Skip to content

Documentation

Enterprise Workflow, Proxy/Session & Database Support Inventory

Verified inventory of workflow/approval enforcement, reverse-proxy and session behaviour, and per-engine database support on `next` (post-PR #566). Every row carries path:line; unlocated capabilities are recorded as 'not found' with the search that proves it.

11/11/2026
26 min read Edit on GitHub

Status: current audit of next. The baseline snapshot was efa57c55e on the (now deleted) fix/enterprise-readiness-20261110 branch; 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 Per-finding status is listed below — entries marked fixed are no longer current defects and must not be read as open risks.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 their not found result 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
TRUSTED_PROXIES matched only exact strings — documented CIDR blocks never matched Fixed in PR #566 (commits 0040f073d, f2281aeaf) isTrustedProxyIp() in hook-utils.ts; tests/unit/security/trusted-proxy-cidr.test.ts
Workflow publish gate unreachable for normal collections (plan.hasWorkflows never set) Corrected; dispatch/worker fixed; schema writer still missing (commit 5dafc3a5d; worker gate 2026-11-11) hasWorkflows is derived in execution-plan.ts:80-83 from workflow/reviewWorkflow/hasWorkflow/plugins, so the single-update gate is reachable; the bulk path calls assertPublishAllowed (tests/unit/services/bulk-publish-workflow-gate.test.ts). Content-package import is gated in applyImport (content-package-service.ts:1014-1067, gate :1056), again in the worker before any write (jobs/import-jobs.ts:292-334 → assertJobPublishAllowed, called at :486 for the content-package lane and :687 for the direct lane), and in the synchronous namespace lane before its write (assertCollectionImportPublishAllowed, import-jobs.ts:347-366, called from misc-namespaces.ts:308). Replace mode’s deleteMany runs after the gate. The worker fails closed with PERMANENT_FAILURE when the actor is missing (import-jobs.ts:306-310) or unresolvable (:315-319). Still open: nothing in src/ writes the workflow/hasWorkflow schema flag the gate reads (§1.3).
Bulk write paths bypassed assertPublishAllowed Fixed (bulk update/batch status) collectBulkRejections gates every publishing row on bulk update; workflow-less collections stay off the lazy import and the query; system writes and non-publish updates untouched. Bulk create still initialises the workflow after insert and has no gate (see §1.3).
Tenant/global session revocation had no distributed marker path; now read, not published Partially fixed — read path is distributed; no production publisher for tenant/global Readers now compare a session’s issued-at against user/tenant/global markers through getDistributedRevocationEpoch() (revocation-publish.server.ts:149-179, 5 s memo :34): warm LRU and the shared gate in handle-authentication.ts:305-321 / :340-348 (warm call :613), the session:<tenant>:<id> snapshot branch :709-726 (gate :715), and the store layer session-manager.ts:53-64 (:94-97, :117-120, :283-289, :321-327). publishUserRevocation() is called from auth-service.ts:1198, handlers/auth.ts:827 and auth-namespace.ts:416; publishTenantRevocation()/publishGlobalRevocation() (:100-116, :77-94) have no production caller. Remaining limits: the store branches check tenant/global but not the distributed user marker, and the memo admits up to 5 s of staleness. Tests: tests/unit/auth/revocation-read-path.test.ts (8), tests/unit/auth/revocation-distributed.test.ts (6).
No versioned DB migrations; MongoDB has no transactions; README claims MySQL without DB_TYPE=mysql Still open (re-verified 2026-11-11) find src scripts tests -name "*.sql" → 0 files; the only migrations dirs are src/plugins/{smart-importer,unified-data-hub}/migrations. Mongo declares supportsTransactions: false (mongodb/adapter-core.ts:76) and returns TRANSACTIONS_DISABLED (mongo-db-adapter.ts:225-234). README advertises “MariaDB/MySQL” (README.md:42, :63); createAdapterInstance() accepts only sqlite \| postgresql \| mariadb \| mongodb (db-init.ts:101-116). The MariaDB DDL uses CREATE INDEX IF NOT EXISTS (mariadb/adapter-core.ts:728, :2272; mariadb/schema-sync.ts:201-249) and JSON_UNQUOTE(JSON_EXTRACT(...)) expression indexes (:2248-2272), which MySQL 8 rejects.

This page records what the source actually enforces on branch next (baseline snapshot efa57c55e). It is an inventory, not a design document: each row points at the line that does the work, and each gap points at the line where a caller can step around it.

Evidence rule. Every statement carries path:line. A capability that could not be located is written as not found, together with the search that was run. A doc sentence, a config key or a unit test with a hand-built fixture is not evidence of enforcement — only a line on the request/write execution path counts. Method (2026-11-10). Read-only audit of the worktree. No source file was changed, nothing was committed, and no suite was executed (the host was running a performance benchmark). Test coverage below is read from the test files and from package.json / .github/workflows/ci.yml, not from a run.


1. Workflow and approval enforcement

1.1 What the model actually is

There are two independent status systems, and they are not synchronised:

System Values / shape Where
Entry status archive, draft, publish, unpublish, schedule, clone, delete — no review value src/content/status.ts:14-22
Workflow instance Free-form FSM states (states[], transitions[], gatePublication) owned by workflow_definitions / workflow_instances src/types/workflow-types.ts:6-59

A review workflow therefore lives only in the workflow tables. No code maps a workflow state onto entry.status: transition() writes the instance and the audit log (src/services/background/workflow-service.ts:371-419) and never touches the content row. Publishing is a separate write that is (conditionally) checked against the instance.

1.2 The single gate function and its call sites

assertPublishAllowed() (src/services/background/workflow-service.ts:503-549) is the whole approval mechanism:

  • admins/super-admins short-circuit (:509);
  • no definition, or gatePublication !== true → allow (:512);
  • create-path call without entryId → always reject (:514-520);
  • otherwise the instance must exist and sit in a state with isFinal (:522-548); a missing instance is lazily enrolled and the write rejected (:526-539).

Call sites (seven):

Caller Line
SDK collections.create src/services/sdk/namespaces/collections-namespace.ts:963
SDK collections.update src/services/sdk/namespaces/collections-namespace.ts:1214
SDK bulk update / batch status src/services/sdk/namespaces/collections/bulk-operations.ts:188
Scheduler status transition src/services/scheduler.ts:328
Scheduled-publish sweep src/services/background/jobs/scheduled-jobs.ts:155
Content-package planning src/services/core/content-package-service.ts:1056
Background import worker src/services/background/jobs/import-jobs.ts:277

1.3 Capability matrix — workflow / approval

Capability Status Implementation (file:line) Existing tests Gap
Workflow definition CRUD complete src/services/background/workflow-service.ts:70-131 (save), :136-147 (delete), topology validation :638-653 tests/unit/services/workflow-service-completion.test.ts:205-300; tests/e2e/routes/config/workflows.spec.ts:59-104 saveWorkflow only writes workflow_definitions; it never sets the collection flag the gate reads (see next row).
Publish gate reachable for a normal collection missing Gate is consulted only when plan.hasWorkflows is true (collections-namespace.ts:957, :1207); that flag is derived solely from schema.workflow \| reviewWorkflow \| hasWorkflow \| plugins.includes("workflow") (src/services/sdk/namespaces/collections/execution-plan.ts:79-84) tests/unit/services/execution-plan.test.ts:122-135 — uses a literal { workflow: true } cast to Schema grep -rn "hasWorkflow\|reviewWorkflow" src tests → only execution-plan.ts:79-84 (readers) and collections-namespace.ts:957,1207 (use). No writer exists: nothing in src/ or src/routes/(app)/config/collectionbuilder sets those fields, and workflow is not a member of Schema (grep -n "workflow" src/content/types.ts → 0 hits). A definition in workflow_definitions alone therefore never gates publishing. Admin-only saveWorkflow (workflow-service.ts:78-80) cannot fix this.
Gates skipped by presence cache partial src/services/background/workflow-presence.ts:14-24 (60 s note), consumed at collections-namespace.ts:960, :1211 via !workflowKnownAbsent(...) tests/unit/services/workflow-presence.test.ts Cache is per process with a 60 s TTL (workflow-presence.ts:17, :47); a peer instance keeps publishing for up to 60 s after a definition is created.
FSM transition validity complete src/services/background/workflow-service.ts:300-310; no instance → 404 :288-293 tests/unit/sdk/workflow-namespace.test.ts:156-165 Instance must pre-exist; only post-write.ts:280-307 creates one, and only on create (see initialisation row).
Transition RBAC (workflow:transition, transition.requiredRole) complete (service) / unreachable (REST) Service check workflow-service.ts:314-338; permission registered :26-32; catalogue bit src/databases/auth/permission-catalog.ts:382 tests/unit/services/workflow-service-completion.test.ts:213-300 PATCH /api/workflows — the only transition route (src/routes/api/[...path]/+server.ts:118) — requires admin for every non-GET method (src/routes/api/[...path]/handlers/system/system-workflows.ts:33-34). A non-admin reviewer can never transition, so requiredRole and the base permission are dead paths over REST.
Assignee lock (requiresAssignee) complete (service) workflow-service.ts:343-369; handoff :431-488 tests/unit/services/workflow-service-completion.test.ts:213-300 Same REST/admin obstacle as above; the SDK path (src/services/sdk/namespaces/workflow-namespace.ts:60-80) bypasses the HTTP guard for in-process callers.
Publish-on-create (REST) partial src/services/sdk/namespaces/collections-namespace.ts:946-990 (isPublishing → assertPublishAllowed) tests/unit/services/workflow-service-completion.test.ts:175-182 Whole block is inside if (plan.hasWorkflows) (:957), and system: true bypasses it (:952-953). Handler never passes system (handlers/collections.ts:664-676).
Publish-on-update (REST/GraphQL/SDK) partial collections-namespace.ts:1196-1261 tests/unit/services/workflow-service-completion.test.ts:138-203 Gate runs only for an explicit status: "publish"\|"published" payload (:1199-1200, :1207). A PATCH without status on an already-published entry never consults the workflow (:1206), so content can be edited after approval without re-review.
Bulk create (publish) missing src/services/sdk/namespaces/collections/bulk-operations.ts:207-313 (executeBulkCreate) — no assertPublishAllowed call in the create path; the workflow is only initialised after insert (:305-320). The only gate call in this file is the bulk-update collector (:188). tests/integration/databases/bulk-operations-contract.test.ts (CRUD only); no workflow case POST /api/collections/:id/bulk and array-bodied POST /api/collections/:id insert status: "publish" rows with no gate and no required-field-for-publish check. Search: grep -n "assertPublishAllowed" src/services/sdk/namespaces/collections/bulk-operations.ts → 1 hit, the bulk-update collector (no create call).
Bulk update / batch status fixed (bulk update/batch status) bulk-operations.ts:136-201 — collectBulkRejections calls assertPublishAllowed per publishing row; required-fields-for-publish checked first (:177-184) tests/unit/services/bulk-publish-workflow-gate.test.ts Reached from PATCH /api/collections/:id/bulk, POST /api/collections/:id/batch {action:"status"} and PATCH /api/collections/:id/:entryId/status with {entries:[…]}. Workflow-less collections stay off the lazy import (workflowKnownAbsent).
Scheduled publish job (publish-scheduled handler) partial src/services/background/jobs/scheduled-jobs.ts:135-209; gate at :153-168; actor must exist :138-152 none found (grep -rn "scheduledPublishHandler" tests → 0) Publishes with db.content.nodes.update(node.path, …) (:182-187), i.e. outside the SDK write lane: schema validation, required fields, hooks, revisions and the widget pipeline never run. Cache invalidation is manual and best-effort (:199-202).
Scheduler status transition partial src/services/scheduler.ts:289-347; gate for publish only :325-329 tests/unit/services/scheduler-job.test.ts:10-54 (payload acceptance only; no execution test) When the job carries entryPath, the status is written directly (:330-334) — unpublish/draft/delete bypass validation/hooks/revisions entirely, and even the publish path never re-validates required fields. Actor is pinned server-side (:42-68, :302-305), which prevents impersonation but not the bypass.
Content-package import fixed at planning/dispatch and in the worker src/services/core/content-package-service.ts:1014-1067 (assertImportPublishAllowed), gate :1056; the content-package job payload is { planId, tenantId, userId, operations } and carries no duplicateStrategy (:885-894), because the plan already resolved create/update/skip/conflict. The worker re-gates both lanes before any write — content-package import-jobs.ts:396, direct :507-512 — including before replace mode’s deleteMany (:97-104), and fails closed with PERMANENT_FAILURE when payload.userId is absent (:258-262) or unresolvable (:267-271). Actors: content sync passes the operator id persisted in the job payload (never the "system" sentinel) and turns a gate block into an explicit job error naming the channel (src/services/core/content-sync-service.ts:692-751); the SDK resolves params.userId → options.userId → options.user._id (src/services/sdk/namespaces/data-operations.ts:76-85). tests/unit/services/content-package-workflow-gate.test.ts (9 cases: non-admin rejected, admin allowed, ungated collection unaffected, update→publish gated, background dispatch blocked, draft unaffected, unresolved/no actor fail closed); tests/unit/services/import-worker-workflow-gate.test.ts (9 cases: non-admin, replace-wipe ordering, admin, workflow-absent, missing actor, unresolvable actor, content-package round trip, pre-write gate, update upsert/skip tally); tests/unit/services/import-actor-resolution.test.ts (9 cases) Imported status: "publish" rows are now checked against the same gate as a normal write in both lanes, and a gate rejection writes nothing. All three former gaps are closed: ImporterNamespace.importData() resolves the actor from options.userId → options.user._id (src/services/sdk/namespaces/misc-namespaces.ts:248-260, dispatch :281-282); the backup create/restore payloads are typed (import-jobs.ts:88-116) and the restore branch re-gates before the write (import-jobs.ts:562-608); a plan conflict now fails the job permanently (import-jobs.ts:404-409) instead of counting as a skip. The synchronous namespace lane is gated too (assertCollectionImportPublishAllowed, import-jobs.ts:347-366, called misc-namespaces.ts:308; tests tests/unit/services/import-lane-gaps.test.ts). Import still bypasses hooks/enrolment/revision.
Smart Importer (bulk import) missing src/plugins/smart-importer/index.server.ts:405-414 (crud.updateOne / crud.insert), :520-522 → src/plugins/smart-importer/performance.ts:314-336 (insertMany) tests/unit/plugins/smart-importer/* (parser/mapping only) Same bypass, plus insertMany skips per-row validation. Remote import is a first-class admin feature, so the practical effect is “import anything, including published rows”.
Draft → publish promotion complete src/services/sdk/namespaces/drafts-namespace.ts:380-394 delegates to collections.create/update, so the update/create gate applies tests/unit/sdk/drafts-promotion-atomicity-sqlite.test.ts; tests/unit/databases/mongo-drafts-parity.test.ts On MongoDB promotion is not atomic (drafts-namespace.ts:251-258, :396-412) — see §3.
Workflow enrolment of new entries partial src/services/sdk/namespaces/collections/post-write.ts:280-307; needs workflowApplies(...) (a prior lookup) and !skipReturning tests/unit/sdk/collections-post-write.test.ts Entries created by the import paths above, by system: true callers (forms.ts:423) or with skipSideEffects never get an instance, so a later gate check has to lazily enrol them (workflow-service.ts:526-539) — the first publish attempt after enabling a workflow always fails once.
Workflow reset when approved content changes missing — none found grep -rn "resetWorkflow\|revertToDraft\|invalidateWorkflow\|demote" src → only cache invalidation (workflow-service.ts:61-65). Editing an entry whose instance sits in a final state leaves it publishable with no re-review.
Admin override of gate/assignee complete (by design) workflow-service.ts:37-39, :509, :314-315, :345 tests/unit/services/workflow-service-completion.test.ts:196-203 Documented escape hatch; not signalled in the audit record when used.
Transition audit trail complete workflow-service.ts:390-409, assign :469-485 tests/unit/services/workflow-service-completion.test.ts (mock asserts call) Audit is written after the instance update; a failure between :382-387 and :390 leaves an unaudited transition. No test for that ordering.
FS, state and gate unit coverage on a real engine missing — all gate tests use a fake adapter: tests/unit/services/workflow-service-completion.test.ts:5,29-40 grep -rln "gatePublication\|assertPublishAllowed" tests/integration → 0 files. Nothing proves the gate against SQL JSON round-trips, tenant scoping or Mongo behaviour.

2. Reverse proxy and session behaviour

2.1 Request-trust chain as implemented

  1. getClientIp() reads the transport peer (event.getClientAddress(), src/utils/hook-utils.ts:286-295), then — only if that peer is in TRUSTED_PROXIES — reads forwarded headers in the order cf-connecting-ip → x-real-ip → first hop of x-forwarded-for (:301-315). Result is memoised on locals.__clientIp (:282-284, :317).
  2. TRUSTED_PROXIES is parsed in src/databases/config-state.ts:439-460 ("all"/"*" sentinel, comma string or array, cached against a config stamp).
  3. The peer match is isTrustedProxyIp() (src/utils/hook-utils.ts:395-434): exact addresses and IPv4/IPv6 CIDR, IPv4-mapped IPv6 normalised (::ffff:a.b.c.d), malformed entries degrade to a harmless exact string (fail closed), and the compiled rules are cached against the parsed allowlist.
  4. CSRF compares Origin (or Referer) against the Host header (src/utils/security/csrf-utils.ts:126-167), else requires the X-CSRF-Token cookie/header pair.
  5. With MULTI_TENANT, the tenant id is derived from the request hostname (src/utils/tenant.ts:35-48, called from src/hooks/handle-security.ts:220).

2.2 Capability matrix — proxy, sessions, multi-instance behaviour

Capability Status Implementation (file:line) Existing tests Gap
Trusted-proxy allowlist complete (except all) src/databases/config-state.ts:439-463; src/databases/private-config-schema.ts:71-75; matcher src/utils/hook-utils.ts:395-441 tests/unit/security/trusted-proxy-cidr.test.ts; tests/unit/databases/config-state-trusted-proxies.test.ts:28-58 (cache invalidation) Entries are exact addresses or CIDR (10.0.0.0/8, 2001:db8::/32), IPv4-mapped IPv6 normalised, malformed entries fail closed and are cached. TRUSTED_PROXIES=all/* still trusts forwarded headers from any peer, which re-opens client-IP spoofing — keep it to development.
Forwarded-IP spoofing resistance (HTTP) complete src/utils/hook-utils.ts:299-315 (headers ignored unless peer trusted); fallback 0.0.0.0 not loopback (:286-295) tests/unit/hooks/rate-limit.test.ts:183 (“XFF does not split a real client”); tests/unit/hooks/adversarial.test.ts When TRUSTED_PROXIES="all" (or a shared LB IP), the client-supplied X-Forwarded-For first hop is taken verbatim — spoofing returns for rate-limit bucketing. No test asserts behaviour for "all".
Forwarded-IP resistance (WebSocket) complete src/hooks.ws.ts:100-115 transport only; asserted in :265-291 tests/unit/hooks/ws-ratelimit.test.ts:42-74 normalizeUrl() still trusts x-forwarded-proto unconditionally (hooks.ws.ts:145-147) when rebuilding the upgrade URL; used for scheme, not for the origin comparison (:163-173), so impact is limited to isSecureCookieContext inputs at :213.
Host / origin validation partial WS: src/hooks.ws.ts:163-173 (origin host must equal request host; empty origin allowed :164); HTTP: src/utils/security/csrf-utils.ts:126-167; framework-level ORIGIN patched in vite.config.ts:1270-1285 tests/unit/services/scheduler-job.test.ts:56-70 (websocketOriginAllowed); tests/unit/hooks/adversarial.test.ts No allowlist of legal Host values anywhere (grep -rni "allowedHosts\|trustedHosts\|x-forwarded-host" src → 0 hits). The Host header is simultaneously the CSRF reference and the tenant selector (handle-security.ts:220 → tenant.ts:44-46), so a spoofed Host: acme.example.com picks the acme tenant scope for the request.
Session cookie flags complete src/routes/api/[...path]/handlers/auth.ts:250-267 (httpOnly:true, sameSite: strict\|lax, secure, 24 h); __Host- naming src/databases/auth/constants.ts:78-79; secure decision constants.ts:43-62 tests/unit/auth/sessions/session-cookies.test.ts:40-155 (incl. subdomain cookie-tossing :139) Non-secure context downgrades to SameSite=Lax and drops the __Host- prefix (auth.ts:264, constants.ts:79) — a proxy that terminates TLS without url.protocol=https (missing PROTOCOL_HEADER) yields a lax, non-Secure cookie. isSecureCookieContext mitigates in production for non-loopback hosts (constants.ts:59-62), but there is no test for “proxy + http protocol + production host”.
CSRF enforcement on mutating APIs complete REST dispatcher src/routes/api/[...path]/+server.ts:549-569; GraphQL src/routes/api/graphql/+server.ts:859-877; migration import src/routes/api/migration/import/+server.ts:43-45; preview …/preview/authorize/+server.ts:19-21 tests/integration/api/security-negative.test.ts; tests/unit/api/* Same-origin fast path accepts origin === //host (csrf-utils.ts:129) and treats loopback origin+host pairs as same-origin (:136-140); acceptable, but it means the check is a Host-header comparison, not a configured-origin comparison.
Session store / multi-instance sharing partial Default is in-memory (src/databases/auth/session-manager.ts:505-513); the Redis-backed store is only chosen when a Redis client is passed (:262-298, :300-340, :505-513) tests/unit/auth/session-rotation.test.ts; tests/unit/databases/relational-auth-session-mutations.test.ts getDefaultSessionManager() calls createSessionManager() with no argument (session-manager.ts:520-528), so the default process always gets the in-memory store even when Redis is configured; cross-instance invalidation then relies on DB session rows and the revocation markers below.
User-level session revocation (multi-instance) complete src/databases/auth/revocation-publish.server.ts:20-29 (marker prefixes, 30 d TTL) and :53-71 (publishUserRevocation); read on the warm LRU via getDistributedRevocationEpoch() (handle-authentication.ts:305-321, call :586-592) and in the session stores (session-manager.ts:94-97, :117-120, :283-289, :321-327) tests/unit/auth/session-rotation.test.ts; tests/unit/auth/revocation-read-path.test.ts; tests/unit/auth/revocation-distributed.test.ts Fail-open: a marker write failure only logs (revocation-publish.server.ts:65-69), and workers that never saw the marker still honour the DB row delete.
Tenant-wide / global session revocation (multi-instance) read path shipped; publish unwired Publishers exist (revocation-publish.server.ts:77-94 global, :100-116 tenant) but grep -rn "publishGlobalRevocation\|publishTenantRevocation" src returns only their definitions and the doc comments in revocation-epoch.ts. Readers are wired: handle-authentication.ts:305-321 (:586-592), session-manager.ts:53-64. tests/unit/auth/revocation-read-path.test.ts (8); tests/unit/auth/revocation-distributed.test.ts (6) A tenant lockout or “log out everywhere” still only reaches other nodes once a caller publishes the marker — none does today. The hook’s session:<tenant>:<id> snapshot branch (:652-671) has no revocation check at all; the store layer checks tenant/global but not the distributed user marker; the read memo admits up to 5 s of staleness (revocation-publish.server.ts:34).
Rate limiting (HTTP) complete (single instance) / partial (cluster) src/services/security/response-service.ts:537-604: Redis primary :578-587, memory fallback :586,589, key :571 tests/unit/hooks/rate-limit.test.ts:150-183; tests/unit/security/* Fallback is fail-open (:582-586 logs and continues on memory), so a Redis outage multiplies the effective limit by the instance count. The skip list excludes /login (:551-560), leaving password brute-force to the per-account lockout (src/databases/auth/auth-service.ts:932-949) rather than the WAF.
Rate-limit storage inspection complete src/services/security/state-store.ts:35-160 (svelty:sec: keys, L1 + L2 read) tests/unit/security/* Incidents are truncated to the last 100 per tenant (state-store.ts:132) — telemetry loss under attack, not an enforcement gap.
WebSocket connection limit across instances partial src/hooks.ws.ts:266-291: per-process activeConnections set (:44-52) tests/unit/hooks/ws-ratelimit.test.ts:76-90 RATE_LIMITER_WEBSOCKETS_REDIS_PREFIX is read (:268-270) but only used in a log message (:286); there is no Redis counter, so N instances allow N × maxConnsPerIp.
Cache invalidation across instances complete (with Redis) / partial (without) src/databases/cache/cache-service.ts:357-365 (subscribe), :428-436 and :988-999 (publish, guarded by isL2Ready()), node self-filter :365-366 tests/unit/databases/cache-service-l2-contract.test.ts:91-233 (explicit two-instance pub/sub test :196) Without L2 the publish branch only purges the CDN (:433-435), so a peer’s L1 keeps stale entries until its own TTL (content 120 s, entry 60 s, collection 300 s — cache-service.ts:43-57).
Background-job coordination partial Atomic claim by filtered update src/services/scheduler.ts:217-228; reaper for orphaned running rows src/services/background/jobs/job-queue-service.ts:310-348 (15 min lease); single-writer gate src/hooks.server.ts:371-393 (SVELTY_BACKGROUND_MODE=disabled\|child) tests/unit/services/scheduler-job.test.ts:10-54; no reaper test No leader election: every instance polls and relies on the claim. The scheduler’s own claim has no lease/heartbeat (scheduler.ts:217-226) — a crash leaves the row running until the job-queue reaper happens to run, and grep -rn "reapStaleRunningJobs" src shows the only caller is the queue’s own interval (job-queue-service.ts:391).
Schema-migration coordination complete src/databases/migration-lock.ts:23-59 (PG advisory lock, MariaDB GET_LOCK, SQLite file lock, Mongo lock doc); used at src/databases/postgresql/postgres-adapter.ts:105, src/databases/mariadb/mariadb-adapter.ts:111, src/databases/sqlite/adapter-core.ts:1418, and for content migrations src/services/core/migration-engine.ts:555 tests/unit/databases/in-memory-provision-lock.test.ts; tests/unit/databases/adapter-wrap-hot-path.test.ts Unknown adapter types run migrations without a lock (migration-lock.ts:51-56) — irrelevant today because only four types exist, but it is a silent degrade.
Fast-lane request identity partial vite.config.ts:1344 builds the lane request with clientAddress: a.socket?.remoteAddress \|\| "127.0.0.1" tests/integration/api/lane-transport-identity.test.ts The "127.0.0.1" fallback is a loopback address, which handle-security.ts:182-191 treats as local; combined with IS_TEST_MODE it short-circuits the WAF. In production the socket address is the proxy IP, so IP bucketing collapses to one bucket unless TRUSTED_PROXIES is set.

3. Database support: claim versus reality

3.1 Engine matrix

Engine Status Adapter (file:line) DDL / migrations Existing tests
SQLite complete SQLiteAdapter selected at src/databases/db-init.ts:102-104; implementation src/databases/sqlite/sqlite-adapter.ts, core …/sqlite/adapter-core.ts No versioned migrations. bootstrapSystemSchema("sqlite", …) guarded by withMigrationLock (sqlite/adapter-core.ts:1404-1418); per-collection DDL in …/sqlite/schema-sync.ts:93-192 tests/integration/databases/sqlite-adapter.test.ts; contract.test.ts; adapter-parity.test.ts; differential-parity.test.ts; transaction-contract.test.ts; tests/unit/databases/sqlite-schema-bootstrap-exec.test.ts; sqlite-statement-cache-lru.test.ts
PostgreSQL complete PostgreSQLAdapter at src/databases/db-init.ts:105-107; src/databases/postgresql/postgres-adapter.ts Same bootstrap path: postgres-adapter.ts:105; DDL spec src/databases/system-schema-spec.ts:37-147; column/index sync …/postgresql/schema-sync.ts; PG-only GIN indexes system-schema-spec.ts:1225-1231 tests/integration/databases/postgresql-adapter.test.ts; contract.test.ts; differential-parity.test.ts; tests/unit/databases/postgresql-adapter-lifecycle.test.ts; postgresql-index-usage.test.ts; postgresql-fts-adapter.test.ts; postgresql-wire-etag.test.ts
MariaDB complete MariaDBAdapter at src/databases/db-init.ts:108-110; …/mariadb/mariadb-adapter.ts (mysql2 driver :88) bootstrapSystemSchema("mariadb", …) under lock (mariadb-adapter.ts:107-113); …/mariadb/schema-sync.ts:93-249 tests/integration/databases/mariadb-adapter.test.ts; contract.test.ts; tests/unit/databases/mariadb-insert-coalescer.test.ts; mariadb-wire-etag.test.ts
MongoDB complete (documents) / partial (relational features) MongoDBAdapter at src/databases/db-init.ts:111-113; …/mongodb/mongo-db-adapter.ts No DDL at all: models registered on connect (…/mongodb/model-registration.ts:25-40), indexes created lazily per collection (…/mongodb/collection-module.ts:355-440) tests/integration/databases/mongodb-adapter.test.ts; contract.test.ts; tests/unit/databases/mongo-drafts-parity.test.ts; mongo-facet-list.test.ts; mongo-map-query.test.ts; mongo-unordered-bulk.test.ts; mongo-upsert-one-rt.test.ts
“MySQL” not-found createAdapterInstance() accepts exactly sqlite \| postgresql \| mariadb \| mongodb, everything else throws Unsupported database type (src/databases/db-init.ts:101-116) — no MySQL-specific suite; grep -rn "mysql" src/databases/mariadb/*.ts returns mysql2 driver/dialect usage only (e.g. adapter-core.ts:48-51, mariadb-adapter.ts:88,97) — no MySQL adapter or DB_TYPE=mysql

Claim under test. README.md:42 and :63 advertise ”MariaDB/MySQL”. Reality: there is no DB_TYPE=mysql, and the MariaDB path emits syntax MySQL 8 rejects — CREATE INDEX IF NOT EXISTS (src/databases/mariadb/adapter-core.ts:728, :2272; …/mariadb/schema-sync.ts:201-249) and JSON_UNQUOTE(JSON_EXTRACT(...)) expression indexes (adapter-core.ts:2248-2272). MySQL compatibility is therefore unverified and likely broken at first index creation; nothing in tests/ runs against a MySQL server (tests/docker-compose.yml profiles are mongodb, mariadb, postgresql and redis — SQLite runs in-process — and the four-entry matrix is generated in .github/workflows/db-matrix.ts:36-69).

3.2 Capability differences per engine (declared)

Capability PostgreSQL MariaDB SQLite MongoDB Evidence
Multi-document transactions yes yes yes no (replica set required) postgresql/adapter-core.ts:87; mariadb/adapter-core.ts:72; sqlite/adapter-core.ts:1824; mongodb/adapter-core.ts:76 + mongo-db-adapter.ts:221-234 (TRANSACTIONS_DISABLED)
RETURNING writes yes no yes n/a (driver returns doc) postgresql:96, mariadb:81, sqlite:1833, mongodb:84
Native JSON column + queries JSONB JSON TEXT + json_extract BSON system-schema-spec.ts:145-147; sqlite/adapter-core.ts:901,1975-2032; mariadb/adapter-core.ts:184-196,330-350; supportsNativeJson flags above
JSON-path secondary indexes GIN (PG only) expression indexes (MariaDB 10.6+) none wildcard/compound system-schema-spec.ts:1225-1231; mariadb/adapter-core.ts:2248-2272; sqlite has none (grep -rn "create.*Index.*json" src/databases/sqlite → 0)
Aggregation pipeline subset subset subset native supportsAggregation: true for all four; unsupported stages fail closed at src/databases/core/aggregation-translator.ts (comment postgresql/adapter-core.ts:90-92)
Streaming / wire streaming yes yes no (supportsStreaming:false) yes / no wire postgresql:93,99; mariadb:78,84; sqlite:1830,1836; mongodb:82,87
Full-text search yes yes no (supportsFullTextSearch:false although src/databases/sqlite/fts-adapter.ts exists) yes postgresql:89; mariadb:74; sqlite:1826; mongodb:79
Max batch size 1000 1000 100 1000 postgresql:101; mariadb:86; sqlite:1838; mongodb:75
Row-level security opt-in, unwired; fails closed none none none postgresql/adapter-core.ts:2231-2298 implements enforceTenantPolicy, which now re-reads relrowsecurity/relforcerowsecurity/ownership and refuses to report success when the policy cannot filter; no production caller wires it. Tenant isolation is application-level (src/databases/crud-tenant-guard.ts). See the tenancy inventory §3.
Atomic draft promotion yes yes yes no — create-then-delete fallback src/services/sdk/namespaces/drafts-namespace.ts:251-258, :396-412
Read replicas adapter-agnostic adapter-agnostic adapter-agnostic adapter-agnostic src/databases/db-init.ts:137-171 (DB_READ_REPLICAS, ReplicaRouterAdapter)

3.3 How the matrix is actually exercised

Concern Where Notes
CI matrix .github/workflows/ci.yml:457-553; generator .github/workflows/db-matrix.ts:36-80 4 jobs (sqlite, postgresql, mariadb, mongodb), each runs the whole tests/integration/ tree (ci.yml:531-535) with DB_TYPE set.
Local parity runner scripts/runs/run-local-db-matrix.ts:41-50 (5 suites), :67-72 (targets), :221-231 (env); package.json:162 Skips offline engines unless --require-all (:288-295); the header notes the pre-push hook only proves SQLite (:8-9). The hinted scripts/runs/run-all-db-matrix.ts does not exist (ls scripts/runs → run-local-db-matrix.ts only).
Per-suite engine gating tests/integration/databases/adapter-test-env.ts:39-60 Suite runs only when DB_TYPE matches and the Docker container is up (or CI=true).
Transaction contract gating tests/integration/databases/transaction-contract.test.ts:71 Title says “All Adapters” but there is no DB_TYPE guard and the assertions require a working commit/rollback (:77-129). On the MongoDB job the adapter returns TRANSACTIONS_DISABLED (mongo-db-adapter.ts:225-234), so either the suite fails or it silently never exercises the engine — no in-repo exclusion exists (grep -rn "transaction-contract" scripts .github → only the smart-test list in scripts/test/test-smart.ts:336).
Versioned DB migrations not found re-verified 2026-11-11: find src scripts tests -name "*.sql" → 0 files; find . -type d -name migrations → only src/plugins/*/migrations (plugin data migrations). Core schema is create-if-missing + column sync (system-schema-spec.ts, */schema-sync.ts), and MigrationEngine (src/services/core/migration-engine.ts:363) migrates content schemas, not the database.

4. The five most important missing tests for data safety

Each case is written to fail on today’s code. Expected failure modes are the ones implied by the cited lines, not guesses.

T1 — Bulk publish bypasses the publication gate

  • Target file: tests/integration/api/workflow-bulk-gate.test.ts (new)
  • Precondition: collection posts with a workflow definition where gatePublication: true and the entry’s instance sits in a non-final state (draft). Editor user with collections:update, no admin. plan.hasWorkflows must be true (set the schema flag the gate reads, per execution-plan.ts:79-84).
  • Action: PATCH /api/collections/posts/bulk with [{ id, data: { status: "publish" } }]; repeat with POST /api/collections/posts/batch { action: "status", entryIds: [id], status: "publish" }; repeat with POST /api/collections/posts/bulk carrying status: "publish" rows.
  • Expected behaviour (now implemented for bulk update): PATCH /api/collections/posts/bulk and POST /api/collections/posts/batch {action:"status"} now call assertPublishAllowed per publishing row (bulk-operations.ts:186-200) and return 403 WORKFLOW_PUBLISH_GATE. Bulk create (POST /api/collections/posts/bulk with status: "publish" rows) still inserts without the gate — that half of the test remains open.

T2 — A workflow definition alone does not gate publishing

  • Target file: tests/integration/api/workflow-gate-reachability.test.ts (new)
  • Precondition: fresh database; a collection created through the normal collection-builder payload (no workflow, hasWorkflow, reviewWorkflow or plugins: ["workflow"] fields); POST /api/workflows with gatePublication: true and a non-final initial state.
  • Action: as a non-admin editor, POST /api/collections/posts with status: "publish", then PATCH /api/collections/posts/:id { status: "publish" }.
  • Expected failure mode: both writes succeed and the row is published. The gate is guarded by plan.hasWorkflows (collections-namespace.ts:957, :1207), which is false because nothing writes the flag (execution-plan.ts:79-84; no writer found by grep -rn "hasWorkflow\|reviewWorkflow" src). The test should demand 403 WORKFLOW_PUBLISH_GATE after saving a workflow — i.e. it encodes the invariant “a saved definition is enforced, or saveWorkflow refuses to save”.

T3 — Editing approved content republishes without re-review

  • Target file: tests/integration/api/workflow-stale-approval.test.ts (new)
  • Precondition: gatePublication: true; entry published through the workflow (instance in a final state); non-admin editor with collections:update.
  • Action: PATCH /api/collections/posts/:id with a body that changes title/body copy and omits status; then GET the entry as an anonymous reader.
  • Expected failure mode: the public read returns the edited copy while the workflow instance is still in the final state — no transition, no re-review, no audit entry. Cause: gating only runs for an explicit publish (collections-namespace.ts:1199-1207), and transition()/update() never reset or re-open the instance (workflow-service.ts:371-419; post-write.ts:280-307 initialises on create only). The test should require either a state reset or a rejection until the instance returns to a final state.

T4 — Scheduled transition and import paths write around the write lane

  • Target file: tests/integration/api/workflow-out-of-band-writes.test.ts (new)
  • Precondition: collection with a required: true field, a JSON field and revision: true; a status-transition job whose payload contains entryPath (scheduler.ts:35 accepts it), scheduled by a real user id (:302-305).
  • Action: let the scheduler claim and execute the job (scheduler.ts:212-272), then separately import a content package with a publish row (content-package-service.ts:885-894 dispatch; adapter writes at :925-950) and a Smart Importer batch (smart-importer/index.server.ts:405-414).
  • Expected failure mode: all three persist an entry that violates the required field / skips the revision snapshot / skips hooks and widget processing, because they call db.content.nodes.update or crud.insert directly instead of the collections namespace (scheduler.ts:330-334; scheduled-jobs.ts:182-187; content-package-service.ts:925-950). The content-package import is now gated at planning (content-package-service.ts:1014-1067) and again in the worker before any write (import-jobs.ts:244-286, called :396/:507-512), but scheduled transitions/imports still bypass the write lane. The namespace now dispatches the actor id (misc-namespaces.ts:248-260), a plan conflict fails the job permanently (import-jobs.ts:404-409), and the synchronous lane is gated too (tests/unit/services/import-lane-gaps.test.ts). Assert (a) no revision row, (b) the invalid required-field row is readable, (c) no workflow_instances row was created for the imported entries.

T5 — MongoDB loses atomicity on draft promotion and bulk writes

  • Target file: tests/integration/databases/mongo-draft-promotion-atomicity.test.ts (new)
  • Precondition: DB_TYPE=mongodb, collection with a unique index on a business key, an existing draft whose promotion target already exists; also a 20-row bulkCreate where row 11 violates a NOT NULL-equivalent schema rule.
  • Action: call drafts.promoteDraft(...) for the conflicting draft; then collections.bulkCreate with the mixed batch.
  • Expected failure mode: on Mongo the promotion runs the non-transactional path (drafts-namespace.ts:251-258, :396-412): the live write lands first, the draft delete fails, and the caller receives “promoted but could not remove it” (:402-412) — a duplicated live row plus a stale draft. bulkCreate inserts through batch.bulkInsert with no transaction (bulk-operations.ts:260-276), so a partial failure leaves the earlier rows committed while the caller sees an error. Correct behaviour: document/enforce all-or-nothing (or expose supportsTransactions:false to the caller and require idempotency keys). Note transaction-contract.test.ts:71 currently claims “All Adapters” while this engine returns TRANSACTIONS_DISABLED (mongo-db-adapter.ts:225-234).

Appendix A — searches behind the “not found” rows

Claim Search
No writer for schema.workflow / hasWorkflow / reviewWorkflow grep -rn "hasWorkflow\|reviewWorkflow" src tests; grep -n "workflow" src/content/types.ts; grep -rn "workflow" src/routes/(app)/config/collectionbuilder
No workflow reset on content edit grep -rn "resetWorkflow\|revertToDraft\|demote\|invalidateWorkflow\|currentState =" src
No gate in bulk paths grep -n "assertPublishAllowed" src/services/sdk/namespaces/collections/bulk-operations.ts src/services/sdk/namespaces/collections/bulk-validation.ts
No integration test for the gate grep -rln "gatePublication\|assertPublishAllowed" tests/integration
No CIDR proxy matching grep -n "isProxyTrusted" -A8 src/utils/hook-utils.ts; grep -rn "10.0.0.0/8" tests/unit/databases/config-state-trusted-proxies.test.ts
No host allowlist grep -rniE "allowedHosts\|trustedHosts\|x-forwarded-host" src
No distributed tenant/global revocation grep -rn "revokeAllSessions\|revokeTenantSessions" src tests
No MySQL adapter sed -n "101,116p" src/databases/db-init.ts; grep -rn "mysql" src/databases/db-init.ts
No versioned DB migrations find src scripts tests -name "*.sql"; find . -path ./node_modules -prune -o -type d -name migrations -print
RLS never applied grep -rn "enforceTenantPolicy" src; `grep -rniE “row.level security CREATE POLICY” src`
Runner name in the brief ls scripts/runs → no run-all-db-matrix.ts (only run-local-db-matrix.ts)
workflowapprovalreverse-proxysessionsdatabaseinventory
Was this page helpful?