Ekklesia/docs/development/authorization.md

Authorization

Phase 4C establishes Core's authorization foundation and wires it into apps/api over HTTP. Phase 5A extends that same system so an installed module's declared permissions are enforced and grantable through the identical mechanism as Core's own. This page covers all of it: the schema/evaluator/store/registry Core owns, the platform-owner bootstrap operation, the enforcement/endpoints apps/api exposes, and how module manifests contribute to the one platform permission registry. See Identity for the identity foundation this builds on, API for the HTTP transport it is mounted into, Module development for the manifest shape that declares module permissions, and ADR 0013 / ADR 0014 for the durable architectural decisions this page describes in more implementation detail.

What this is, and is not

An operator can bootstrap exactly one platform owner for a deployment; a platform owner (or anyone they delegate to) can create, list, and revoke explicit authorization grants over HTTP for any registered permission, Core-owned or module-owned; a module can register its own HTTP routes and have them mounted, enforced, and reachable in production (see API); and every enforced route fails closed on infrastructure trouble instead of silently allowing or denying. It is not possible to manage roles, groups, teams, or tenants; there is no user-deletion endpoint, owner-management HTTP endpoint, or owner-recovery workflow (revoking or re-granting platform-owner status is not exposed at all, and deleting the last owner does not reactivate bootstrap); there are no service accounts or API tokens, so only cookie-authenticated browser callers can reach mutation routes, Core-owned or module-owned; and relationship-derived grants (for example, automatic access from a roster assignment) cannot yet be created by anything in this repository — the generic grants API structurally cannot create or revoke them either.

Core authorization schema and store

Established as a prior checkpoint on this branch (ekklesia_core_platform_owner, ekklesia_core_grant, ekklesia_core_authorization_audit; AuthorizationEvaluator, AuthorizationStore; @ekklesia/contracts's PermissionId/GrantId/AuthorizationScope branded types). This page does not restate that design — read core/src/authorization/schema.ts, evaluator.ts, and store.ts directly for it. This page covers what pass 2 added on top: the permission registry, the bootstrap operation, and HTTP enforcement — plus the Phase 5D pass 3 integrity rules recorded in ADR 0019.

What user deletion removes, and what it does not

Deleting an identity user still cascades to that user's live authorization rows: their grants (ekklesia_core_grant) and their platform-owner row (ekklesia_core_platform_owner) are removed, so nothing continues authorizing a principal that no longer exists.

It does not remove security records. ekklesia_core_authorization_audit.subject_user_id is a retained, non-null historical snapshot with no foreign key to the user table, and ekklesia_core_platform_owner_bootstrap has no user foreign key at all. Both survive user deletion. Previously the audit subject cascaded, so deleting a user destroyed exactly the record of what they had been granted, by whom, and when it was revoked.

Only opaque identifiers are retained — a subject id and an actor id. No name, email, or other profile data is ever copied into an authorization table.

Append-only is an application convention, not a database guarantee. Core's authorization store exposes no update or delete path for audit rows. Nothing in PostgreSQL enforces that: no role, trigger, or rule prevents a database administrator, a future migration, or anything else holding the same credentials from modifying them. They are not cryptographically immutable.

Deploying these changes

All of the above depends on Core migration 0002_authorization_integrity, which creates ekklesia_core_platform_owner_bootstrap, drops the cascading audit-subject foreign key, and backfills consumed state. Migrations remain explicit operator work and never run at application startup — apply 0002 with pnpm migrations:apply before starting the new code, exactly as with every other migration. It is classified {"mode":"none"} in core/drizzle/meta/_ekklesia.json: it creates a table, removes one referential constraint specifically to retain data, and inserts one backfill row; it drops no table, column, or row. Migrations 0000_identity_core_tables and 0001_authorization_core_tables are byte-identical and retain their existing checksums.

Permission registry

PermissionDeclaration (packages/contracts/src/authorization.ts) is the one authoritative, strict Zod-validated shape of a registered declaration — exactly {id, baseline, description}, nothing more — shared by Core's authorization registry and a module manifest's permissions array. Core does not declare its own separate interface for this; core/src/authorization/permissions.ts imports and re-exports the contracts type so existing @ekklesia/core/authorization consumers keep importing it from the same place. id is a branded PermissionId, baseline is public/authenticated/restricted, and description is required, trimmed, non-empty operator and audit-facing text up to 240 characters — informational only; nothing in authorization reads or interprets it.

core/src/authorization/permissions.ts defines the smallest possible registry on top of that shape: an immutable, composition-time-built PermissionRegistry distinguishing a registered declaration from a merely well-formed PermissionId string. Two properties matter:

• Duplicate declarations fail at composition time. createPermissionRegistry throws PermissionRegistryError (PERMISSION_REGISTRY_DUPLICATE_DECLARATION) if the same permission id is declared twice — this can only happen while wiring the registry itself, not at request time.
• Unknown permissions fail closed at lookup time. registry.get(permissionId) and registry.has(permissionId) both re-validate their input (safeParsePermissionId internally) and return undefined/false for anything malformed or simply not registered — there is no "trust this permission id because it looks right" path.

Permission matching is always exact string equality, both in the registry and in the evaluator's grant lookup — there is no wildcard, prefix, or namespace-inheritance matching anywhere; the required dot-namespacing in a PermissionId (core.authorization.grants.manage, calendar.events.read) is purely a naming convention for readability and collision-avoidance.

CORE_PERMISSION_REGISTRY is Core's own registry and remains what Core-only code and existing tests use directly. CORE_PERMISSION_DECLARATIONS is the same declarations exported separately in deeply immutable form, for the module composition helper below to build on without reconstructing Core's own list:

Permission	Baseline	Description
core.authorization.grants.manage	restricted	Create, list, and revoke another user's explicit authorization grants.
core.modules.ui.read	authenticated	List the installed module UIs a browser host may mount.
core.dashboard.navigation.manage	restricted	Organize the Dashboard: create, rename, order, and delete navigation groups.
core.modules.lifecycle.manage	restricted	Read and change which installed modules are enabled.

Every Core permission is a platform permission and can never be declared, gated, or influenced by a module — core.* is a reserved namespace (see below). core.modules.lifecycle.manage gates both reading and changing module enablement (ADR 0028): the read is not a milder version of the write, because the lifecycle view is the only place a disabled module, a blocked module, and the required-dependency edges between them are visible at all, whereas core.modules.ui.read deliberately covers only what is mountable. Holding it confers no authority inside any module.

Enablement never changes how a permission is enforced. A disabled module's permissions stay registered and grantable — composition spans the whole discovered inventory, not the eligible subset — so a grant survives an enable/disable cycle and can be issued ahead of enabling a module. Nothing enforces those permissions while the module is disabled, because it mounts no routes at all.

Module permission composition

A module manifest's permissions array (manifestVersion: 4, Module development) declares permissions in the exact same PermissionDeclaration shape — but declaring one there only proves it is well-formed and correctly namespaced under the module's own id; it does not, by itself, register anything. composePlatformPermissionRegistry (core/src/modules/permissions.ts) is the trusted composer: given an already-discovered module list (for example discoverModules's result), it builds one combined PermissionRegistry from CORE_PERMISSION_DECLARATIONS followed by every module's declarations, in discovery order, using the same createPermissionRegistry constructor — not a second registry implementation. Declaration order is deterministic: Core first, then each module in the order it appears in the supplied list, then each module's own declarations in manifest-array order. A duplicate exact permission id anywhere in that combined list fails composition the same way a duplicate would fail createPermissionRegistry directly — since a permission id must start with its own module's id, this is only reachable across two entries that share the same module id, which real discovery already rules out (MODULE_ID_DUPLICATE); composition still re-derives the check rather than trusting that guarantee alone.

core.* is reserved for Core's own permissions. A discovered module contributing any core.-prefixed permission id fails composition with ModulePermissionCompositionError (MODULE_PERMISSION_RESERVED_NAMESPACE) before duplicate detection ever runs — since every Core-owned id is itself core.-prefixed, this is what actually catches a module attempting to collide with a Core-owned id specifically, regardless of that module's own id and regardless of whether the id happens to exactly match an existing declaration.

Composition itself performs no filesystem, environment, database, network, or module-code operation — it is pure, synchronous, in-memory composition over data a caller already discovered and validated (core/src/authorization and core/src/modules remain independent of each other in the other direction: authorization never imports module discovery or runtime, only the module integration layer depends on authorization).

API startup: discovery, composition, and enforcement

apps/api builds its production permission registry, and starts the module runtime, from the exact same installed-module inventory every time it starts, in a fixed order, before anything else:

1. Resolve module-discovery configuration (the repository's modules/ directory, and the platform/Module SDK versions from their own package.json files).
2. Discover installed modules exactly once (discoverModules).
3. Compose the platform permission registry from that exact discovered list (composePlatformPermissionRegistry).
4. Construct the shared PostgreSQL pool.
5. Construct Identity, the authorization store, and the evaluator.
6. Start the module runtime from that same discovered list (startDiscoveredModuleRuntime) — modules are never rediscovered for this step.
7. Create the Hono application, injecting the composed registry and mounting the started module runtime's published routes (see API).
8. Bind the HTTP server.

apps/api/src/platform-permissions.ts's prepareStartupModuleInventory performs steps 1–3, returning both the discovered module list and the composed registry so step 6 can start the module runtime from that identical list rather than rediscovering modules a second time. (composeStartupPermissionRegistry remains available as a registry-only convenience wrapper around it, for callers — Doctor, modules:validate, existing tests — that only need the registry.) Discovery and composition happen strictly before any pool, Better Auth instance, module entrypoint import, or HTTP server exists, so a failure there — an invalid manifest, a broken dependency graph, a module claiming the reserved core.* namespace, or a duplicate permission declaration — leaves nothing to close: it propagates directly out of startApi(), and startup never falls back to CORE_PERMISSION_REGISTRY or any Core-only registry. The repository root is resolved from this compiled/source file's own location (two directories up from apps/api/dist/*.js or apps/api/src/*.ts, which sit at identical depth), never process.cwd() — a package-scoped test run has apps/api as its working directory, which would otherwise point discovery at the wrong modules/ directory.

The resulting registry is a static, immutable snapshot taken once at startup. There is no hot installation, removal, or dynamic mutation of installed modules or their permissions while the process is running — an installed-module change (add, remove, or edit a manifest) takes effect only on the next API restart. createApiApp requires this registry as part of its injected authorization options; requirePermission and the explicit-grant routes (GET/POST /api/v1/grants, POST /api/v1/grants/:grantId/revoke) consult exactly the same registry Core's own core.authorization.grants.manage permission is registered in, so a module-owned permission and Core's own are enforced, and grantable, through the identical mechanism — there is no module-specific enforcement path. Removing a module from the next startup's discovered inventory makes its permission unknown to new authorization checks and new grants (a well-formed but unregistered permission denies exactly as before, including on POST /api/v1/grants, which checks the registry explicitly before calling the store); it never deletes, revokes, or otherwise modifies a historical grant or audit row already recorded for that permission. A previously granted, still-unrevoked explicit grant for a now-uninstalled module's permission remains visible through GET /api/v1/grants and simply stops being something a restricted-baseline authorization check can newly succeed against. Listing and explicit revocation (POST /api/v1/grants/:grantId/revoke) are unaffected either way: neither route consults the registry for a grant that already exists — revocation is keyed to the grant's own id and source kind, not to whether its permission is currently registered — so a manager can still revoke a historical grant for a now-unregistered permission.

A module's own mounted HTTP routes (/api/v1/modules/<module-id>/*) are enforced through this identical registry and requirePermission middleware — see API for the route-mounting contract. Relationship-derived grant creation remains a deferred, unrelated slice — this work only makes a module's declared permission usable through Core's existing, generic explicit-grant HTTP API, enforcement middleware, and now its own mounted routes.

Doctor and modules:validate

pnpm doctor's modules.permissions check composes the platform permission registry from the same already-discovered module list modules.discovery/modules.dependencies already produced — it never re-discovers modules itself. A successful composition reports a pass with a safe declaration count (for example 4 permissions registered with zero installed modules, since Core's own declarations are always present); zero installed modules still composes successfully. If module discovery itself failed, the check is skipped, matching the existing schema/migration checks' pattern; if discovery succeeded but composition failed, schema and migration checks still run normally — they are independent diagnostics that do not depend on the permission registry. A reserved-namespace violation (ModulePermissionCompositionError, MODULE_PERMISSION_RESERVED_NAMESPACE) and a generic duplicate declaration (PermissionRegistryError, PERMISSION_REGISTRY_DUPLICATE_DECLARATION) are reported as distinct failure details — a reserved core.* claim is never described as an ordinary duplicate, even though (see above) a module attempting to exactly collide with a Core-owned id always surfaces as the reserved-namespace failure specifically. Composition requires no database and never imports or executes a module's server entrypoint. pnpm modules:validate composes the same registry from its own already-discovered modules for the identical reason: a manifest that is otherwise structurally valid but claims the core.* namespace, or duplicates a permission id, now fails validation before that module is considered installable.

Two layers, not one: shape versus registration

createAuthorizationEvaluator (core/src/authorization/evaluator.ts) has no registry and is never given one — it independently re-validates permission, baseline, and scope as a pure shape check (safeParsePermissionId/permissionBaselineSchema/safeParseAuthorizationScope) and denies a malformed value under every baseline, including public and authenticated, before any baseline is allowed to return an allowed decision. This is defense in depth for a future caller that reaches the evaluator directly, not a substitute for the registry: the evaluator has no way to tell a well-formed-but-unregistered permission id (for example a plausible-looking calendar.events.read that nothing has ever declared) from a genuinely registered one — it can only reject what is not even shaped like a PermissionId at all.

Registration — "does this permission actually exist" — is exclusively the registry's responsibility, consulted by every trusted caller before the evaluator ever runs: requirePermission (apps/api/src/authorization-middleware.ts) looks up the caller's own required permission (for example core.authorization.grants.manage, always a hardcoded route constant, never client-supplied) in the registry before resolving a principal or evaluating anything; POST /api/v1/grants (apps/api/src/grants-routes.ts) separately looks up the target permission being granted — the one named in the request body — in the same registry, after body shape-validation and before ever calling store.createGrant, so a caller cannot create a dormant grant for an invented permission id that might later become effective the moment someone actually registers it. A malformed target permission (fails PermissionId shape validation) still resolves to the body-level 400 {"code":"INVALID_REQUEST"} documented below, unchanged; a well-formed but unregistered target permission instead receives the identical sanitized 403 {"code":"AUTHORIZATION_DENIED"} a genuinely denied caller would, so that response never reveals whether the caller was denied or the permission was simply never registered.

Platform-owner bootstrap

There is no "first user becomes owner" inference anywhere in this repository, and no HTTP endpoint exposes bootstrap. The only way to establish a platform owner is the local CLI command:

pnpm migrations:apply   # the database must already be migrated
DATABASE_URL=postgresql://... pnpm authorization:bootstrap-owner -- \
  --user-id <existing-ekklesia-user-id> \
  --confirm-user-id <same-id>

DATABASE_URL is read only when the command actually executes (never merely by importing the CLI module — proved by core/tests/cli-bootstrap-platform-owner.test.ts), and it is never printed, logged, or included in any error. --user-id and --confirm-user-id must match exactly, checked before the id is even parsed as a valid EkklesiaUserId — this command never infers a target from signup order or "the only user in the database"; the operator must name the exact, opaque id twice.

Bootstrap is permanently single-use

core/src/authorization/bootstrap.ts's bootstrapPlatformOwner(db, {userId}) is a dedicated, Hono-independent Core operation — deliberately not a composition of "check for an owner, then call the existing grant function."

"Already consumed" means the singleton row exists in ekklesia_core_platform_owner_bootstrap — a dedicated Core-owned table with id pinned by a check constraint to the literal 'singleton' and no foreign key to any user. Core never deletes from it.

This is deliberately no longer "an owner row exists". That marker cascaded on user deletion, so deleting the bootstrapped owner's user row removed the owner row and its audit evidence, after which the one operation that establishes a platform's first administrator could be run again. The operational consequence of the fix is explicit and intended: deleting the last platform owner does not reactivate bootstrap, and neither does revoking one. Re-establishing an owner after that requires a separately designed owner-management flow, which does not exist yet. There is no reset command, deletion path, HTTP endpoint, environment override, or force flag.

Upgrading an existing installation is handled by Core migration 0002_authorization_integrity, which inserts the singleton if any platform-owner row already exists — active or revoked — so an existing deployment cannot gain a fresh first use. It records no consumed_for_user_id, because a backfill cannot know which of several historical owner rows the original single use targeted, and guessing would fabricate history. An installation that never had an owner is left untouched and keeps its one first use.

One upgrade limitation is irrecoverable: if an operator manually deleted the bootstrapped owner's user row before this migration, the old cascade already erased both the owner row and its audit evidence, and the migration has nothing left to detect — such an installation backfills as never-bootstrapped. At the current checkpoint no supported Better Auth or API user-deletion path exists that could produce this state; it requires direct database access.

The sentinel claim, the target user's existence check, the owner-grant insert, and its audit entry are all one PostgreSQL transaction. Consumption is therefore tied to success: a missing target user, a constraint violation, or any other failure rolls the sentinel back with everything else, leaving the one legitimate first use available. The very first statement inside that transaction is a transaction-scoped advisory lock:

select pg_advisory_xact_lock(1500000001, 1500000002)

This lock key pair is distinct from the migration runner's own advisory lock (core/src/migrations/postgres-database.ts), so bootstrap serialization can never collide with or be blocked by unrelated migration locking. pg_advisory_xact_lock releases automatically on commit or rollback, so no manual unlock path is needed even when the transaction fails partway. Two concurrent bootstrap attempts serialize on this lock: the second caller blocks until the first's transaction commits or rolls back, then re-evaluates the "does any owner row exist" check against now-current state — so exactly one of two concurrent attempts ever succeeds, proved live against real PostgreSQL in core/tests/authorization-bootstrap.integration.ts (Promise.allSettled on two simultaneous calls; asserts exactly one fulfilled, exactly one rejected with PLATFORM_OWNER_BOOTSTRAP_ALREADY_CONSUMED, exactly one owner row, exactly one audit row).

A failed target-user lookup, an audit-insert failure, or losing the concurrency race all roll back the entire transaction — there is no partial state where an owner row exists without its audit entry, or vice versa.

The audit entry for a bootstrap grant always uses a fixed, sanitized, non-user actor identity — never derived from any caller input — so an auditor can tell a given platform_owner.grant row came specifically from this command rather than from a future owner-management flow:

Field	Value
actor_type	system
actor_id	authorization-bootstrap-cli

Error codes

PlatformOwnerBootstrapError never carries the underlying database error, SQL text, host, port, path, or stack — only a stable code and a fixed message:

Code	Meaning
PLATFORM_OWNER_BOOTSTRAP_ALREADY_CONSUMED	The durable bootstrap-consumed singleton exists; bootstrap has already been used and cannot be used again.
PLATFORM_OWNER_BOOTSTRAP_USER_NOT_FOUND	--user-id does not match an existing identity user.
PLATFORM_OWNER_BOOTSTRAP_FAILED	Any other failure (database/infrastructure).

HTTP authorization enforcement

apps/api/src/authorization-middleware.ts exports requirePermission(options, permissionId), a Hono middleware factory reused by every protected route:

1. Looks up permissionId in the trusted, server-side PermissionRegistry (never a client-supplied baseline) — an unregistered permission id fails closed with 403 before the principal is even resolved. In today's wiring only hard-coded registered constants are ever passed by route code, but the check stays live per-request as defense in depth rather than being trusted once at composition time.
2. Resolves the principal exclusively through the same server-side identity runtime /api/v1/principal already uses (EkklesiaIdentityRuntime.resolvePrincipal) — nothing here ever trusts a client-supplied claim about identity.
3. Calls the injected AuthorizationEvaluator with the registry's baseline and permission for that principal (platform owners pass through the evaluator's owner bypass here — this is how delegation to core.authorization.grants.manage recipients also works, without a second code path).
4. On success, stores the resolved EkklesiaPrincipal (Core's narrow {kind, userId} — never a Better Auth session/user object) on the Hono context via c.set, readable by the route through getAuthorizedPrincipal(c). This is a deliberately narrow exception to the "no generalized session/user object in Hono context" concern ADR 0011 raised — only this already-public-safe principal shape is ever set.

A denied decision and an unregistered permission id both return a sanitized 403 {"code":"AUTHORIZATION_DENIED"}. A genuine infrastructure failure — principal resolution, evaluation, or the database underneath either — is never folded into that 403; it returns a sanitized 503 instead (PRINCIPAL_RESOLUTION_FAILED or AUTHORIZATION_EVALUATION_FAILED), mirroring /api/v1/principal's existing rule that "could not be determined" must never look like "no permission." Neither response ever reveals whether a grant or owner row exists for the caller.

Origin enforcement for mutations

requireTrustedOrigin(trustedOrigins) requires an exact match against the same EKKLESIA_IDENTITY_TRUSTED_ORIGINS set the API already uses for CORS and Better Auth's own checks (see API) — a missing or untrusted Origin header is rejected with 403 {"code":"UNTRUSTED_ORIGIN"} before the permission check or the route handler ever runs. This exists because CORS response headers alone are not CSRF protection: they only stop a browser from reading a cross-origin response, not from sending the state-changing request in the first place. Every grants mutation route (POST /api/v1/grants, POST /api/v1/grants/:grantId/revoke) requires this; the read-only list route does not. Non-browser, non-cookie automation (service accounts/API tokens) is out of scope for this slice and has no way to satisfy this check yet.

Explicit-grant HTTP API

Mounted at the API's root by apps/api/src/grants-routes.ts. Every route requires core.authorization.grants.manage. All three routes only ever act on explicit grants — the generic HTTP API structurally cannot create or revoke a relationship-derived grant.

Route	Origin required	Notes
GET /api/v1/grants?userId=<id>&limit=<1..100>&cursor=<opaque>	No	One bounded page of the subject's explicit grants only (sourceKind: "explicit" filter applied server-side on every page).
POST /api/v1/grants	Yes	Creates a new unscoped explicit grant.
POST /api/v1/grants/:grantId/revoke	Yes	Revokes an explicit grant, and every unrevoked equivalent of it, without deleting anything.

:grantId must be the canonical lowercase hyphenated UUID PostgreSQL emits (8-4-4-4-12 lowercase hexadecimal). Uppercase, whitespace-padded, shortened, overlong, and non-UUID values are rejected with 400 {"code":"INVALID_REQUEST"} after authorization but strictly before any store or database call — previously such a value reached PostgreSQL's own UUID comparison and returned an indistinguishable 503. A canonical id that does not exist still returns the sanitized 404.

POST /api/v1/grants request body

Only three keys are accepted; any other key present (scope, actor, owner, isOwner, sourceKind, sourceRef, id, createdAt, revokedAt, createdByActorId, effectivePermissions, or anything else) makes the entire body invalid rather than being silently dropped, so a caller cannot smuggle a server-derived field through a request that otherwise looks well-formed:

Field	Required	Notes
subjectUserId	Yes	Must parse as a valid EkklesiaUserId.
permission	Yes	Must parse as a valid PermissionId and be registered in the running API's composed platform registry (Core's own declarations plus every installed module's — see API startup). A malformed value is a 400; a well-formed but unregistered one is a 403.
expiresAt	No	Exactly YYYY-MM-DDTHH:mm:ss.sssZ, byte-for-byte equal to the parsed date's own toISOString(), and strictly in the future.

The server always derives, and the request body can never supply: the grant id, sourceKind (always "explicit"), the actor (always the authenticated caller, via createUserActor(principal.userId)), and all audit/timestamp fields.

The generic HTTP surface creates unscoped explicit grants only. A scope field is rejected in every form — including null, empty, malformed, and otherwise valid scope strings — with 400 {"code":"INVALID_REQUEST"} before any store mutation. No HTTP route middleware supplies a scope to the evaluator today, so a scope accepted here would produce a grant that persists, appears in listings, and can authorize nothing. Core's AuthorizationScope, the evaluator's exact-scope semantics, and the store's scope column are unchanged and remain available to trusted internal callers; listing still shows a scope on historical or internally created grants. Scopes cannot authorize a module HTTP route today, and no route-parameter-to-scope resolver was added.

expiresAt deliberately accepts one representation and nothing else. Rejected: offsets (2027-01-01T00:00:00.000+02:00), missing milliseconds (2027-01-01T00:00:00Z), excess fractional precision, lowercase separators, whitespace padding, date-only (2027-01-01), locale forms (March 5, 2027), extra trailing text, normalized impossible dates (2026-02-31T00:00:00.000Z, which bare new Date() silently rewrote to March 3rd), and any instant at or before the acceptance clock. Core's store independently rejects an invalid or non-future Date before any SQL, so a trusted internal caller cannot create a dead-on-arrival grant either. A rejected expiration creates no grant and no audit row. There is no maximum lifetime, renewal workflow, scheduler, or background expiration job.

Duplicate creation and all-equivalent revocation

Two grant issuances are the same authorization source when all five of these are equal, with NULL equal only to NULL: user_id, permission_id, scope, source_kind, source_ref. expires_at, creation metadata, actor metadata, and the grant id do not distinguish them.

Creating a grant while an equivalent one is currently effective (revoked_at IS NULL and expires_at IS NULL OR expires_at > now()) fails with 409 {"code":"AUTHORIZATION_GRANT_ALREADY_EXISTS"} and Cache-Control: no-store. The response names no existing grant id, actor, scope, source reference, expiry, database detail, SQL, or stack — only the code. A rejected duplicate creates no grant and no audit row. A revoked or already-expired equivalent does not block a fresh issuance. This is a conflict rather than an idempotent return of the existing grant: the caller asked to record a new authorization with their own actor and expiry, and answering "done" while recording neither would hide that nothing happened.

Enforcement is database-backed, not process-local: every create and revoke touching one equivalence tuple takes a transaction-scoped pg_advisory_xact_lock keyed by a hash of that tuple, so multiple API processes serialize identically. Its single-key lock space is disjoint from the two-key space used by the migration runner and platform-owner bootstrap.

One revocation revokes every unrevoked equivalent, including expired-but-unrevoked legacy rows, in one transaction, writing one grant.revoke audit row per grant actually revoked, each carrying that row's own grant id. Grants with a different subject, permission, scope, source kind, or source reference are untouched. This replaces behavior where revoking one of two equivalent grants returned success while the other kept authorizing the same user. The public success response remains 200 {"success": true}.

Historical duplicates are not silently revoked by any migration. The store now prevents creating another effective equivalent, and an operator's explicit revocation resolves the whole equivalence class.

Response shapes

A grant, wherever returned:

{
  "id": "...",
  "userId": "...",
  "permission": "core.authorization.grants.manage",
  "scope": null,
  "sourceKind": "explicit",
  "createdAt": "2026-07-23T00:00:00.000Z",
  "expiresAt": null,
  "revokedAt": null
}

POST /api/v1/grants → 201 {"grant": {...}}. POST /api/v1/grants/:grantId/revoke → 200 {"success": true}.

GET /api/v1/grants → 200 {"grants": [...], "nextCursor": <string|null>}.

Listing is bounded keyset pagination

GET /api/v1/grants?userId=<id>&limit=<1..100>&cursor=<opaque>

Parameter	Required	Notes
userId	Yes	The subject whose explicit grants to list.
limit	No	Canonical base-10 integer, 1 through 100. Defaults to 50.
cursor	No	An opaque position string returned as a previous page's nextCursor.

Unknown query keys, duplicated userId/limit/cursor keys, empty values, non-canonical numbers (0, 101, -1, +5, 01, 1.5, 1e2, whitespace-padded), and malformed cursors are all rejected with 400 {"code":"INVALID_REQUEST"} before any store or database work. Permission enforcement completes first, so query validation never discloses a validation distinction — or the existence of any subject or grant — to an unauthorized caller.

Rows are ordered created_at DESC, id DESC, with a strict keyset boundary on both values, so grants sharing a created_at can neither be skipped nor repeated across pages. At most limit + 1 rows are read internally and at most limit returned. The final page returns "nextCursor": null. Relationship grants never appear on any page. There is no total count, offset pagination, search, or sorting option.

The cursor is a small versioned payload containing only the subject user id, boundary creation timestamp, and boundary grant id, encoded as canonical unpadded base64url. Decoding is strict — exact version, exact keys, a subject matching the requested one, a canonical toISOString() timestamp, a canonical GrantId, and a re-encoding that reproduces the supplied string byte-for-byte — and a rejected cursor reveals nothing it decoded.

It is a pagination position, not an authorization token. It is deliberately unsigned: the permission check and the server-side subject filter remain authoritative on every page, so a forged cursor can at most name a different position within grants the caller may already list. It carries no permission, scope, source reference, actor data, credential, or secret, and it is not a consistency snapshot — rows created or revoked between pages are seen or not seen according to the ordering.

Status-code mapping

Status	When
200 / 201	Success.
400 {"code":"INVALID_REQUEST"}	Malformed query/body/param (including a disallowed extra field).
403 {"code":"AUTHORIZATION_DENIED"}	Caller lacks core.authorization.grants.manage; the caller's own required permission id was unregistered; or, for POST /api/v1/grants, the well-formed target permission named in the body is not registered.
403 {"code":"UNTRUSTED_ORIGIN"}	Missing or untrusted Origin on a mutation route.
404 {"code":"AUTHORIZATION_GRANT_NOT_FOUND"}	The canonical grant id does not exist as an unrevoked explicit grant — this is also what revoking a real relationship-derived grant's id looks like, indistinguishably.
409 {"code":"AUTHORIZATION_GRANT_ALREADY_EXISTS"}	An equivalent grant is already currently effective for the subject.
503	Principal resolution, evaluation, or store failure (PRINCIPAL_RESOLUTION_FAILED, AUTHORIZATION_EVALUATION_FAILED, or an AuthorizationStoreError code).

Every response sets Cache-Control: no-store.

Shared database composition

apps/api/src/database.ts's createApiDatabase(databaseUrl, options) is the one place apps/api touches PostgreSQL connection setup. It constructs exactly one validated, timeout-bounded pg.Pool (createValidatedPostgresPool, @ekklesia/core/database) and wraps it with three independent typed Drizzle views — createIdentityDrizzleView(pool), createAuthorizationDrizzleView(pool), and createModuleDatabaseView(pool) (the ModuleDatabase facade module code queries through context.services.database; see API and Module development) — rather than three separate pools. close() on the returned handle is the only thing that ever calls pool.end(); no domain view owns or closes the pool itself. apps/api/src/server.ts calls createApiDatabase once during startApi(), builds the identity runtime from .identity, the AuthorizationStore/AuthorizationEvaluator from .authorization, and the frozen ModuleRuntimeServices from .module off that same handle, and closes the handle exactly once — on graceful shutdown, and on any startup failure after the pool was already constructed (mirroring the existing identity-only failure-cleanup path from API).

The existing per-domain factories (createIdentityDrizzleDatabase, createAuthorizationDrizzleDatabase), which each construct and own a private pool, are unaffected and remain available for standalone/testable use — the CLI bootstrap command uses createAuthorizationDrizzleDatabase directly for exactly that reason, since it only ever needs the authorization schema and has no reason to share a pool with anything else.

Both the shared pool and the per-domain factories accept the same optional transport policy — Ekklesia's single PostgreSQL TLS decision, described in API § Database transport (TLS). Every executable that opens a connection (the API, both migration commands, the platform-owner bootstrap CLI, and pnpm doctor) resolves it the same way, so production connections are TLS-verified regardless of which factory built them. Omitting it is a development/test affordance only: no ssl option is set and the connection is plaintext, which a production configuration refuses outright.

apps/api/tests/lifecycle.test.ts proves this composition explicitly: exactly one new Pool(...) call for a full startApi() lifecycle (covering identity, authorization, and module database usage), and exactly one pool.end() call regardless of how many concurrent shutdown() callers there are or whether the failure happened during startup.

Tests

• core/tests/authorization-permissions.test.ts — registry construction, duplicate-declaration rejection, fail-closed lookup, CORE_PERMISSION_REGISTRY/CORE_PERMISSION_DECLARATIONS contents and immutability.
• packages/contracts/tests/authorization.test.ts — PermissionDeclaration shape validation: every baseline, unknown-property/missing-baseline/invalid-baseline/blank-or-oversized-description rejection, malformed permission id rejection.
• packages/contracts/tests/module-manifest.test.ts — manifestVersion: 4 acceptance and version 2/string-array rejection, per-declaration validation (mirroring the contracts-level cases above), module-namespace enforcement, duplicate permission id at its exact permissions[index].id issue path, and default-empty-list isolation across separate parses.
• core/tests/module-permissions.test.ts — composePlatformPermissionRegistry: Core-only and Core-plus-modules composition, deterministic ordering, exact baseline/description preservation, duplicate rejection, reserved core.* namespace rejection, unknown-permission denial once a module is absent from the supplied list, and input/output immutability with no I/O.
• core/tests/database-pool.test.ts — shared pool validation/construction (mirrors the existing per-domain database tests).
• core/tests/authorization-bootstrap.test.ts — import safety, PlatformOwnerBootstrapError behavior.
• core/tests/cli-bootstrap-platform-owner.test.ts — the CLI never constructs a pool merely by being imported.
• core/tests/authorization-bootstrap.integration.ts (live PostgreSQL) — successful bootstrap and its audit row, nonexistent-user rollback, second-bootstrap rejection, a revoked historical owner still blocking reuse, and the concurrency scenario described above. Pass-3 scenarios add: exactly one consumed row plus one owner row plus one audit row in one transaction; deleting the owner's user row keeping the sentinel and the audit history while removing the live owner row and not reactivating bootstrap; a grant subject's create/revoke audit rows surviving that subject's deletion while their live grant cascades; a forced failure after the sentinel was claimed rolling everything back so the one first use is not burned; and migration 0002 backfilling consumed state for an active and a revoked existing owner while leaving a never-owned installation with exactly one first use.
• core/tests/authorization-store.integration.ts (live PostgreSQL) — pass-3 grant-integrity scenarios: sequential and six-way concurrent equivalent creation yielding exactly one grant and one create audit row; a differing expiry not bypassing the conflict; different subject/permission/ scope/source-kind/source-ref being non-equivalent; fresh issuance after a revoked or expired equivalent; one revocation clearing every unrevoked equivalent (including an expired-but-unrevoked one) with per-row audit entries while leaving differently scoped or sourced grants untouched; a concurrent create/revoke race serializing without an unexplained duplicate; a forced transaction failure rolling back every grant and audit change; and bounded keyset pagination traversing seven rows that all share one created_at with no duplicates or omissions.
• apps/api/tests/authorization-middleware.test.ts — allow/deny/unregistered/infrastructure- failure paths for requirePermission, and requireTrustedOrigin's allow/reject behavior; a dedicated block proves requirePermission enforces module-owned public/authenticated/ restricted declarations identically to Core's own, that the evaluator receives the registry's exact id/baseline, and that an absent module permission denies before principal resolution or evaluator access — the shared mechanism, not a module-specific code path.
• core/tests/doctor.test.ts / core/tests/doctor-report.test.ts — the modules.permissions check: passing count (including zero installed modules), reuse of the already-discovered module list (no rediscovery), distinct reserved-namespace vs. generic-duplicate failure detail, an unexpected-exception fallback to a fixed generic detail, schema/migration checks continuing independently after a composition failure, and the exact documented terminal report shape.
• apps/api/tests/grants-routes.test.ts — all three routes' allow/deny paths, disallowed-field and malformed-field rejection (including every form of a scope field), canonical-GrantId and canonical-future-expiresAt rejection with the store never called, the sanitized 409 conflict, the full pagination contract (default/boundary limits, non-canonical limit rejection, unknown and duplicated query keys, multi-page traversal, and every malformed or foreign cursor), a well-formed but unregistered target permission rejected with 403 and store.createGrant never called, exact server-derived-argument assertions, sourceKind-filtered revoke returning 404 for a non-explicit grant, origin enforcement, and store-error mapping.
• apps/api/tests/module-http-routes.test.ts — the production route-mounting contract: exact prefix/path joining (including the non-trailing-slash root case), every supported method, independent same-path routes across two modules, deterministic static-versus-parameterized resolution by published order, request/handler adaptation (raw Request, frozen context, decoded and frozen path parameters, empty frozen params for static/root routes), the registry's exact baseline for public/authenticated/restricted module permissions, denial/infrastructure- failure paths never invoking the handler, ModuleRouteConfigurationError for an inconsistent route/registry snapshot, trusted-origin enforcement for state-changing methods, CORS (advertised methods, no reflection for an untrusted origin, no authorization or handler invocation on preflight), fixed sanitized 500 for a throwing/rejecting/non-Response-returning handler with no raw leakage, a valid module-owned 4xx/5xx response passing through unchanged, forced Cache-Control/Vary with a handler's own Vary preserved, stripped module-supplied CORS headers, HEAD traversing the same permission middleware as GET with no body, and unmatched paths/zero-route/no-handler-during-construction behavior.
• apps/api/tests/lifecycle.test.ts — shared-pool-closes-exactly-once coverage described above, plus module-permission composition at startup: discovery/composition run before any pool, Better Auth instance, or HTTP server is constructed; a discovery failure, a reserved-namespace failure, and a generic duplicate-declaration failure each create zero pools/Better Auth instances/servers and remain distinguishable from one another; zero installed modules still succeeds; and import safety for platform-permissions.ts. A dedicated block covers module runtime startup and shutdown against real, working (non-throwing) fixture modules: permission composition and module runtime startup share the exact same discovered inventory, route mounting order matches discovery order, a load/register/start failure each close the pool exactly once and bind no server (preserving Core's own rollback for a start failure), an app-construction failure and a bind failure both stop the module runtime before closing the pool, normal shutdown stops the server/module-runtime/pool in that exact order, a module stop failure still runs pool cleanup before the shutdown rejection surfaces, and a direct discoveredModulesOverride drives both composition and runtime startup identically.
• core/tests/authorization-evaluator.test.ts — includes a malformed permission id denying under every baseline (public, authenticated, restricted), never just restricted, and never consulting the grant or platform-owner ports for public/authenticated.
• apps/api/tests/api-http.integration.ts (live PostgreSQL) — end-to-end: anonymous/non-owner denial, real bootstrap against the running API's own database, owner authority over the HTTP grants API, the owner delegating core.authorization.grants.manage to another account and that account then managing grants as a non-owner, a well-formed but unregistered target permission rejected with 403 (with no grant or audit row written) even for a caller with full grant-management authority, a create/list/revoke cycle with audit-trail verification (using core.authorization.grants.manage itself as the granted permission), disallowed-field rejection, missing/untrusted-origin rejection, a real relationship-derived grant proven invisible to LIST and unreachable by REVOKE, and an infrastructure-failure scenario (database dropped out from under a live session) asserting a 503 rather than a false 403 or 200.
• apps/api/tests/api-http.integration.ts, "installed module permission end-to-end" (live PostgreSQL) — a real temporary module fixture (structurally valid manifest; a real, working server entrypoint recording register/start/stop markers instead of throwing, and declaring a restricted GET /events/:eventId route plus a restricted, state-changing POST /events route against its own manifest permission) proves the complete chain against a real API process and real Doctor run: Doctor recognizes the installed module and passes permission composition without executing its entrypoint (an empty marker set); a successful startApi() discovers the same fixture, imports/registers/starts its runtime exactly once, and makes its declared permission enforceable; a bootstrapped platform owner creates a real explicit grant for the module's permission with an atomic audit row recorded under its exact declared id; the module's own mounted route denies an anonymous caller and then serves the newly granted owner with the decoded path parameter and authenticated principal; the module's state-changing route requires a trusted Origin exactly like the platform's own mutation routes; a well-formed permission absent from the composed registry is rejected with no grant or audit write; and, after a clean shutdown (module stop invoked exactly once) and restart with the module absent, its former route returns the ordinary 404, the same permission can no longer be newly granted, and the earlier, still-unrevoked grant remains listed — with ordinary identity HTTP behavior (sign-up, sign-in, principal resolution) unaffected throughout.

Run the focused suites:

pnpm --filter @ekklesia/core test
pnpm --filter @ekklesia/api test
pnpm test:migrations:integration   # requires Docker

Next phase

Module HTTP route registration and production mounting (see ADR 0014 for the deferral this closes) is now implemented — see API and Module development. Deliberately deferred, unrelated slices remain: a module runtime-context authorization service, relationship-derived grant creation itself (a roster or similar module operation, not the generic grants API), reinstall/reactivation policy for a permission whose owning module was previously removed, roles/groups/teams/tenants, an owner-management HTTP endpoint, and service accounts/API tokens for non-browser automation.