Enterprise Identity & Audit-Logging Inventory
Verified inventory of SAML, OIDC, SCIM, WebAuthn and audit-logging behaviour in the source on `next` (post-PR #566). Every row carries path:line; unimplemented capabilities are recorded as 'not found' with the search that proves it.
On this page
Status: current audit of
next. The baseline snapshot wasefa57c55eon the (now deleted)fix/enterprise-readiness-20261110branch; PR #566 is merged and every row below was re-verified against the current tree, so a reference that moved cites the symbol. Per-finding status is listed below — entries marked fixed are no longer current defects and must not be read as open risks. Rows without a reachable call site keep theirnot foundresult and the search behind it. Re-verified 2026-11-11 on the current tree: the status and evidence of every row this pass touched were checked against the code again; references that moved cite the symbol.
| Finding | Status | Evidence |
|---|---|---|
SCIM active:false did not deprovision (wrote isActive; auth enforces blocked) |
Fixed — PATCH, PUT and POST now set isActive and blocked; non-boolean values are rejected with 400 |
parseScimActive() (src/utils/scim-utils.ts:167), applyScimPatchOps (:180, :199-205), PUT (handlers/scim.ts:229-233), POST (:164-190); tests/unit/api/scim-deprovisioning.test.ts, tests/unit/api/scim-put-endpoint.test.ts |
| IdP-driven logout deleted SSO metadata only; the session stayed valid | Fixed — all three OIDC paths now call terminateSessionAndMetadata() |
terminateSessionAndMetadata() (src/databases/auth/sso-session.ts:487); tests/unit/auth/sso-logout-revocation.test.ts |
IdP-driven logout could only find sessions created on the same worker (ssoSessionMetadata is a process-local map) |
Fixed — every logout lookup now resolves through a distributed index + directory | Directory sso:directory:<sessionId> and buckets sso:index:{sid,sub,nameid}:<value> written by setSsoSessionMetadata() (sso-session.ts:535, helpers :215-297), read by collectIndexedSsoSessions() (:299), used by front-channel (:1247), back-channel (:1316) and findSamlSessions() (:601). resetSsoSessions() (:655) exists so tests can model a second node. Cross-node tests: tests/unit/auth/sso-distributed-logout.test.ts (8: two users, two sessions per user, two tenants sharing a provider id, SAML NameID, index pruning, cache-down fallback). |
Legacy /login/oauth callback does not verify state (login CSRF) — and the OIDC login start accepted any redirect_uri |
Partially fixed in PR #566 (commit 5dafc3a5d): redirect_uri is same-origin only |
handleOidcLoginStart in handlers/auth.ts; tests/unit/auth/oidc-login-redirect.test.ts (3 of 4 cases fail on the old handler). The legacy /login/oauth state check is still open. |
| Audit chain verifier read a hash nothing wrote; buffered writes lost rows on crash/outage | Fixed — writer and verifier share one canonical payload; fork guard + tail checkpoint, with limits | canonicalAuditPayload() / computeChainHash() (src/services/audit-chain.ts:224, :246); auditChainService.append() (:299); chainScope = tenantId ?? "__global__" (chainHashScope, :55); unique (chainScope, previousHash) index (system-schema-spec.ts:1350-1367); re-chain after a rejected fork (audit-service.ts:38, :352-400); checkpoint (audit-chain.ts:432-483, verified :628-660). Tests: tests/unit/services/audit-chain-writer.test.ts, tests/integration/databases/audit-chain-integrity.test.ts, audit-chain-fork-guard.test.ts. Limits: the DB checkpoint is written after the batch and audit rows are not in the business transaction. An independent anchor is now written outside the database (AUDIT_ANCHOR_PATH → audit-anchor-<scope>.json, atomic temp→rename) and compared by verifyChain().anchor, so deleting the tail and the DB checkpoint row no longer verifies; anchor write failures are counted (getAnchorStatus()) and logged as ANCHOR_WRITE_FAILED. A failed checkpoint write is no longer silent: it is counted per tenant, logged at error level (CHECKPOINT_WRITE_FAILED), reported by getCheckpointStatus() (src/services/audit-chain.ts) and returned by the admin verification as checkpointPersistence (tests/unit/services/audit-chain-checkpoint-health.test.ts); it is still not retried, and the file fallback is not the verified ledger. |
OIDC redirect_uri unvalidated at login; no at_hash; SAML signature checks are library-side and mocked in tests |
Partially fixed — redirect_uri same-origin; SAML IdP-initiated + SLO shipped; OIDC logout now scoped |
redirect_uri same-origin http(s) only; SAML_ALLOW_IDP_INITIATED (saml-auth.ts:120), handleSAMLACS (:585), handleSAMLLogoutRequest (:773), SLO route (handlers/auth.ts:1667); OIDC sid/sub stored at login (handlers/auth.ts:900-913) and matched on logout (sso-session.ts:1247 front, :1316 back); tests/unit/auth/saml-slo.test.ts, sso-logout-scoping.test.ts. Still open: at_hash validation, a signed LogoutResponse requires SAML_SP_PRIVATE_KEY, and SAML logout session lookup is distributed (local map + shared index). |
This page records what the source does, not what a settings screen or a prose doc promises. It was
built by reading the tree on next. Every statement below is a source citation, and every
“not found” is paired with the exact search that returned nothing.
Status vocabulary:
| Status | Meaning |
|---|---|
| Implemented + tested | Code path exists and a named test exercises the behaviour described. |
| Implemented, no test | Code path exists; no test asserts it. |
| Partial | Code exists but a named part of the contract (validation, revocation, persistence) is absent. |
| Not found | No code path exists. The search that proves it is given. |
The repository has an HTTP-mutation log and a separate database audit service. They are not the same thing and neither is transactional. Read §3 before relying on either for compliance evidence.
1. Identity protocols
1.1 Protocol capability table
| Capability | Status | Implementation (file:line) | Validation enforced | Existing tests | Gap |
|---|---|---|---|---|---|
| SAML 2.0 SP (SP-initiated) | Implemented + tested (state/InResponseTo), signature delegated and untested | Route dispatch src/routes/api/[...path]/handlers/auth.ts:1652-1745 (handleSAMLRoutes; case "acs" :1661, login :1697, config :1733); engine src/databases/auth/saml-auth.ts (handleSAMLACS :585, processSAMLResponse :468, getSAMLConfig :130); library @node-saml/node-saml saml-auth.ts:12 (^5.2.0, package.json:210) |
CSRF state bound to a 5-min HttpOnly cookie; InResponseTo compared to a one-time cookie (assertInResponseTo); response processed by validatePostResponseAsync (saml-auth.ts:491) with idpCert (buildSAMLOptions :177), wantAssertionsSigned: true (:153, :178) and signatureAlgorithm: "sha256" (:179); audience falls back to the SP issuer in node-saml; assertion NotOnOrAfter enforced by the library |
tests/unit/api/saml.test.ts:81-259 (config gate, state mismatch 403, missing request-id cookie 403, InResponseTo mismatch 403); tests/unit/auth/sso/saml.test.ts:125-259 (metadata parse, XXE/DOCTYPE reject, tenant binding, InResponseTo); tests/unit/auth/sso/saml-sso-federation.test.ts:30-48 |
No test performs real XML-DSig verification: the library is mocked at tests/unit/api/saml.test.ts:13 and tests/unit/auth/sso/saml.test.ts:59. A broken signature path, an expired assertion, or a wrong audience would not fail CI. |
| SAML IdP-initiated | Implemented + tested (opt-in) | isIdpInitiatedLoginAllowed() src/databases/auth/saml-auth.ts:120; handleSAMLACS :585: an unsolicited response is accepted only when SAML_ALLOW_IDP_INITIATED (or its samlAllowIdpInitiated alias) is on; a response carrying InResponseTo is never treated as unsolicited (:468-483) |
Default is off (fail closed); enabling it is an explicit operator decision. CSRF flow cookies still win when present (:617-627). |
tests/unit/auth/saml-slo.test.ts:115-215 (accept when on, alias, reject when off, reject InResponseTo, reject foreign flow cookie) |
Opt-in only; an unsolicited assertion is not bound to a browser that started a login, so it is not enabled by default. |
| SAML Single Logout (SLO) | Implemented + tested (IdP-initiated request + SP-initiated response) | handleSAMLLogoutRequest() src/databases/auth/saml-auth.ts:773; handleSAMLSLOEvent() :981 (GET Redirect + POST binding); route case "slo" src/routes/api/[...path]/handlers/auth.ts:1667-1669 |
node-saml validates the LogoutRequest signature/issuer/NotBefore before anything is destroyed; sessions are located by NameID (+ SessionIndex when present) and removed via terminateSessionAndMetadata() (:829-833); a request matching no session still succeeds (idempotent). |
tests/unit/auth/saml-slo.test.ts:217-430 (bad signature destroys nothing, session + metadata deleted, missing/foreign SessionIndex, no match, SP-initiated request, both bindings) |
findSamlSessions() (sso-session.ts:601) resolves candidates through collectIndexedSsoSessions() (:299) — the local map plus the shared sso:index:nameid:<NameID> bucket — so a logout handled by another node still finds the session (tests/unit/auth/sso-distributed-logout.test.ts). A signed LogoutResponse requires SAML_SP_PRIVATE_KEY. |
| OpenID Connect (authorization code + PKCE) | Implemented + tested | Start handlers/auth.ts:626-674; callback :679-911; engine src/databases/auth/sso-session.ts:850-951; JWT verify :737-845 |
state cookie compare handlers/auth.ts:704-706; nonce compare :731-737; PKCE S256 :640-642,664 + sso-session.ts:105-112; exact iss sso-session.ts:756-758; asymmetric-alg allowlist :761-764; exp :768-773; nbf :774-781; iat :782-789; aud contains clientId :795-802; azp :803-805; strict kid select :820-828; JWKS mandatory (fail closed) :811-816; signature :836-839; discovery/JWKS fetch egress-guarded :551,611,914 |
tests/unit/auth/sso/oidc-jwt-verification.test.ts:141-345 (RS256, ES256, PS256, expired, aud, azp, HS256 reject, missing kid, wrong key, nbf, iat); tests/unit/auth/sso/oidc-pkce-jit.test.ts:19-172; tests/unit/auth/sso-session.test.ts:328-416; tests/integration/auth/oidc-live.test.ts:154-340 (live server: PKCE mismatch, wrong redirect URI, tampered signature) |
at_hash / c_hash are not checked: grep -rn 'at_hash\|c_hash' src/ → 0 hits. Not required for the code flow, but it means an id_token is never bound to the access token. |
OIDC redirect_uri validation at login |
Implemented (same-origin) | handleOidcLoginStart handlers/auth.ts:640-655: the requested URI must be a valid http(s) URL with the same origin as the request, and it is then stored in the authorize request and the state cookie |
A foreign origin or non-http(s) scheme is rejected with 400 (:644-655). This is not a per-client registered-URI allowlist: the post-logout allowlist (sso-session.ts, allowedRedirectUris) is not consulted here. |
tests/unit/auth/oidc-login-redirect.test.ts (3 of 4 cases fail on the old handler) |
OIDC Core 3.1.2.3 asks the RP to require a pre-registered redirect URI; same-origin is a meaningful narrowing (no open redirect to a foreign host) but not that check. The code exchange reuses the cookie value, so PKCE still protects the code. |
| OIDC RP-initiated logout | Implemented + tested | handlers/auth.ts:540-620; sso-session.ts:450-527 |
Redirect allowlist sso-session.ts:389-416; local session terminated handlers/auth.ts:578-580 |
tests/unit/auth/sso-session.test.ts:363-416 |
— |
| OIDC back-channel logout | Implemented + tested (session-terminating, scoped) | handleBackChannelLogout() src/databases/auth/sso-session.ts:1316; route src/routes/api/[...path]/handlers/auth.ts:192-193; sid/sub recorded at login :900-913 |
iss; event claim; iat ≤ 300 s; nonce forbidden; aud; JWKS signature. The token’s sid narrows the match, falling back to sub only when no sid is present; a token that names no local session terminates nothing (:1199-1226). |
tests/unit/auth/sso-logout-scoping.test.ts:154-229 (sub-only, sid preferred, no match); tests/integration/auth/oidc-live.test.ts:292-340 |
The session store walked is in-process, so a logout delivered to another node does not terminate sessions created here. Session validation now reads the distributed user/tenant/global markers (handle-authentication.ts:305-321, :586-592), but a single-session logout publishes none of them, so a peer’s warm LRU entry survives until it re-validates. |
| OIDC front-channel logout | Implemented + tested (session-terminating, scoped) | handleFrontChannelLogout() src/databases/auth/sso-session.ts:1247; route src/routes/api/[...path]/handlers/auth.ts:190-191; oidcSid recorded at login :900-911 |
Matches the OP sid against the session’s stored oidcSid; an unmatched (or absent) sid terminates nothing, and the response is still 200 so the OP iframe completes (:1061-1068). |
tests/unit/auth/sso-logout-scoping.test.ts:123-154 (matching sid only; unmatched terminates nothing) |
The endpoint is unauthenticated (the message is unsigned by spec) and walks the in-process session map, so a logout on another node can miss sessions here. |
SCIM 2.0 (/api/scim/v2) |
Implemented + tested (CRUD + deprovisioning) | Dispatcher src/routes/api/[...path]/handlers/scim.ts:34-79; map +server.ts:70,108; Users scim.ts:124-277; Groups :220-277; Bulk :288-432 |
Bearer token looked up hashed (src/databases/core/relational-auth.ts:1635-1652 — expiry and consumed filtered in the query), type === "scim", not blocked, not expired src/utils/scim-utils.ts:273-282; tenant taken from the token scim.ts:50; admin-session fallback scim-utils.ts:294-300. active:false sets both isActive and blocked on PATCH/PUT/POST; non-boolean values are rejected 400 (parseScimActive, scim-utils.ts:167). |
tests/unit/api/scim.test.ts:113-287; scim-deprovisioning.test.ts; scim-put-endpoint.test.ts; tests/unit/services/scim.test.ts:25-249 |
active:false now cuts access because blocked is what authentication reads (handle-authentication.ts), but SCIM still deletes no session row, so an already-issued session survives its expiry window. No audit record is written. externalId is stored but never used for lookup, so an IdP rename can create a second user (dedupe is by email only). |
| WebAuthn / passkeys | Implemented + tested | src/databases/auth/webauthn/webauthn-service.ts:213-302; login src/routes/login/auth.remote.ts:1025-1165 |
Challenge equality + type webauthn-service.ts:174-187; origin :188-190; RP-ID hash :98-101,265-267; assertion signature :274-282; user-present flag :288; sign-counter replay :290-292; challenge single-use + 300 s TTL :20-46 |
tests/unit/auth/webauthn/webauthn-ceremony.test.ts:73-88 (foreign origin, no-UP flag, happy path); webauthn-attestation.test.ts; webauthn-cbor.test.ts |
User verification is not required (userVerification: "preferred" :119,148; only the UP bit is asserted :288), so a bare touch satisfies a passkey login. Attestation is not verified (attestation: "none" :117; credentialDeviceType hard-coded "multi-device" :243). The login path emits no audit record, is not rate-limited, and creates a session without tenantId (auth.remote.ts:1144-1147). |
Legacy Google / GitHub OAuth (/login/oauth) |
Implemented, state unverified | src/routes/login/oauth/+page.server.ts:390-620; state signing src/databases/auth/google-auth.ts:72-112; redirect URI google-auth.ts:22-31 |
Google branch exchanges the code server-side +page.server.ts:538-566; GitHub branch :480-525 |
tests/unit/auth/google-auth.test.ts:54-96 tests the signer as a helper; tests/e2e/routes/login/oauth.spec.ts:44-151 covers the mock happy path |
The callback never calls verifyOAuthState: it reads state and decodeURIComponents it into the registration token (+page.server.ts:410-418). The verifier exists (google-auth.ts:83-112, constant-time at :107) and is used on the other callback (src/routes/login/+page.server.ts:357-361), but getOAuthRedirectUri() points the OAuth client at /login/oauth (google-auth.ts:26-30, consumed at +page.server.ts:539). email_verified is never inspected: grep -rn 'email_verified\|verified_email' src/ → 0 hits. Account is selected by email alone (+page.server.ts:196), with no tenant argument. |
| LDAP / Active Directory bind | Not found | — | — | — | grep -rn -i 'ldap' src/ --include=*.ts --include=*.svelte --include=*.mdx → 12 hits across four files, all LDAP-injection scanners (src/services/security/threat-scan.ts:592-637, src/services/security/auth-guard.ts:83-84, src/services/security/response-service.ts:462-463, src/services/security/types.ts:37). No bind, no directory client, no ldapjs/ldapts dependency in package.json. |
1.2 Account linking and identity collision
| Capability | Status | Implementation (file:line) | Validation enforced | Existing tests | Gap |
|---|---|---|---|---|---|
| SAML identity binding | Partial | Lookup by email + tenant src/databases/auth/saml-auth.ts:441-446; JIT create :448-461 |
Tenant scoping on the lookup and the insert :440,456 |
tests/unit/auth/sso/saml.test.ts:210-223 (tenant binding) |
The stable subject (NameID, samlId) is never persisted or matched. samlId exists in the model (src/databases/auth/types.ts:66) and getUserBySamlId exists (auth-service.ts:237-243), but the JIT insert at saml-auth.ts:449-457 sets no samlId and the lookup at :441 uses email. Renaming a user in the IdP orphans the local account; an IdP that asserts admin@… logs into the existing local admin. No domain allowlist, no SAML_JIT_PROVISIONING role mapping (role is hard-coded "user" :454, unlike OIDC handlers/auth.ts:750). |
| OIDC identity binding | Partial | Lookup by email + tenant handlers/auth.ts:752; JIT :760-787; SSO metadata sso-session.ts:328-347 |
Tenant scoping :752; role mapping resolveJitRole :750, sso-session.ts:114-137 |
tests/unit/auth/sso/oidc-pkce-jit.test.ts:46-134 (role mapping only) |
Account linking still uses email only: the id token’s sub is now persisted as oidcSubject on the session metadata (handlers/auth.ts:900-913) but is used for logout scoping, not for lookup. Linking is therefore still email-based, so a subject change in the IdP re-links by whatever email it now asserts. |
| Legacy OAuth linking | Partial | +page.server.ts:196 (auth.checkUser({ email }), no tenant), :204-241 (existing user → login) |
None beyond the TLS-verified userinfo response | e2e mock only tests/e2e/routes/login/oauth.spec.ts:81-151 |
First registered user becomes admin (+page.server.ts:245-270); the same email arriving from GitHub (:518-525) or Google is linked without an email_verified check. |
| Collision handling on create | Partial | SCIM scim.ts:159-160 (409 on duplicate email); OIDC handlers/auth.ts:784-786 (500 on unique-constraint failure) |
Email uniqueness only | tests/unit/api/scim.test.ts:140-172 |
No cross-provider collision policy: two IdPs asserting one email both authenticate into the same row, and externalId/sub are not consulted. |
1.3 Session termination and revocation
| Capability | Status | Implementation (file:line) | Validation enforced | Existing tests | Gap |
|---|---|---|---|---|---|
| Local logout | Implemented + tested | handlers/auth.ts:512-528; src/services/sdk/namespaces/auth-namespace.ts:634-647 |
Row deleted via the DB auth adapter auth-namespace.ts:646; SSO metadata removed :640-641; cookie cleared with matching attributes handlers/auth.ts:524, src/utils/security/session-cookie.ts |
tests/unit/routes/cookie-logout.test.ts:9-36 |
Logout terminates one session. invalidateAllUserSessions is not called on this path: grep -rn 'invalidateAllUserSessions' src/ → 15 hits, none in handleLogout. Other devices stay signed in by design. |
| Global revocation primitive | Implemented + tested | src/databases/auth/auth-service.ts:1191-1227 (publish revocation epoch, drop L0/L1/L2) |
Epoch + distributed marker :1198-1201; DB sweep :1224. Read path: warm LRU handle-authentication.ts:305-321 (call :586-592); store session-manager.ts:53-64 |
tests/unit/auth/revocation-epoch.test.ts; revocation-read-path.test.ts (8); revocation-distributed.test.ts (6) |
publishUserRevocation() is wired to password change (src/utils/server/user-attribute-update.server.ts:188-198 → auth-service.ts:1198), OIDC role sync (handlers/auth.ts:827) and SDK attribute updates (auth-namespace.ts:416). The hook’s session:<tenant>:<id> snapshot branch now runs the shared gate (handle-authentication.ts:709-726, gate :715) and the warm-LRU path calls it at :613, so a user/tenant/global marker is honoured on every session read. Remaining limits: publishGlobalRevocation()/publishTenantRevocation() still have no production caller (nothing in src/ triggers a tenant lockout or “log out everywhere” yet); the session store’s get/getSessionData branches check tenant/global but not the distributed user marker; the memo admits up to 5 s of staleness (revocation-publish.server.ts:34). |
| Deprovisioning: block / delete user | Implemented, no direct test | auth-namespace.ts:848-917 (batchAction delete/block/unblock) |
Cache purges per affected user :884-913; enforcement is per-request in src/hooks/handle-authentication.ts:704-707, :763-766 |
No test calls cms.auth.batchAction; grep -rn 'batchAction' tests → 6 textual hits (mock stubs/comments) |
Relational deleteUsers removes session + token rows in one transaction (src/databases/core/relational-auth.ts:795-830); Mongo deleteUsers deletes the user docs only (src/databases/mongodb/auth-user.ts:391-414), leaving session rows. block/unblock purge caches only (:884-913) and depend on the DB blocked check on the next request. publishUserRevocation is not called by batchAction. No audit record (grep -n -i audit auth-namespace.ts → role paths :713, :753, :770, :808 only). |
| Deprovisioning: SCIM | Implemented, still no session kill | scim.ts:164-190 (POST), :229-233 (PUT), scim-utils.ts:199-205 (PATCH) |
isActive: false and blocked: true, which authentication enforces; non-boolean rejected 400 |
tests/unit/api/scim-deprovisioning.test.ts; scim-put-endpoint.test.ts |
No session revocation and no audit: an already-issued session survives until expiry even though the account is now blocked. |
| Deprovisioning: IdP-initiated | Implemented (distributed) | SAML SLO saml-auth.ts:773; OIDC front sso-session.ts:1247; back :1316 |
All three call terminateSessionAndMetadata(), which deletes the session row; candidates come from the local map plus the distributed index (collectIndexedSsoSessions, :299) |
tests/unit/auth/saml-slo.test.ts; sso-logout-scoping.test.ts; sso-logout-revocation.test.ts; sso-distributed-logout.test.ts |
Cross-node lookup is solved; two limits remain. A single-session logout publishes no revocation marker, so a peer’s warm session cache entry is only dropped when it re-validates. And the index is a cache write: when the cache is unavailable the lookup silently degrades to the local map (logged at debug, never thrown). |
2. Where the identity code is NOT wired to audit
The audit subsystem is reachable only by explicit calls. These four modules contain zero audit
references (case-insensitive grep -n -i 'audit' on each file returned no lines):
| Module | Consequence |
|---|---|
src/routes/api/[...path]/handlers/auth.ts |
API login, logout, OIDC login/logout, SAML SSO, 2FA, role updates are unaudited. |
src/routes/api/[...path]/handlers/scim.ts |
Every SCIM provision/deprovision/deprovision-failure is unaudited. |
src/databases/auth/saml-auth.ts |
SSO authentication, JIT provisioning and IdP reconfiguration are unaudited. |
src/databases/auth/sso-session.ts |
Provider configuration changes, RP-initiated, front- and back-channel logout are unaudited. |
The only identity audit call sites in the tree are:
- form login —
src/routes/login/auth.remote.ts:648-659(fire-and-forget,.catch()swallows the failure) - role matrix write —
src/services/sdk/namespaces/auth-namespace.ts:808-816 - individual role create/update —
auth-namespace.ts:770-784 - GDPR erasure —
src/services/security/gdpr-service.ts:241-248
src/routes/login/auth.remote.ts:1081-1165 (passkey login) contains no audit call either.
3. Audit logging
3.1 What exists (two independent pipelines)
| Capability | Status | Implementation (file:line) | Validation enforced | Existing tests | Gap |
|---|---|---|---|---|---|
| HTTP mutation log (pipeline 1) | Implemented + tested, not persisted | src/hooks/handle-audit-logging.ts:153-189; queue :44-45; drain :64-110 |
Fields: timestamp, method, path, status, userId, tenantId, ip, durationMs, success, error :31-42,137-150; actor = locals.user._id ?? "anonymous" :131 |
tests/unit/hooks/audit-logging.test.ts:38-199 (success entry, failure entry, test-bypass, env-disable, non-API/read pass-through, thrown-status capture) |
Successful mutations are only emitted at debug level (:80-83); production runs at error, so successes are discarded. Nothing is written to a database — the sink is logger (:80-86); only a leaf hash feeds the in-memory rollingMerkleAccumulator (:88-94, rolling-merkle.ts, never persisted). The ring buffer drops the oldest record past 2048 entries (:44,104-107). No target object, no before/after, no actor email. |
| Audit records (pipeline 2) | Implemented + tested | src/services/security/audit-service.ts — buffer (:185), log() (:671), _flushOnce() (:329), insertBatch() (:476) |
Full field set on AuditLogEntry; chainScope set from chainHashScope() (:710); secret redaction (redactSensitiveDetails, :54, applied :708) |
tests/unit/services/audit-service-redaction.test.ts:22-96; tests/unit/services/audit-log-microbatch.test.ts:16-85; tests/unit/security/audit-chain-verification.test.ts:18-149; tests/unit/security/audit-ledger-integrity.test.ts:13-109; tests/unit/auth/roles/rbac-audit-trail.test.ts:11-17 |
Atomicity: none. insertBatch() calls a standalone crud.insertMany (:476-495) with no transaction, although the adapter exposes one. Durability by default is a 5 s flush / 15 ms micro-batch (:202-203) and the default flag is chainSync: false (src/utils/security/audit-flags.ts:44). |
| Durability on DB outage | Partial | _flushOnce catch audit-service.ts:405-428; file sink :572-575 |
Buffer restored and capped at maxTotalBuffer = 200 (:201, :421-426) |
tests/unit/services/audit-log-microbatch.test.ts:85 (500 concurrent writes, no drop — with a healthy mock DB) |
On a failed write the batch is appended to logs/audit.log and re-buffered; once the restore exceeds 200 entries the excess is dropped with only a logger.warn (recordDroppedEntries, :236, :421-426). Neither sink fsyncs. |
| Durability on crash | Partial | log() returns after enqueue unless chainSync (audit-service.ts:745-760) |
— | none | A SIGKILL/OOM between push (:745) and the flush loses every buffered entry — up to maxBufferSize = 50 on the healthy path (:200, :770) and maxTotalBuffer = 200 when flushes are failing. AUDIT_CHAIN_SYNC=true closes this window but is off by default (audit-flags.ts:44). |
| Tamper-evidence (writes) | Implemented + tested (with limits) | AuditChainService.append() (src/services/audit-chain.ts:299): reads the tenant head under a per-tenant in-process lock, hashes with computeChainHash() (:246) over canonicalAuditPayload() (:224), writes chainScope = tenantId ?? "__global__" (:55). audit-service.ts:721-731 persists chainHash (+ hash alias). |
One SHA-256 per entry; previousHash comes from the DB head (cached per tenant), not process genesis. |
tests/unit/services/audit-chain-writer.test.ts; tests/integration/databases/audit-chain-integrity.test.ts; audit-chain-fork-guard.test.ts |
withTenantLock serialises writers inside one process. Multi-process forks are rejected by the unique (chainScope, previousHash) index (system-schema-spec.ts:1350-1367; Mongo partial index in mongodb/adapter-core.ts), and the flush re-chains onto the winning head (3 tries). Pre-existing unchained rows stay unverifiable. |
| Tamper-evidence (verification) | Implemented + tested | src/services/audit-chain.ts:502-680 (verifyChain); exposed via src/routes/(app)/audit-history.remote.ts |
Writer and verifier call the same canonicalAuditPayload()/computeChainHash() — no parallel field list. verifyChain also checks the tail checkpoint (:628-660) and accepts second-precision timestamps for legacy rows, counted as legacySecondPrecisionEntries (:118, :621). |
tests/integration/databases/audit-chain-integrity.test.ts (7 tests, 4 adapters); tests/unit/services/audit-chain-writer.test.ts |
Truncation is signalled by the checkpoint head hash (counts are per-writer lower bounds) when the head is no longer in the chain. Limitation: the DB checkpoint itself is not an independent anchor — but the out-of-DB anchor (AUDIT_ANCHOR_PATH) is compared on every verification (#, above), so deleting the tail and the DB checkpoint row is now detected as long as the anchor file survives. A checkpoint that could not be written is reported as checkpointPersistence (unhealthy, with the last error) instead of the row silently ageing; it is not retried. Pinned by tests/unit/services/audit-chain-checkpoint-health.test.ts. |
| Field mapping to schema | Partial | Service still writes actorIp and a hash alias (audit-service.ts:700-731); schema declares ipAddress (system-schema-spec.ts:1277), chainHash (:1298) and chainScope (:1305); sessionId, userAgent, correlationId, errorDetails are schema columns the service does not set |
— | none | No shared mapping layer exists between AuditLogEntry and the table: actorIp and hash have no declared column of the same name, so whether the IP survives a relational insert depends on adapter behaviour rather than a schema contract. The chain fields (previousHash, chainHash, chainScope) are written and projected correctly. |
| Immutability vs. a normal admin | Partial — no guard | No append-only enforcement found for auditLogs; the only collection-level protection is the plugin sandbox src/plugins/sandbox.ts:52-60,93 |
— | none | Two internal paths delete audit rows: GDPR erasure gdpr-service.ts:202-207 (deleteMany("audit_logs", { actorId: userId })) and tenant deletion src/services/core/tenant-service.ts:119-126. No DB trigger protects the table — the only triggers in the schema are on content_nodes (system-schema-spec.ts:1730-1737). Search for an immutability guard: grep -rn 'immutable\|appendOnly\|protectedCollections' src/ --include=*.ts → hits are HTTP cache headers, widget definitions and comments only. |
| Audit suppression switch | Implemented, no test | audit-flags.ts:52-84 (DISABLE_AUDIT_LOGS, AUDIT_CHAIN_SYNC resolved from env or DB settings, 5 s TTL) |
— | tests/unit/hooks/audit-logging.test.ts:113 covers the env path only |
Whoever can write those system-settings keys can silence both pipelines at runtime without a deploy, and the resolution module performs no permission check of its own (audit-flags.ts:69-83). |
| Read access to audit records | Partial | handlers/dashboard.ts:591-600 (case "audit"), :548-589 (case "logs"); src/routes/(app)/audit-history.remote.ts:9-46 |
Dashboard gate is dashboard:read or system:dashboard (+server.ts:346-367) |
tests/unit/api/dashboard-license-gate.test.ts (license only) |
Not admin-only: every seeded editor role carries system:dashboard (+server.ts:346-350). audit-history.remote.ts:23 uses getAuthenticatedUser, not an admin guard, and returns rows for any targetId in the tenant (:27-31). |
| Event coverage | Partial | Enum audit-service.ts:55-84 |
— | — | Emitted: USER_LOGIN (form path only), ROLE_MUTATED, ROLE_DELETED, USER_ROLE_CHANGED, DATA_DELETION, DATA_EXPORT, UNAUTHORIZED_ACCESS (1 hit each). Never emitted: USER_LOGOUT, USER_CREATED, USER_DELETED, USER_UPDATED, PASSWORD_CHANGE, USER_LOGIN_FAILED, PRIVILEGE_ESCALATION, API_KEY_CREATED — grep -rn "AuditEventType.<NAME>" src/ outside audit-service.ts returns 0 for each of those eight. |
| Per-row data audit | Implemented, off by default | audit-service.ts:281-347 (registerHook) |
— | none | Guarded by AUDIT_INSERT_HOOK !== "true" (:317-321), so by default a content insert writes no audit row; the only coverage is the non-persistent HTTP log (§3.1 row 1). |
Documented limitations of the chained ledger (visible on purpose — do not read the chain as a complete evidence store):
- The checkpoint shares the audit database.
systemPreferencesis tenant-scoped KV in the same database asaudit_logs. An actor who deletes the tail and the checkpoint row leaves a shorter chain that still verifies; evasion needs a store outside the CMS database (see the roadmap’s external-anchoring candidate). - The checkpoint is written after the batch, not with it. A checkpoint failure only warns (
AuditChainService.commitCheckpoint,audit-chain.ts:462-483); the entries are already durable, so this under-counts rather than losing data, and verification treats extra entries as a normal append. - Audit rows are not committed in the business transaction. The audit insert is a standalone
crud.insertMany(audit-service.ts:476-495), sequenced after the mutation; a crash between the two can leave a business change without its audit row. - The file fallback is not the verified ledger. When the audit table is unavailable the batch is appended to
logs/audit.log(audit-service.ts:569-575) and counted as dropped from the chain; those entries can never be verified or linked.
3.2 The outbox is not an outbox
src/services/outbox/outbox-service.ts:18-24 states it directly: “This is NOT a transactional outbox
— there is no atomic same-transaction write… a crash between the content commit and the queue flush can
drop the delivery intent.” Enqueue is a synchronous in-memory push with a 25 ms debounce
(:27-28,506-521). Audit records do not use it at all; they use the separate buffer in
audit-service.ts.
Net answer to “is the audit write atomic with the business change?” No. Three independent writes
exist — the business row, the buffered audit row (audit-service.ts:233-238), the buffered outbox
event (outbox-service.ts:213) — sequenced after the fact, each with its own failure mode.
4. The five highest-value tests for data safety
Each was written as a failing-then-passing case. T1, T2 and T4 have since shipped — their “expected” behaviour is now the implementation and is pinned by the named tests. T3 and T5 remain open.
T1 — Deprovisioning must kill the session, not just the metadata
- Precondition: OIDC provider configured with a JWKS; user
u1with one active sessionsess-1(setSsoSessionMetadata("sess-1", {provider: "p1"}, "tenant-a"));db.auth.validateSession("sess-1")returnsu1. - Action:
handleBackChannelLogout(validLogoutToken)then re-validatesess-1; repeat the same sequence withhandleFrontChannelLogout(issuer, sid)for the front-channel endpoint. - Expected behaviour (now implemented):
terminateSessionAndMetadata()deletes the session row through the adapter, invalidates the in-process session cache and drops the SSO metadata, sovalidateSession("sess-1")returnsnulland the next API call is 401. Pinned bytests/unit/auth/sso-logout-revocation.test.tsandtests/unit/auth/sso-logout-scoping.test.ts. - Residual gap: the lookup is distributed (local map + the shared
sso:index/sso:directorywritten at login,sso-session.ts:215-297, read bycollectIndexedSsoSessions():299), so a logout delivered to another worker still finds the session — pinned bytests/unit/auth/sso-distributed-logout.test.ts(including two tenants that share a provider id). What remains: the logout publishes no revocation marker, so a peer’s warm session cache entry is only dropped when it re-validates; and with the shared cache down the index read degrades to the local map instead of failing the logout.
T2 — active: false from SCIM must lock the account
- Precondition: SCIM bearer token for
tenant-a; useru2created through SCIM; one live session and a valid session cookie. - Action:
PATCH /api/scim/v2/Users/u2with{Operations:[{op:"replace", path:"active", value:false}]}, then call an authenticated write endpoint with the existing cookie. - Expected behaviour (now implemented):
applyScimPatchOpswritesisActive: falseandblocked: true(src/utils/scim-utils.ts:199-205), which is what the per-request gate reads (handle-authentication.ts), so the next write is rejected; non-booleanactivevalues are rejected with 400 (parseScimActive,:167). Pinned bytests/unit/api/scim-deprovisioning.test.ts. - Residual gap: the session row is not deleted, so an already-issued session survives until expiry (the block is enforced on the next re-validation, not by revocation).
T3 — Audit records must survive a mid-request crash
- Precondition:
AUDIT_CHAIN_SYNCunset (defaultfalse,audit-flags.ts:44); DB adapter with a settablefailNextInsert; audit table empty. - Action: call
auditService.log(...)for a privileged action (e.g.USER_ROLE_CHANGED), do not await a flush, and simulate process death (vi.resetModules()after clearing the module registry, or spawn a child thatprocess.exit(1)s immediately afterlog()resolves); then boot a fresh service and query the table. - Expected failure mode (today):
log()resolves as soon as the entry is in the in-memory array (audit-service.ts:421), so the record is absent after the restart while the API reported success. The test must assert either the row exists, or thatlog()rejected — never “success, no row”. - Companion assertion in the same test: with the DB insert failing, after 201 queued entries the
oldest are dropped silently (
:263-267) — assert an explicit, observable drop counter rather than alogger.warn.
T4 — The verifier must fail only when the chain is actually broken
- Precondition: write three entries through
auditService.logwithAUDIT_CHAIN_SYNC=true; no tampering. - Action: call
auditChainService.verifyChain(tenantId)— the exact call the UI makes (audit-history.remote.ts:54). - Expected behaviour (now implemented): a clean chain verifies. Writer and verifier share
canonicalAuditPayload()/computeChainHash(), so the hash the service writes is the one the verifier recomputes. Deleting a middle row is detected by the linkage check, and deleting the tail is detected by the per-tenant checkpoint (head hash no longer present). Pinned bytests/integration/databases/audit-chain-integrity.test.ts(7 tests, all four adapters). - Remaining limit: the checkpoint lives in the same database (
systemPreferences), so an actor who deletes the tail and the checkpoint row evades truncation detection.
T5 — Auditing must not be switchable off by the actor it records
- Precondition: admin session;
DISABLE_AUDIT_LOGSabsent; one existing audit row. - Action: as an admin, write the value
trueto theDISABLE_AUDIT_LOGSsystem setting, perform a privileged mutation (role change + user delete), then read the audit table. - Expected failure mode (today):
audit-flags.ts:69-83reads the key 5 s later and both pipelines short-circuit (handle-audit-logging.ts:129,166,audit-service.ts:221-224,380,388), so the mutation leaves no record and no marker that logging was off during that window. Also assert the second half: GDPR erasure must not remove the trail of the erased subject’s privileged actions —gdpr-service.ts:202-207deletes everyaudit_logsrow whoseactorIdis the erased user (i.e. exactly those actions), and theDATA_DELETIONrecord written afterwards (:241-248) survives only because it is logged withid: "system".
5. Method and negative results
Tree: next (baseline snapshot efa57c55e).
Commands that produced the “not found” rows, verbatim:
grep -rn 'SingleLogout\|LogoutRequest\|logoutRequest' src/ tests/ # 34 hits in src/ (SAML SLO shipped); no longer a "not found" row
grep -rn -i 'ldap' src/ --include=*.ts --include=*.svelte --include=*.mdx # 12 hits in 4 files, all injection scanners
grep -rn 'verifyOAuthState' src/ # google-auth.ts:83 (def) + login/+page.server.ts:357 (1 caller)
grep -rn 'email_verified\|verified_email' src/ # 0 hits
grep -rn 'at_hash\|c_hash' src/ # 0 hits
grep -n -i 'audit' 'src/routes/api/[...path]/handlers/auth.ts' # 0 hits
grep -n -i 'audit' 'src/routes/api/[...path]/handlers/scim.ts' # 0 hits
grep -n -i 'audit' src/databases/auth/saml-auth.ts # 0 hits
grep -n -i 'audit' src/databases/auth/sso-session.ts # 0 hits
grep -rn 'createLog\|auditChainService\.' src/ # 1 hit: verifyChain caller only
grep -rn 'CREATE TRIGGER' src/databases/ # content_nodes only
Paths in this repository are not under src/lib/server or src/routes/api/auth/: the auth engine
is src/databases/auth/*, the HTTP surface is the single dispatcher
src/routes/api/[...path]/handlers/*, and the request pipeline is src/hooks/*. A search in the
conventional SvelteKit layout returns nothing and must not be read as a missing feature.
6. Related reading
docs/reference/security/authentication-system.mdx— narrative description of the same components (this page is the evidence layer under it).docs/reference/architecture/transactional-outbox.mdx— the delivery queue, including its own non-transactional caveat.docs/reference/security/secrets-inventory.mdx— where the IdP client secrets and certificates live.