The FK conversion (teams.owner_id, commissioners.user_id → UUIDs) and
email_verified update were inside runClerkMigration() which only ran when
CLERK_SECRET_KEY was set. If the key wasn't in the environment at deploy
time, all Clerk IDs were left unconverted, causing UUID parse errors at
runtime.
Split into two functions:
- convertForeignKeys(): always runs, idempotent WHERE clauses
- importClerkOAuthAccounts(): only runs when CLERK_SECRET_KEY is set
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Migrate authentication from Clerk to BetterAuth (#322)
Replaces @clerk/react-router with self-hosted better-auth to eliminate
the external Clerk dependency and keep all user/session data in our own
PostgreSQL database.
**What changed**
- New: auth.server.ts (BetterAuth config w/ Drizzle adapter, bcrypt, Resend), auth-client.ts, api.auth.$.ts handler
- New: /login and /register pages with email+password and Google/Discord OAuth; open-redirect guard on redirectTo param
- New: UserMenu component replacing Clerk's UserButton
- Schema: sessions, accounts, verifications tables; emailVerified column; clerkId made nullable
- Migrations 0081 (BetterAuth tables) and 0082 (accounts extra columns for v1.6.9)
- All ~30 route files: getAuth → auth.api.getSession, isUserAdminByClerkId → isUserAdmin
- root.tsx: isAdmin read directly from session.user.isAdmin (no extra DB query)
- useDraftAuthRecovery: removed Clerk JWT refresh logic; replaced with cookie-session check
- models/user.ts: removed findUserByClerkId, findOrCreateUser, updateUserByClerkId (webhook pattern)
- Deleted: app/routes/api/webhooks/clerk.ts; uninstalled @clerk/react-router, @clerk/themes, svix
- scripts/migrate.mjs: extended with idempotent Clerk → BetterAuth data migration (FK conversion, email_verified, OAuth accounts)
- scripts/migrate-clerk-passwords.mjs: one-time script to import bcrypt hashes from Clerk CSV export
- BETTERAUTH_MIGRATION.md: dev and production runbooks
- All test mocks updated: vi.mock('~/lib/auth.server') instead of @clerk/react-router/server
- Test fixtures: added emailVerified field
**Follow-up (post-stable)**
- Rename actor_clerk_id column → actor_user_id in commissioner_audit_log
- Drop clerk_id column from users once migration confirmed
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Add .npmrc with legacy-peer-deps for better-auth/drizzle peer dep conflict
better-auth@1.6.9 declares peerOptional deps on drizzle-orm ^0.45.2 and
drizzle-kit >=0.31.4, but we run drizzle-orm ~0.36.3 / drizzle-kit ~0.28.1.
The adapter works correctly at runtime with our versions — the peer dep is
only for stricter type checking. This unblocks npm ci in CI without a risky
drizzle major-version upgrade.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Disables prepared statements when PGBOUNCER=true (required for PgBouncer
transaction mode). Adds DATABASE_DIRECT_URL so migrations always use a
direct connection, bypassing the pool.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
client.end() in the finally block was throwing (postgres driver already
cleaned up connections internally during migrate()), causing an unhandled
rejection and exit code 1 despite the migration succeeding.
- Use explicit process.exit(0)/exit(1) instead of relying on implicit exit
- Wrap both client.end() calls with .catch(() => {}) to tolerate cleanup errors
- Add onnotice: () => {} to suppress noisy NOTICE messages (schema/table
already exists) that are normal on every idempotent re-run
https://claude.ai/code/session_01ReaqH3o9NVH4QU4qE9WMMQ
* Use Docker Compose init service pattern for database migrations
Replaces the fragile double-migration approach (server.ts startup +
drizzle-kit CLI in CI) with a one-shot migrate service in Docker Compose
that the app depends on via condition: service_completed_successfully.
- Add scripts/migrate.mjs: programmatic drizzle-orm migration (uses
production deps, not drizzle-kit which is devOnly and absent from image)
- Dockerfile: copy scripts/ into image, remove drizzle.config.ts (only
needed by drizzle-kit CLI)
- server.ts: remove runMigrations() entirely; migrations are now handled
by the migrate container before the app starts
- deploy.yml: remove explicit docker run migration step; replace sleep 10
health check with a poll loop and explicit migrate exit-code check
Production server's docker-compose.yaml needs a one-time manual update
to add the migrate service — see plan for exact config.
https://claude.ai/code/session_01ReaqH3o9NVH4QU4qE9WMMQ
* Move drizzle-kit to devDependencies; use docker compose wait in deploy
drizzle-kit is a dev-only tool (schema generation and local migrations).
Now that production migrations run via the programmatic drizzle-orm API
in the migrate init container, drizzle-kit has no runtime role.
Also replaces the polling health check loop with docker compose wait,
which blocks until the migrate service exits and returns its exit code
cleanly — no sleep or manual status inspection needed.
https://claude.ai/code/session_01ReaqH3o9NVH4QU4qE9WMMQ
---------
Co-authored-by: Claude <noreply@anthropic.com>
Mirrors how DATABASE_URL is loaded, so users can store both in their
local .env instead of passing PROD_DATABASE_URL inline each time.
https://claude.ai/code/session_01EiYh5ZiuWAnpo1PND45Yi1
Co-authored-by: Claude <noreply@anthropic.com>
* Add prod-to-dev database sync script
Adds scripts/sync-prod-db.sh and npm run db:sync-prod to safely copy
production data to the development database. After restoring, the script
nulls all Discord webhook URLs and anonymizes non-admin user PII (email,
name, avatar) to prevent notification leaks and reduce PII exposure on
developer machines. Admin accounts are preserved intact.
https://claude.ai/code/session_01EiYh5ZiuWAnpo1PND45Yi1
* Fix six code review issues in sync-prod-db.sh
- Add upfront check for pg_dump/pg_restore/psql availability (#5)
- Fix password masking to use sed instead of broken bash glob (#6)
- Replace DROP/CREATE DATABASE with pg_restore --clean --if-exists,
which works on managed cloud databases without superuser access (#2)
- Remove dead ADMIN_DB assignment that was immediately overwritten (#3)
- Handle pg_restore exit code 1 (non-fatal warnings) gracefully
instead of aborting via set -e (#1)
- Count webhook URLs before the UPDATE for an accurate cleared count (#4)
https://claude.ai/code/session_01EiYh5ZiuWAnpo1PND45Yi1
---------
Co-authored-by: Claude <noreply@anthropic.com>
