The inline node -e '...' blocks inside the YAML run: block had shell
quoting conflicts (nested parens, single-quote escapes) that caused
act/local tooling to fail with 'syntax error near unexpected token ('.
Split into two dedicated scripts:
- scripts/preflight-check.js — neon_auth schema check + 0001_init.sql tracking repair
- scripts/postflight-check.js — admin_users table verification after migrations
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The "Run migrations" job (and thus the whole deploy) was failing on every push after the first successful bootstrap with:
✗ 0001_init.sql failed: relation "admin_users" already exists
Root causes:
- 0001_init.sql used plain CREATE TABLE/INDEX/TRIGGER despite the header comment claiming "CREATE TABLE IF NOT EXISTS".
- The _migrations tracking row for 0001_init.sql was never recorded on the prod DATABASE_URL (tracking logic landed after initial apply, or apply happened outside the runner).
Fix (3 layers of defense):
* db/migrations/0001_init.sql: every CREATE TABLE now uses IF NOT EXISTS (63 tables), CREATE INDEX uses IF NOT EXISTS (49), all CREATE TRIGGER wrapped in safe DO $$ IF NOT EXISTS (pg_trigger join check) THEN ... END IF; $$, removed the file's own BEGIN/COMMIT so the runner's transaction is authoritative. RLS drop+create and CREATE OR REPLACE FUNCTION were already safe.
* scripts/migrate.js: added ensureTracked() right after loading the applied set. For 0001/0002, if the core objects (admin_users table, password_hash column) already exist in information_schema but the tracking row is absent, INSERT the filename (ON CONFLICT DO NOTHING) and log "repaired tracking". Then the normal skip path handles it.
* .gitea/workflows/deploy.yml: added an inline node -e pre-repair (same admin_users existence check → force _migrations row) immediately before `npm run migrate:one`. This protects even on an older checkout of the runner. The neon_auth preflight and post-apply "admin_users must exist" hard gate are untouched.
* Minor: cleaned 0002_admin_password.sql tx wrapper for consistency; updated MEMORY.md with the incident + resolution.
Result: deploys (and any server-side `node scripts/migrate.js` recovery runs) on an already-initialized prod DB will repair/skip 0001 cleanly, pass the verification query, and continue to build/deploy. The scp of scripts/ + db/migrations/ in the Deploy step now carries the fixed versions.
See the plan note in deploy.yml and MEMORY.md for background.
- getAdminUser now properly supports platform_admin with 0 brand links and loads full brand_ids
- createAdminUser action now inserts into admin_user_brands join table
- Admin layout surfaces the signed-in email on Access Denied
- AdminAccessDenied links to /login instead of dead-end /admin
- Main dashboard uses direct pool query instead of dead supabase shim
- Improved provision-admin.ts script for prod bootstrap (loads .env.production too)
The membership query in getAdminUser() was selecting role from
adminUserBrands.adminUserId, which is wrong - admin_user_brands has no
role column. Role is stored in admin_users. Also added try/catch
around withPlatformAdmin to prevent DB errors from throwing.
- Add MinIO/S3-compatible storage client (src/lib/storage.ts) with uploadObject,
deleteObject, presigned URL helpers, and BUCKETS constant
- Wire product images, brand logos, and water log photos to MinIO via the
new storage client
- Migrate forgot-password to Neon Auth (remove Supabase /auth/v1/recover call)
- Migrate send-scheduled cron to direct Postgres + Resend (remove Supabase Edge
Function proxy)
- Add logoUrl to email types (OrderReceipt, Welcome, PasswordReset) and pass
brand_settings.logo_url from all call sites
- Update email templates to use dynamic logoUrl instead of hardcoded Supabase
bucket URLs
- Remove hardcoded Supabase URLs from TuxedoVideoHero, TuxedoAboutPage,
TimeTrackingFieldClient; use brand_settings props + local public/ fallback
- Download brand logos (3) and tuxedo-hero.mp4 (36MB) from Supabase bucket to
public/ for local development
- Add MinIO env vars to .env.example (endpoint, access key, secret, buckets)
- Fix TimeTrackingFieldClient to destructure logoUrl and brandAccent props
- Fix admin/users.ts logoUrl type (null → undefined for optional string)
- Remove stale sb- cookie from wholesale-auth
- Migrate tuxedo/about page to remove supabase import and use pool query for
wholesale_settings lookup
The products list page at /admin/products still queried the legacy Supabase
mock client, which returns an empty array. Combined with the SaaS schema
renaming brands → tenants, the platform admin's brand picker for new
products was silently failing (getBrands SELECTed from a non-existent
'brands' table), making 'Add Product to catalog' look like a no-op.
- src/app/admin/products/page.tsx: replace supabase.from('products') with
a Drizzle query using withPlatformAdmin (cross-tenant) or
withTenant(brandId, ...) (scoped). Map new columns back to the legacy
Product shape the existing UI consumes:
* price_cents (integer) → price (dollars)
* tenant_id → brand_id
* product_images LEFT JOIN for first image per product
* default type='pickup', is_taxable=false (columns not in SaaS schema)
- src/actions/admin/users.ts (getBrands): query tenants table instead of
the legacy brands table.
Verified locally against the seeded dev DB: 8 products across 2 tenants
now render in the catalog, and the brand picker populates with both
tenants for platform admins.
Previously, any non-empty AUTH_GOOGLE_ID + AUTH_GOOGLE_SECRET would
trigger the Google button. That broke dev/CI setups where the env
vars are placeholder strings (e.g. 'dummy-google-client-id') — the
button would render and immediately 401 from Google with 'invalid_client'.
Real Google OAuth client IDs always end in '.apps.googleusercontent.com'.
Gate hasGoogle on that suffix so the login page falls back to the
credentials form (or the 'not configured' message) when the values
aren't real.
The seed script previously imported hashPassword from src/lib/passwords.ts,
which has 'import "server-only"' at the top. That guard throws when the
file is loaded outside a Next.js server runtime (which is what 'tsx db/seed.ts'
is — plain Node). Inline the same scrypt format (matches verifyPassword)
so the seed runs from a CLI.
Gitea deploy failed with:
Module not found: Can't resolve '@stripe/stripe-js'
Module not found: Can't resolve '@stripe/react-stripe-js'
These were imported by src/components/storefront/StripeExpressCheckout.tsx
(transitive: src/lib/stripe-client.ts → src/app/checkout/CheckoutClient.tsx)
but never declared in package.json. Add them so the Gitea build can resolve
the imports.
- playwright.config.ts: add testMatch pattern so playwright only picks up
*.spec.ts (skipping tests/unit/*.test.ts which are vitest's domain).
Before this, 'npx playwright test' would fail trying to run vitest tests.
- .gitignore: exclude test-results/ and playwright-report/ (generated by
'npx playwright test', not source files).
Final wave of the Full Supabase migration. All 114 files with Supabase
REST calls have been converted to Drizzle ORM + raw pg.Pool queries.
Verification:
- npx tsc --noEmit: clean (0 errors)
- npm run test: 22/22 pass
- npm run build: succeeded
Zero @supabase imports or rest/v1 references remain in src/.
- analytics.ts: rewrite getReportsSummary, getRevenueChart, getSalesByProduct,
getContactGrowth, getRecentOrders, getConversionFunnel against pool + new
orders/customers schema. Drops retired columns (subtotal, pickup_complete)
and re-implements the SQL by hand.
- import-orders.ts: bulk import via withTx using orders + orderItems + customers
Drizzle tables, computes total_cents from current product prices.
- import-products.ts: rewrite to use withTenant(brandId) and Drizzle products
table.
- products/create-product.ts, update-product.ts, upload-image.ts: switch to
withTenant + Drizzle; image_url moves to product_images table.
- reports.ts: rewrite against pool + new orders schema.
- route-trace/lots.ts: stub functions (route-trace feature retired from SaaS
rebuild — harvest_lots table not in db/schema). Uses discriminated union
return types so consumer narrowing works in both branches.
- settings/features.ts: switch to withTenant + Drizzle brandSettings.
- shipping.ts: switch to pool + Drizzle orders/orderItems.
- api/v1/referrals/route.ts: fix typecheck (referred_user_id undefined → 'anonymous').
Typecheck: clean. Tests: 22/22 pass. Build: succeeds.
Next.js 16 enforces 'one of middleware.ts OR proxy.ts' - both can't
coexist. Our src/middleware.ts (Auth.js v5) is the canonical one;
src/proxy.ts was the legacy dev_session auto-issuer.
- db/seed.ts: upserts admin@route-commerce.local with a scrypt password
hash (default password 'admin', override with SEED_ADMIN_PASSWORD)
and grants them the platform_admin role on the Tuxedo tenant so
getAdminUser() resolves it for the Credentials provider's authorize
function
- src/app/login/page.tsx: reads ?error=... from the URL and passes
hasCredentials + seededEmail + error to the client so the form
pre-fills in dev and surfaces 'Invalid email or password' cleanly
- src/app/login/LoginClient.tsx: adds the email + password form below
the Google button, with a divider, dev-mode pre-fill, and local error
handling for client-side failures (server-side failures come back
through the ?error=... param)
- src/auth.config.ts: edge-safe Auth.js config (no pg, no Credentials)
imported by the middleware; exports authConfig + isDevLoginEnabled
- src/lib/auth.ts: extends authConfig with the Credentials provider
(Node-only) for email + password sign-in via users.password_hash
- src/middleware.ts: uses authConfig (edge-safe) instead of the full
src/lib/auth.ts so the edge bundle stays small
- src/actions/auth-actions.ts: adds signInWithCredentials that calls
the Credentials provider; signInWithGoogle stays the production path
- Migration 0002 adds nullable password_hash to users (idempotent)
- src/lib/passwords.ts: encode/verify with self-describing format
(algo$N$salt$hash) so we can migrate to a stronger KDF later
- Schema adds passwordHash column; OAuth-only users leave it null
- Create src/lib/db.ts (shared pg.Pool, query/withTx helpers, server-only)
- Add @types/pg devDep
- Migration 204: add email, auth_provider, auth_subject columns to
admin_users; backfill from auth.users; new upsert_admin_user accepts
multi-provider args; new get_admin_user_for_session RPC resolves
Auth.js session id (UUID or Google sub) to an admin row
- Refactor getAdminUser() to use pg + new RPC; auto-provisions on first
Google sign-in using session.user.email
- Refactor updatePasswordAction to call update_user_password via pg
(drops the Supabase REST hop)
- Delete orphaned src/actions/login.ts (replaced by auth-actions.ts)
- Drop remaining DEV_FORCE_UID references in users.ts; dev path now
uses dev_session cookie (the only cookie the dev flow can set)
- Update AdminUser type: user_id is now string | null (Google users
have no Supabase auth id); add email + auth_provider fields
- Fix downstream type errors: StopProductAssignment.callerUid accepts
null; pickup.ts performedBy widens to string | null
- Bump vitest config, tests/, and other earlier cleanup changes
The /admin page was throwing a Server Components render error (digest
4266906817) in production. The exact throw point was not visible from
the browser (Next.js hides it for security) and server logs were not
available at debugging time.
The layout and page both call data-loading functions (getActiveBrandId,
listBrandsForAdmin, supabase brands lookup) that could throw on a
transient DB/network failure, and these calls had no try/catch. A
single failed call would crash the entire admin shell.
Wrap each call in try/catch with console.error so:
1. The page renders with sensible defaults if a call fails
2. The actual error is logged server-side (visible via the digest in
the admin error boundary or PM2/Docker logs)
3. The admin shell stays functional even if a single data source is
down
No behavior change on the happy path.
Cleanup after Auth.js v5 became the only sign-in path. The platform
had three overlapping auth modes (dev cookie, legacy rc_auth_uid, Auth.js
JWT) and a pile of dead-code pages/routes that only existed to support
the legacy path.
What changed:
* getAdminUser() now has only two auth paths:
1. dev_session cookie (auto-issued by src/proxy.ts for /admin/* when
ALLOW_DEV_LOGIN is enabled)
2. Auth.js v5 JWT (the encrypted cookie + auth() lookup)
The legacy rc_auth_uid/rc_uid branch and the Supabase REST fetch
against admin_users are gone.
* The signIn callback in src/lib/auth.ts enforces ADMIN_ALLOWED_EMAILS
when set. Unset = open mode (backward compatible with demo/dev). Dev
credentials provider is exempt. The new env var is wired through
.env.example and .gitea/workflows/deploy.yml (read from
secrets.ADMIN_ALLOWED_EMAILS, written to the server .env file).
* change-password/page.tsx now uses auth() server-side instead of
fetching the deleted /api/auth/uid endpoint. The form is split into
page.tsx (server component, auth check) + ChangePasswordForm.tsx
(client component, form state). updatePasswordAction now reads the
user id from auth() instead of the rc_auth_uid cookie.
* Deleted 14 dead-code files:
- Pages: login2, logout, auth/callback, admin/debug-auth,
admin/test-auth
- API routes: api/login, api/logout, api/auth/uid, api/force-admin,
api/set-auth-cookie, api/debug-cookie, api/debug-me,
api/debug-auth
- Actions: src/actions/login.ts
These were the old email/password login, the old Supabase OAuth
callback, the old /api/auth/uid probe, and a pile of debug endpoints
that have been superseded by the new proxy + the new /login page.
* next.config.ts: set outputFileTracingRoot: '.' to silence the
Next.js 16 lockfile-inference warning. Without this the build
walked up from package.json looking for a lockfile, found the
homelab runner's stale act cache at /home/tyler/.cache/act/.../package-lock.json,
and warned on every build. '. resolves to the project root in both
dev and CI, so it's the right answer.
Out of scope (deferred):
* src/actions/admin/users.ts still uses rc_auth_uid internally for its
dev-bypass logic. It works (the rc_auth_uid branch is gated on
NODE_ENV != 'production' and DEV_FORCE_UID), but it's now genuinely
unreachable in production. Clean up in a follow-up.
Pre-flight:
* npx tsc --noEmit: clean
* npm run lint (touched files): clean
* npm run build: clean — proxy picked up, no lockfile warning, all
93 static pages generated.
Next.js 16 renamed middleware.ts to proxy.ts and only allows one of the
two. Delete src/middleware.ts and rewrite src/proxy.ts to do the dev
auto-login + /login bounce + /admin gate explicitly (no NextAuth auth()
wrapper — auth.handler() returns a NextMiddleware taking (req, event),
which makes wrapping it awkward when we also need to set cookies on the
response).
Drop the now-dead 'authorized' callback from auth.config.ts (it was only
fired by the NextAuth wrapper, which we no longer use).
Migration 209: add a defensive unique constraint on admin_users.user_id
so the ON CONFLICT (user_id) clause in the new RPC resolves regardless
of whether the Supabase-dashboard-created table shipped with one.
The 'cp -f deploy/docker-compose.yml ...' was sitting AFTER
'docker compose down' in the step. 'docker compose down' reads
and validates the compose file on the server — so the old copy
(with the dead nextjs service and its env_file: ../.env.production
reference) was still being read, and docker compose bailed on the
missing .env.production file before the copy could overwrite it.
Moved the config-file seeding to the top of the step, right after
'mkdir -p $APP_DIR'. Now the new compose file is in place before
either 'docker compose down' or 'docker compose up' runs.
The 'Start Docker stack' step used '[ -f ... ] || cp' for
docker-compose.yml, which only copied the file on the first
deploy. Subsequent deploys kept the stale copy on the server.
The stale copy still had the dead 'nextjs' service with
'env_file: ../.env.production', which docker compose validates
on every 'up' and bailed because .env.production is written
later by the 'Deploy' step.
Changed to unconditional 'cp -f' so the server always has the
latest compose file.
Build was failing on the 'Start Docker stack' step with two issues:
1. PGRST_DB_URI not set — the env var was only in the 'Deploy' step,
which runs after PostgREST has already started. PostgREST booted
with a blank DB URI and the step exited 1.
2. docker-compose.yml had a 'nextjs' service with
env_file: ../.env.production, but .env.production is written
later by the 'Deploy' step. docker compose validates the entire
compose file on 'up' and bailed because the path didn't exist
yet.
The 'nextjs' service is dead code anyway: PM2 runs Next.js
directly from $APP_DIR, never through docker. Removed it.
Also fixed: 'docker compose up -d db postgrest minio minio_init'
referenced services that don't exist in the compose file (Postgres
runs on the host, not in docker). Changed to just 'postgrest', and
the pg_isready check now uses host psql directly instead of
'docker compose exec -T db'.
Changes:
- deploy/docker-compose.yml: drop nextjs service, keep only postgrest
- .gitea/workflows/deploy.yml:
- Add PGRST_DB_URI / PGRST_DB_ANON_ROLE / PGRST_SERVER_PORT to
the 'Start Docker stack' step env
- Write them to $APP_DIR/.env so docker compose picks them up
- 'docker compose up -d postgrest' (was: db postgrest minio minio_init)
- pg_isready check uses host psql (was: docker compose exec -T db)
Tyler
openclaw
tyler