opaque other_details strings (only revision was converted). They now land in their ADL 2 homes: build_uid becomes the archetype’s build identifier; original_namespace, original_publisher, custodian_namespace, custodian_organisation and licence become the matching description attributes; and references / ip_acknowledgements become keyed lists, one entry per line with surrounding whitespace stripped. Each converted item is removed from other_details; a value that does not match its documented syntax (e.g. a build_uid that is not a GUID) is left in other_details untouched rather than guessed at, and the guide’s “other items” (MD5-CAM-1.0.1, current_contact, review_date, responsible_organisation) pass through unchanged as before.

  • ADL 1.4 archetypes using the chapter’s custom constraint forms now upload. Two constructs of ADL 1.4 §Customising ADL (master09) were rejected outright:

    • the inline dADL C_CODE_PHRASE <…> section — the chapter’s own worked example — is now read and lowered to exactly the constraint its compact [local:: at0039, at0040] twin produces (the chapter presents them as two spellings of the same constraint), including an assumed_value CODE_PHRASE; a block that is not a C_CODE_PHRASE instance (no terminology, no or empty code list, an attribute the type does not define) is refused with a syntax error naming the defect instead of being guessed at;
    • the openEHR-profiled ordinal shorthand 0|[local::at0005], 1|[local::at0006] — ubiquitous in real 1.4 scores and scales, and optionally carrying a ; assumed value — now parses in ADL 1.4 and is lowered to the generic DV_ORDINAL [value, symbol] tuple that ADL 2 names as its replacement. ADL 2, which removed the form, still refuses it.

    Both forms’ codes take part in validation exactly as the equivalent standard spellings do (an undefined at-code still raises VATDF). C_DV_STATE, the remaining custom constrainer, has no shape in any openEHR specification and stays a loud refusal naming the type.

  • ADL 1.4 archetypes whose section keywords are not all-lowercase now upload. The ADL 1.4 lexical specification (master08 §Symbols) spells every section keyword case-insensitively, so ARCHETYPE (adl_version=1.4), Specialise, CONCEPT, DEFINITION and ONTOLOGY are valid headers; they were previously rejected with “expected an artefact keyword” / “expected a section header”. ADL 2 keeps the exact lowercase spelling its own grammar defines.

  • Old-form ADL 1.4 archetypes with no language section now upload. Where the archetype carries primary_language/languages_available in its ontology section instead — the form master08 §Language Section and §Ontology Header Statements tell tools to accept and upgrade — the language is now lifted into original_language plus one translation entry per other available language, instead of the upload failing with “no language section found”. An archetype with neither a language section nor a primary_language is still rejected.

  • A missing or undefined ADL 1.4 concept section is now reported instead of passing silently. The 1.4 grammar makes the concept section mandatory and master08 §Validity Rules VARCN requires its term to exist in the archetype ontology: an archetype with no concept section (or a concept clause that is not a term-code reference) is refused with a SACO syntax error, and a concept term missing from term_definitions now raises VARCN on the 1.4 validation path.

  • ADL 1.4 assertion expressions now accept the chapter’s full operator set: the ADL 1.4 inequality spelling <> and the symbolic existential quantifier ∃ applied to a path (equivalent to the exists keyword per the assertion-language symbol table) both parse in invariant/rules sections instead of being rejected.

  • ADL 1.4 assertions now accept every path form the ADL paths grammar defines: movable path patterns (leading //) and single-segment relative paths with a node predicate (items[at0001]) parse instead of being rejected; absolute/relative multi-segment paths and the position/meaning/at-code predicate forms were already accepted.

  • ADL 1.4 use_node internal references with a target path that does not resolve within the definition are now reported as a VDFPT validation error (path validity in definition) instead of passing silently; a resolving target stays clean.

  • A server-side failure of the password-verification task no longer masquerades as 401 Invalid credentials: if the blocking Argon2 verification task itself fails (panic/cancellation), the API now returns 500 Internal Server Error. A wrong password still returns 401.

  • Basic-authentication credentials are now decoded by the pinned RFC 4648 base64 crate instead of a hand-rolled decoder. Canonical (padded) and unpadded credentials are accepted as before; malformed base64 with excess or interior padding is now rejected outright (it previously decoded to garbage and failed the user lookup — the response stays 401 either way).

  • ADL 1.4 archetypes carrying a revision_history section now parse and upload. The section is defined by the ADL 1.4 specification (master08 §Revision History Section) but was not recognised, so an entire spec-valid archetype was rejected with a “expected a section header” syntax error. It is now read and preserved on the 1.4 source model. (ADL 2 removed the section, so it has no ADL 2 counterpart; a 1.4 upload stores the source verbatim, so nothing is lost.)

  • The dADL/ODIN reader accepts the leaf and structure forms the ADL 1.4 specification defines (master04 §dADL), which were previously rejected: the <...> empty-section marker at any level; the whole partial date/time family (yyyy-MM-ddT??:??:??, yyyy-MM-??T??:??:??, yyyy-??-??T??:??:??); integers written with an exponent (29e6); booleans in any case (TRUE, fAlSe); infinity / -infinity / * as unbounded interval endpoints; and local term codes ([at0200]) as leaf values.

  • Values that were silently mis-read are now read correctly or rejected loudly. A (TYPE)-cast section value is read through instead of being dropped; an |N +/- M| domain constraint becomes the interval [N-M, N+M] instead of collapsing to the centre; a duplicate sibling attribute is a typed error naming the attribute instead of the last one silently winning (rule VDATU); duplicate keys in the language section are now reported (VOKU); and an interval used as a _default value is rejected instead of silently becoming a null default.

  • Multi-line string values drop their continuation-line indentation, as the specification requires (master04 §String Data), instead of carrying the source file’s leading whitespace into the value.

  • An inline ADL 1.4 domain constraint is no longer lowered to the wrong data type. (TYPE) <…> domain blocks (ADL 1.4 master09 §Customising ADL) other than C_DV_ORDINAL were all turned into a DV_QUANTITY, so e.g. a C_CODE_PHRASE coded-term constraint silently became a quantity constraint. C_DV_QUANTITY and C_DV_ORDINAL are lowered as before; any other domain type is now a syntax error naming the type instead of a wrong answer.

  • An inline ADL 1.4 domain constraint’s assumed_value is no longer dropped. It is now carried onto the constrained leaves (and an assumed value that satisfies none of the block’s own list rows is rejected instead of silently ignored).

  • ADL 1.4 coded-term constraints are checked for definedness. The dominant ADL 1.4 spelling of a value set ([local:: at0004, at0005; at0004], master09 §Custom Syntax) bypassed the term-definedness rules entirely, so an archetype referring to codes its own ontology never defines uploaded cleanly. Every listed code — and the assumed-value code — is now checked (VATDF/VACDF, ADL 1.4 master08 §Validity Rules); codes of an external terminology are correctly exempt.

  • The ADL 1.4 cardinality/occurrences rule VCOC is enforced (master05 §Occurrences): a container whose children’s occurrences cannot fit inside its cardinality is now rejected. The ADL 1.4 default occurrences {1..1} and existence {1..1} (master05 §Occurrences / §Existence) are applied when evaluating it, and the default occurrences is now written out explicitly by the ADL 1.4 → 2 conversion, where an unstated occurrences means something different.

  • use_node type conformance is enforced (VUNT). An internal reference whose declared type is neither the referenced node’s type nor an ancestor of it is now a validation error instead of passing silently. It is also reached by an ADL 1.4 upload: VUNT is a rule of the ADL 1.4 specification itself (master05 §Internal References), so the 1.4 validity check now runs the reference-model pass for that rule instead of stopping before it.

  • cADL keywords are recognised in any case. MATCHES, Occurrences, CARDINALITY, Is_In, TRUE and every other keyword are now lexed as keywords, as both the ADL 1.4 specification’s own lexical rules (master05 §Symbols) and the normative ANTLR grammars require; previously only the lower-case spelling worked, so an upper-case archetype failed to parse.

  • infinity is accepted as an interval bound in cADL constraints. rate matches {|0..infinity|} — the ADL 1.4 specification’s own worked example (master05 §Interval of Integer) — parsed in a dADL section but was rejected in a cADL constraint. -infinity and * are accepted likewise, and each yields a genuinely unbounded endpoint.

  • The ^…^ regular-expression delimiter survives a re-print. A constraint written {^km/h|mi/h^} (master05 §Regular Expression) was normalised to {/km/h|mi/h/}, which no longer re-parsed; the inner delimiters are now escaped, so parse → print → parse is lossless.

  • More date/time constraint-pattern forms are accepted: patterns with literal date/time numbers substituted for the placeholder fields (1995-??-XX, master05 §Patterns), the ASCII timezone modifiers +hh:mm / +hhmm / -hh (previously only the literal ± character worked), and the space-separated date/time pattern (yyyy-mm-dd hh:mm:XX), which the specification’s own assumed-value example uses.

  • Character constraints are accepted (color_name matches {'r','g','b'}, master05 §Constraints on Character), as are the , ... list-continuation marker on every primitive list and the exclusive-lower interval spelling the ADL 1.4 chapters write (|0>..<1000|).

  • An inline ADL 1.4 domain block containing a one-sided interval now parses. A C_DV_QUANTITY <… magnitude = <|>0.0|> …> block was cut short at the interval’s own >, so the archetype was rejected as invalid dADL.

  • Defects inside an ADL 1.4 listed term constraint are now reported: a repeated code (STCDC) and an assumed-value code that is not a member of the list (STCAC).

  • A date/time interval must use a timezone on both endpoints or on neither (master05 §Intervals), so two endpoints that cannot be compared are rejected instead of silently accepted.

  • Operators the ADL 1.4 chapter names but no grammar defines are refused with a message that says so — the negated ~matches / ~is_in / ∉ family and the =~ / !~ regex-match operators — instead of being read as their affirmative counterparts, which would invert the constraint.

3.14.0 - 2026-07-30

Fixed

  • AQL VERSION coded-field predicates respect their sub-paths. Predicates on commit_audit/change_type and lifecycle_state compared every sub-path against the stored numeric code, so the rubric form (…/value='creation') silently never matched. The three defined sub-paths now compare correctly (defining_code/code_string against the code, value against the terminology rubric, terminology_id/value against openehr); other suffixes are clean invalid-query rejections.
  • AQL SELECT DISTINCT with ORDER BY executes correctly. Sorting a DISTINCT projection by one of its selected columns previously failed with a database error surfaced as HTTP 500; it now orders by the output column. Sorting a DISTINCT projection by an expression that is not selected is a clean invalid-query rejection (the AQL specification defines no semantics for it) instead of a 500.
  • AQL date/time functions work in temporal comparisons. Comparing a temporal path against NOW()/CURRENT_DATE_TIME etc. previously failed with a database type error surfaced as HTTP 500; function operands now join the comparison in the same coercion space as literals.
  • Comma-fraction ISO 8601 timestamps compare correctly in AQL. Canonical DV_DATE_TIME values using the ISO-permitted comma decimal sign (21:22:19,501+00:00) were silently excluded from temporal comparisons (and their promoted index column stored NULL). Both the write-time promotion and the query-time casts now normalize the comma form.
  • AQL coded-name node predicates match correctly. The name term-code shortcut ([at0002, snomed_ct(3.1)::313267000], and the terminology::code|informational text| form) was compared as one raw token against code_string and could never match. It now decomposes per the AQL specification’s canonical expansion: code_string and terminology_id/value are compared separately, the informational |…| tail is ignored, and a bare at-code name operand asserts the archetype’s local terminology.

Changed

  • TOP n BACKWARD is now rejected with rewrite guidance. The deprecated direction variant previously returned the first n rows silently. The server now refuses it as an invalid query whose message shows the recommended rewrite (ORDER BY <path> DESC LIMIT n). Plain TOP n and TOP n FORWARD are unchanged.

3.13.0 - 2026-07-30

Added

  • The ISO 8601 date/time/duration types implement their computational functions (BASE foundation_types/master06-time_types.adoc §Computational Functions + the four Iso8601_* class definitions): the DEFINITE add/subtract/diff on dates, times and date/times — a duration reduced to exact seconds with the Time_definitions Average_days_in_year/Average_days_in_month lengths — and the NOMINAL add_nominal/subtract_nominal, which advance the calendar to the same day-of-month and clamp it down where the target month is shorter (29 Feb ++ P1Y → 28 Feb, 31 Jan ++ P1M → 28/29 Feb). Durations gain add/subtract/multiply/divide/negative. Also added across the four types: as_string (the value in extended format), is_extended, is_decimal_sign_comma and has_fractional_second. Arithmetic on a partial value, or a result outside the representable 0000–9999 year range, is reported as no result rather than an invented one.

  • openEHR path expressions support general comparison predicates (BASE architecture overview §Paths and Locators, “Other Predicates”): path predicates of the form [at0007 and time >= '2005-06-24T09:30:00'] or [value/defining_code/code_string = 'A04'] — a relative attribute path, an operator (=, !=, <, <=, >, >=), and a quoted-string or numeric literal — now parse and evaluate everywhere RM paths are resolved, including ehr: URI resolution. Strings compare lexically (ISO 8601 date/times order temporally), numbers numerically, with XPath existential node-set semantics; predicate text outside the grammar is still rejected loudly. Previously these spec-defined forms were refused as unsupported.

3.12.0 - 2026-07-29

Changed

  • The conformance verdict model no longer has excused capability states. The published reports and certificates previously carried two non-verdict evidence tokens — unrealized (every case excused by a register citation) and no_cases (a claim the catalogue named no case for). Both are deleted: the catalogue gates now refuse those shapes before any server is assessed, and every capability a party claims is reported as exactly one of passed / failed / inconclusive / not-evidenced. A required capability without passing evidence now fails its tier with no excuse arm, for every assessed party alike; both committed records and the published comparison were re-derived under the stricter model (no tier verdict changed for either party).

  • An empty TDD batch answers 200 with [] instead of 201. POST /message/tdd/{ehr_id}/batch with an empty array creates nothing, and 201 Created reported a creation that did not happen. Batches with members are unaffected.

  • EXTRACT_SPEC.extract_type now accepts every code the openEHR Reference Model names. POST /message/export previously refused openehr-synchronisation and openehr-generic — two of the five extract types the RM’s EHR Extract chapter lists by example — as out of group. Both are accepted now, alongside openehr-ehr, openehr-demographic, generic-emr and the catch-all other.

  • Conformance: a product is no longer excused from 186 test cases for declaring an older REST release (#635). Every conformance case may declare the openEHR release its behaviour needs, and systems declaring an earlier release are skipped for it. That declaration had been copied onto 343 cases as authoring boilerplate, which quietly wrote off most of the EHR, COMPOSITION, DIRECTORY, CONTRIBUTION, QUERY and template surface for any product declaring ITS-REST 1.0.3 — behaviour those products do implement and should be judged on. Each case was re-derived against the released amendment record, and the requirement is kept only where the released text actually dates it: ITEM_TAGs, the Demographic API, admin EHR deletion, the Simplified Formats media types, Prefer: return=identifier, the audit-details system_id, the reserved aql query name, the template /example sub-resource, and SMART on openEHR. Everything else is now judged for every product, with the two genuinely release-dated header rules (the weak W/ ETag form and the read/delete Location restriction) still applied only to the release that introduced them. No test was removed or weakened; the comparison against other openEHR products now covers what they really implement.

  • The conformance statement no longer claims nine capabilities it cannot demonstrate (#623). ADL 1.4 and ADL 2 archetype provisioning, the admin Activity Report, EHR dump/load, EHR and demographic archiving, EHR Extract, TDS, and the MESSAGE API were all being claimed while every one of their test cases was excused: openEHR’s released REST API publishes no endpoints for them, and EHRbase-rs exposes none of its own either — the underlying service methods exist, but nothing reaches them over HTTP, so a conformance runner has nothing to drive. Claiming a capability is the obligation to prove it, so those claims are withdrawn until the routes exist. Nothing was removed from the product; what changed is that the published statement now only claims what can be demonstrated.

  • Helm chart and operator-facing comments cite durable references (#322). The chart’s values.yaml, Chart.yaml and post-install NOTES pointed at internal design documents that no longer exist (the deleted design and enterprise doc trees) and at retired decision-record numbers. Each is now either the official upstream documentation it was standing in for (the PostgreSQL docs for the unprivileged app role and the lock_timeout migration wrapper, the Kubernetes Pod Security Standards for the container security posture), an explicit “our own extension, no openEHR spec governs it” flag on the optional integrations, or the rationale written out inline — so an operator reading the chart is never sent to a dead path. No default value, template, or rendered manifest changed.

  • The published per-chapter outcome bars are now a two-level chart with no Other bucket (#613). The single bar per schedule chapter hid the EHR chapter’s hundreds of cases behind one rectangle and swept the System API and anything unrecognised into an Other row. The chart now renders a chapter header carrying the chapter’s total above one scaled bar per band — the surface a case actually exercises (EHR resource / EHR_STATUS / COMPOSITION / DIRECTORY / CONTRIBUTION / item tags / revision history, ADL 1.4 vs ADL 2 vs stored queries, ad-hoc vs stored query execution, parties vs relationships vs versioned party, and so on) — with the exact passed / FAILED / errored / cited-N-A counts printed beside every row, so a small band never loses its numbers to a short bar. Cited-N/A segments carry a hatch texture so “not executed, with a citation” can read as neither a pass nor a failure. The taxonomy is total: every case id maps to a named band and an unmapped id fails the render naming the id, rather than landing in a silent bucket. Both published SUTs render the same bands — a band with no case shows as an explicit no cases row — so the comparison page reads band-for-band.

  • The conformance pipeline now exercises BOTH claimed version-signing modes in every run, in the one committed record (#609). openEHR defines a version signature at two depths of one mechanism — a plain digest (an integrity check) and an openPGP RFC 4880 signature (which additionally authenticates the author) — and a running server does one or the other. The product claims both, so scripts/conformance.sh now brings up a second deployment of the same built image in the openPGP posture alongside the standard stack (its own compose project, host port 8081), and the party’s ixit.json declares it as an extra instance carrying its own signing block. The openPGP signature cases address that instance; one run, one results.json, and the Signing capability’s evidence covers both modes. Consequently the CONF_SIGNING_MODE=pgp environment switch and the separate ixit.pgp.json party file are removed — there is nothing left to select, both modes always run. A conformance target that declares no such instance (upstream EHRbase) has the openPGP cases recorded not-applicable with that citation instead of failed, which is also now true for any case addressing an instance a party does not declare: it is excused at selection time rather than surfacing as an inconclusive row.

  • The conformance suite proves EHR-scoped querying against two EHRs, not one (#604). The four cases that check a query is confined to the EHR named in the openehr-ehr-id request header used to run against a server holding a single EHR, so a server that ignored the header returned the same rows and passed. Each now creates a second EHR with its own content first: an unscoped answer carries the extra row and fails the case. The behaviour being checked is unchanged; the check can no longer be satisfied by accident.

  • The conformance suite now names a malformed request and invalid content differently everywhere (#605). Fifteen more conformance cases used to report a rejected request as a content-validation failure when what the request actually broke was its own syntax — an unparseable template upload, a path segment that is not an identifier, a version_at_time outside the ISO 8601 form the specification mandates, or a tag list sent as something other than a list. Those now report as malformed requests. Nothing changes on the wire (all fifteen answered 400 Bad Request before and after) and no server passes or fails differently; the published conformance report and case records simply name one rejection law one way, so a reader can tell the two families apart.

  • Two behaviours the conformance suite used to treat as optional are now required of every server (#556). openEHR publishes its REST specification as normative prose and as OpenAPI files, and the prose is silent on more than it looks. Where the prose says nothing, those OpenAPI files are now read as part of the specification rather than set aside — so behaviours previously recorded as “the specification does not say” turn out to be specified after all, and the suite stops excusing them. Two change how a server is judged. Uploading an operational template under a template_id that already exists must answer 409 Conflict; it was previously a declared choice between refusing and silently replacing the stored template, and a server could opt out of the refusal. Updating a COMPOSITION whose request body carries a uid naming a different version container than the URL must be rejected; the mismatch was previously reported without affecting the verdict. Both are now gating conformance cases. The published conformance artifacts and the ambiguity register record the specification sentence behind each.

  • The published Conformance Statement now declares the non-openEHR surface this server serves (#527). A new “Additional non-openEHR surface” section lists every extension route family — health, status, the OpenAPI/Swagger meta routes, management, terminology, event subscriptions, multi-tenancy, the FHIR R4 connector and its mapping store, the ITI-81 audit read, PARTY_RELATIONSHIP, the bare stored-query list, the admin template/query/config routes and SMART discovery — with the routes it serves and the configuration that enables it. The section states plainly that none of it is part of any conformance claim: no openEHR specification governs these routes, no conformance case exercises them, and no verdict depends on them. A reader of the statement no longer has to discover the extension surface on the wire.

  • Canonical-XML support is now declared per resource family in the Conformance Statement, instead of being assumed for every resource (#572). The openEHR release publishes an XML document element for only eight names — composition, version, items, template, extract, extract_request, versioned_object, archetype — while its REST API addresses application/xml to the whole resource surface. For a resource with no published document (EHR, EHR_STATUS, the directory FOLDER, the demographic party types, CONTRIBUTION) the specification therefore neither requires a server to serve XML nor forbids it, so the suite no longer asserts either answer: the statement declares, per family, whether this server offers canonical XML there, and the conformance run judges the matching branch — the XML read, or the 406 Not Acceptable refusal the specification designates for an Accept a service cannot fulfil. This server declares XML support for EHR, EHR_STATUS, directory and the party families and declares it unsupported for CONTRIBUTION reads, exactly as it behaves. The full per-resource classification, its citations and the upstream report asking openEHR to reconcile the two inventories are recorded in the conformance ambiguity register as AMB-167 / UPR-127.

  • Stored top-level objects now carry their copied uid at commit time (#439). The full three-part OBJECT_VERSION_ID is stamped into the canonical body before it is decomposed, signed, and stored, so the contained object served inside an ORIGINAL_VERSION envelope, the bare resource reads, AQL projections, and EHR Extract exports all carry the identical uid value (ITS-REST overview Resources §Identifier types). Previously the uid was injected only on some read paths; clients now see one consistent shape everywhere. Imported (EHR Extract) content is exempt — its bodies are preserved verbatim.

  • EHR.ehr_status references the version container by its HIER_OBJECT_ID (#426). The served EHR body’s ehr_status OBJECT_REF (typed VERSIONED_EHR_STATUS per the RM invariant) previously carried an OBJECT_VERSION_ID naming one version — inconsistent with the sibling ehr_access ref and with the RM’s container semantics (OBJECT_REF.id is the id of the referenced object; the referenced object is the VERSIONED_EHR_STATUS, whose uid is a HIER_OBJECT_ID). Both refs now carry the container id. Clients that read the current EHR_STATUS version uid from the EHR body must fetch GET /ehr/{ehr_id}/ehr_status and use its own uid instead.

Fixed

  • ADL 1.4 archetypes with anonymous archetype slots are accepted. The ADL 1.4 specification writes archetype slots without a node id in its own examples (allow_archetype OBSERVATION occurrences matches {0..1} …), and published CKM archetypes use that form — but the parser demanded [atNNNN] and refused such sources with a syntax error, so POST /definition/archetype/adl1.4 answered 422 for spec-valid archetypes. Both the anonymous and the identified slot forms now parse; ADL 2 sources still require the node id, as ADL 2 defines.

  • An empty TDD batch aimed at a non-existent EHR is now refused. POST /message/tdd/{ehr_id}/batch verified the target EHR once per document, so a batch with no documents answered success without checking the EHR at all. The target is now verified for every batch, empty ones included, and an unknown EHR answers 404.

  • An activity-report request whose time interval runs backwards is now refused instead of answered. GET /admin/report/* accepts an optional time_interval=<lower>/<upper>; a pair bounded on both sides with the lower bound after the upper one is not an interval at all (the openEHR BASE Interval invariant requires lower <= upper), and the server used to run it anyway and hand back the empty count such a range selects — a truthful-looking answer for a window nobody asked for. It is now 400. Equal bounds remain a legitimate single-instant interval.

  • A corrupt dump archive no longer reports as an internal server fault. POST /admin/load against a location whose manifest or segment is mangled or truncated used to surface as an unexpected-exception 500, while a location holding no archive at all reported the openEHR service model’s file_not_writable. Both are the same fact — the location does not hold a readable archive — and both now report file_not_writable, with nothing loaded either way.

  • An export requesting a format this server does not implement now answers 501 Not Implemented instead of 400. openehr_canonical_xml and the 7z compression format are valid values in the openEHR service model that this server does not build; reporting them as malformed requests was wrong, and the response now says the functionality is unsupported.

  • The template list collapses to the latest version of each template when the version parameter is absent (#614). GET /definition/template/adl1.4 (and the ADL2 twin) used to return every stored version regardless; the released openEHR REST API says an absent version returns “only the latest version”. Pass version=* to list every stored version — the admin console’s template inventory does exactly that, so its view is unchanged.

  • Conformance runner: a requirement the openEHR specification dates to a release is now judged only against the servers that claim that release (#627, #628). Two rules the ITS-REST overview introduces at Release 1.1.0 — that an ETag carrying a resource identifier must be weak (W/"…"), and that Location is no longer returned on reads and deletes — used to be enforced or waived by accident, depending on whether a case happened to carry a version floor for some unrelated reason. Each rule now carries the floor itself, so a target declaring an earlier ITS-REST release is still driven for the operation and still judged on everything else, while a target declaring 1.1.0 or later faces the rule everywhere it applies; a test derives the affected set from the committed catalogue so no future binding can escape it. Separately, the query resultSet’s ETag is no longer required to be PRESENT: the specification names the header without any requirement keyword and its only strength anywhere is a SHOULD, so a server that omits it is not failed — while a server that emits it must still emit it in the weak form. Conformance verdicts for this product are unchanged (it declares ITS-REST 1.1.0 and returns the header).

  • Conformance runner: one operation is now sent one way (#629). The runner built request headers in two places — once for a case’s own steps and once for the preconditions a case needs — and the two disagreed, so the same operation went on the wire differently depending on which path reached it (a template upload was refused as a case and accepted as a precondition). There is now a single header-construction path; a binding declares the Accept it intends; and a refusal of that Accept is recorded as a named outcome instead of vanishing into an unmapped status. The published conformance report also gains a per-capability table showing how many cases passed, failed, and came back inconclusive, so a divergence can no longer hide behind an inconclusive exchange. A binding may now also declare the release its wire first appeared in, and cases driving it are recorded not-applicable — with that citation — for targets that declare an earlier one.

  • Conformance runner: a create asked for a minimal response may answer either 201 Created with an empty body or 204 No Content (#630). The specification says an empty-bodied response SHOULD use 204, and its machine-readable artifacts declare 201 for creates; both are therefore conformant, and the suite now judges both instead of leaving one of them an inconclusive result. Updates are unchanged (204 only, where both sources agree).

  • Web Template: a template that narrows a party to PARTY_RELATED now describes its party fields (#600). When an operational template pins a party slot — a subject, a composer, a participation performer — to PARTY_RELATED, the generated Web Template used to describe that node as an empty container: none of the |name, |id, |id_scheme and |id_namespace fields the Simplified Formats specification gives every party appeared, so a form builder reading the Web Template could not offer them even though the server has always accepted and returned them. The four fields are now described, alongside the relationship sub-path the narrowing adds. The same held wherever a party node also constrained an attribute; the fields survive that too. Nothing changes for stored data or for the FLAT and STRUCTURED wire.

  • FLAT/STRUCTURED: the specification’s other spelling of a related party’s relationship is accepted on input (#589). The Simplified Formats mapping table for a PARTY_RELATED writes the relationship sub-path …/_relationship, while every example block in the same section — and the participation-performer table — writes …/relationship|code. Only the example form was accepted, so a producer that followed the table row had its composition rejected with an unknown-path error. Both spellings are now read, and either one makes the party a PARTY_RELATED. What the server emits is unchanged: always the example spelling, so stored data and round-trips look exactly as before.

  • Three FLAT/STRUCTURED mapping gaps: an entry’s subject, null-flavoured elements, and the event-context paths (#532, #533, #534). An entry’s subject now travels over the wire in both directions: a composition whose OBSERVATION, EVALUATION, INSTRUCTION, ACTION or ADMIN_ENTRY names someone other than the record subject emits …/subject|name, …|id, …|id_scheme, …|id_namespace plus the /_identifier:i and /relationship sub-paths, and the same keys are accepted on input — previously the subject was dropped on the way out and rejected as an unknown suffix on the way in, so the information was lost in both directions. A “self” subject carrying an external reference is marked |_type: PARTY_SELF so it comes back as itself rather than as an identified party. Second, an element that records why a value is missing (a null flavour, which the reference model makes mutually exclusive with the value) now keeps /_null_flavour and /_null_reason through a full round trip; the flattener reached elements only through their value, so a null-flavoured element vanished entirely. Third, the event-context fields the specification also spells as paths — …/context/start_time and …/context/setting — are honoured on input instead of being silently discarded in favour of the ctx/ defaults, as are an entry’s …/language and …/encoding; a bare …/context/setting|code resolves against the openEHR setting value set exactly as ctx/setting does. Paths the specification does not define are still rejected with a clear error rather than ignored.

  • The SMART discovery endpoint is fully described in the published API reference (#535): the application/json requirement, the required org.openehr.rest service with its absolute baseUrl, the capability-honesty rules, the public pre-auth posture, and a worked document example — previously a one-line declaration.

  • The FLAT/STRUCTURED mapping of INSTRUCTION_DETAILS and INTERVAL_EVENT.sample_count (#521). An ACTION’s instruction details now travel over the wire exactly as the Simplified Formats specification maps them — three suffixes on the _instruction_details field itself: |path, |composition_uid and |activity_id. Previously the server emitted a nested _instruction_details/instruction_id field with generic object-reference suffixes, so |composition_uid was never produced, the instruction path sat one level too deep, and two suffixes the specification does not define were emitted; clients that sent the specified form had the details silently dropped. Both directions are now symmetric, so a composition round-trips through FLAT and STRUCTURED without losing the reference. Separately, an interval event’s |sample_count (the count of samples the interval summarises) is now both emitted and accepted; it was previously ignored in both directions.

  • The template documentation no longer advertises a build flag that does not exist (#521). The templates-and-validation page described an ehrbase-quirks build that renumbers duplicate node names and accepts two vendor-only DV_QUANTITY suffixes; that feature was removed long ago. There is one behaviour — the one the specification prescribes.

  • The published API reference is now honest about the non-openEHR surface, and follows a non-default base path (#526). Every operation this server serves outside the standardised openEHR ITS-REST resource set — the management, terminology, event-subscription, multi-tenancy and FHIR R4 groups, plus the IHE ITI-81 audit retrieval — now states in its own published description that no openEHR specification governs it (the flag previously lived only in source-module comments the document never carried), and each disabled-group 404 now says that an unauthenticated caller is answered 401 first, which is what the server actually does. The /status, openapi.json, Swagger-UI and System-OPTIONS declarations now follow a configured server.base_path instead of always printing the default deployment’s paths. The twelve per-family ehrbase-{family}.openapi.json documents are documented, the ITI-81 and admin-extension operations now appear in a family document (previously in none), and the document itself declares a servers block, a link to the implemented ITS-REST release, descriptions for every tag it uses, and the implemented ITS-REST contract version as x-openehr-its-rest (distinct from info.version, which is the product version).

  • SMART App Launch conformance (#519). The discovery document’s services.*.baseUrl values are now absolute URLs built from the new required smart.public_base_url origin (the specification requires absolute URLs); the openehr-permission-v1 capability is advertised only in fail-closed mode (require_smart_scopes = true) so advisory deployments no longer over-claim fine-grained enforcement; operators can advertise the HL7 base capabilities via smart.endpoints.capabilities; enabling SMART now boot-validates the origin plus the authorization/token endpoints; the published OpenAPI’s discovery path follows a configured platform_base_url; and — the substantive gap — the template and AQL scope families now enforce: in fail-closed mode a token without a matching template-…/aql-… scope is denied 403 on the template and query routes (previously only the composition family was gated).

  • The published API reference now describes the admin endpoints in full, and the disabled-admin answer is documented correctly (#513). The five admin operations — the released DELETE /admin/ehr/{ehr_id} and DELETE /admin/ehr/all, plus the template-delete, stored-query-version- delete and effective-config extensions — gained the branches they actually answer (400 for a malformed EHR id, 401/403 from the admin role gate, 404, the template 409), the mandatory empty Allow header on every disabled-group 405, and worked request/response examples. They now carry the released operation text verbatim, including the permanent-physical- delete cascade and its data-protection (GDPR) sentence, the development/testing note on the bulk route, and the fact that this server deletes synchronously (204 only — the specification’s optional asynchronous 202 is never returned). The bulk route documents both accepted query forms (?ehr_id=a&ehr_id=b and ?ehr_id=a,b) and that an absent or empty list deletes every EHR; the three extension routes are flagged plainly as our own, governed by no openEHR operation. Reference documentation and configuration docs that claimed a disabled admin API answers 404 were corrected: it answers 405 Method Not Allowed with an empty Allow header.

  • Demographic ITEM_TAG collections honour the released dual-form addressing (#509). A version-addressed uid_based_id on the demographic tag routes now reads, replaces, and deletes that VERSION’s own distinct tag collection (previously every form reached the container’s set); the tags GET and DELETE now answer 404 for a nonexistent, wrong-kind, or cross-space target (previously an empty 200 list); both openehr-item-tag and openehr-version-item-tag request headers are accepted on party create AND update, each landing on its own target’s collection with its own response-header echo; a tag’s target is now the bare RM UID_BASED_ID (an OBJECT_VERSION_ID for version targets) and its owner_id follows the released examples’ local/SYSTEM shape; and the PARTY_RELATIONSHIP extension’s stale-delete 409 now echoes the latest version_uid in ETag like the party delete it mirrors.

  • The published API reference now describes the demographic item-tag endpoints in full (#510). The sixteen ITEM_TAG operations — the person / agent / group / organisation / role tags read, replace and delete-by-key, plus the space-wide GET /demographic/tags — gained the status branches they actually answer (400, 404, 406, 415, 422), the Prefer / Content-Type / Accept request headers, the Preference-Applied echo on both replace branches, and worked ITEM_TAG examples. They now state plainly that a version-addressed uid_based_id and a container-addressed one name two DISTINCT tag collections, that an empty list on the replace clears every tag, that deleting by key alone removes every tag under that key whatever its target_path, and that a tag collection is never change-controlled — so no ETag, Last-Modified or Location is offered anywhere on the family. The space-wide list is documented for what it is: the one tag route with no scoping parameter at all, no paging, and 200 (an empty array when nothing matches) or 400 as its only outcomes.

  • Demographic header echoes (#388). The stale-version party DELETE’s 409 Conflict now returns the latest version_uid in ETag (the released response requires it); and the demographic CONTRIBUTION read now carries the weak ETag (the contribution uid) and Last-Modified from the committal instant, matching its EHR sibling.

  • The published API reference now describes the demographic endpoints in full (#505). The 26 person / agent / group / organisation / role, versioned-party and demographic-contribution operations in the served OpenAPI document gained the response headers they actually send (ETag, Location, Last-Modified, Preference-Applied and the two openehr-*-item-tag headers), the status branches they actually answer (204, 400, 406, 409, 412, 415, 422), the committal (openehr-version / openehr-audit-details), Prefer, If-Match, Accept and Content-Type request headers, worked PERSON / VERSIONED_PARTY / CONTRIBUTION examples, and a spec citation on every branch. Reads and deletes no longer suggest a Location header they never send, and the party routes now state plainly that Simplified (FLAT / STRUCTURED) media types are refused because a demographic party is not templated. The eight PARTY_RELATIONSHIP operations are labelled for what they are — an extension of this server, with no openEHR REST operation behind them — and the group carries a note that the openEHR Demographic API is itself a DEVELOPMENT-state specification.

  • Stored-query stores answer honest Locations and validate the version segment (#498). The version-less PUT /definition/query/{name} now always names the version it actually wrote (…/1.0.0) in Location — previously, when a higher version already existed, the header pointed at that untouched neighbour. The versioned PUT /definition/query/{name}/{version} now requires an exact numeric major.minor.patch and rejects prefix, pre-release, or malformed version segments with 400 Bad Request — previously any string was stored verbatim, and a single non-numeric version (e.g. 1.0.0-rc.1) broke every later stored-query list and retrieval on the server. Both store forms also now refuse a payload declaring a media type other than their single text/plain body type with 415 Unsupported Media Type (an absent Content-Type remains accepted).

  • The published API reference now describes the stored-query endpoints in full (#499). The four stored-query operations in the served OpenAPI document gained the Location response header both stores actually send, the bodyless shape of their 200, the status branches they actually answer (400 everywhere, 406 on the reads, 409 on the versioned store, 404 on the version read), the qualified-name and version grammars (including the reserved aql name and the read-side prefix-resolution rule), request and response examples, and a spec citation on every branch. The bare “list every stored query” route is now labelled for what it is — an extension of this server, not a released openEHR operation.

  • Template rejection statuses are coherent across both upload routes (#493). An ADL2 source with grammar-level syntax errors now answers 400 Bad Request (the released “syntactically invalid … content” branch) instead of 422; AOM2 validation-phase failures on a parsed source keep answering 422 with the rule codes in validationErrors. On the ADL 1.4 side, an AOM2 artefact-validity violation on a successfully parsed OPT now answers 422 with the rule code in validationErrors (previously 400) — syntax gates 400, semantics gate 422, on both routes.

  • Template-upload rejection statuses follow the released split (#489). An ADL 1.4 OPT upload whose body is not well-formed XML now answers 400 Bad Request (the released “syntactically invalid … content” branch) instead of 422; well-formed XML that is not a valid OPT stays 422. The ADL2 template upload now refuses a payload declaring a media type other than its single text/plain body type with 415 Unsupported Media Type (an absent Content-Type remains accepted), mirroring the ADL 1.4 guard.

  • The published API reference now describes the template endpoints in full (#490). The nine ADL 1.4 / ADL 2 template operations in the served OpenAPI document gained the response headers they actually send (ETag, Location, Preference-Applied), the status branches they actually answer (400, 406, 415, 422), request/response examples, and a spec citation on every branch; “Get template at version” is now marked deprecated, as the openEHR REST specification marks it.

  • Stored-query POST bodies accept {}; the query POSTs accept the URL parameter forms (#481). The three body members of the stored-execute body are optional (the docs text gives offset a default and makes fetch implementation-default — the stalled required-list loses), so a parameterless stored query executes with an empty body; and all three POSTs now accept offset/fetch/named $parameters from the URL (the docs-text SHOULD-list draws no GET/POST distinction), with a body-vs-URL disagreement rejected 400.

  • Tag GET/DELETE verify the addressed target; empty target_path normalizes to absent (#474). GET/DELETE on the per-target tag routes now answer 404 for a nonexistent, foreign-EHR, or wrong-kind uid_based_id (the released trigger: “when the uid_based_id does not exist”; previously the GET answered 200 [] and the DELETE was not kind-checked), and a target_path: "" on the tag PUT is normalized to the absent path so "" and absent are one (key, target_path) identity. The EHR-wide tag listing likewise answers 404 for an unknown ehr_id (previously 200 []).

  • Contribution change-type mismatch statuses follow the released assignment (#467). A non-creation change_type committed as the FIRST version of a versioned object (the released 400_CONTRIBUTION trigger: “the modification type does not match the operation - i.e. first version of a MODIFICATION”) now answers 400; a 249|creation| member carrying a preceding_version_uid — the unassigned mirror case — moves to 422. Previously the two were inverted.

  • The CONTRIBUTION GET serves ETag + Last-Modified (#463). GET …/contribution/{contribution_uid} now carries the contribution-uid weak ETag (the same identity the 201 already carries) and a Last-Modified derived from the contribution audit’s commit instant (ITS-REST overview Requests and responses §“ETag and Last-Modified”).

  • Directory by-version reads verify the full addressed identity; the directory DELETE 204 carries the deleted version’s identity (#456). GET …/directory/{version_uid} now answers 404 when the addressed creating_system_id does not match the stored identity (ITS-REST overview Resources §Identifier types), and DELETE …/directory answers 204 with the NEW 523|deleted| version’s weak ETag + Last-Modified (previously header-less), matching the composition DELETE.

  • COMPOSITION update body-uid mismatch is 422, not 400 (#451). A PUT whose body COMPOSITION.uid names a different versioned object than the request path is now rejected 422 Unprocessable Entity — the body is well-formed and the contradiction is semantic (ITS-REST Requests and responses §HTTP status codes, the 422 row; no released sentence assigns the rejection — register-documented).

  • Versioned-composition version-by-id reads are container-scoped (#449). GET …/versioned_composition/{versioned_object_uid}/version/{version_uid} previously ignored the container segment and served any version the version_uid named; a version_uid whose object_id does not match the path’s container now answers 404 (ITS-REST overview Resources §Identifier types; RM Owner_id_valid).

  • EHR creation mints RM-valid EHR_STATUS and EHR_ACCESS objects; the RM archetype-root invariants are enforced on client bodies (#423). The bootstrap defaults carried an archetype-HRID archetype_node_id with no archetype_details — violating RM Is_archetype_root (unconditional on both classes) with Archetyped_valid (“is_archetype_root xor archetype_details = Void”). Both defaults now carry the ARCHETYPED block (archetype_id = the node id, rm_version 1.2.0), and a client-supplied EHR_STATUS/EHR_ACCESS violating Archetyped_valid (a root without archetype_details, or a mismatching archetype_id) or Links_valid (an explicit empty links list) is rejected with 422. Clients that previously committed root objects without archetype_details must now supply it.

  • Imported and archive-loaded EHRs are now complete, first-class EHRs (#425). An EHR-Extract clone (import_ehr) created no EHR_ACCESS, so a source extract that carried none produced an EHR permanently violating the RM invariant Ehr_access_valid (EHR.ehr_access is 1..1) whose served GET /ehr/{ehr_id} body simply omitted the mandatory reference; the clone now commits the same default EHR_ACCESS the create path uses (RM ehr master04 §EHR Creation — a root EHR object, an EHR Status object and an EHR Access object), in the import’s own transaction. Neither the import nor the admin archive load promoted the EHR’s subject, so imported/loaded EHRs were invisible to GET /ehr?subject_id=…&subject_namespace=… and exempt from the one-EHR-per-subject 409; both paths now derive the subject from the landed EHR_STATUS. Consequences: importing or loading an EHR whose subject this repository already holds is now refused — 409 for an import (naming the subject and the EHR that holds it), and, for the archive load, a per-record DUMP_LOAD_FAIL_REPORT entry that skips just that EHR exactly like a duplicate EHR id, leaving the rest of the archive to load.

  • Both EHR creates now accept and merge the committal request headers (#422). ITS-REST docs/overview/Requests_and_responses.md §“openehr-version and openehr-audit-details” makes it a MUST that a service accept openehr-version / openehr-audit-details on the direct PUT/POST/ DELETE commits of change-controlled resources and merge “whatever is provided … with the default VERSION and VERSION.audit_details attributes on commit runtime”. Creating an EHR commits its EHR_STATUS and EHR_ACCESS in a contribution (RM ehr master04 §EHR Creation), but POST /ehr and PUT /ehr/{ehr_id} ignored both headers — while the served OpenAPI already claimed they were merged. They now are: the supplied description, committer, and system_id land on the creating contribution and on both committed versions’ commit_audit, and openehr-version: lifecycle_state.code_string sets the new EHR_STATUS version’s lifecycle state. change_type is constrained to 249|creation| (a create commits a first version): restating 249 is accepted, another group code is 400, and a token outside the audit_change_type group is 422. The OpenAPI document now also lists both headers as documented parameters on the two create operations.

  • Served OpenAPI: the four EHR-resource operations now document the whole wire (#427). GET /ehr, POST /ehr, GET /ehr/{ehr_id} and PUT /ehr/{ehr_id} under-described what the server actually does. The two creates now declare their 415 (an unprocessable request Content-Type, including a Simplified Format, which is defined only for templated COMPOSITION content) and 406 (an Accept that canonical JSON/XML cannot satisfy) branches, both a MUST in the REST spec’s format sections; the two reads declare 406 as well, and GET /ehr/{ehr_id} declares the 400 it returns for a malformed (non-UUID) ehr_id. Every success response now carries a header block: ETag/Location/Last-Modified/ Preference-Applied on the 201s, ETag on the reads — where the absence of Location and Last-Modified on a read is now stated explicitly rather than left unsaid. The Prefer-conditional 201 body is documented as a named example pair (representation — the full RM EHR; identifier — the single-uid object), the Prefer header enumerates its three tokens and its default, and the request body, the read bodies, and the subject query parameters carry real served-shape examples. Corrected false claim: PUT /ehr/{ehr_id} described ehr_id as any HIER_OBJECT_ID with “a UUID strongly recommended”, while the server accepts UUIDs only — which is what the abstract service model types the argument as, and every UUID is a valid HIER_OBJECT_ID root.

  • Served OpenAPI: the System API’s OPTIONS operation is now documented (#418). The Options-and-Conformance endpoint (OPTIONS on the API base path) was served but absent from the generated OpenAPI document and Swagger UI, because its route mounts outside the documenting router (above the CORS layer, deliberately). A documented twin now carries the full operation description — the Options manifest schema with field documentation and example, the Allow/Content-Type response headers, and the 406 negotiation branch.

  • version_at_time now accepts a datetime without a timezone, interpreting it in the server’s local timezone (#401). ITS-REST docs/overview/Resources.md §“Datetime format” requires the extended ISO 8601 form for datetime query parameters and states that “Timezone SHOULD be only supplied when needed, otherwise the local timezone is assumed” — so ?version_at_time=2016-06-23T13:42:16 is a valid request. It was answered 400 Bad Request, because both at-time parsers (the EHR group’s and the DEMOGRAPHIC group’s duplicate) required an offset. A single shared decoder now backs every at-time read — EHR_STATUS, COMPOSITION, DIRECTORY, the versioned_* version reads, the demographic party/relationship reads, and the contribution time_range — and resolves an offset-less value against the server’s system timezone (TZ, else the platform’s local-time configuration); a value falling inside a daylight-saving fold or gap resolves to the earlier and the later instant respectively. Genuinely malformed input is unchanged: the basic ISO 8601 format (20160623T134216Z), a date without a time (2016-06-23 — the parameter is specified as “a given time”, and reading it as midnight would silently serve a version the caller never asked for), a timezone-less value carrying an [Area/Location] annotation, and anything unparseable are all still 400.

  • Every 405 Method Not Allowed now carries an Allow header, and the 408/413 transport refusals use the openEHR error body (#400). ITS-REST docs/overview/Requests_and_responses.md §“HTTP Methods” answers a method a resource does not serve with 405, over RFC 9110 — the authority that section names — whose §15.5.6 requires that “the origin server MUST generate an Allow header field in a 405 response containing a list of the target resource’s currently supported methods”. The router’s 405 already carried it; the 405 returned when the admin API is disabled did not, because it comes from a matched handler and so never reached the router’s allow-header machinery. It now sends the empty field value RFC 9110 §10.2.1 defines for exactly this case (“An empty Allow field value indicates that the resource allows no methods, which might occur in a 405 response if the resource has been temporarily disabled by configuration”). Separately, a request that times out (408) or declares a body over the 16 MiB limit (413) was answered by the middleware with an empty or text/plain body; both now render the same { "error", "message" } JSON every other error path emits. Finally, the deviation behind all of this is now recorded rather than silent: the same spec section also SHOULDs 501 Not Implemented for an unrecognized method, and we answer 405 there too — the two SHOULDs overlap for any method outside the tabulated subset, 405 is a predefined non-conflicting code in the spec’s own status table, and a blanket 501 fallback would misreport unknown paths that are owed 404. 501 is still returned for a recognized but unimplemented operation. Adjudicated in the conformance ambiguity register as AMB-60, with the wire-surface boundary registered alongside it.

  • The openehr-ehr-id request header now scopes GET query execution too, and a scope named twice must agree (#399). ITS-REST docs/query/Request.md §“About the ehr_id parameter” lets clients supply the single-EHR scope “as a query parameter ehr_id or alternatively as a request header named openehr-ehr-id”, and §“Common Headers and Query Parameters” applies that to “all query execution requests”. Only the POST forms honoured the header: GET /query/aql, GET /query/{name} and GET /query/{name}/{version} read the scope from the query string alone, so a header-scoped GET silently ran as a population query across every EHR. All six execution operations now resolve the scope through one seam. The released text never says what a request carrying both forms means, so the handling is adjudicated and registered (ambiguity register AMB-59): both forms naming the same EHR execute normally, and both forms naming different EHRs are rejected 400 Bad Request rather than silently picking one — a request that names two EHRs cannot be answered correctly (docs/overview/Requests_and_responses.md §“HTTP status codes”, row 400). An empty header value carries no identifier and neither scopes nor conflicts. The deprecated openEHR-EHR-id spelling keeps working (HTTP field names are case-insensitive). The header is now also declared on every query operation in the served OpenAPI.

  • Prefer / representation polish: return=identifier is structurally never 204, item-tag echoes are per-target, and Preference-Applied is emitted from one seam (#398). Three divergences from ITS-REST overview Requests_and_responses.md:

    • Prefer: return=identifier could fall through to the empty (possibly 204 No Content) minimal response while still claiming Preference-Applied: return=identifier. The identifier branch now carries the identifier it renders, so it is unreachable without one — §“Prefer only identifier”: “the status will be 201 Created or 200 OK, never 204 No Content”. A write that genuinely produces no identifier applies, and declares, the default return=minimal instead of claiming an unapplied preference.
    • The openehr-item-tag and openehr-version-item-tag response echoes on a change-controlled write merged both targets’ tags into one list and repeated it under both header names. Each header now confirms only its own target’s stored list — §“openehr-item-tag and openehr-version-item-tag”: “openehr-item-tag applies to VERSIONED_OBJECT targets” while “openehr-version-item-tag applies to a specific target VERSION”, each confirming “the actual list of ITEM_TAGs stored”. A header the request did not send is not echoed. (The demographic surface still emits both headers from one set, because its tags are stored against the VERSIONED_OBJECT only, so the two targets coincide there.)
    • Preference-Applied was emitted only by the canonical RM / JSON write helpers. It is now declared by every write path through the same seam — the demographic party, relationship and CONTRIBUTION writes, both ADL 1.4 and ADL 2 template uploads, the ITEM_TAG collection writes, and the Simplified-Formats (FLAT/STRUCTURED) COMPOSITION commit — always naming the preference the response actually applied, including the applied default return=minimal when no Prefer was sent. Demographic party writes additionally honour Prefer: return=identifier ({uid} body), which they previously ignored.
  • ADL 1.4 template negotiation: the response type mirrors Accept, and a non-XML OPT upload is 415 (#397). Two divergences from ITS-REST overview Resources.md:

    • GET /definition/template/adl1.4/{template_id} with Accept: application/json was answered Content-Type: application/openehr.wt+json — a type the client never accepted. It now returns the same Web Template document under Content-Type: application/json (§JSON Format: “Proper header Content-Type: application/json MUST be present in the response of the service unless the response has no content body”). Accept: application/openehr.wt+json keeps the Web Template media type and Accept: application/xml the canonical OPT, both unchanged. (The released source is internally inconsistent here — the operation description names only XML + wt+json while its Accept/Content-Type enumerations include application/json with no schema — so serving the Web Template body is the recorded fixed handling, not a 406.)
    • POST /definition/template/adl1.4 accepted any Content-Type and failed a JSON payload with 400 from the OPT parser. A request declaring a non-XML payload type is now refused 415 Unsupported Media Type before parsing (§XML Format: “If the service cannot process the request payload as XML format, it MUST respond with HTTP status code 415 Unsupported Media Type”). application/xml and text/xml upload as before, and an absent Content-Type still reads as the operation’s single body type (the header is a client MAY).
  • Last-Modified and ETag completion on the EHR and DEFINITION surfaces (#396). ITS-REST overview Requests_and_responses.md §“ETag and Last-Modified” requires both headers on “VERSION, VERSIONED_OBJECT, or other resources that have versioning or unique state identifiers”, with Last-Modified “derived from VERSION.commit_audit.time_committed.value”. Only the ETag half shipped previously:

    • Last-Modified (IMF-fixdate) is now emitted on every VERSION read (…/versioned_composition/{uid}/version[/{version_uid}], …/versioned_ehr_status/version[/{version_uid}]), on all COMPOSITION and EHR_STATUS reads and writes (including the delete 204 and the FLAT/STRUCTURED representations, whose version identity is serialization-independent), and on the EHR create 201. The value is the served version’s commit instant — read off the VERSION envelope where the body carries one, and off the version row / commit result for the bare COMPOSITION and EHR_STATUS representations, which have no commit_audit of their own.
    • GET /ehr/{ehr_id} and GET /ehr?subject_id=… now carry the weak ETag built from EHR.ehr_id.value — the source the spec section itself names. (No Last-Modified: the RM EHR root is not a VERSION, and time_created is not a last-modification instant.)
    • The ADL2 template responses (POST /definition/template/adl2, GET …/adl2/{template_id}, GET …/adl2/{template_id}/{version}) now carry the weak ETag their ADL 1.4 siblings already emitted. The value is the resolved ARCHETYPE_HRID, so addressing a template by a partial id or major-version prefix still yields an ETag that changes when the served artefact does. CONTRIBUTION creation still omits Last-Modified (the commit instant is not carried out of the version-set commit yet) and is marked TODO in the service layer.
  • Committal request headers: client change_type honoured, DELETE accepts the headers, deprecated openEHR-AUDIT_DETAILS spelling restored (#395). Three divergences from ITS-REST overview Requests_and_responses.md §“openehr-version and openehr-audit-details” + §“Deprecated headers”:

    • A client-supplied AUDIT_DETAILS.change_type (e.g. change_type.code_string="250" for an amendment) is now merged into the commit instead of being silently replaced by the operation default — the spec lists change_type first among the client-suppliable attributes and requires “whatever is provided it MUST be merged”. The value is validated against the openEHR audit_change_type group (out-of-group → 422, AUDIT_DETAILS.Change_type_valid) and against the operation (a contradicting code such as 249|creation| on an update is rejected; the exact status is spec-unassigned — see ambiguity AMB-54 — and returns 400). Applies to the direct COMPOSITION/EHR_STATUS/DIRECTORY commits and the demographic party/relationship commits alike.
    • DELETE /composition/{id} and DELETE /directory now accept openehr-version/openehr-audit-details and merge the supplied description/committer/system_id into the 523|deleted| commit audit — the spec requires the headers accepted on PUT, POST and DELETE.
    • The bare deprecated header name openEHR-AUDIT_DETAILS (the exact spelling in the spec’s deprecation table, which is a different HTTP header name than openehr-audit-details) is accepted again alongside the 1.0.3 dotted forms and the current name; the current name still wins on conflict. The audit_change_type constant set now mirrors the complete TERM group (all nine codes), locked by a test against the terminology bundle. Five new CNF catalogue cases pin the merge family end-to-end.
  • Demographic API: response-header discipline and If-Match handling (#394). Three MUST/SHOULD-level divergences from the ITS-REST overview (Requests_and_responses.md) are corrected on the /demographic surface:

    • Location is no longer emitted on reads, deletes, or 409/412 error responses. The header now rides create/update writes only, per §Location (“MUST NOT be used to indicate an alternate representation of an existing resource”; “MUST ONLY be used for resource creation … or redirect responses”) and §“Deprecated headers”, which deprecates it on GET and DELETE. Those responses keep the weak ETag (and Last-Modified where known), so a client reading the version identity is unaffected; a client that was following the Location of a GET/DELETE must use the request URL it already has.
    • If-Match now accepts the weak W/"…" form the server itself emits as the ETag, alongside the bare-quoted and unquoted forms — previously echoing the server’s own ETag back was rejected as a malformed precondition (400). The full OBJECT_VERSION_ID is compared case-insensitively (BASE composite-identifier semantics), so a case-variant creating_system_id no longer raises a spurious 412. A syntactically invalid If-Match remains a 400 and is never silently ignored.
    • The versioned_party / versioned_party_relationship reads (the container, its revision history, and the version reads) now carry the weak ETag and, where the served body exposes the commit instant, Last-Modified — both SHOULD-present on VERSION/VERSIONED_OBJECT responses.

Added

  • Several terminology servers can now serve one instance at the same time. Every entry under [terminology.external.providers] is started up, not just default, and a new [terminology.external.routes] map sends each terminology to the server that serves it — the key is a terminology id (SNOMED-CT) or a system URI (http://snomed.info/sct), matched case-insensitively, and the value names a provider. A terminology with no route goes to the provider named default (or to the sole configured one). Routing applies to the whole terminology surface: the /terminology/* extension API, AQL TERMINOLOGY(…), and composition validation. So SNOMED CT can live on one server while LOINC or ICD live on others — the deployment reality openEHR’s terminology chapter describes. Configuring a route to a provider that does not exist is now a startup error instead of a silent fallback.

  • Terminology servers can require OAuth2. A provider’s oauth2_client key now does something: it names an entry under [terminology.external.oauth2_clients] (token endpoint, client id, client secret or client_secret_file, optional scopes, refresh_leeway_secs, and client_secret_basic / client_secret_post), and the CDR obtains a client-credentials access token and sends it as a bearer credential on every request to that server. The token is cached and renewed shortly before it expires, so a validation burst costs one token request per token lifetime. A refused grant fails the call with a clear error — a request is never sent unauthenticated as a fallback.

  • Terminology servers can require a client certificate (mutual TLS). A provider takes three new keys — client_cert_path, client_key_path and ca_bundle_path — so the CDR presents a client certificate to that terminology server and verifies the server against that server’s own trust anchors. The identity is per provider because a client certificate is issued by the peer’s PKI: a deployment enrolled with a national SNOMED CT service, a commercial value-set server and an in-house server holds three different certificates, and repeating the same paths covers the case where one identity really does serve them all. ca_bundle_path replaces the default trust anchors for that provider, so a privately-issued terminology server is pinned to its own CA instead of also accepting the whole public web PKI. There is no option to skip verification — server-certificate and hostname verification stay on for every provider; the bundle changes which anchors are trusted, never whether the server is checked. Anything broken (one half of an identity, an unreadable PEM, a key file holding no key, a CA bundle holding no certificate) fails at startup, never at the first validated code.

  • Composition commits can now check archetype value-set bindings against a live terminology server. When a template binds an ac code to an external terminology query, and [terminology.external] is enabled, committing a COMPOSITION resolves that query and requires the coded value to be a member of the value set it returns. A non-member is a 422 naming the path, the code and the bound query. If the value set cannot be resolved at all (server down, error response, no server configured for that terminology), the existing fail_on_error switch decides: false (the default) accepts the commit and logs a warning, true rejects it. With [terminology.external] disabled — the shipped default — nothing is resolved, no request is made, and commit behaviour is exactly as before.

  • An opt-in terminology Compose profile runs a real FHIR R4 terminology server beside the CDR. docker compose --profile terminology -f docker-compose.yml -f docker/sut-terminology.yml up starts a digest-pinned HAPI FHIR JPA server (host port 8090, EHRBASE_TERMINOLOGY_PORT) plus a one-shot container that seeds it — over the server’s own FHIR API — with two synthetic test code systems and their value sets, one SNOMED-CT-shaped and one LOINC-shaped, and verifies $validate-code and $expand before exiting. The overlay switches on the [terminology.external] providers now shipped (disabled) in docker/ehrbase.dev.toml, so the plain quickstart is unchanged. No licensed terminology content is distributed: the fixtures live under the reserved example.test domain, and the SNOMED-CT-shaped and LOINC-shaped codes are invented for the test corpus.

  • The conformance record now covers the terminology-routed surface. scripts/conformance.sh composes the terminology profile for every ehrbase-rs run, and the catalogue gained eight cases: AQL TERMINOLOGY() resolved through the routed server (the Boolean validate form answering true and false, and the expand filter over committed data), the two-simultaneous-servers routing proof, and commit-time archetype constraint-binding validation (a member code accepted, a non-member refused, and the unresolvable value set under each declared posture). A party’s ixit.json gains a terminology block declaring its terminology servers, the namespaces each answers for, and its fail-open/fail-closed posture; a party that declares none has those cases recorded not-applicable with that citation instead of failed.

  • POST /admin/dump now serves the openehr_canonical_xml logical format, which used to answer 501. Both openEHR export formats are available: the default openehr_canonical_json keeps each version’s content inline in the archive’s segment files, while openehr_canonical_xml writes each version to its own versions/<version_uid>.xml entry — a complete ORIGINAL_VERSION document under the openEHR-published <version> root, readable by any tool that speaks canonical openEHR XML. The archive’s own bookkeeping (manifest.json, the segment files) stays JSON in both formats, because openEHR publishes no XML document form for it. POST /admin/load is unchanged for callers: it still takes only a location and now reads the logical format out of the archive’s manifest, exactly as it already detected the container. Both formats round-trip in all three containers (loose, archive.zip, archive.7z) and reproduce every record byte-for-byte. A single unreadable versions/*.xml entry is reported against the one EHR it belongs to and skipped, while the rest of the archive loads.

  • The definition and messaging extension routes now document their refusal branches. Every ADL 1.4 / ADL 2 archetype route, every /message route and every PARTY_RELATIONSHIP route declares 401 (no valid principal) and — on the writes — 403 (a principal holding the configured read-only role) in the served OpenAPI, so a client can see the whole answer set of an endpoint before it calls it. The TDD batch additionally documents its 413 boundary: the batch has no cardinality limit of its own, only the server-wide request-body limit.

  • Conformance: the admin extension batteries now test what must be REFUSED, not only what must work. A server that accepts what the contract forbids is as non-conformant as one that refuses what it must accept, and the activity-report, EHR/demographic archive and dump/load batteries proved only their happy paths. They now also drive the unauthenticated (401) and non-administrative (403) probes on every route of each family, every argument-type refusal (a service outside the enumeration, the three ways a time interval can be malformed, a malformed id in an archive list, an unknown export format, a non-positive segment size), the empty-selection and repeat-archive boundaries, and the zip / 7z / loose container-detection round trips on load — with the duplicate-report body now asserted rather than assumed. The branches that cannot be driven from a client (the admin-disabled 405, which needs a differently configured deployment, and the corrupt-archive 5xx, which needs bytes placed on the server’s own file system) are recorded as explicit boundaries and covered by in-process tests instead of being silently absent.

  • The documentation site renders mathematics. Formulas on the performance pages (the open-loop arrival schedule, the population-anchored write-rate derivation) are now typeset with KaTeX, pre-rendered to static HTML at build time — pages stay self-contained with no client-side script and no CDN request; the KaTeX stylesheet and fonts are served by the site itself.

  • Conformance: ADL 1.4 archetype provisioning is now tested rather than excused. openEHR’s released REST API defines no ADL 1.4 archetype resource, so the capability used to be reported as “excused — unrealized on this technology profile” even though this server serves archetype routes of its own design. Six conformance cases now execute against /definition/archetype/adl1.4 — upload with source read-back, an unparseable-source refusal, listing, and the get/delete branches including their not-found halves. Because openEHR gives the capability no wire, the published certificate marks the row extension and it no longer gates the CORE profile — a conscious, register-recorded departure from the conformance profiles book, which requires a capability the release gives no wire for.

  • The admin dump/load archive now supports 7z compression. POST /admin/dump accepts compression_format: "7z" alongside zip and the uncompressed form, packing the same archive entries into one archive.7z; POST /admin/load detects and reads all three container forms without being told which one it was given. (The openehr_canonical_xml logical format remains a declared 501 boundary — the archive’s XML form is a design of its own, tracked separately.)

  • Repository dump/load and the whole messaging surface are now HTTP routes — the last service capabilities that had no wire. Under the existing EHRBASE__ADMIN__ENABLED gate and ADMIN role, POST /admin/dump writes an archive of every EHR to a location on the server’s file system and POST /admin/load populates the repository from one; both answer 200 with the per-entity report the openEHR service model defines (empty when everything succeeded), and a load into a non-empty repository reports each already-present EHR rather than failing. Under the ordinary clinical authentication — these are not admin routes — the new /message group serves EHR Extract export (GET /message/export/{ehr_id}, POST /message/export by specification) and import (POST /message/import for a whole-EHR clone, POST /message/import/{ehr_id} to add content to an existing one), plus Template Data Document import (POST /message/tdd/{ehr_id} and its all-or-nothing /batch sibling). Like the admin extensions, all of these are ehrbase-rs extensions: the openEHR service model defines the operations, the released REST API surfaces no endpoint for them, and no openEHR conformance claim rests on the URLs — see the book’s Operations page and the served OpenAPI document, which flags every one of them.

  • The dump archive can be written as a ZIP. POST /admin/dump accepts the service model’s zip compression format, packing the manifest, segments and multimedia blobs into a single archive.zip instead of loose files. Load takes no format argument and detects the container, so an archive always reads back the way it was written.

  • The admin API gained an activity report and archiving, and the definition API gained archetype provisioning — service capabilities that had no HTTP route until now. Under the existing EHRBASE__ADMIN__ENABLED gate and ADMIN role: GET /admin/report/contribution[/count], GET /admin/report/versioned_composition/count and GET /admin/report/composition_version/count report CONTRIBUTION and COMPOSITION-version activity per service over an optional ISO 8601 time interval, and POST /admin/archive/ehrs / POST /admin/archive/parties mark a named set of EHRs or demographic parties archived (a read-neutral, idempotent, all-or-nothing marker — never a delete). Alongside the released template routes, the definition API now serves the ADL 1.4 archetype store (POST/GET /definition/archetype/adl1.4, GET/DELETE /definition/archetype/adl1.4/{archetype_id}) and the ADL 2 archetype/artefact views (GET /definition/archetype/adl2[/count], GET /definition/artefact/adl2[/count], DELETE /definition/artefact/adl2/{artefact_id}). All of these are ehrbase-rs extensions: the openEHR service model defines the operations, the released REST API surfaces no endpoint for them, and no openEHR conformance claim rests on the URLs — see the book’s Operations page and the served OpenAPI document, which flags every one of them.

  • The measured hospital-simulation workload now exercises every claimed capability (#625). The performance run used to touch about a third of the capabilities the conformance statement claims, and the rest were listed as “not yet exercised” catalogue gaps. Sixteen new operations joined the measured workload — demographic registration (person create/read/amend plus relationship churn), template example and ADL 2 definition polls, advanced and terminology-backed AQL reads, Simplified-FLAT commit and read-back, version-provenance (signature) reads, the System API options probe, the SMART service-discovery fetch, and the two access-control refusals — so the published Workload Coverage table now answers “yes” for every claimed capability except eleven that carry a per-capability, register-linked reason: either the operation would destroy the measured population mid-run (physical deletion and the released admin delete API), or openEHR defines no wire and this product serves no route for it, leaving the load instrument nothing to send. No row is left undecided, and a future journey that lands one of those capabilities is forced to delete its exclusion.

  • Measured runs can now drive a SMART-secured deployment: the load client mints the scope-limited access token its ixit principal declares (once per token lifetime, never per request), so a deployment running the SMART resource-server posture — the standard EHRbase-rs conformance posture — is measurable at all. Boundary probes address the read-only and unauthenticated principals a deployment declares; a deployment that declares none simply runs the workload without those journeys.

  • Conformance: the last two untested capabilities now carry real executed batteries (#624). “Demographic archetype validation” and “Bulk EHR load” were the two capabilities the conformance report listed with no cases — named in openEHR’s conformance profiles book, but never actually exercised. Both are now tested against the released REST wire. Demographic archetype validation gets eight isolated cases over the party-commit endpoints: a committed PERSON/ROLE is refused when it is not archetype-rooted, when its root archetype identifier contradicts its own archetype details, when an optional list (contacts, roles, capabilities) is present but empty, and when an identity’s value is missing or carries the wrong openEHR type — plus an accept case proving a fully archetyped party, contacts, addresses and languages included, is stored and read back intact. Bulk EHR load is verified as what it actually is on released wire — a population loaded through the ordinary EHR and composition endpoints — with one case covering breadth (eight EHRs, one composition each, all identities distinct and every document read back unchanged) and one covering depth (four commits into a single EHR, each independently addressable, with an AQL query over that EHR returning exactly the loaded set). Both capabilities are now claimed in the published conformance statement, and their case-count floors are recorded so the coverage can only grow.

  • Conformance: the PARTY_RELATIONSHIP capability is now tested rather than excused (#623). openEHR’s released REST API defines no PARTY_RELATIONSHIP resource, so the six relationship operations used to be reported as “excused — unrealized on this technology profile” even though EHRbase-rs serves them. They are now driven for real: fifteen conformance cases execute against the /demographic/party_relationship routes this product serves of its own design, covering create, read, read-at-time, read-at-version, update and delete plus their refusal branches. The certificate marks the row extension, which is a promise as much as a label — no openEHR profile result may rest on a route openEHR does not specify, and the runner now fails validation if that line is ever crossed (a new realization-scope gate, with the binding’s route required to appear in the published extension-surface declaration). Such cases are also skipped, with a cited reason, for any system under test whose conformance statement does not claim the capability — a route openEHR does not specify is an offer only the party making it answers for, so the published comparison against other products never charges them for routes they never offered.

  • Conformance runner: a certification claim can no longer be hollow (#622). cnf-runner validate now reads the committed party statements beside the artifact root and relates every claim to the catalogue, so three new gates fail before any system under test is even composed. claim-completeness rejects a claimed capability with no verdict-bearing case at all, and requires a capability whose every case is excused (because the openEHR release publishes no wire for it) to name the register entry that adjudicated that — an excuse that outlives the missing wire is a finding too. capability-depth gives every capability a min_cases floor so one token case can never certify it; floors only ever ratchet up. workload-coverage requires every claimed capability the measured hospital-simulation workload does not exercise to carry an adjudicated exclusion, which the conformance certificate now prints with its reason in place of the previous bare “NO — catalogue gap” cell. The certificate’s Profile Report also gains a Realization column saying whether a capability was verified over released ITS-REST wire or over routes this product serves of its own design (the latter can never gate an openEHR profile tier).

  • Conformance runner: the SMART on openEHR boundary is now executed, not declared (#538). Three behaviours that were previously carried as statement-level claims are real conformance cases: the /.well-known/smart-configuration discovery document (served from the Platform base URL as application/json, advertising the required org.openehr.rest service at an absolute base URL), the resource-scope grammar that lets a granted scope reach exactly the operation it names, and the 403 refusal of a request the granted scopes do not permit. Because SMART is off by default, they run in their own lane: CONF_SMART_MODE=1 bash scripts/conformance.sh boots the server with the SMART resource-server posture enabled (docker/sut-smart.yml), drives the SMART group, and writes to docs/conformance/<sut>-smart/; the default lane is untouched and remains the published baseline. To exercise scopes at all the runner now mints its own short-lived access tokens against a committed test issuer (tools/cnf-runner/party/smart/ — public test key material for the harness, never usable for anything else), because a CDR validates tokens and never issues them and the conformance stack runs no Authorization Server. A conformance target that does not run the SMART role simply does not declare the lane in its ixit.json, and these cases are recorded not-applicable with that citation rather than failed.

  • Conformance runner: two more wire behaviours are now measured, not excused (#539, #569). The bulk admin delete’s subset selector is exercised in the repeated ?ehr_id=a&ehr_id=b form the openEHR path template asks for, proving every named EHR is deleted rather than only the first; and the rule that a server stamps its OWN configured system identifier into a commit audit when the client supplies none is now checked against that identifier, not merely against “some non-blank value”. The identifier is a deployment fact no openEHR operation exposes, so a conformance target declares it in its ixit.json ("system_id": "…"); a target that declares none has those cases recorded not-applicable with that citation instead of being checked against a guess. Both behaviours were previously carried as cited coverage gaps.

  • Conformance coverage: calling a resource with the wrong HTTP method is now measured (#596). The openEHR REST specification says a method the specification recognizes but the addressed resource does not serve should be answered 405 Method Not Allowed, and the HTTP standard it defers to requires that answer to carry an Allow field listing the methods the resource does support. The conformance suite now proves both on a real resource — a DELETE to the EHR collection, which the specification serves only under POST and GET — instead of recording the behaviour as an untestable gap. The Allow check asserts that both specified methods are listed while tolerating any order and any additional methods a server chooses to support.

  • Canonical XML: choose the openEHR schema namespace per request (#196). openEHR publishes its XML schemas in two lineages that differ only by the namespace a document declares — http://schemas.openehr.org/v1 (the stable release) and http://schemas.openehr.org/v2 (the newer, trial release). You can now pick one with a version parameter on the XML media type: Accept: application/xml; version=2 returns the v2 namespace, and Content-Type: application/xml; version=2 declares a v2 request payload. A v2 response is labelled Content-Type: application/xml; version=2. Nothing changes for existing clients: omitting the parameter (or sending version=1) serves the v1 namespace under a plain Content-Type: application/xml, exactly as before, and request payloads in either namespace have always been accepted. Asking for a namespace the server does not serve is 406 Not Acceptable on Accept and 415 Unsupported Media Type on Content-Type. Operational-template XML (…/definition/template/adl1.4/{template_id}) is always v1 and ignores the parameter. The parameter is an EHRbase-rs extension — the openEHR REST specification predates the two lineages and defines no way to select one.

  • Conformance coverage: the ITEM_TAG routes are now measured (#288). All twenty-three released tag operations — the EHR-wide and demographic-wide listings, the COMPOSITION and EHR_STATUS families, and the five demographic party families — are enumerated by the conformance instrument for the first time; they have no openEHR service-model interface, so they were previously invisible to its coverage derivation. Thirty-two new cases turn the five tag laws into executed wire assertions: tag identity is the (key, target_path) pair, a container’s tag collection and a version’s tag collection are disjoint on read, write and delete alike, ITEM_TAG.target is served as the bare openEHR identifier, every typed tag route answers 404 for a uid of another kind (within the EHR space, within the demographic space and across the two), and the openehr-item-tag / openehr-version-item-tag request headers on a commit land in their own separate collections. Tag support is reported under a new ItemTags capability at the OPTIONS tier, matching the specification’s own statement that a server need not support ITEM_TAGs.

  • Conformance coverage: the COMPOSITION, CONTRIBUTION and PARTY resources are now exercised in canonical XML and in the Simplified Formats, not only in canonical JSON (#288). Eighteen new CNF cases drive Accept: application/xml reads and Content-Type: application/xml commits across composition create/update/latest/at-time/at-version, the VERSIONED_COMPOSITION container, the composition and contribution existence probes, and the whole PERSON create/update/read family, plus FLAT and STRUCTURED reads of a composition at latest and at time and FLAT/STRUCTURED composition updates. Each row asserts the negotiated response media type the specification makes a MUST, and the XML commits are compared against the canonical-JSON twin of the same resource, so a format-specific data loss shows up as a failure rather than a silent difference. One branch stays deliberately unexercised and is now recorded with its full derivation: the openEHR release declares application/xml for the CONTRIBUTION commit but publishes no XML form of the commit envelope, which is reported upstream rather than invented locally.

  • Served OpenAPI: complete documentation for the six Query operations (#482). The two ad-hoc and four stored AQL executions now document what the wire actually does. Every 200 declares the weak RESULT_SET ETag (an identifier of the result set — ours is a deterministic content digest, since the released ResultSet schema carries no id field) and carries a canonical RESULT_SET example: columns[] with the #N unaliased-column convention, rows whose cells are JSON primitives and canonical _type-tagged RM objects, and the optional meta (_type, _schema_version, _created in extended ISO 8601, and _executed_aql = the parameter-SUBSTITUTED text, with q keeping the query as submitted). The parameters now carry the released semantics: the named-$parameter binding law and its un-prefixed rule, the ehr_id duality (query parameter or openehr-ehr-id header, deprecated MixedCase spelling accepted, a conflict 400), offset’s default of 0 and fetch’s implementation-defined default with the one released prohibition (fetch cannot be combined with AQL TOP), the qualified-query-name grammar including the reserved aql, and the version exact/prefix matching law. Also declared: 415 on the three POSTs, request-body examples, and the Prefer-scope reason no query response carries Location or Preference-Applied. Where the released text is silent the declarations say so explicitly — the reserved protocol keys that never bind as AQL parameters, REST paging composing over AQL LIMIT/OFFSET, the URL-vs-body precedence on the POSTs, and the ehr_id-scope 404. Document only — no wire change.

  • Served OpenAPI: complete documentation for the seven EHR ITEM_TAG operations (#475). The EHR-wide read, the two per-target reads, the two collection replaces and the two key-scoped deletes now document what the wire actually does. The dual-form uid_based_id is spelled out with the released version/container sentence and the disjointness it implies (a tag has exactly one target, so container tags and version tags are separate collections and neither read sees the other). The PUT bodies are described as what they are — a bare JSON array of UPDATE_ITEM_TAG (key required, value/target_path optional, target/owner_id server-assigned from the route and ignored if sent), with [] quoted as the clear-all form, (key, target_path) as the identity, last-wins on a duplicate pair, an empty target_path normalizing to absent, and the 200/204 Prefer split (204 by default, 200 carrying the full RESULTING list, return=identifier resolving to minimal because an ITEM_TAG has no uid). The deletes document their SET semantics (every tag under the key on the addressed collection) and the released third 404 trigger that makes them deliberately non-idempotent. Every operation now declares the target guard’s 404s (unknown, foreign-EHR, wrong-kind or missing-version target), the JSON-only reality (406 for an XML Accept, 415 for an XML Content-Type — no ITEM_TAG type exists in the canonical XML ITS), the RM-invariant 422 family on the writes, the ehr_tags_get filter semantics (AND-combined, exact, case-sensitive, scalar, unbounded), and real ITEM_TAG examples including a VERSION-targeted tag. Also recorded: no tag route serves ETag/Last-Modified or accepts If-Match — a tag has neither a version nor a uid — and the released-text defects met on the way (the aggregate read’s COMPOSITION-typed response schema, the _updated responses’ copy-pasted “retrieved” wording, tag_key vs the key path parameter, and the “(logically) deleted” wording on a non-versioned resource). Document only — no wire change.

  • Served OpenAPI: complete documentation for the three CONTRIBUTION operations (#464). The native change-set commit now declares the whole NewContribution envelope — versions[] of UPDATE_VERSION (preceding_version_uid, signature, lifecycle_state, attestations, data, commit_audit) plus the change-set audit, the accepted _type spellings (UPDATE_AUDIT / AUDIT_DETAILS / omitted), the server-set time_committed, the honoured-if-unused client uid, and the committer/system_id copy-down — with a canonical two-member example (a COMPOSITION creation plus an EHR_STATUS modification) and the SPECITS-84 rule quoted: the envelope stays canonical JSON, only each versions[i].data takes the FLAT/STRUCTURED form. Every branch is documented: 201 with the weak ETag carrying the contribution uid (not a version uid), Location, Preference-Applied and the Prefer-conditional bodies (the representation lists the minted version OBJECT_REFs, the identifier body the contribution uid, minimal an empty 201); 400 with the released first-version-of-a-MODIFICATION trigger; 404; 406; 409 (client uid in use — released — plus the non-modifiable EHR, duplicate singletons and an EHR_STATUS delete member, flagged as ours); 412 for a stale member preceding_version_uid; 415; and the full 422 family (empty versions, out-of-group change types, data on a delete/attestation member, missing data, template and RM-invariant failures). The by-uid GET documents the plain-UUID contribution_uid, Prefer: return=representation, resolve_refs (members resolved to full ORIGINAL_VERSIONs, which is also what makes a simplified Accept meaningful), its 200 headers and canonical example (members as OBJECT_REFs, full AUDIT_DETAILS with optional description), and 400/404/406. The contribution-list route is prominently flagged as our own extension with no openEHR spec behind it, and its offset/fetch clamping (0 / 20, capped at 100 — never a 400) and row shape are now described accurately. Document only — no wire change.

  • Served OpenAPI: complete documentation for the five DIRECTORY operations (#457). Every response now declares its headers (weak ETag, Last-Modified, Location, Preference-Applied, item-tag echoes) and the reads and writes carry canonical FOLDER examples (nested folders, items as OBJECT_REFs); the writes document the If-Match precondition — carried in the header because these routes have no version segment, so a stale value is 412, never 409 — plus Prefer, the openehr-version / openehr-audit-details committal headers and the item-tag headers, and the canonical-JSON/XML-only request bodies (a Simplified-Format Content-Type is 415, an unfulfillable simplified Accept 406: a FOLDER is not templated). The version_at_time and path query parameters are described with the released sentence plus our register-documented resolution rules (root-implicit, leading-slash tolerant, folders-only, first-match; a future time serves the latest version, a time before the first commit is 404), and every branch the wire serves is documented — the deleted-directory 204 on both reads, the DELETE’s 204 carrying the new deleted version’s identity, the 404s (including an EHR with no directory), the 412s with the latest-uid ETag, 400/406/415/422, and the 409s that are our own design (creating a directory when one already exists, and a non-modifiable EHR), each flagged as such. Document only — no wire change.

  • Served OpenAPI: complete documentation for the eight COMPOSITION and VERSIONED_COMPOSITION operations (#450). Every response now declares its headers (weak ETag, Last-Modified, Location, Preference-Applied, item-tag echoes) and a canonical example; the commits document the openehr-version / openehr-audit-details / openehr-template-id request headers and the four negotiable media types (canonical JSON/XML plus application/openehr.wt.flat+json and application/openehr.wt.structured+json); and every branch the wire actually serves is described — the GET’s deleted-version 204 for all addressing forms, the DELETE quartet (204 carrying the NEW deleted version’s identity, 400 already-deleted, 404, 409 not-latest with the latest-uid ETag), 412/415/406/422, and the 409s that are our own design (duplicate live persistent COMPOSITION per template, and a non-modifiable EHR), each flagged as such. Document only — no wire change.

  • Served OpenAPI: complete documentation for the seven EHR_STATUS and VERSIONED_EHR_STATUS operations (#443). Every response now declares its headers (weak ETag, Last-Modified, Location, Preference-Applied, item-tag echoes), canonical examples, and the 406/415 negotiation branches; the EHR_STATUS update documents the openehr-version / openehr-audit-details committal headers and the Prefer: return=identifier response shape. Document only — no wire change.

  • Last-Modified on VERSIONED_OBJECT container and revision-history reads (#442). GET …/versioned_ehr_status, GET …/versioned_composition/{uid}, and both …/revision_history reads now carry Last-Modified derived from the newest held version’s commit instant, alongside the existing container-uid weak ETag (ITS-REST overview Requests and responses §“ETag and Last-Modified”: both headers SHOULD accompany a VERSIONED_OBJECT response).

  • [server] system_id — the deployment’s own openEHR system identifier is now configurable (#424, EHRBASE__SERVER__SYSTEM_ID, default unchanged at ehrbase-rs.local). The value is stamped into EHR.system_id at EHR creation (RM EHR Information Model §EHR Identifier Allocation: the identifier “that would normally be used for locally created EHRs”), into AUDIT_DETAILS.system_id whenever the client supplies none through openehr-audit-details (the REST API requires the server to “set it to its own configured system identifier”), and into every minted OBJECT_VERSION_ID.creating_system_id. Previously it was a hard-coded constant that no configuration could change. Choose it before the first EHR is created and keep it stable — the value is stored per EHR and per version, so a later change affects only newly authored data and never rewrites existing identifiers. It is distinct from [server.identity], which is only the OPTIONS manifest’s display identity. An empty value, or one containing the OBJECT_VERSION_ID separator ::, is refused at boot.

Removed

  • The bare-root OPTIONS / alias of the System API endpoint (#420). The System API defines exactly one location for the Options-and-Conformance operation — the API base-path root (OPTIONS {base_path}, e.g. /ehrbase/rest/openehr/v1); the extra bare-root mount was our own duplication and answered identically. Clients probing OPTIONS / must use the base path.

3.11.0 - 2026-07-26

Added

  • Admin console: EHR_STATUS editing and a status version history (#306). The EHR detail screen’s Status tab is no longer read-only: an Edit status card toggles is_queryable and is_modifiable and edits other_details (canonical-JSON ITEM_STRUCTURE; blank removes it), committing a new EHR_STATUS version conditionally on the version the screen loaded. Every other attribute — the subject included — is sent back exactly as the CDR served it, so an edit can never drop what the form does not show; a non-object other_details is refused before anything is sent, and a rejected document keeps the CDR’s own diagnostic on screen beside the form. If another client committed a new status meanwhile, the write is refused rather than overwriting it, and the console says so with what to do next. A new Status history tab adds the versioned view: the VERSIONED_EHR_STATUS container plus the selected version’s envelope facts, the revision history newest-first, a date-and-time lookup that resolves the version extant at that instant, and any version’s document opened by its own OBJECT_VERSION_ID. A non-queryable EHR’s warning now points at the toggle that fixes it.

  • Admin console: SMART scope previewer + effective identity (#299). The user menu’s “View scopes” drawer no longer prints a raw list of scope strings. It now states who you are and what decides what you may do — the authenticated principal and the policy source behind it (a Basic session replays its CDR account and carries no SMART scopes; an OIDC session’s roles and permissions come from the same access token whose scopes are listed) — and renders every scope as its parsed grant: the compartment it delegates to (patient/user/system), the resource family and id pattern it reaches, the create/read/update/delete/search operations it permits, and a broad access marker on a bare *. Launch contexts and identity claims are labelled as such, and an unrecognised scope stays visible verbatim instead of vanishing. A new previewer field takes any scope string — or a whole space-separated claim — and renders the same reading, with an actionable explanation when a resource-shaped scope is malformed (a bad compartment, a missing or invalid .<permission> tail, an unknown resource). The drawer also states plainly that scopes narrow access and never grant it: the CDR remains the enforcer. The reading comes from the same scope grammar the CDR’s own SMART gate enforces with, so the console’s explanation cannot drift from the server’s behaviour.

  • Admin console: grouped multi-series result charts (#296). The results pane (both the point-and-click builder and the raw AQL editor) now charts every numeric result column instead of only the first one: one line per column, named by the column’s own alias, with a legend whose entries switch a series on and off — the last visible series stays on, so the chart never empties itself. When a column holds ISO-8601 date/times it is offered as the X axis and used by default, giving a real time scale in which the points sit at their true distance apart whatever order the rows arrived in; the row order remains available as the fallback axis. A single numeric column still draws as one plain line with no legend. The Table | Chart toggle is now offered for every non-empty result set, and a result set with nothing to chart (no numeric column, or a single row) explains that in the chart pane instead of showing a blank box.

  • Admin console: EHR-detail and System-panel completions (#315). The EHR detail screen now opens with a summary header read from the EHR resource itself (id, creating system, creation time, current EHR-status reference), so an unknown or mistyped EHR id is reported once at the top of the screen instead of once per tab. The Create EHR card takes an optional EHR id: supply a UUID to create that exact EHR (a non-UUID is refused before anything is sent, and an id already in use comes back as the CDR’s own conflict with what to do next), or leave it blank as before. The composition viewer gains Delete composition — the openEHR logical delete of the latest version behind a confirmation dialog, which returns to the EHR’s composition list on success and, if the version moved on meanwhile, says so instead of deleting the wrong one — and a Versioned object card reading the versioned composition and the selected version directly (lifecycle state, preceding version, contribution, signature, whether the version still carries content). The contributions tab opens with a contribution activity timeline of writes per day. On System, a conformance manifest card shows what the CDR advertises about itself through the openEHR System API (product, vendor, claimed conformance profile, and the API groups it actually mounts), and the served-OpenAPI card gains a per-family document selector whose choice lives in the URL (/system?openapi=query), so a family document is shareable and survives a reload.

  • Admin console: run a stored query with its parameters, at the version form you choose (#295). A stored-query row now offers Run, which opens a runner screen for that query: it shows the stored AQL, prompts one field per $parameter the query declares, and executes it on the CDR as a real stored query (POST /query/{name}[/{version}] carrying query_parameters) rather than re-sending the text as an ad-hoc query. The results land in the same results pane as everywhere else, with paging — except when the query sets its own LIMIT/TOP, which the screen says instead of fighting. All three openEHR version-resolution forms are selectable and labelled with the exact request they will send: latest (no version), a version prefix like 1 or 1.2 (the CDR picks the latest match), or an exact 1.2.0. A parameter value that reads as JSON is sent as that type (38.5 as a number, true as a boolean); anything else is sent as text, and quoting forces text ("0123"). A field left blank is not sent at all.

  • Admin console: open a stored query in the query builder (#295). Stored queries and the raw editor now offer Open in builder beside Open in editor: a query that fits the point-and-click builder’s model is loaded back into it — template, conditions, output shape, ordering and limit — with the next version proposed for saving, so a stored query can be revised visually instead of by editing text. The load is never lossy: the builder only accepts a query it can reproduce byte for byte, and anything else (a parameterised query, a hand-written shape the builder has no controls for) opens with a notice naming exactly what it could not express and a link to work on it in the raw AQL editor.

  • Admin console: the stored-query and template tables are paged (#298). Both listings now carry the console’s shared pagination footer — which rows are on screen out of how many (26–50 of 137 templates), previous/next, and a rows-per-page choice (25/50/100). The page and the window size live in the address bar (?page=/?size=), so a page is shareable and survives a reload, the browser’s back/forward walk the pages, and the controls work before the console’s WebAssembly bundle has loaded. The templates filter still narrows the rows client-side; the footer counts what the filter left. Deleting the last row of the last page lands on rows rather than on a blank table, and a hand-typed window size is clamped to a sane range.

  • Admin console: a real document viewer (#297). Every pane that shows a wire document — the composition viewer, the EHR status tab, the directory raw mode, a contribution, a template’s OPT and example tabs, a stored query — now offers three views of it plus a Copy button. Highlighted (the default) shows the byte-exact document with JSON/XML syntax highlighting from a pure-Rust tokenizer (no JavaScript, no new dependency; a very large document is shown unstyled instead of tokenized), Raw shows the same text unstyled, and Rendered shows a template-free clinical reading of a canonical openEHR JSON document: RM section headings with their type and archetype node id, and one label/value row per ELEMENT — quantities with their units, coded text with its terminology code, a null-flavoured leaf saying so. The rendered view needs no operational template, so a composition whose template was since removed still reads normally; it folds away the bookkeeping (language, territory, category, uid) that the raw views keep in full, and is read-only — nothing is stored anywhere.

  • Admin console: stored-query versions are reachable (#336). Both save surfaces (the point-and-click builder and the raw AQL editor) now carry an optional Version field beside the namespace and name, and state under the fields exactly which store a click will perform: leaving it empty stores at the server-assigned version and replaces what is there, while a major.minor.patch version stores a new immutable version and is refused with the CDR’s own message if that pair already exists. Open in editor now keeps the version it loaded and proposes the next minor one, so editing a stored query publishes a new version instead of colliding with the one it came from, and a partial pattern (1, 1.0) is refused in the save field with an explanation — that form selects the latest matching version when reading a query, and is not something to file a definition under.

  • Admin console: an Operations panel (/operations) over the CDR’s operational surfaces — dependency health, build and specification provenance, the metric registry, and runtime log control. The health card reads the public readiness probe (GET /health/readiness) and renders the aggregate plus one row per indicator, explaining on screen how that differs from the topbar status pill; the build card reports the CDR’s version, git commit, rustc, PostgreSQL target and openEHR specification pins; the metrics card shows four headline tiles plus a browser over the whole registry, with the selected metric in the URL so a view is shareable; and the log card changes the live log filter (and resets it to the boot value) behind a confirmation dialog that names the consequence, re-reading the CDR’s answer so the panel shows what the server confirmed. The screen appears in the sidebar only when the CDR serves its management surface (the console probes GET /management/info); a deployment with it switched off sees no Operations entry at all, and an individual endpoint left off renders as a stated absence rather than an error. The redacted effective configuration is deliberately not duplicated here — the CDR serves the same snapshot on both its management surface and its admin API, so the panel links to the one viewer on the System screen.

  • Admin console setting cdr.management_base_url (EHRBASE_ADMIN__CDR__MANAGEMENT_BASE_URL): the CDR’s management surface including its base path, for deployments that serve it on a separate internal listener (management.port) or under a renamed management.base_path. Unset, the console derives {cdr.base_url}/management.

  • The compose quickstart enables the CDR’s management surface (docker/ehrbase.dev.toml): info/metrics at private, prometheus public, env/loggers admin_only — so the console’s Operations panel works out of the box on the dev stack. The surface remains off by default on the bare binary and in the Helm chart.

  • Admin console: delete templates, stored queries and EHRs when the CDR’s admin API is enabled. The Template Manager list rows and the template detail screen can delete an operational template; the stored-query rows can delete a query version from the CDR store (labelled “Delete from CDR”, clearly separate from the console-local “Remove group”); and the EHR detail screen can physically delete an EHR, returning to the EHR list on success. Every action confirms in a modal dialog naming the exact object (the query-group removal now does too), and every failure names the object and the next action — a template still referenced by a committed version, or a session without the ADMIN role, is refused by the CDR and reported as such. The console first asks the CDR which API groups it serves (the openEHR System API conformance manifest, OPTIONS on the API base path) and renders no delete affordance at all when the admin group is not among them.

  • Public health probes (/health/liveness, /health/readiness): the server now always serves a complete health family on its main HTTP port, unauthenticated and independent of every configuration switch — GET /health (unchanged: constant 200 OK, plain-text OK), GET /health/liveness (an identical alias under the orchestrator-conventional path), and GET /health/readiness (the indicator-backed probe: database ping, migrations applied and the in-memory component flags, 200 when the aggregate is UP/DEGRADED, 503 when a required component is DOWN, with the full per-indicator JSON body). They are mounted outside the API’s authentication and overload-shedding layers, so they answer without credentials and are never shed on a saturated server. This family is now the only health surface (see Removed).

Changed

  • ETag/Last-Modified on every versioned read (#368). The VERSIONED_COMPOSITION and VERSIONED_EHR_STATUS container reads, the VERSION-by-id reads, and both revision-history reads now carry the versioning headers (container/version uid as the ETag; the commit instant as Last-Modified where the body carries one) — previously only the at-time variants did.

  • Unqualified stored-query names are one identity everywhere (#366). A query stored without a namespace (PUT /definition/query/my_bp/1.0.0) now lands under the openEHR-assumed misc namespace — the same identity the by-name GET, the listing, the SM calls, and the admin delete address — so a bare-named query is no longer invisible to the admin delete (and vice versa). Descriptors return the canonical misc::-qualified name; a bare-name listing pattern also matches its misc:: composition.

  • Query GETs bind the spec’s named parameters (#364). AQL $parameter binds on GET /query/aql and GET /query/{name}[/{version}] now arrive as ordinary named query-string parameters (?temperature_from=36&…), exactly as the REST API documents them — values are typed JSON-first with string fallback, a $ prefix is tolerated, and the previous JSON-object query_parameters= form remains accepted (a named parameter wins a collision).

  • Version identity is the full three-part version_uid, compared case-insensitively (#367). Deleting a composition (and reading a version by id) with a fabricated creating_system_id is now refused (409 / 404) — previously only the version number was compared, so a made-up system id could delete the latest version. Conversely, a version_uid or If-Match differing only in case is accepted as the same identifier, per the openEHR composite-identifier case rule.

  • Item tags follow the spec’s identity and target model (#365). Two tags sharing a key on different target_paths now coexist (the ITEM_TAG identity is the key + target_path pair, per the ITS-REST item-tag prose) instead of silently collapsing; a version-addressed tag (…/composition/{version_uid}/tags) now tags THAT VERSION, disjoint from the container’s tags, instead of being folded onto the container; the tag’s target is returned in the RM shape (a bare HIER_OBJECT_ID or OBJECT_VERSION_ID, replacing the former OBJECT_REF wrapper — the released RM wins over the stalled OAS schema); tag routes now 404 when the addressed object is of the other kind; and the openehr-item-tag / openehr-version-item-tag commit headers write to their own distinct collections. Deleting by key removes every path under that key in the addressed collection (the wire has no path selector).

  • Admin console: accessibility and empty-state polish (#302). Table header cells are now announced as column headers by screen readers, and every icon-only control in the query builder (the catalog’s expand/collapse chevrons, the remove buttons on conditions, groups, columns and sort rules) plus the unlabelled column-alias, sort-path and sort-direction controls state what they do. Data regions that used to come back as a line of grey text — template usage, the served OpenAPI list, the commit-activity chart, an EHR’s compositions, the directory version history, a version’s audit, the query builder’s conditions and result rows, and a template filter that matches nothing — now render the console’s standard empty state: an icon, what is empty, and what to do about it. The user menu’s popover matches the rest of the console’s panels instead of the widget kit’s stock chrome, and the modal backdrop is a theme token, so it dims correctly in dark mode.

  • Contribution list shows the change type’s display rubric (#304). The EHR contribution-list extension (GET /ehr/{ehr_id}/contribution) now carries change_type_rubric beside the raw change_type group code — resolved from the openEHR audit_change_type terminology group by the CDR itself, so clients never maintain a local code table. The admin console’s contributions tab displays the rubric (code on hover). The SM-catalog delete_opt service path now also refuses with the same friendly 409-and-reference-count as the admin template delete while committed versions still reference the template, instead of relying on the raw foreign-key error.

  • openEHR BASE spec pin refreshed (#341). The vendored BASE 1.3.0 spec text and BMM codegen input now track upstream specifications-BASE master e4879576 (24 commits: the SPECBASE-48 RESOURCE_DESCRIPTION invariants, the SPECAM-82 CODE_PHRASE package move into base_types, SPECPR-426/386/460 corrections). No wire or validation behaviour changes: every behaviour-relevant item was verified already satisfied by the implementation; the regenerated crates differ only in documentation text and the CODE_PHRASE module location.

  • Readiness moved from /management/health/readiness to the public /health/readiness (and liveness to /health/liveness). Database-backed readiness no longer hides behind management.enabled + a probe switch — nothing has to be enabled for an orchestrator to probe the server. Point existing probes at the new paths.

  • Helm probes use the public paths on the main HTTP port: the chart’s httpGet liveness/startup probes hit /health/liveness and readiness hits /health/readiness on the http port, with no prerequisite — the previous render-time failure demanding config.management.enabled=true + config.management.probes_enabled=true is gone. Prometheus scrape annotations still point at the management surface (and its separate port when configured), unchanged.

  • Admin console: the System screen’s activity tile now links to the audit browser (#301) instead of stating that the CDR exposes no audit read surface — it does (the IHE ITI-81 retrieval the /audit screen has been browsing all along). The tile carries a one-line description of the trail and an Open audit browser button.

  • Admin console: every write now reports its failure as prominently as its success (#301). Uploading a template, creating an EHR, committing or updating a composition, saving a stored query or query group, and creating, saving, restoring or deleting a directory all raise a failure notification naming the object, what the CDR objected to (its diagnostic verbatim), and the next action to take — a stale version to reload, a role to sign in with, an unreachable CDR to check. Where the diagnostic is worth reading line by line (template validation, a rejected composition body) it also stays on screen beside the form as before. Previously a failed write showed a quiet inline message that was easy to miss after a run of green success toasts.

  • Admin console: query groups are now derived from the stored-query namespace instead of being named sets kept in a console-local file. A stored query is identified by a qualified name — namespace::name, the namespace optional and, per the openEHR REST specification, a reverse domain name whose purpose is separation of stored queries by team or organisation — so the console groups by exactly that: the Queries screen’s right-hand panel and the Dashboard’s cohort tiles are both derived live from GET /definition/query, and queries saved without a namespace collect under unqualified. The grouping consequently lives in the CDR: it is visible to every openEHR client, survives a console restart, needs no backup, and is identical across console replicas. The group create/edit/remove controls are gone — a query joins a group by being saved under that namespace — and both save surfaces (the query builder and the raw AQL editor) now offer a first-class Namespace field beside the query name, showing the exact qualified name the save will write. Existing local groups are not migrated: re-save a query under the namespace you want it grouped by.

  • Admin console: the EHR Directory tab creates an empty root folder, then the structured tree editor builds the hierarchy. The console no longer ships or stores named folder shapes to start from.

Removed

  • GET /ehrbase/rest/status/health is removed. It was a third name for the constant liveness answer already served at /health and /health/liveness, with no consumer anywhere in the product (no probe, no client, no documentation pointed at it). Point any caller at /health (load balancers, container HEALTHCHECK) or /health/liveness (orchestrator probes); GET /ehrbase/rest/status — the product status document, a different contract — is unchanged.
  • /management/health, /management/health/liveness, and /management/health/readiness are removed; the management surface is now ops introspection only (info, prometheus, metrics, env, loggers). The aggregate component view is the body of the public /health/readiness.
  • The management.probes_enabled and management.endpoints.health configuration keys are removed. Configuration is strict, so a config file or EHRBASE__MANAGEMENT__PROBES_ENABLED / …__ENDPOINTS__HEALTH environment variable still setting them fails at boot with an unknown-key error — delete the keys; the probes are always on.
  • Admin console: folder templates are removed. The named FOLDER-tree shapes the Directory tab could start from (and their admin-ui-folder-templates.json store) are gone; create the empty root and build the hierarchy in the tree editor, which commits it as ordinary directory versions the CDR owns.
  • Admin console: both console-local JSON stores are removed — admin-ui-groups.json (query groups) and admin-ui-folder-templates.json (folder templates). The console now keeps no local domain state at all: it has no database and writes no files, so every fact it shows lives in the CDR and reads the same for every client and every replica. Delete the files; nothing reads them.
  • The admin console’s groups_file configuration key (EHRBASE_ADMIN__GROUPS_FILE) is removed. Console configuration is strict, so a config file or environment variable still setting it fails at boot with an unknown-key error — delete it.

Fixed

  • The observability Compose overlay boots again (#321). Every server variable in docker-compose.observability.yml was written in a single-underscore form (EHRBASE_MANAGEMENT_*, EHRBASE_OTEL_*, EHRBASE_LOG_FORMAT) that the strict boot-time sweep of the reserved EHRBASE_ namespace rejects, so docker compose -f docker-compose.yml -f docker-compose.observability.yml up failed at startup with unknown-variable errors instead of starting the server. The overlay now uses the documented EHRBASE__… grammar (EHRBASE__TELEMETRY__OTLP_ENDPOINT, EHRBASE__LOG__FORMAT, EHRBASE__MANAGEMENT__ENABLED, EHRBASE__MANAGEMENT__PORT, EHRBASE__MANAGEMENT__ENDPOINTS__{INFO,METRICS,PROMETHEUS}), with unchanged intent: OTLP traces to the bundled collector, JSON logs, and the management surface public on internal port 9464 for the bundled Prometheus to scrape. A test now runs the real sweep over every variable the shipped Compose files set on the server service, so this class of drift fails in the test suite rather than at docker compose up.
  • The compositions_committed_total metric now counts (#332). The counter was declared and scraped but never incremented, so dashboards over it rendered a permanently empty series. Every commit route that lands a COMPOSITION version — create, update, delete, and a CONTRIBUTION commit — now increments it once per committed version, labelled change_type with the openEHR audit_change_type code recorded on that version’s audit (249/251/523/…). The increment happens after the transaction commits, so a rolled-back write is never counted. In the same audit of the metric registry, five metrics that were emitted but not registered (version_signature_invalid_total, authz_cedar_decisions_total, authz_remote_pdp_calls_total, atna_audit_rejected_total, atna_audit_reaped_total) now carry their # HELP/# TYPE descriptions in the /management/prometheus exposition.
  • Admin console: find-an-EHR-by-id works without JavaScript (#301). The finder is now a plain GET form: submitting it before (or without) the browser app loading redirects to the EHR’s detail screen server-side, and /ehrs?find=<ehr_id> is a shareable shortcut to any EHR. With the app loaded the lookup is unchanged — one client-side navigation, no page reload.
  • Admin console: template links and query-string values are now percent-encoded via the standard codec (#293); template ids containing reserved characters no longer produce broken links. The console’s hand-rolled percent encoder is gone — every internal link (the template detail link and its tab links, the stored-query “Open in editor” link, and the query builder’s “Open in raw editor” link) builds its path segment and query-string values with the urlencoding crate.
  • Admin console: deleting a document reference from a directory folder no longer risks row state attaching to the wrong sibling (#292). The item rows of the directory tree editor are now identified by a stable per-item identity instead of their position in the folder, so removing one reference leaves every remaining row bound to its own reference.

3.10.0 - 2026-07-25

Added

  • CNF total wire-surface coverage gate (#271): a new surface-coverage machine gate in the CNF runner (cnf-runner validate) fails on any spec-defined wire behaviour with no covering case and no adjudicated exception — enforcing breadth, not just pass rate (.claude/rules/testing.md §CNF coverage). It measures three axes against the RELEASED spec sources only (the SM platform interfaces + the ITS-REST docs text, never the vendored OAS): (1) every SM operation of the platform interfaces has an its-rest binding or a cited boundary; (2) every realized binding’s declared outcome and format branch is exercised by a case or excepted; (3) the cross-cutting wire behaviours (conditional headers ETag/Location/Last-Modified/ Prefer/If-Match, JSON+XML negotiation, the 406/415 families, the error-body and deprecated-media families) map to covering cases or exceptions. The authored, spec-cited exception ledger is a new committed artifact (tools/cnf-runner/artifacts/vocab/wire_surface.yaml, with a published JSON Schema); cnf-runner validate --specs … refreshes a deterministic coverage report at docs/conformance/coverage-report.md. Coverage only ratchets up.

  • CNF catalogue content deepening (coded-text value dimension, deferred grounds, spec-authored corpus) (#278): the coded-text content cases (CONT-DV_CODED_TEXT-validate_local_codes / -validate_ext_term) gain an acceptance-direction value dimension — value = the bound rubric, value ≠ rubric (an arbitrary label), and value = the raw code are all accepted (no RM invariant requires value to equal the coded rubric — the “must be the rubric” text is dv_coded_text.adoc Description prose, registered as AMB-55), while an empty value is rejected (the sole value invariant, RM dv_text.adoc §Invariants Valid_value: not value.is_empty); the synthesized OPTs now bind component-ontology rubrics for their local constraint codes. New functional coverage: a template-example round-trip (the generated example commits back cleanly) and a deprecated-media Accept → 406 response-side case (the ICS-conditional companion to the existing request-side 415, under AMB-39). Registered as spec-silent boundaries: Accept q-value negotiation strictness (AMB-56 — ITS-REST defines only “unfulfillable Accept → 406”, nothing about q-value weighting) and a simplified-inner-data CONTRIBUTION surface (AMB-57 — ITS-REST 1.1.0 commits CONTRIBUTIONs canonical-only). All remaining corpus-manifest “structural placeholder” markers are replaced with spec-authored fixtures or cited boundaries.

  • Version-signature conformance breadth (CNF): a distinct-signature-per- version case (the signature is computed over the version’s canonical form, which includes uid — two versions can never share a signature; RM common master06 §Digital Signature), backed by a new distinct_from fact on the runner’s signature assertion and a signature capture on the version-envelope read binding. DIRECTORY (FOLDER) version-signature cases land SM-anchored as N/A-with-citation on ITS-REST 1.1.0 (no versioned_directory resource — AMB-24), activating automatically if a later ITS release adds the endpoint. The runner’s binding-completeness gate now mirrors the interpreter’s variant-based binding selection.

  • ADL2/OPT2 templates are full FLAT/STRUCTURED peers of OPT 1.4 (#269): a FLAT (application/openehr.wt.flat+json) or STRUCTURED (application/openehr.wt.structured+json) composition commit keyed to an ADL2-registered template now resolves and is validated against that template’s archetype constraints, exactly as an ADL 1.4 commit is. Two behaviours were brought to parity: the am24 (OPT2) Web-Template builder now populates the archetype-conformance constraints (existence, cardinality, closed-attribute sibling sets, archetype slots, structural stubs) that composition validation reads — so an ADL2-template instance is archetype-constraint-checked, not only RM- and terminology-checked — and the runtime template resolver falls back to the ADL2/OPT2 store when a template id is not an ADL 1.4 template (previously a commit against an ADL2-registered template returned 422 “operational template not known”).

  • Citation metadata (CITATION.cff): the repository is now citable in research papers — GitHub renders a “Cite this repository” button (APA + BibTeX) from the new CFF 1.2.0 file (author with ORCID, Apache-2.0, abstract, keywords, release version/date). A citation-guard CI job schema-validates the file and enforces that its version matches the workspace version; the release procedure bumps version/date-released on every cut.

Changed

  • AQL engine: post-streaming optimization rungs (measured, one change per rung): the streaming shape’s dead root LATERAL is elided when the root is unreferenced (one fewer pk_node probe per version row; a bare uid/value projection now runs with zero node probes), and the archetype predicate column is case-folded at write (BASE base_types master05 §Composite Identifiers and Case) so archetype equality is plain indexed equality — LOWER() disappears from every containment hop. Measured on the seeded 10k bench: ward statement execution −11.3%, buffer reads −10.7%, planning −10.3%; stress knee re-measured at the committed 512 arrivals/s. The aql-probe instrument now attributes planner time per statement (pg_stat_statements.track_planning).

  • Docker / Compose deployment rework (#282): a from-the-ground-up rebuild of the container surface for smaller images, faster builds, and a production-grade posture on every build.

    • One Dockerfile per image, two targets, zero drift: docker/Dockerfile and docker/admin-ui/Dockerfile now each expose runtime-from-source (what docker compose build uses) and runtime-prebuilt (what CI uses), both sharing a single runtime stage — so the compose-built and published images can no longer diverge. The separate *.runtime Dockerfiles are removed.
    • Faster rebuilds: dependency compilation is split into its own cargo-chef layer, so editing application code no longer recompiles dependencies, and CI now reuses that layer across runs via an exported build cache.
    • Debian 13 + digest pinning: builder and runtime moved to Debian 13 (“trixie”); the runtime is distroless/cc-debian13 (non-root user 65532). Every base image is now pinned by immutable digest, and the bundled versions are refreshed — PostgreSQL 18.4, Keycloak 26.7.0, SeaweedFS 4.40, Grafana otel-lgtm 0.29.2.
    • Compose: the optional services are now opt-in behind profiles (--profile s3 for SeaweedFS, --profile keycloak for Keycloak); every service declares memory/CPU limits mirroring the Helm chart; Keycloak has a real healthcheck; and there is no hard-coded project name, so the dev, conformance, and E2E stacks no longer collide.
    • Build provenance no longer reads .git from the build context (which is now excluded, shrinking the context and stabilising the cache): the /management/info commit SHA flows through the standard REVISION build argument (the same value as the org.opencontainers.image.revision label) and degrades to unknown when unset — never a failed build.
  • Simplified Formats folded into openehr-its (#268): the FLAT / STRUCTURED / Web-Template implementation moved from the standalone openehr-flat crate into openehr-its as the openehr_its::flat module, mirroring the openEHR ITS component decomposition (Simplified Formats is a STABLE ITS-REST 1.1.0 sub-specification, alongside canonical JSON, XML, and the REST contract this crate already houses). Pure packaging refactor — no change to the FLAT/STRUCTURED/Web-Template wire behaviour.

Fixed

  • ADL2 filler-root naming in the projected WebTemplate: a use_archetype-filled archetype root resolved its display rubric in the component (filled archetype) terminology first, so the template-side slot id could false-positively match an unrelated internal id of the constituent (e.g. a filled OBSERVATION surfacing as “history”). The slot rubric now resolves in the introducing template’s own terminology first (ADL2 obliges the introducing artefact to define its node ids), with the component scope as last resort — FLAT paths over filled ADL2 templates carry the template-declared names.
  • CNF runner: openehr-template-id for in-flow-provisioned templates: the simplified-format commit header now resolves from the committed data set’s own manifest-declared template_id (falling back to the case’s provisioned template list), so cases that upload their template inside the flow (the ADL2 FLAT pair) drive the commit correctly instead of omitting the header.

3.9.0 - 2026-07-24

Added

  • Content structural conformance cases from the official schedule: the master15 COMPOSITION content×context tables and the master16 ENTRY-family tables (OBSERVATION, HISTORY, EVENT, ITEM_STRUCTURE) are now encoded under their verbatim official ids, replacing the ad-hoc structural cases that had been authored on the false claim that those chapters were empty; derivable catalogue extensions beyond the official cells survive as flagged addition cases.

  • Dual POC measured records on the v3.8.0 build, both directions published: ehrbase-rs earns class POC (normative hour at 2.03/s offered, worst p99 108 ms, 0 errors / 7,320 requests); upstream EHRbase 2.34.0 on the identical instrument, corpus, and resource floor does not (ward-dashboard AQL p99 10.9 s vs the 1 s ceiling, 2.4% errors). Comparison page and all measurement visuals derive from the committed runner artifacts.

Changed

  • Version-signature read verification is now strict by default (#273): with signing enabled and signing.verify_on_read unset, the server now recomputes the signature of every version it served and returns a 500 integrity fault on a mismatch, instead of the previous silent-pass (off) default that signed every version and then never checked it. Set signing.verify_on_read explicitly to warn (log + meter, still serve) or off (never check) to opt out. Client-supplied signatures (an author’s own signature, or one carried by an imported version) are tracked as such and are always stored verbatim and never re-verified, so strict-by-default never rejects a legitimately-stored foreign signature. Our-own-design integrity hardening — no openEHR spec governs server-side verify-on-read timing (RM common master06 §Digital Signature).

  • CNF catalogue audited case-by-case against the official spec text (#231): every case in every chapter re-verified across grounds, expectations, citations, fixtures, captures, and register linkage, with the findings applied directly to the catalogue and register (the durable record is the register + closed issues + git history). Highlights: spec-overreaching rejection rows removed (AQL TERMINOLOGY operation strictness; the mixed-precision interval rows now report-only under the SPECPR-380 openness); the SEC-BASIC proposal citations corrected; stale stub-era template ids fixed; the delete-latest-version OPT case realigned to the official version-less ground; the wrong-template update ground rebased onto a fixture that is valid against its own template; the physical-EHR-delete binding accepts the OAS-enumerated async 202; eight new ambiguity-register entries pin previously prose-only adjudications; and every phantom REQUIREMENTS.md pointer now carries its real anchor.

Fixed

  • Conformance-runner commit provisioning fails loud: a requires.commit key resolving to a plain composition fixture was silently skipped, leaving the case’s committed-state precondition unestablished; a single object now commits as a one-item set and any other shape is a provisioning error.

  • The measured-window driver accepts the spec-legal 204 No Content minimal-return form on create-family writes (ITS-REST: with Prefer: return=minimal a service SHOULD use 204 when no body is returned) — previously every upstream journey commit was falsely counted an error; and the upstream comparison stack’s database now gets the same shm_size floor as the ehrbase-rs stack (Docker’s 64 MB default starved its PostgreSQL during maintenance settling).

3.8.0 - 2026-07-24

Added

  • CNF catalogue: stored-query name-grammar cases — three new definition_query cases pin the ITS-REST Qualified_query_name grammar: a plain unqualified name and a namespace-less dotted name (the dot is part of the query-name character set, not a namespace separator) both store and read back, and the reserved query-name aql is rejected case-insensitively.
  • cnf-runner stress-compare — the cross-SUT stress overlay: both systems’ latency-throughput curves on one canvas, rendered deterministically from the two committed stress.json reports (driven by scripts/render/comparison.sh); both directions on equal footing.
  • Measured runs record resource telemetry: each measurement in results.json now carries an optional, schema-published resources block — per-container (server and database separately) CPU, resident memory, block-device and network I/O sampled every 10 s across the whole window (run-clock offsets, warmup/measured/drain phase stamps), plus the database volume’s on-disk size at four anchors (empty → scale seed → ward seed → after the window) with the derived bytes per committed composition. Sampling is enabled by the new optional containers block in the ixit (compose container names); without it a run records no resources and the report says so — telemetry never influences a class verdict. Two new rendered assets (the resource time-series and the disk-growth chart) join the perf-assets family and the book’s Performance chapter, drift-guarded in CI like every published number.
  • cnf-runner aql-probe — the seeded-corpus AQL optimization probe: fires the measurement machinery’s own AQL vocabulary against a freshly seeded server, records wire-latency percentiles per probe, and attributes the database-side cost per statement (pg_stat_statements through the ixit containers capability, degrading honestly without it). Report schema published (aql-probe.schema.json); exploration evidence only — never a conformance record.
  • Stress steps carry resource telemetry — every load-ladder rung records the same per-container CPU/memory/I/O series as the measured class runs over its own warmup+hold window, so a breached rung shows where it saturated; the stress progress stream now logs each rung’s verdict live (stable/BREACHED with the sustained rate, resource peaks, and named breaches) plus a ladder recap, and measured class runs log their verdict evidence at window end.
  • A diurnal day-curve arrival option for the extended 8/12-hour measured holds (ITU-T E.500 busy-hour semantics: the class floor is the busy-hour rate).
  • The conformance certificate gains a Workload Coverage section: claimed capabilities vs the set the measured hospital simulation actually exercised, with untouched claimed capabilities listed explicitly as journey-catalogue gaps.
  • scripts/generate-ckm-examples.sh — regenerates the committed CKM example payload skeletons from a running SUT’s example endpoint; scripts/vendor/ckm-templates.sh now vendors the runner’s journey template pack.
  • Conformance visuals: the capability-matrix heat grid (one cell per claimed capability, grouped by profile tier, evidence encoded as a CVD-safe color AND a glyph) and per-chapter outcome bars, rendered deterministically from the committed verdicts/results by the new cnf-runner conformance-assets subcommand (scripts/render/conformance-assets.sh, CI regenerate-and-diff guarded) and embedded on the book’s conformance and comparison pages (both SUTs) and the landing page.

Changed

  • --skip-seed and the sidecar corpus index are retired (CLI flags on perf/stress, the CONF_PERF_SKIP_SEED pipeline variable): every measurement instrument now always seeds a freshly composed, empty server and the stack is torn down afterwards — seed reuse bred stale-state confusion.
  • Measurement instruments settle database maintenance deterministically (vacuumdb --analyze through the DB container) after seeding and before every measured window and stress rung — a stale-statistics plan after the million-row seed cost a measured ~9× on the ward-worklist query; settling moves that debt outside every measured window, identically for every SUT.
  • The CNF measured-performance workload is now a full hospital simulation: the class cases (PERF-hospital_sim-*, renamed from PERF-mixed_load-*) schedule clinical journeys — ADT admission/discharge, vitals rounds, the medication loop, medicines reconciliation, asynchronous laboratory/imaging order-to-result pipelines, specialist/registry reporting, public-health notifications, chart review, ward dashboards with a registered stored query, versioned corrections, contribution audit review, workflow tagging, logical deletion, and template polling — expanding into 22 measured operation kinds instead of 4, each with its own HDR-V2 record. The population-anchored envelope is unchanged and now validator-enforced (the expanded write share must reconcile to the derivation’s 10:1..50:1 read:write band); journey payloads commit against 15 COMPOSITION-rooted openEHR CKM templates vendored with provenance.

Removed

  • The transitional benchmark lab (tools/benchmark, scripts/benchmark.sh, docker/benchmark/, the manual benchmark workflow, and the committed docs/benchmarks/** artifacts): all measurement is native to the CNF runner — measured class runs, the stress ladder, the AQL probe, and the cross-SUT stress overlay — and the comparison page now derives its performance side from the committed docs/conformance/<sut>/stress.json reports (upstream shown as “not measured yet” until its report lands, never a one-sided claim).
  • The completed ECC→CNF cutover comparison lane: the generated docs/conformance/cnf-comparison.md, the cnf-runner compare-ecc subcommand, the drift gate, and the preserved ECC catalogue/map (all in git history; the five deferred grounds are re-registered on the catalogue-deepening tracker). The docs/conformance/CATALOG.md pointer stub is gone with it, and the CNF 2.0 design record moved to docs/conformance/cnf-design.md as a permanent reference document.

Fixed

  • Storing a query under the reserved name aql is now rejected with 400, case-insensitively and whether or not a namespace is supplied (ITS-REST Qualified_query_name §NOTE — the name would collide with the ad-hoc /query/aql route). A three-part ns::aql::name name keeps working: its middle segment is the formalism, not the query-name.
  • A coded value whose text is not the template-bound rubric is now rejected at commit (422 naming the path, the committed value, and the bound rubric): RM DV_CODED_TEXT — “value must be the rubric from a controlled terminology” — enforced wherever the template itself is authoritative for the rubric (archetype-local at-codes and explicitly bound external term definitions, any bound language); openehr- terminology codes stay unchecked (the terminology ships official translations the template cannot enumerate), and a bound code with no rubric stays accepted. The once-accepted code-as-value instance is a pinned rejection.
  • Coded-text example values now carry the template-bound rubric: the Web Template builder resolved display labels only for local at-codes, so an external code’s rubric (OPT term_definitions keyed TERMINOLOGY::code, e.g. SNOMED-CT bindings) was lost and generated examples emitted the raw code as DV_CODED_TEXT.value — spec-invalid instance data (RM: “value must be the rubric from a controlled terminology”). The qualified key now resolves; the covid19 example regenerates with rubrics; every pack example commits clean on strict validators.
  • Child-assembled DV_INTERVAL values now carry the mandatory boundary flags: an interval built from lower/upper sub-path children (the FLAT builder’s container path — template examples included) previously omitted lower_unbounded/upper_unbounded/lower_included/ upper_included, making every half-open interval spec-invalid (BASE Interval: the flags are mandatory and Limits_consistent is unevaluable against an absent bound); the flags now derive from bound presence, an explicit datum flag wins, and the committed CCTA example is regenerated. Strict validators (upstream EHRbase) rejected the old instances with 422.
  • Population AQL with LIMIT now streams instead of materializing the corpus: a LIMIT-bearing, unordered, non-DISTINCT, non-aggregate population query lowers to a streaming FROM shape (the current-version spine with LATERAL node probes), so PostgreSQL stops at the LIMIT instead of building an archetype-anchor bitmap over every matching node first — measured on a million-composition corpus, the cross-EHR ward worklist drops from ~113 ms to ~2 ms per execution (~40× fewer buffer reads); ordered/aggregate/EHR-scoped queries keep the previous plan shape, and result semantics are unchanged. A version-field projection of uid/contribution_id/lifecycle_state no longer joins the audit table it never reads.
  • AQL cross-EHR queries with LIMIT no longer collapse under corpus scale: predicates on multi-valued (anchored) paths now lower as existential semi-joins (EXISTS — the predicate holds when ANY matched node satisfies it; deterministic where the previous first-match pick was plan-dependent), the archetype anchor index leads with the RM type so the whole CONTAINS-class + archetype boundary is one index probe, and queries that never touch audit fields no longer join the audit table. The measured ward-dashboard profile (p99 5.8 s at class-POC scale) drops to milliseconds-per-request territory.
  • The template example generator no longer collapses DV_INTERVAL wrappers onto a single constrained bound: interval-valued elements keep their interval identity (bounds as /lower//upper sub-paths per the Simplified Formats mapping), fixing generated examples the platform’s own validation rejected (the CKM CCTA report OPT); the CNF journey catalogue re-commits the CCTA imaging report.

3.7.0 - 2026-07-22

Added

  • The conformance pipeline assesses upstream EHRbase as a second system under test: CONF_SUT=ehrbase scripts/conformance.sh composes the official ehrbase/ehrbase:2.34.0 + ehrbase-v2-postgres images on fresh volumes (docker/sut-ehrbase.yml, readiness probed externally — the official image carries no in-container health tooling) and runs the same committed catalogue with upstream’s own committed party set (tools/cnf-runner/party/ehrbase/). The public comparison (docs/conformance/COMPARISON.md + the website comparison page) is fully generated from the two committed results/verdicts sets — profile verdicts, the 39-capability evidence matrix, and failure tables in both directions.

  • The conformance runner performs ISO/IEC 9646-style ICS-driven test selection: cnf-runner run --statement excuses option-gated cases whose register branch the party statement does not declare as N/A with citation (previously they ran and recorded spurious failures the verdict pipeline then excused).

  • Conformance badges carry measured amounts: per-tier badges read e.g. PASS 10/10 capabilities, the overall badge CORE+STANDARD PASS · 323/323 cases — derived from verdicts.json + the capability matrix, never hand-typed.

  • Read-only role support in RBAC: a principal carrying the configured authz.rbac.readonly_role (default READONLY) is refused with 403 on every write operation — creating an EHR, committing a composition, uploading a template, and any update/delete — even when it also holds granting roles such as ADMIN. Reads and AQL queries stay permitted, so a READONLY account is an authenticated, view-only principal. The dev compose stack ships an ehrbase-readonly account (password ehrbase) for evaluation.

  • CNF 2.0 reference runner, third increment — the executor and both verdict machineries: the data-driven flow interpreter under the five interpreter laws (per-row re-provisioning, step-mismatch row abort, errored-vs-failed classification, fixed temporal resolution, aggregates-after-last-row) with the live HTTP driver realized purely from the operation bindings, the reference resolver (corpus/recipes/rows/captures with normative sentinel semantics), the normative RESULT_SET equivalence comparator, content-case execution via the synthesized generate→commit→expect flow, the party artifacts (statement/results/ixit with schema validation and mandatory N/A citations), the pure verdict pipeline + deterministic report/statement/certificate renderers, the runner-verification pack (committed transcript + player: adjudicated verdicts reproduced, broken runners rejected), and the performance machinery (class cases with the published population-anchored floors, re-checkable HDR V2 measurement records, the earned/not-earned pure verdict). Nine published JSON-Schema families, drift-guarded. Live-SUT runs (the earned-class measurement and pack part 2) execute against a composed SUT via the new run/verdicts CLI once cutover lands.

  • CNF 2.0 reference runner, second increment: the complete CNF 2.0 catalogue authored from the framework — 347 cases across every schedule chapter (EHR, EHR_STATUS, COMPOSITION, CONTRIBUTION, DIRECTORY, ADL 1.4 + ADL2 definitions, stored queries, demographic, admin, messaging, AQL, content data-type and structural validation, simplified formats, Security SEC-BASIC + Signing) with 84 per-operation ITS-REST bindings (every status/header mapping cited to its OAS source; wire gaps are typed unrealized declarations, not silent absences), the ambiguity register grown to 38 adjudicated entries, and the ECC↔CNF comparison gate CLEAN: all 394 active rows of the old harness’s catalogue adjudicated (350 covered, 5 deferred to the simplified-formats deepening, 18 dropped with justification, 9 out of scope, 12 ADL2 rows covered) in the committed map with the generated report at docs/conformance/cnf-comparison.md (drift-guarded). Old-harness retirement follows the owner’s report review with the executor/emission workstreams so an acceptance instrument runs continuously.

  • CNF 2.0 reference runner (tools/cnf-runner), first increment: the typed schedule-artifact model (case cores, per-ITS operation bindings, outcome + selector vocabularies, the capability→family→tier matrix, corpus manifest, ambiguity register — every closed vocabulary a Rust enum/newtype), a published JSON-Schema set for all seven artifact families (committed under tools/cnf-runner/schemas/, drift-guarded, vendorable by any runner), a full cross-artifact validator (id uniqueness, SM-operation and spec-ref resolution against the vendored specs, binding completeness, corpus integrity, reference/sentinel and decision-table grammars, capability-tier consistency), the cnf-runner CLI (emit-schemas, validate), and the eight pilot case encodings as the first schedule artifacts. The existing ECC (tools/conformance) is unchanged and remains the acceptance instrument until the comparison gate.

  • Performance conformance, measured end to end: a cnf-runner perf run plays an open-loop offered-load schedule against a composed server at a population-anchored volumetric class (proof-of-concept, small, large, regional), records re-checkable HDR histograms into the conformance results, and earns — never declares — a class verdict recomputed by the verdict pipeline. CONF_PERF_CLASS=<class> scripts/conformance.sh runs it as a pipeline stage; the earned classes flow into the verdicts, report, certificate, and a performance badge. Published SVG assets (the class ladder and per-class latency charts) plus a generated summary are rendered from the committed measurement records by scripts/render/perf-assets.sh and guarded against drift in CI, and a new Performance chapter on the documentation website explains the class ladder, the floors’ derivation from official activity statistics, how a coordinated-omission-free run works, and how to reproduce it.

  • The sustained-window ladder: cnf-runner perf --hours 1|2|4|6|8|12 (pipeline: CONF_PERF_HOURS) extends a class run’s measured window beyond the normative hour — a longer hold of the same offered load is a stricter demonstration and persists like any measured run. There is deliberately no shortened run.

  • A step-load stress instrument, distinct from conformance: cnf-runner stress climbs short intense load steps (geometric doubling, ~two-minute holds, bisection refinement) to the maximum sustainable throughput inside a latency budget, over the same seeded corpus and workload mix as the class runs. The report (stress.json, schema-published, environment-bound, per-step re-checkable histograms) earns no class and never touches the conformance results; the class floors appear as context only. A latency-throughput curve SVG renders from the committed report through the same drift-guarded asset pipeline, and the documentation’s Performance chapter tells the two-instrument story.

Changed

  • The conformance acceptance instrument is now the CNF 2.0 reference runner (tools/cnf-runner) end to end: scripts/conformance.sh composes the SUT on fresh volumes, executes the committed machine-readable catalogue, computes verdicts through the pure pipeline, and writes results/verdicts/report/statement/certificate + badges per SUT. The ECC harness (tools/conformance) is retired — its final inventory is preserved at tools/cnf-runner/comparison/ecc-catalog.tsv and the reviewed cutover record is docs/conformance/cnf-comparison.md; the previous ehrbase comparison artifacts are frozen as historical data. Committed per-SUT party sets (ixit + statement) live under tools/cnf-runner/party/.

  • Verdict semantics: a REQUIRED capability whose every selected case is excluded by a schedule-registered ambiguity (an unrealized wire on the technology profile, e.g. ADL 1.4 archetype provisioning under ITS-REST 1.1.0 — AMB-41) is now recorded as an explicit unrealized scope exclusion on the certificate instead of silently failing the tier; the API-presence capabilities (EHR/DEFINITION/QUERY API) are evidenced by chapter exemplar cases.

  • The benchmark harness converged onto the conformance runner’s corpus, recipes, and ixit topology, so both instruments seed identical clinical documents through the public write path. The performance numbers in the README and on the website are no longer hand-typed: they derive from committed run artifacts (the benchmark comparison charts and the CNF measurement records), and the site stale-numbers guard now also rejects a hand-typed rate, latency, or footprint in the sources.

  • OPT-1.4 → ADL2 conversion fidelity: DV_ORDINAL/DV_QUANTITY constraints now convert to real AOM2 attribute tuples ([value, symbol], [units, magnitude(, precision)]) instead of loose unconstrained nodes; slot include/exclude assertions are carried (both retained 1.4 slots and the filled-slot include naming the embedded archetype); OPT default_values are carried and serialized as the ADL2 _default pseudo-attribute; temporal constraints keep both the ISO8601 pattern and the range plus assumed values; referenceSetUri becomes an ac-code term binding; CONSTRAINT_REF resolves against the merged 1.4 constraint_definitions/constraint_bindings; and everything a decomposed root cannot express (out-of-scope bindings, tuple assumed values, DV_STATE machines, unconvertible assertions) is reported in the converted archetype’s RESOURCE_DESCRIPTION.conversion_details. The whole vendored OPT corpus now converts, validates and re-parses as the standing test gate.

Fixed

  • OPT 1.4→2 decomposition now emits phase-1-clean ADL2 sources for every template in the corpus: a --specialised embedded root (whose differential lineage a flattened OPT cannot resolve) is emitted as an unspecialised depth-0 archetype with every dotted code renumbered into the flat code space, and 1.4 node codes legitimately reused across sibling subtrees re-mint archetype-wide-unique ADL2 ids — terminology definitions and bindings follow in both cases, and every remap is recorded in the converted archetype’s conversion_details provenance.

  • The ATNA Audit Record Repository no longer loses records under a sustained write load: the audit drain now takes queued events in batches and persists each batch in one multi-row INSERT (the previous per-event round trips saturated far below write-path rates, filling the bounded queue and fail-open dropping the tail). Drop warnings are rate-limited to one per interval carrying the count since the previous warning instead of one log line per dropped record (the exact count stays on the atna_audit_dropped_total metric), and the default audit.queue_capacity rises from 1024 to 8192 for burst headroom.

  • Composition validation closes eight archetype-constraint enforcement gaps the CNF content chapter exposed: C_STRING list/pattern constraints on DV_IDENTIFIER.issuer/assigner/type (only id was checked); DV_MULTIMEDIA.size against C_INTEGER list and range constraints (previously unvalidated); C_ATTRIBUTE existence 1..1 on OBSERVATION.state/protocol, HISTORY.summary, and EVENT.state now rejects the absent attribute; DV_SCALE value/symbol value-set constraints (generic C_REAL list + C_CODE_PHRASE code list — AOM 1.4 has no C_DV_SCALE) are enforced, including on DV_INTERVAL bounds; timezone_validity on C_TIME/C_DATE_TIME (mandatory and prohibited) is honoured; half-open (one-side-unbounded) temporal range constraints reject out-of-range values; a DV_PROPORTION of kind fraction or integer-fraction with a non-zero precision is rejected (Fraction_validity); and a partial DV_TIME such as 10 is no longer over-rejected against HH:??:??/HH:XX:XX patterns (optional and not-allowed fields both admit an absent field).

  • A DV_TIME/DV_DATE_TIME literal carrying a fraction on the hours or minutes component (e.g. 10.5, 10:05.5) is now rejected: openEHR supports fractional seconds only (BASE time types §ISO 8601 semantics not included).

  • A DV_URI whose value has no URI scheme (e.g. xyz, www.example.org) is now rejected on commit per the CNF content schedule’s RFC-3986 rule; plain-text URI content after the scheme remains accepted per the RM’s plain-text allowance.

  • A COMPOSITION create (201) or update (200) whose response is negotiated as a Simplified Format (Accept: application/openehr.wt.flat+json or …wt.structured+json) now returns the ETag and Location headers, matching the canonical (application/json/application/xml) response. Previously a FLAT/STRUCTURED commit body omitted both version-id headers, so clients could not read the new version uid or resource URL from a simplified-format commit.

  • Composition validation now rejects a DV_DURATION whose value carries a decimal fraction on any component other than seconds (e.g. P1Y3M4DT2.5H or PT2H14.5M). openEHR permits a fraction only on the seconds component (BASE time types: “in openEHR, only fractional seconds are supported”), so such a value now fails its RM Value_valid invariant with 422 instead of being accepted.

  • Composition validation now enforces a DV_QUANTITY constraint that fixes a measurement property (with no enumerated unit list): the committed units must be a unit of that physical property (per the openEHR measurement property↔unit table). A quantity constrained to length committed with a mass unit such as mg is now rejected with 422 instead of being accepted.

  • Composition validation now rejects a coded value whose terminology is foreign to a C_CODE_PHRASE constraint that explicitly binds the archetype-local terminology with a closed code list. Committing a DV_CODED_TEXT whose defining_code uses, e.g., SNOMED-CT against a local-scoped closed list now yields 422 instead of being accepted.

  • The AQL ehr_id execution scope now also binds bare FROM EHR e sources: a scoped query without a CONTAINS chain previously ran over the whole population instead of the single EHR context the ehr_id parameter selects (ITS-REST query Request.md §Common Headers and Query Parameters).

  • A CONTRIBUTION delete member targeting the EHR_STATUS is now refused with 409 Conflict: EHR.ehr_status is mandatory (RM ehr, EHR class, 1..1), so deleting the only status would leave the EHR violating its own invariant.

  • FLAT/STRUCTURED commits: spec-listed direct RM-attribute paths that an operational template leaves unconstrained are no longer rejected as unknown paths. ACTION/ism_transition (current_state/transition/careflow_step

    • _reason:i) and ACTION/time, plus INSTRUCTION/narrative, OBSERVATION/history_origin, ACTIVITY/timing + action_archetype_id, and INTERVAL_EVENT/width + math_function, are now built from their datum parts per the ITS-REST Simplified-Formats master05-rm_mapping.adoc per-type tables, and emitted symmetrically on the reverse (RM → FLAT) direction so round-trips stay lossless. Previously a client-supplied ism_transition was rejected with “unknown simplified path” and the ACTION state fell back to the synthesized initial default.
  • AQL paging: the REST fetch/offset parameters now page over the result set the AQL LIMIT/OFFSET clauses define instead of being rejected with 400 when combined. Per ITS-REST query Request.md, only pairing fetch with the deprecated AQL TOP modifier is prohibited — that rejection remains. Negative fetch/offset values are now rejected explicitly.

  • Spec version identity is now derived from the openehr-* crate versions instead of hand-typed literals, fixing the stale values those literals had drifted to: the startup banner advertised ITS-REST 1.0.3 (now 1.1.0), and the AQL RESULT_SET meta._schema_version was still emitted as 1.0.3 (now 1.1.0, the implemented ITS-REST release). Every openehr-* spec crate exposes a SPEC_VERSION constant (= its crate version; the AM crate also exposes per-generation am14/am24 constants from the BMM schemas), and the shared provenance constants behind the banner, /status, OPTIONS / (System Options), and /management/info read those, so a future pin bump propagates everywhere at compile time. The served restapi_specs_version/openehr_rest_api_version identity is now the plain version string 1.1.0 (matching the System API OAS example) instead of the tag-styled Release-1.1.0.

  • SM call-status fidelity: service-layer “does not exist” failures now carry their granular CALL_STATUS_TYPE (ehr_id_does_not_exist, composition_does_not_exist, template_does_not_exist, object_version_does_not_exist, …) end-to-end instead of resurfacing as the generic versioned_object_does_not_exist after crossing the service boundary. HTTP status codes are unchanged (every does-not-exist status was and remains 404); some 404 body messages are now the precise construction-site text.

3.5.0 - 2026-07-21

Changed

  • Conformance: zero skipped outcomes. The former 35 skips are eliminated — 11 cases now execute against the documented ehrbase-rs extension surfaces (contribution listing, admin template deletion, bare stored-query listing), 6 more execute via new composed-stack wiring (an OpenPGP-signing sibling instance and a hermetic FHIR terminology fixture with fault injection) and loaded-database AQL golden support, and 18 native-API-only service operations are now first-class not-applicable verdicts carrying their SM citation and native-test evidence.

Added

  • ADL 2 archetype validation now enforces VETDF (external term-binding validity): a term bound to an external terminology (SNOMED CT, LOINC, …) that the configured terminology service reports as absent is rejected 422 with the VETDF rule code. Bindings the service cannot verify (no external provider configured, an unknown terminology, or a transport fault) are not raised, per the spec’s “subject to tool accessibility” carve-out; archetype-internal (local/openehr) bindings are unaffected (covered by VTTBK/VTCBK key validity).
  • ISO 8601 temporal ordering on the openEHR BASE time types (Iso8601_date/_time/_date_time/_duration): comparison with honest incomparability (partial-date range semantics, UTC normalization for zoned values, duration ordering via the spec’s own to_seconds reduction with the Time_definitions average constants). ADL 2 archetype validation now enforces assumed-value interval containment for temporal constraint types (previously undecidable and skipped); an incomparable pair never raises a violation.

3.4.0 - 2026-07-20

Changed

  • The implemented openEHR REST API is ITS-REST Release-1.1.0 (published upstream 19-Jul-2026). The server was already built against the pre-release text of this release — the regenerated REST contract is byte-identical at the release tag — so wire behaviour is unchanged; the advertised API identity moves from 1.0.3/development to 1.1.0 everywhere (documentation, OpenAPI metadata, conformance artifacts), and the openehr-its spec crate is now versioned 1.1.0. Conformance reports state the tested edition as release-1.1.0 (formerly development; the old label remains accepted as a CLI/config alias).

3.3.0 - 2026-07-20

Added

  • ADL2 templates are now compiled and validated by the full ADL2 engine. POST /definition/template/adl2 runs the complete openehr-adl pipeline — parse, then the AOM2 validity catalogue (phase 1 basic integrity, reference- model conformance, and specialisation conformance against an already-loaded parent) — in place of the former source-subset probe. An invalid artefact is a 422 whose Error.validationErrors list the offending rule-code mnemonics (S-codes for an unparseable source, V-codes for a validation-phase failure). GET /definition/template/adl2/{template_id} now serves the application/json OperationalTemplateV2 projection alongside the text/plain source, and resolves a partial template_id to the latest matching version; the previously 501 …/{template_id}/{version} (versioned get, marked deprecated in the spec) is implemented, and template list rows now carry concept and archetype_id. GET …/{template_id}/example now generates an example COMPOSITION from the compiled operational template (an ADL2 → Web Template front end feeding the shared example generator), served across the four Accept_LOCATABLE representations (canonical JSON/XML, openehr.wt.flat, openehr.wt.structured) with type (input/output) + detail_level (required/medium/complete) query parameters, and 400/404/406 exactly as the ADL 1.4 example endpoint. An Accept naming only application/xml on the plain template GET is a 406 (the operation declares no XML response body).

  • ADL 1.4 archetypes are now validated by the ADL 1.4 engine, and can be migrated to ADL 2. An ADL 1.4 source archetype (the I_DEFINITION_ADL14 archetype surface) is now parsed and validated as ADL 1.4 by the openehr-adl engine — the subset of the phase-1 catalogue that corresponds to the ADL 1.4 / AOM 1.4 standalone validity rules (VARID, VARDT, VARCN, VATID, VDSEV/VDSIV, …), replacing the former structural probe. An invalid source is a 422 naming the offending rule-code mnemonic. A new service capability migrates a stored ADL 1.4 archetype to ADL 2 source (adl14_convert_to_adl2); no openEHR spec governs 1.4 → 2 conversion (our own design/extension) and the ITS-REST contract declares no conversion operation, so it is a library capability with no REST endpoint. The ADL 1.4 operational-template (OPT) REST surface (/definition/template/adl1.4) is unchanged.

  • RM terminology-backed invariant validation. Composition (and any RM value) validation now enforces the openEHR terminology-service and code-set RM class invariants at the wire-boundary dispatcher, unified into a single hook (openehr-its) that every validation consumer inherits. The 30 wired invariants (each audited clean against the whole corpus before enforcement): COMPOSITION category/language/territory, EVENT_CONTEXT setting, ELEMENT null-flavour, ISM_TRANSITION current-state/transition, PARTICIPATION + EXTRACT_PARTICIPATION function/mode, INTERVAL_EVENT math-function, TERM_MAPPING purpose, AUDIT_DETAILS change-type, ATTESTATION reason, PARTY_RELATED relationship, VERSION lifecycle-state, ENTRY/DV_TEXT language + encoding, DV_MULTIMEDIA media-type/charset/language/compression/integrity algorithms, DV_PARSABLE charset/language, DV_ORDERED normal-status, and the AUTHORED_RESOURCE / RESOURCE_DESCRIPTION_ITEM / TRANSLATION_DETAILS original-language. An out-of-vocabulary openEHR code is a 422 naming the violated RM invariant; HTTP status codes are unchanged.

  • Admin console: the Directory tab is now a complete directory experience — a structured folder-tree editor (add/rename/remove sub-folders, attach and remove composition item references with a picker), version history with read-only views and one-click restore, a version_at_time time-travel control, a sub-folder path query, and directory deletion with confirmation — on top of the existing create-from-template flow (raw JSON editing stays available as an advanced mode).

Changed

  • RM validation invariant messages now carry the spec’s (BMM) invariant names. Three class-invariant violation messages were reconciled from their inherited archie spellings to the openEHR BMM invariant names, so a 422 validation payload reporting one of them changes text: Accuracy_valid → Accuracy_validity (DV_AMOUNT and its descendants — DV_QUANTITY, DV_COUNT, DV_DURATION, DV_PROPORTION), Is_archetypeRoot → Is_archetype_root (the ENTRY subtypes — OBSERVATION, EVALUATION, INSTRUCTION, ACTION, ADMIN_ENTRY), and Location_validity → location_valid (EVENT_CONTEXT). The check logic and HTTP status codes are unchanged; only the invariant name inside the Invariant <name> failed on type <TYPE> message differs.

  • Canonical-JSON codec cutover. The openEHR spec types are now (de)serialized to/from canonical JSON entirely by a native emitted ToJson/FromJson codec in openehr-its — the spec types (openehr-base, openehr-rm, openehr-am, openehr-term, openehr-lang) no longer carry a serde derive, and the openehr-derive proc-macro crate is removed. The wire bytes are unchanged (proven by the R0 determinism manifest + the byte-hazard gates); the only externally visible difference is the error-message shape on a malformed JSON request body — the codec’s parser reports expected … at line N column M / missing field … on … diagnostics instead of the previous serde phrasing (the HTTP status codes are unchanged: still 400/422). A present-but-null array field is now rejected as a type error (was silently treated as an empty array), matching the strict tolerance contract.

  • The served OpenAPI document now describes the COMPLETE wire for every operation (162 declarations across all API groups): every path/query parameter, request headers (Prefer incl. return=identifier, required If-Match forms, the committal headers), every reachable status code with its exact trigger, and the load-bearing response headers (weak ETag, Location, Last-Modified) — audited operation-by-operation against the vendored ITS-REST specification (both the operation definitions and the normative overview rules). A structural completeness test now gates the document.

  • A disabled Admin API now answers 405 Method Not Allowed (the status the ITS-REST specification declares for a disabled admin operation) instead of 404.

  • COMPOSITION and EHR_STATUS tag updates now honour the Prefer header as the specification defines: the default (return=minimal) returns 204 No Content; return=representation returns 200 with the stored tag list. Previously the stored list was always returned with 200.

  • Demographic responses now carry Last-Modified (from the version’s commit time) alongside the weak ETag; PARTY_RELATIONSHIP create/update honour Prefer: return=identifier.

Fixed

  • Template example generation now produces fully-valid compositions. GET /definition/template/adl1.4/{template_id}/example populated only a skeleton for many templates (issue #94) and could emit out-of-range or wrongly-typed values. The generator now synthesizes spec-valid values for every constrained field — quantities inside their magnitude ranges (with dimensionless empty units preserved), proportions satisfying their kind’s invariants inside the archetype’s numerator/denominator ranges, durations inside their declared range, coded text from closed value lists, URIs and parsables honouring their pattern constraints, and the archetype-constrained container/event types (ITEM_LIST/ITEM_SINGLE/INTERVAL_EVENT) instead of abstract defaults — and every generated example at the committable detail levels (required, medium) passes the server’s own full composition validation. Generation is byte-deterministic.

  • Archetype-conformance validation no longer demands archetype_node_id on reference-model types that cannot carry one. EVENT_CONTEXT (and any other non-LOCATABLE type) inherits PATHABLE, which the RM gives no archetype_node_id; a template archetyping /context[at…] therefore could never be satisfied by canonical data and such compositions were wrongly rejected on commit. Non-LOCATABLE nodes now match structurally by their attribute position (per the RM inheritance graph); LOCATABLE nodes keep strict node-id matching.

  • Admin console: text typed into the EHR finder and create-EHR fields before the app finished loading is no longer silently wiped (the inputs are now hydration-safe, like the login form); success toasts no longer intercept clicks on buttons beneath them in the e2e battery.

  • GET /ehr/{ehr_id}/directory/{version_uid} now honours the path query parameter (slash-separated FOLDER names selecting a sub-folder subtree), as the ITS-REST directory_get_by_version_id operation specifies; an unresolved path returns 404. Previously the parameter was accepted but ignored and the full tree was always returned.

  • The served OpenAPI now documents the full DIRECTORY wire contract (version_at_time/path parameters, Prefer including return=identifier, If-Match, and the complete status ladders including 204/400/409/412).

3.2.0 - 2026-07-18

Added

  • GET {base}/admin/config — the redacted effective configuration (an ehrbase-rs extension; the openEHR admin API defines only EHR deletes). Returns the merged effective configuration (file + EHRBASE_* env + --set overrides) as a JSON tree with every secret-bearing value redacted structurally by its secret type — passwords, password hashes, HMAC/signing secrets, and S3 secret keys render as ***, and connection URLs (database, AMQP) mask their embedded credentials while keeping host and path; non-secret identifiers (usernames, roles, OIDC issuer) stay visible. Shares the admin gate and authorization of the admin deletes (EHRBASE__ADMIN__ENABLED=true, ADMIN role); disabled admin API answers 404.

  • ehrbase-admin-ui — the admin console, a new standalone web application (its own binary and OCI image, ghcr.io/rubentalstra/ehrbase-rs-admin-ui) that manages any ITS-REST-1.0.3 CDR strictly over its REST API. Pure Rust end to end (Leptos SSR + WASM, zero hand-written JavaScript). Feature set: dual Basic + OIDC login (credentials held server-side in the BFF), a dashboard (count tiles, query-group tiles, a commit-activity trend chart), a Template Manager (list/filter/upload OPTs with the CDR’s validation diagnostics verbatim; per-template path-catalog tree, raw-OPT view, and format-switchable generated example), an EHR browser (finder, status/directory/compositions/contributions, and a composition viewer with canonical JSON/XML + FLAT/STRUCTURED toggle, version history, and audit details), a point-and-click Query Builder that assembles the real AQL AST (typed per-datatype criteria from the template’s constrained value sets, nested AND/OR/NOT groups, projection columns, live AQL preview) and runs it via the Query API, a raw AQL editor with BFF-side grammar validation and parameter bindings, stored-query management with console-local query groups, and a system panel (CDR status, SMART discovery, the served OpenAPI rendered natively). Configured by one ehrbase-admin-ui.toml (+ EHRBASE_ADMIN__* env); ships in the quickstart compose as the ehrbase-admin-ui service on port 3000. The sign-in page is served fully rendered and works with JavaScript disabled (the login form posts and redirects natively), and offers exactly the methods that can work: the console’s configured login modes intersected with the authentication schemes the CDR advertises in its WWW-Authenticate challenge. The console received a full design system (semantic design tokens with lockstep light/dark theming, a teal brand shared by the widget kit, iconified navigation, breadcrumbed page headers, named table headers, empty states, and toast feedback on every mutation) and the complete working feature set: query result export (CSV/JSON, a plain form download that works without WebAssembly), EHR creation (empty or subject-bound) and find-by-subject-id, composition commit (canonical JSON/XML/FLAT with verbatim CDR validation diagnostics) and edit-as-new-version (If-Match concurrency), stored-query open-in-editor, shareable URL-driven tab state on the detail screens, a template identity card (version, languages, UID, archetype id), an EHRs (cohort) query shape (SELECT DISTINCT over the criteria tree), a Table | Chart toggle on numeric result columns, a version timeline strip with a version_at_time picker on the composition viewer, and a contributions table on the EHR detail screen. The Directory tab can now create and edit the EHR folder directory (spec-standard POST/PUT with If-Match), starting from console-local folder templates (two built-ins included); the System panel gained repository usage (per-template composition counts) and a read-only runtime configuration view backed by the CDR’s new redacted GET /admin/config endpoint (secrets redacted structurally by their types — never by key matching). The E2E harness gained an image mode (UI_E2E_IMAGE=1) that runs the identical journey battery against the composed OCI image — including a genuinely end-to-end OIDC journey: the quickstart Keycloak now pins one canonical issuer and the dev CDR config trusts it via standard OIDC discovery, so a bearer-authenticated console session queries the CDR for real. Verified by a Rust-native browser E2E journey suite (merge-gating in CI, screenshots published as artifacts), including journeys over seeded clinical data and a JavaScript-disabled login journey.

  • GET /ehr/{ehr_id}/contribution — a paged contribution list (an ehrbase-rs extension; the openEHR REST API defines only the by-uid read). Returns the EHR’s contributions newest-first as { "rows": [ { uid, time_committed, committer, change_type } ], "total" }, paginated with offset (default 0) and fetch (default 20, capped at 100); 404 for an unknown EHR. Authenticated like the other EHR reads.

  • DELETE /admin/template/{template_id} and DELETE /admin/query/{qualified_query_name}/{version} — admin deletes for operational templates and stored-query versions (ehrbase-rs extensions; the openEHR admin API defines only EHR deletes). Same admin gate and authorization as the EHR deletes: 204 on success, 404 for an unknown id. The template delete additionally returns 409 when a committed version still references the template, so a physical delete never orphans clinical data.

  • ATNA audit — richer DICOM records: every audit record now carries the concrete operation as a DICOM EventTypeCode (login/logout as DCM 110122/110123; REST operations as their ITS-REST operation id under the openEHR-ITS-REST code system), and Bearer-authenticated requests record the token’s jti as the minimal token identity (token contents are never logged).

  • ATNA audit — FHIR R4 AuditEvent rendering (IHE BALP): every audit record also renders as a FHIR R4 AuditEvent conforming to the IHE Basic Audit Log Patterns (Patient*/plain Create/Read/Update/Delete/Query profiles, OAUTHaccessTokenUse.Minimal token agent, profile claims only when genuinely satisfied) — the modern half of the dual ATNA format.

  • ATNA audit — local Audit Record Repository, on by default: audit records are persisted in a new PostgreSQL audit schema (append-only; strictly outside the EHR content; per-sink delivery stamps; configurable retention_days with an hourly reaper). Every deployment now gets a queryable audit trail out of the box with nothing leaving the node.

  • ATNA audit — RESTful ATNA forwarding (ITI-20 ATX:FHIR Feed): opt-in [audit.fhir_feed] sink POSTs each FHIR AuditEvent to an external Audit Record Repository; with the local store on, delivery is outbox-driven — an ARR outage loses nothing and pending records ship on recovery.

  • ATNA audit — per-sink metrics (atna_audit_sent_total{sink=…}, …send_failed_total{sink=…}, atna_audit_rejected_total, atna_audit_reaped_total).

  • ITI-81 Retrieve ATNA Audit Event (GET /fhir/r4/AuditEvent): the official RESTful-ATNA retrieval — a FHIR search over the local Audit Record Repository returning a searchset Bundle of the stored AuditEvent documents. Filters: date (ge/le), patient, agent, entity, outcome, action, plus _count/_offset paging. Admin-only under RBAC; 404 when the local store is disabled.

  • Native TLS + mutual-TLS client authentication ([server.tls]): the main listener can terminate TLS itself (TLS 1.2+ floor per IETF BCP 195) and demand a verified client certificate (client_auth = "off" | "optional" | "required") against an explicit CA — the IHE ATNA ITI-19 node-authentication posture. The management listener stays plain HTTP.

  • A dedicated Audit trail (IHE ATNA) book chapter covering the dual formats, the sinks, the ITI-81 retrieval, fail-mode semantics, and mTLS.

  • Admin console — the Audit log screen (/audit): browse the CDR’s ATNA security audit trail through the standard ITI-81 retrieval, with URL-driven filters (event-time window, patient, principal, outcome, action), pagination, and a per-row view of the full stored FHIR AuditEvent. Admin-only under RBAC; a disabled local audit store and a no-matches filter each render their own first-class state.

Changed

  • The ITS-REST template list (GET /definition/template/adl1.4) now reports the optional version field of each TemplateMetadata, derived from the template id’s version axis (the spec documents the value as “taken from template_id”); it is omitted when the id carries no version.
  • Audit configuration redesigned: [atna] is now [audit], on by default with only the local store active, and sink-structured: [audit.store] (local repository), [audit.syslog] (classic DICOM-over-syslog feed; keys host/port/transport/tls_ca_file/ tls_identity_cert_file/tls_identity_key_file replace the old repository_host/repository_port/tls_*_path), [audit.fhir_feed] (RESTful ATNA). resolve_subject now defaults to true. A configuration still using [atna] fails at boot with did-you-mean guidance (strict loader; no silent aliasing).
  • Fail-closed auditing got stronger: with fail_mode = "closed" and the local store enabled, a store that stops accepting writes makes every subsequent auditable operation answer 503 Service Unavailable until a write succeeds again — no un-audited PHI access.

Fixed

  • ATNA audit — IHE/DICOM conformance corrections (IHE ITI TF-2 ITI-20 / DICOM PS3.15 §A.5.1): the syslog MSGID is now the mandated IHE+RFC-3881 (was IHE+DICOM); AQL query execution uses the dedicated DICOM EventID 110112 “Query” (was 110110); EHR-Extract communication uses the direction-coded EventIDs 110106 “Export” / 110107 “Import”; authentication events (genuine logins and rejected 401/403 attempts) use EventID 110114 “User Authentication” with EventTypeCode 110122 “Login” (were generic Application Activity); and 1xx/3xx responses (e.g. 304 Not Modified) are now recorded as success instead of minor failure.
  • Admin console — icon-only chrome and small polish: every emoji and typographic glyph in the UI is replaced by a proper SVG icon (folder tree, status capability badges, remove buttons, disclosure carets, upload trigger, pagination arrows); the Audit log screen highlights its own navigation entry; and the documentation screenshots now cover every EHR detail tab — including the directory tab both before (create from a folder template) and after the directory exists — plus the audit raw-record view.

3.1.1 - 2026-07-17

Fixed

  • The release pipeline attaches the per-architecture server binary tarballs again: since the crate consolidation the binary is produced by the ehrbase-server package (the executable is still named ehrbase), but the release asset build still compiled the ehrbase platform library and failed — v3.1.0 published without binary assets. Container images were not affected. Use v3.1.1 for downloadable binaries.

3.1.0 - 2026-07-17

Added

  • External terminology providers cache their FHIR operation results ($validate-code/$expand/$subsumes/$lookup) for a configurable TTL ([terminology.external.providers.<name>] cache_ttl_secs, default 300 s, 0 disables; cache_capacity, default 10000) — a validation burst over the same codes costs one remote round trip per window instead of one per code.
  • A new atna_audit_serialize_failed_total metric counts ATNA audit records dropped because the message failed to serialize, so audit loss is always metered.

Changed

  • The FLAT and STRUCTURED (Simplified Formats) layer was rewritten against the official openEHR ITS-REST Simplified Formats specification: exact node-id generation, per-type attribute suffixes, the full ctx/ vocabulary with its documented defaults, |raw embedding, and the |other open-value-set rules (invalid combinations are now rejected with 422 instead of being silently ignored). Unknown field identifiers in a simplified payload are now rejected rather than dropped.

  • Format selection is done exclusively via the Accept and Content-Type headers on every endpoint that supports the simplified media types (application/openehr.wt.flat+json, …wt.structured+json, and application/openehr.wt+json for template rendering), with proper RFC 9110 q-value negotiation, 406/415 answers naming the supported formats, and simplified support on CONTRIBUTION payloads (versions[].data) with the envelope staying canonical.

  • Committing a composition in a simplified format now requires the openehr-template-id request header (422 without it, previously 400); the undocumented template_id query parameter is no longer read.

  • Content negotiation is strict everywhere: an Accept header that none of an endpoint’s supported formats can satisfy is answered with 406 (previously some JSON-only endpoints leniently returned JSON), and the server’s own generated OpenAPI now advertises the simplified media types on the composition, contribution, and template endpoints.

  • Release builds now abort on integer arithmetic overflow instead of silently wrapping (overflow-checks enabled in the release profile) — a corrupted-value class of fault becomes a crash-and-restart instead of wrong clinical data.

  • The application is consolidated to two library crates plus a thin binary (ehrbase — the platform, ehrbase-rest — the ITS-REST adapter, ehrbase-server — the binary): the ehrbase-sm trait catalog is gone, the REST adapter calls the concrete platform service directly, and the full configuration tree ([server], [auth], [authz], [smart], [management], [tenancy], [admin]) is defined in the platform crate. The served wire, the ehrbase.toml schema, and the container entrypoint (ehrbase) are unchanged.

  • Bundle-backed terminology lookups and template/query validity checks are now synchronous in-process calls (no behaviour change on the wire).

  • Every versioned write now commits through the single folded audit+contribution+version statement even with digest signing enabled (the commit instant is read up front with the placement, so the signature is computed before any insert); version-tree placement is one read instead of three, and contribution commits batch their target pre-reads. Fewer round trips per write, identical wire behaviour and stored semantics.

  • The OpenAPI documents (the composed openapi.json and the twelve Swagger spec-selector family documents) and the SMART .well-known/smart-configuration discovery document are now built once at server startup instead of being regenerated on every request. No change to the document content.

Removed

  • The ehrbase-quirks cargo feature and its vendor-specific behaviours (alternate duplicate-id spelling, the non-standard |unit_system / |unit_display_name quantity suffixes) — the specification-defined behaviour is now the only behaviour.

Fixed

  • A tenant-resolution failure (tenant registry unreachable) now fails the request with 503 instead of silently serving it under the default tenant; unknown tenant keys keep the documented unscoped behaviour and are negative-cached.
  • Audits for authenticated writes that carry no committal headers are now attributed to the authenticated user (Basic username / token subject, with the mechanism recorded as the identifier type) instead of the generic system identity.
  • Multi-tenant deployments now actually run on the tenant-scoped connection pool: with tenancy.enabled = true every database connection carries the request’s tenant for the row-level-security policies. Previously the binary always built the plain pool, so all requests fell through to the default tenant regardless of configuration.
  • Multi-tenancy: a connection freshly opened by the pool while serving a request (pool growth under load) could miss the tenant stamp and run as the reserved default tenant — reads returning nothing and writes landing outside the caller’s tenant. The tenant-scoped pool now stamps ehrbase.tenant_id both when a connection is opened and on every checkout, so every connection carries the caller’s tenant. Deployments with tenancy.enabled = true should upgrade.
  • The demographic APIs (party and relationship writes) now honour the openEHR-VERSION.* / openEHR-AUDIT_DETAILS.* committal headers exactly as the EHR APIs do — a caller-supplied committer, description, and system id are merged into the stored version’s audit.
  • Direct COMPOSITION create/update/delete now honour the ITS-REST committal headers (openEHR-VERSION.* / openEHR-AUDIT_DETAILS.*): a caller-supplied committer, audit description, change type, lifecycle state, signature, and attestations are merged into the stored version exactly as on the CONTRIBUTION path (previously the direct paths discarded them and always committed server defaults).
  • The template store no longer double-reads the OPT XML when generating an example for a cold template, and template upload is a single atomic statement (the duplicate-check race window is gone).
  • The event-outbox publisher declares its AMQP topology only on connect or subscription change (previously every poll cycle re-declared each queue), and the FHIR outbound emitter parks a persistently failing row after a bounded retry budget instead of blocking the stream forever.
  • A FLAT/STRUCTURED composition body that parses as JSON but does not conform to its target template now returns 422 Unprocessable Entity instead of 500 Internal Server Error — such an input is client data, not a server fault. Output conversion of stored compositions remains a 500 on failure.
  • Panicking request handlers and audit fail-closed (503) responses now carry the standard openEHR { error, message } JSON error body (the audit 503 also carries Retry-After), instead of a plain-text body.
  • A malformed If-Match header on a state-changing request is now rejected with 400 Bad Request instead of being silently ignored — an unparseable precondition previously ran as if no If-Match was sent, opening a lost-update window. If-Match: * and valid version ids are unaffected.
  • Database constraint and serialization/deadlock failures now surface as 409 Conflict, and connection-pool exhaustion under load as 503 Service Unavailable with Retry-After, instead of collapsing every database error to 500 Internal Server Error.
  • Stored-query and template metadata list/read endpoints no longer silently blank a field when a database column fails to decode; a decode failure now surfaces as 500 with a real error instead of an empty value.

3.0.3 - 2026-07-16

Changed

  • The served OpenAPI documents now categorize operations the way the official ITS-REST reference documents do: standard-group operations are tagged by resource (EHR, EHR_STATUS, COMPOSITION, DIRECTORY, CONTRIBUTION, ITEM_TAG; PERSON, AGENT, GROUP, ORGANISATION, ROLE, VERSIONED_PARTY; ADL 1.4, ADL 2, Query) instead of one flat tag per API group, and the Swagger UI spec selector offers one document per API family — the five standardised openEHR groups and the seven server-extension families — plus the complete composed surface, all filtered from the server’s own generated document.

Fixed

  • Duplicate-template-id fixture resolution in the validation corpus test is now deterministic (sorted path order) instead of OS-dependent read_dir order, fixing a Linux-only CI failure.

3.0.2 - 2026-07-15

Changed

  • The benchmark instrument measures both comparison stacks under a fairer, more deterministic protocol: the databases get a 1 GB /dev/shm floor (Docker’s 64 MB default starved PostgreSQL’s parallel workers mid-run), maintenance debt is settled with VACUUM ANALYZE after seeding and between ladder rungs (autovacuum no longer lands inside measured windows), the ladder drains in-flight backlog between rungs, and the measured cold start no longer includes building the ehrbase-rs container image. Ladder output prints latencies in magnitude-appropriate units (µs/ms/s), and the generated comparison page reports clinical events per minute beside request rates.
  • Configuration is now one ehrbase.toml. The whole server is configured by a single TOML file (sections [server], [db], [log], [telemetry], [auth], [authz], [admin], [tenancy], [smart], [management], [signing], [query], [events], [fhir], [terminology], [multimedia], [atna], [subject_proxy]), discovered from --config, EHRBASE_CONFIG, ./ehrbase.toml, or /etc/ehrbase/ehrbase.toml. Every EHRBASE_* environment variable is now a mechanical per-key override: EHRBASE + the TOML path, upper-cased, with __ between every segment including after the prefix (e.g. EHRBASE__DB__MAX_CONNECTIONS, EHRBASE__AUTH__OIDC__ISSUER). This replaces the previous ~14 independent per-subsystem loaders and their several env-name grammars. Old spellings are not aliased (greenfield — nothing is deployed to migrate): a pre-redesign variable fails at boot with the exact uniform replacement suggested (e.g. EHRBASE_DB_MAX_CONNECTIONS → “did you mean EHRBASE__DB__MAX_CONNECTIONS?”). DATABASE_URL and RUST_LOG remain permanent conventional aliases. New ehrbase config default prints an annotated template and ehrbase config check validates a config (and prints the effective, secret-redacted result) without a database. The compose stack, Helm chart, and docs all move to the new file + spellings; the PostgreSQL-init container variables EHRBASE_DB_USER / _PASSWORD / _NAME were renamed PG_INIT_USER / _PASSWORD / _DB so they no longer collide with the server’s reserved EHRBASE_ namespace.

Removed

  • The nine per-subsystem EHRBASE_*_CONFIG file pointers (EHRBASE_REST_CONFIG, EHRBASE_AUTHZ_CONFIG, EHRBASE_ATNA_CONFIG, EHRBASE_SIGNING_CONFIG, EHRBASE_EVENTS_CONFIG, EHRBASE_FHIR_OUTBOUND_CONFIG, EHRBASE_MULTIMEDIA_CONFIG, EHRBASE_VALIDATION_CONFIG, EHRBASE_MANAGEMENT_CONFIG, EHRBASE_SUBJECT_PROXY_CONFIG): merge each file’s contents into the single ehrbase.toml under its [section].
  • EHRBASE_REST_AUTH__ADMIN_SCOPE: subsumed by authz.rbac.admin_role.

Fixed

  • Unknown or misspelled configuration is now rejected at boot with a did-you-mean suggestion (and the file:line for a file key) — previously a typo’d TOML key or EHRBASE_* variable was silently ignored, so a not-applied security setting could pass unnoticed.
  • The documented EHRBASE__SUBJECT_PROXY__SYSTEMS__<name>__BASE_URL env form now actually binds — the old loader stripped the prefix such that this spelling was dead, so subject-proxy systems could only be set via a file.
  • Unparseable [query] values (query.plan_cache_capacity, query.timeout_ms) now error at boot instead of silently falling back to defaults.
  • The Swagger UI works again and now documents the complete server surface from one natively generated OpenAPI document. …/rest/swagger-ui previously entered an infinite redirect loop (the UI’s trailing-slash redirect fought the server’s path normalization) and its OpenAPI document was an empty stub. The UI now loads directly (documentation URL corrected to /ehrbase/rest/swagger-ui), and its spec selector has a single entry, ehrbase-rest, generated by the server itself (utoipa-axum, one #[utoipa::path] handler per operation, so route and documentation cannot drift): every ITS-REST API group (EHR, COMPOSITION, CONTRIBUTION, DIRECTORY, DEMOGRAPHIC, DEFINITION, QUERY, ADMIN) plus the server’s own extensions (terminology, PARTY_RELATIONSHIP, event-subscription, multi-tenancy, FHIR connector) and its operational endpoints (status/health, management, SMART discovery, the OpenAPI endpoints). No vendored OpenAPI is served. The document also declares the server’s configured authentication scheme so the “Authorize” dialog and per-endpoint padlocks match the running server: HTTP Bearer (JWT) when OIDC is configured, otherwise HTTP Basic, and none when authentication is disabled — never both at once.

3.0.1 - 2026-07-14

Added

  • The server now prints an ASCII-art startup banner to stdout before the structured startup logs: the EHRbase-rs wordmark, the running version, the maintainer credit (Ruben Talstra), the project URL, and the load-bearing spec/platform pins (openEHR RM 1.2.0 · ITS-REST 1.0.3 · AQL 1.1 · PostgreSQL 18). The banner is suppressed under JSON logging (EHRBASE_LOG_FORMAT=json) so machine log consumers see only structured lines.

  • AQL queries are now planned once and cached: a repeated ad-hoc or stored query text reuses its lowered plan instead of re-parsing and re-analysing on every execution, while per-request parameter values, fetch/offset paging, and EHR scope still bind independently. Queries that resolve terminology (matches TERMINOLOGY(…)) are never cached, so their expansion is always current. New configuration knob EHRBASE_QUERY__PLAN_CACHE_CAPACITY (default 256; 0 disables the cache) bounds how many distinct plans are held, and a new aql_plan_cache_events_total metric (event = hit/miss) reports cache activity.

  • Storage migration 0008: a promoted context_start timestamptz column on COMPOSITION root node rows (backfilled from stored data, partially indexed), plus the fail-safe ext.openehr_timestamp conversion function. The AQL engine reads the indexed column for ORDER BY/WHERE on c/context/start_time/value — the measured patient-dashboard hot path — instead of re-extracting JSONB per candidate row; results are unchanged, including NULL placement and the verbatim projected value.

  • Overload backpressure: the REST server now caps the number of API requests it handles concurrently and sheds the excess immediately with 503 Service Unavailable + Retry-After: 1 instead of queueing every request until it runs out of memory. Under sustained offered load beyond database capacity the server now degrades with clean errors rather than being killed. The cap is configurable via EHRBASE_REST_MAX_IN_FLIGHT (concurrent requests, not per second; default 256, raise for high-throughput deployments; 0 disables shedding). The /status, health, and discovery endpoints are never limited, so operators can always probe an overloaded server. (No openEHR spec governs overload behaviour; the 503 follows RFC 9110 §15.6.4.)

  • Conformance framework (tools/conformance) redesigned and rewritten from the openEHR CNF component up (W-10). It now assesses any openEHR CDR: point it at a deployed server (scripts/conformance.sh with CONF_SUT=byo CONF_BASE_URL=…, or the CLI’s --sut byo --base-url …) and receive the full spec-cited artefact set — results.json, a conformance report, a Conformance Statement, a Conformance Certificate (a machine-computed framework assessment, explicitly not an official openEHR certification), and badges, written per SUT. Upstream EHRbase is a built-in target (CONF_SUT=ehrbase) with a committed fairness register; a cross-SUT comparison matrix can be rendered from two or more runs (conformance compare). Assertions carry a spec-edition ladder: the runner tries the newest edition form first (weak W/"…" ETags, RM 1.2.0 wire) and steps down to Release-1.0.3-era forms, reporting the satisfied edition level per case instead of failing a CDR on edition deltas; ehrbase-rs CI runs stay pinned to the development edition so the ladder can never mask a regression.

  • AQL: OR-combined CONTAINS expressions now execute (previously rejected as unsupported), including nested AND/OR/NOT containment trees, and NOT CONTAINS accepts compound operands.

  • ATNA auditing: EHR-Extract export and import operations now emit audit events (object class Extract) when auditing is enabled.

  • Multiple folder hierarchies per EHR (EHR.folders): beyond the /directory hierarchy, additional root FOLDERs can be committed through the CONTRIBUTION endpoint, each versioned independently. The EHR resource now carries the folders reference list (creation order) and directory (always its first member); EHR extract import and admin dump/load carry the hierarchies too. The /directory endpoints behave exactly as before.

  • ehr: URI support: DV_EHR_URI values are parsed against the full openEHR ehr: grammar (EHR / top-level structure by uid or exact version id / interior item paths, absolute and relative forms), and the server can resolve local ehr: references internally (e.g. LINK targets). openEHR path processing now also supports // path patterns and 1-based positional predicates in stored-structure navigation (AQL is unchanged — its grammar defines neither).

  • EHR_ACCESS access-control is now enforced. The spec-mandated, change-controlled EHR_ACCESS object of an EHR (RM ehr §EHR_ACCESS Class) is the foundational access-decision layer, evaluated after authentication and before dispatch on every EHR-scoped route; the enterprise RBAC/ABAC layers compose on top of it. Its settings use the ehrbase.access_control.v1 scheme: a default_access (open/restricted) with a user:/role: access list gating the EHR, per-Composition privacy-level ceilings on Composition reads, and a gate-keeper that guards changes to the settings themselves (403 Forbidden on a denial). Every existing EHR keeps working — the default (no settings) is open.

  • Client-supplied CONTRIBUTION uids are honoured on commit when unused (409 Conflict when already in use; previously silently ignored).

  • Prefer: resolve_refs is honoured on contribution reads: the CONTRIBUTION’s versions are returned as full ORIGINAL_VERSION objects instead of OBJECT_REFs (ITS-REST representation negotiation).

  • AQL single-row functions now execute: LENGTH, SUBSTRING, POSITION, the string CONTAINS, CONCAT/CONCAT_WS, ABS/MOD/CEIL/FLOOR/ ROUND, and CURRENT_DATE/CURRENT_TIME/CURRENT_DATE_TIME/NOW/ CURRENT_TIMEZONE (QUERY master03 §Functions).

  • AQL TERMINOLOGY() Boolean value expressions (TERMINOLOGY('validate'|'subsumes', …) = true) and terminology-URI matches operands (matches { terminology://… }) are now evaluated through the terminology service (previously typed rejects).

  • AQL archetype predicates now honour archetype-specialisation subsumption: a query naming a parent archetype (e.g. [openEHR-EHR-OBSERVATION.laboratory.v1]) also matches data created with any specialisation child (e.g. …laboratory-glucose.v1), scoped to the same RM entity and major version (BASE architecture_overview master10 §Design-time Relationships; AM master07 §Querying). Non-HRID predicates (at/id-codes) keep exact case-folded matching.

  • Version-tree branching and merge provenance (RM common master06 §Version tree / §Distributed versioning / §Version Merging). Branch version ids (trunk.branch.version) are now first-class on every surface: modifying a version that was imported from another system forks a branch with the local creating_system_id (the spec’s mandated rule for local modifications of copied versions) while the imported trunk version stays the container current; branch tips are continued, superseded, read, exported, and re-imported like any version; the container current / LATEST_VERSION (including in AQL) is the latest trunk version. ORIGINAL_VERSION.preceding_version_uid is now stored at commit (previously synthesized) and other_input_version_uids (merge provenance) is accepted on the CONTRIBUTION wire, preserved on import, and served on read. The vo_version storage carries the version tree in explicit columns with per-lineage temporal non-overlap constraints and the spec’s global version-identity uniqueness tuple.

Changed

  • Basic-auth verification no longer re-runs the Argon2 password hash on every request: verified credentials are cached (as a SHA-256 digest, never plaintext) for EHRBASE_REST_AUTH__VERIFIED_CACHE_TTL_SECONDS (default 60 s; 0 disables), and cache misses hash on a background thread. At load this removes roughly a full CPU core of per-request hashing.

  • Composition create/update responses are built from the commit result instead of re-reading the just-written document from the database — one connection acquisition and two queries fewer per write; when version signing is disabled the server also no longer rebuilds the full document it would only have signed. Response bodies and headers are unchanged.

  • Storage: the version table’s two GiST exclusion constraints and two speculative JSONB indexes on the node table (a GIN over every fragment and a magnitude expression index — no query the engine generates could use either) were removed; version-validity non-overlap is unchanged and held by construction (one open row per lineage via unique indexes, atomic close-then-insert writes, and an overlap audit on archive load). This removes the dominant per-commit index-maintenance and lock-contention costs on the write path.

  • Connection-pool defaults changed: EHRBASE_DB_MAX_CONNECTIONS 10 → 20, EHRBASE_DB_MIN_CONNECTIONS 0 → 2, and the per-checkout liveness ping is disabled (a broken connection is detected by its first statement). TCP_NODELAY is now set on accepted sockets, removing Nagle-induced latency on small responses.

  • Composition commits make fewer database round trips: the audit and contribution rows are written in one statement, and the create-path EHR existence + modifiability gates are one read instead of two. Error behaviour is unchanged (a missing EHR is still 404 before a non-modifiable 409).

  • The transactional event outbox is no longer written on every commit when no eventing consumer is configured. The per-commit event_outbox row (and its envelope serialization) is now written only when the AMQP publisher (EHRBASE_EVENTS_ENABLED) or the FHIR outbound emitter (EHRBASE_FHIR_OUTBOUND_ENABLED) is enabled. Consequence: the outbox records commits made while a consumer is enabled (at-least-once, even with zero bound subscribers — the gate is the boot-time config, not the current subscriber set); commits made while every consumer was off are not back-filled if eventing is later enabled.

  • IHE ATNA login (“Application Activity”) records now mark genuine authentication events rather than every authenticated request. A login record is emitted only when the request actually verified credentials (a Basic verified-credential cache miss); a cache hit continues an established session and a Bearer request authenticated out of band at the OIDC provider, so neither mints a per-request login record. Rejections (401/403) are still always audited, and login records remain off by default (EHRBASE_ATNA_SUPPRESS_LOGIN_EVENTS, default true).

  • Per-EHR EHR_ACCESS access-settings are cached as default-open at EHR creation, so the access gate’s first check on a freshly created EHR no longer costs a database lookup (a hospital-day workload creates EHRs constantly). Importing an EHR_ACCESS version into an existing EHR now evicts that cache entry, so the access decision reflects the imported policy immediately.

  • Composition validation is substantially faster with identical outcomes: the RM-invariant pass validates each node directly against the spec-generated Reference Model instead of deserializing every node into its typed struct (falling back to the typed path for anything it cannot vouch for), the archetype-constraint walk reuses constraint paths parsed once per cached WebTemplate instead of re-parsing them on every node visit, and validation error messages are byte-for-byte unchanged (equivalence is pinned by tests across the full corpus). Measured end-to-end: a fully populated International Patient Summary validates in well under half its previous time.

  • Version lifecycle states are now enforced as a state machine (RM common §Version Lifecycle): a commit whose lifecycle_state is not a legal transition from the preceding version’s state (for example incomplete → inactive without completing first) is rejected 422.

  • Template identifiers now compare case-insensitively (case-preserving): lookups accept any casing and uploading a case-variant duplicate is a 409 conflict, backed by a unique index (new migration).

  • AQL MIN/MAX aggregate over non-numeric leaves (text, dates, times) now compares type-appropriately instead of forcing a numeric cast, and mixed-type leaf comparison dispatches numerically for numbers.

  • Contribution commits now verify the target EHR exists (404 otherwise) and honour the EHR_STATUS.is_modifiable = false write guard and versioned-composition invariants on every path, including CONTRIBUTION-wrapped commits. Re-creating an existing directory (a folder hierarchy with the same root archetype and name) via a CONTRIBUTION is a 409 conflict; a hierarchy with a distinct root remains a new EHR.folders member.

  • EHR-index errors now carry the precise SM error names (ehr_id_does_not_exist, subject_id_does_not_exist) instead of a generic not-found.

  • Contribution retrieval now lists versions affected by attestation-only items alongside committed versions for demographic contributions, matching the EHR-scoped behaviour.

  • SMART App Launch resource-server support (openEHR SMART App Launch framework, development edition), config-gated and off by default (EHRBASE_REST_SMART__*): the /.well-known/smart-configuration discovery document, the full resource-scope grammar (compartment/resource.permission with */**/ns::* patterns), and scope + launch-context (ehrId→patient) enforcement composed after RBAC/ABAC.

  • Subject Proxy Service completed (SM I_SUBJECT_PROXY_SERVICE): variables are now tracked over time (a persisted sample history per variable), currency freshness is evaluated (fresh samples are served without re-querying; data-set registration tightens currency), data-set local aliases resolve on reads, using_app_ids lifecycle drops empty data sets, and frames execute with primary→fallback semantics. New FHIR frame executor (config-gated named systems, EHRBASE_SUBJECT_PROXY__*) lets variables be populated from FHIR R4 servers; manual variables gain a notification input channel.

  • System API OPTIONS / conformance manifest rebuilt: reports the live mounted endpoint groups, a single provenance source (the tested development-edition ITS-REST identity), and configurable identity fields (EHRBASE_REST_SYSTEM__*); also mounted at the API base path.

  • Item tags via headers (openehr-item-tag/openehr-version-item-tag): accepted on EHR-group and demographic writes and echoed on responses.

  • Query API: multi-EHR scoping (ehr_ids set), an honest ehr_id_does_not_exist (404) for a well-formed absent EHR id, a weak ETag on RESULT_SET responses, parameter-substituted meta._executed_aql, and an optional query execution timeout (EHRBASE_QUERY__TIMEOUT_MS) mapped to 408.

  • Definition API: template list filtering (template_id glob, concept, version) and pagination are honoured; stored-query query_type is read with an honest unsupported-formalism rejection; ADL1.4 uploads return the JSON TemplateIdentifier under Prefer: return=identifier.

  • FLAT/STRUCTURED (Simplified Formats, now STABLE): the _-prefixed optional RM attribute family (_uid, _link, _feeder_audit, _null_flavour, _mapping, _normal_range, participations, work-flow ids, …) round-trips in both directions; |raw canonical-JSON embedding on write; complete quantity/date-time/multimedia leaf attribute tables; |other open-value-set rules enforced.

  • Development-edition ITS-REST protocol adopted (the server’s tested contract identity, now reported consistently as such): ETag response headers carry the weak W/"…" indicator (bare quoted values are still accepted on If-Match); committal metadata uses the lowercase openehr-version / openehr-audit-details value-form headers (the deprecated openEHR-VERSION.* dotted spellings remain accepted) and a client-supplied system_id is merged into the commit audit; Location is emitted only on resource creation (no longer on reads/deletes); Preference-Applied echoes the honoured Prefer; 405/501 render the openEHR error body.

  • Demographic DELETE follows the published Demographic API: the preceding version id rides in the path; a stale id yields 409 (with the latest version ETag), an already-deleted party 400.

  • Admin DELETE /admin/ehr/all follows the published Admin API: 204 with no body, and an absent ehr_id parameter now means delete ALL EHRs.

  • FLAT duplicate node-name suffixes default to the specification form (name_1); the Better-compatible form (name2) is available behind the ehrbase-quirks feature.

  • The ehrbase-rest and ehrbase-sm crates were restructured specification-first (one folder per ITS-REST spec / SM chapter, all spec-silent surfaces quarantined under extensions/) — no route changes beyond those listed here.

  • PUT …/composition/{uid_based_id} rejects a body whose COMPOSITION.uid does not identify the versioned object addressed by the path (400).

  • AQL semantic analysis is stricter per QUERY master03: duplicate FROM variable names reject, variable references are case-insensitive, LIMIT 0/negative OFFSET reject, SUM/AVG over non-numeric paths reject, scalar-function arity is validated, and LIKE \*/\? escapes now match the literal characters.

  • OPT 1.4 template upload enforces the AOM 1.4 constraint-model invariants (attribute existence bounds, single-attribute occurrences, archetype-id well-formedness and root-type match, slot identifier validity, internal-reference target paths, constraint-reference definedness, boolean satisfiability, assumed-value validity, temporal and duration constraint-pattern validity, duplicate code-list codes) — invalid templates are rejected with 400 carrying the AOM rule code.

  • ADL2 artefact upload (I_DEFINITION_ADL2) now validates sources against the registration-decidable AOM2 catalogue (mandatory sections, header versions, root type/node-id rules, specialisation depth, terminology language consistency, code definedness, value-set validity, term-binding keys) instead of a header-only probe — invalid sources are rejected with 422 carrying the AOM2 rule code.

  • Stricter spec-mandated validation on the commit path: a client AUDIT_DETAILS with an empty system_id, a committer PARTY_IDENTIFIED/PARTY_RELATED with no identity, an empty committer name, or a PARTY_RELATED.relationship outside the openEHR subject_relationship group is now rejected with 422 (previously accepted, or surfaced as a 500 DB error); a non-root RM node carrying archetype_details violates LOCATABLE.Archetyped_valid and is rejected; EHR-Extract versions[] members with a _type other than ORIGINAL_VERSION are rejected on import.

  • AQL VERSION uid values are now built from each version’s stored creating_system_id and version-tree id, not the server’s live system_id configuration.

  • The ehrbase-rs-postgres image now pre-creates the layered group roles (ehrbase_migrator, ehrbase_app, ehrbase_reader), so Compose/dev deployments get the same least-privilege grant topology as hardened deployments instead of roles absent startup notices. Existing data volumes keep working; recreate the volume (or create the roles once by hand) to pick the grants up.

  • Public documentation website at https://rubentalstra.github.io/ehrbase-rs/: a product landing page, a versioned user guide (frozen per release, dev tracking develop), and an offline OpenAPI endpoint reference covering all seven openEHR API groups. Built from website/ and deployed by CI, with link-check and OpenAPI-drift gates.

Fixed

  • The composition validator no longer falsely rejects templates that use the same archetype more than once under one container, differentiated by name: each instance is now routed to the sibling constraint whose name it satisfies, instead of being checked against the first same-archetype sibling’s overlay. Cross-contaminated content (a child from one overlay placed in the other-named instance) is still rejected.

  • Template example generation (GET …/example) at detail_level=medium and complete no longer produces an empty composition for templates whose content is entirely optional: medium now returns a fully-populated single-instance committable example (honouring temporal patterns, C_DURATION field patterns, media-type code lists, and container cardinality bounds), and complete additionally demonstrates a second occurrence of repeating nodes. required (the default) is unchanged.

  • AQL SELECT c/uid/value (and c/uid) on a COMPOSITION — or any versioned-object root — now returns the server-assigned OBJECT_VERSION_ID, version-correct under LATEST_VERSION and ALL_VERSIONS. It previously returned null because the uid was injected only on REST reads, never into stored data. (QUERY master03 lists COMPOSITION.uid.value as a normative identified path.)

  • Composition commits against an already-seen template no longer re-read the stored OPT from the database on every commit — the built WebTemplate cache is now consulted first (measured: 10,206 redundant reads in a 120 s load window, the #2 database statement by total time). Deleting a template now also evicts it from that cache, so a commit racing a delete gets the correct 422 (“template not known”) instead of a foreign-key 500.

  • Template example generation (GET /definition/template/adl1.4/{id}/example) now honours the template’s structural constraints: a missing mandatory ENTRY structure (e.g. ACTION.description) is synthesized with the template’s constrained node (its RM type, archetype_node_id, and name) instead of a blind at0001 placeholder, so generated examples validate and commit against the same template. Surfaced by the official openEHR CKM International Patient Summary template; probed by the new conformance case ECC-TPL-017 (example → commit round-trip).

  • Template list endpoints no longer ignore filter and pagination parameters.

  • The conformance manifest and /rest/status no longer misreport the implemented ITS-REST edition as 1.0.3.

  • Contribution commits: a creation version against an already-existing object, and a modification/deletion/attestation whose preceding_version_uid names an object the server does not hold, now return 400 (the contract’s modification-type-mismatch scope) instead of 422/404 — on POST /ehr/{ehr_id}/contribution, 404 is reserved for an unknown ehr_id.

  • Versioned-object reads (GET …/versioned_composition, …/versioned_ehr_status, versioned directory) now emit the concrete RM class (VERSIONED_COMPOSITION / VERSIONED_EHR_STATUS / VERSIONED_FOLDER) in _type, not the abstract VERSIONED_OBJECT.

  • Demographic API: If-Match preconditions now verify the full OBJECT_VERSION_ID (previously only the version-tree number, which accepted phantom versions); relationship delete now honours the same If-Match preconditions as party delete; demographic ETags are emitted in the weak form (W/"…").

3.0.0 - 2026-07-11

First public release of EHRbase-rs — a pure-Rust openEHR Clinical Data Repository. Version numbering starts at 3.0.0: this project began as a fork of EHRbase (Java, 2.x line) and is released as its next-generation successor; inherited upstream tags/releases were removed from the fork. Published as a pre-release: the platform is feature-complete and conformance-verified, but has not yet run in production.

Added

openEHR platform

  • openEHR REST API (ITS-REST 1.0.3): EHR, EHR_STATUS, COMPOSITION, DIRECTORY/FOLDER, CONTRIBUTION, QUERY, DEFINITION (ADL 1.4 + ADL2), admin and management surfaces, with canonical JSON and XML content negotiation. The wire contract is generated from the official openEHR OpenAPI/BMM/XSD models with a CI drift gate.
  • AQL 1.1 query engine: typed path analysis over a spec-generated Reference Model compiled to PostgreSQL SQL; LATEST_VERSION and ALL_VERSIONS; terminology-backed TERMINOLOGY() expansion; stored parameterised queries.
  • Full change-control semantics: contribution-atomic commits, indelible temporal version history (PostgreSQL 18 WITHOUT OVERLAPS), logical delete, attestations, per-version digital signatures (RFC 8785), point-in-time reads.
  • Templates and validation: OPT 1.4 ingestion with artefact validity checking (AOM2 codes), WebTemplate / FLAT / STRUCTURED simplified formats, deep archetype-constraint validation on every commit.
  • EHR Extract and messaging (SM I_EHR_EXTRACT/I_MESSAGE/I_TDD): whole-EHR export/import preserving distributed version identity, EHR cloning, TDD import.
  • Demographics: versioned party store (PERSON, ORGANISATION, GROUP, AGENT, ROLE) with relationships.
  • Terminology: the bundled openEHR terminology plus pluggable external FHIR terminology servers (validate / expand / subsume).
  • Conformance instrument: the ECC runner executes the full catalogue (341 cases, JSON + XML) against the composed server and computes profile verdicts — CORE: PASS · STANDARD: PASS · OPTIONS: OBTAINED, generating the Conformance Statement + Certificate.

Integration

  • Change events: transactional outbox publishing every contribution commit to AMQP/RabbitMQ — at-least-once, per-EHR ordered, PHI-free envelopes, server-side filterable subscriptions (off by default).
  • FHIR R4 connectors: mapping-driven inbound ingestion (validated compositions with FEEDER_AUDIT provenance), a read façade over AQL, and event-driven outbound resource emission (off by default).
  • S3 multimedia externalization: threshold-based content-addressed offload of DV_MULTIMEDIA to any S3-compatible store with sha-256 integrity verification; SeaweedFS supported out of the box (off by default).

Security & operations

  • Authentication: HTTP Basic (argon2) and OAuth2/OIDC bearer (Keycloak, Active Directory, any standards-compliant IdP).
  • Authorization: RBAC plus ABAC via the embedded Cedar policy engine or a remote PDP.
  • Multi-tenancy: each tenant an isolated logical openEHR system with its own system_id, enforced by PostgreSQL row-level security (off by default — single-tenant mode is unchanged).
  • IHE ATNA system log: DICOM audit messages over (TLS) syslog with build-time operation coverage.
  • Observability: structured logs, OpenTelemetry traces, Prometheus metrics, health probes; identified data never enters telemetry.
  • Layered database roles (migrator / writer / reader) with a hardened PostgreSQL baseline.

Deployment

  • Docker Compose stack (server + PostgreSQL 18) with an optional Grafana LGTM observability overlay.
  • Distroless, non-root, shell-less multi-arch container images (amd64 + arm64) on GHCR.
  • Helm chart with security-hardened defaults (non-root, read-only rootfs, seccomp, default-deny NetworkPolicy) and golden-render validation.