Register GET /health/ready before the /api/* auth middleware so it is
public and reachable by K8s readinessProbe on port 3000 without auth.
On success → 200 {"status":"ready"}; on any DB/schema failure → 503
{"status":"degraded"}. Logs the pg error code; never leaks SQL in body.
A dropped schema (42P01) surfaces as non-200, closing the /health mask
that allowed the GRO-2678 incident to go undetected for ~43h.
Add health-ready.test.ts covering the 200 success path, 503 on schema
drop (42P01), and 503 on connection error; all assert no SQL leakage.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- Register GET /api/readyz after /api/health in src/index.ts (before authMiddleware)
- Runs `getDb().select({id: staff.id}).from(staff).limit(1)` inside try/catch
- Success → 200 {"status":"ready"}; failure → 503 {"status":"degraded","check":"db"}
- Raw driver error is console.error'd but never leaked to the response body
- /health and /api/health are unchanged (K8s probes remain DB-less per CTO decision)
- New unit tests: mock getDb to resolve → assert 200/ready; throw → assert 503/degraded
- Updated UAT_PLAYBOOK.md §4.0 with TC-API-0.2 and TC-API-0.3
Co-Authored-By: Paperclip <noreply@paperclip.ing>
drizzle-orm's migrate() has a high-water-mark (HWM) bug: on a fresh DB,
migration 0000 sets the watermark to 2026-03-17. Migrations 0001, 0003,
0010, 0011 have stale 2025-era `when` timestamps and are silently
skipped. Migration 0003 (recurring_series) is the blocker — its skip
leaves `recurring_series`, `appointments.series_id`, and
`appointments.series_index` missing. A downstream migration inside
migrate()'s single Postgres transaction then fails, rolling back
everything including 0000's `staff` and `services` tables.
Replace the drizzle-orm migrate() call with `pnpm exec drizzle-kit
migrate` (hash-based). drizzle-kit applies every unhashed migration
regardless of `when` ordering, matching the K8s migrate Job exactly.
AbortSignal.timeout(5000) in initAuth's discovery fetch races with
vitest's 5000ms default test timeout, causing intermittent failures on CI push
runs. Stub fetch to return ok:false instantly so tests complete in <1s.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Previously the initAuth() function set authInitPromise to the async IIFE's
Promise but never cleared it on rejection. The index.ts retry loop called
initAuth() up to 10 times, but on attempt 2+ the check
`if (authInitPromise) { await authInitPromise; return; }` would immediately
re-throw the original rejection without performing a real retry.
Fix: wrap the final `await authInitPromise` in try/catch and reset
authInitPromise = null on error. This allows the index.ts retry loop to
create a fresh attempt on each call after a failure.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Previously the initAuth() function set authInitPromise to the async IIFE's
Promise but never cleared it on rejection. The index.ts retry loop called
initAuth() up to 10 times, but on attempt 2+ the check
`if (authInitPromise) { await authInitPromise; return; }` would immediately
re-throw the original rejection without performing a real retry.
Fix: wrap the final `await authInitPromise` in try/catch and reset
authInitPromise = null on error. This allows the index.ts retry loop to
create a fresh attempt on each call after a failure.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Merges feature/GRO-2652-boot-econnreset-resilience into dev. CI passed (run #440). Fixes PROD CrashLoopBackOff caused by bare top-level await initAuth() + unhandled ECONNRESET in auth_provider_config DB query.
Gitea OCI registry sporadically rejects cache blob writes with
"error writing layer blob: unknown" — this fails the build step
even though the image itself was pushed successfully. Adding
ignore-error=true makes cache write failures non-fatal so the
build proceeds regardless of registry-side cache issues.
Fixes recurring CI failure on Build and push Seed image step.
Previously: await initAuth() was a top-level ESM await. Any boot-time
ECONNRESET from Postgres propagated as an uncaught module-evaluation
error and killed the process before the server even started.
Now:
- serve() starts immediately so /health and public routes are available
- initAuth() runs in a retry loop (up to 10 attempts, exponential
500 ms → 30 s); a permanent failure degrades to 503 on auth routes
(existing catch-to-503 in authRouter) rather than crashing the pod
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The auth_provider_config DB query at boot has no error handling;
a transient ECONNRESET causes authInitPromise to reject, propagating
to the top-level await initAuth() and crashing the process (exit 1).
Add up to 5 retry attempts with exponential backoff (1 s, 2 s, 4 s, 8 s)
so a single connection reset does not abort initialization.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The OOBE flow on the web portal calls this endpoint to create a fresh
`clients` row bound to the Better Auth user's email when the SSO
bridge returns 404. Returns 201 on success, 409 if a client with that
email already exists (portal-selection case), 401/503 on auth issues,
400 on invalid body.
The OOBE success path navigates the user back to `/` and lets the
existing `session-from-auth` re-bridge; the new client is now
resolvable by email, so the bridge mints a real portal session.
Tests cover: 401 (no session), 400 (zod), 201 + persisted values
(name trimmed, optional fields normalized to null), 409 (existing
client or unique-constraint race), 503 (auth not configured).
Paired with the web PR on `feature/2357-p2-sso-to-oobe-routing`.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Cosmetic follow-up to GRO-2319 (Phase 4 review by CTO). The synthetic
waitlist card on GET /portal/appointments returned service: {id} only,
so the portal fell back to the literal 'Service' label. CMPO spec did
not call for a service name on the waitlist card, but populating the
real name is non-urgent and closes the cosmetic gap.
- src/routes/portal.ts: include a services SELECT (in addition to
pets and staff) covering both appointment and waitlist serviceIds.
serviceMap feeds a service.name lookup. The synthetic waitlist
card's service object is now {id, name} — same shape the
appointments join returns — so the portal renders the real name.
The appointments join also gains a name (consistent shape, no
regression for the existing path).
- src/__tests__/portal.test.ts: mock the services table and assert
service: {id, name} on both the synthetic waitlist card and the
appointment card.
- UAT_PLAYBOOK.md: TC-API-8.20 covering the waitlist card service
name (TC-API-8.19 retained verbatim for the original GRO-2319
surfacing contract).
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Serialize the entire db:reset chain (DROP → migrate → seed) inside one withSeedAdvisoryLock callback so a concurrent same-PRNG seeder cannot interleave and collide on invoices_pkey. Pool sized max:6 (1 reserved for the lock + work headroom) to avoid the connection-starvation deadlock the CTO caught. Verified with three end-to-end live db:reset runs against a throwaway Postgres.
cc @cpfarhood
Phase 1.2 of Route Optimization. Adds a provider-agnostic geocoding
service layer in the deployed src/ tree:
- GeocodingProvider interface + GeocodeResult type
- NominatimGeocodingProvider (default, free, self-hostable) with an
internal rate limiter enforcing the 1 req/sec Nominatim usage policy
- GoogleGeocodingProvider (optional fallback) keyed by the encrypted
businessSettings.googleMapsApiKey (decrypted via decryptSecret) or
GOOGLE_MAPS_API_KEY env fallback
- resolveGeocodingProvider() selecting on businessSettings.routeOptimizationProvider,
with safe fallback to Nominatim when google is configured but no usable key
- geocodeBatch() throttled batch utility (honors provider rate limit,
captures per-item errors, optional progress callback)
- 20 unit tests covering both providers, selection, throttle spacing, and batch
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Flea Flicker
Flea Flicker