[!CAUTION] Synced from https://github.com/nrjdalal/zerostarter. Customize this skill or remove this note to stop syncing.

Database Migration

PostgreSQL with Drizzle ORM. Schema lives in packages/db/src/schema/, migrations in packages/db/drizzle/. Schema, SQL, and snapshot travel together in one PR.

0. Show the schema and wait

Never generate or apply a migration before the user has seen the schema and said go. Post the table or the altered columns, name the choices that are hard to reverse (nullability, what a null means, foreign keys and their delete behaviour, indexes, whether anything is derived or generated, retention), and stop there.

Two reasons this is a hard stop rather than a courtesy. A shape is cheap to argue about and expensive to change once rows exist, so the review has to happen before the write. And canary and production share one database, so applying reaches production the moment it runs, not when the release ships.

1. Edit the schema

• auth.ts is generated, never edited: declare a plugin or an additionalFields column in packages/auth/src/schema.ts, run bun run auth:schema, then continue below with the regenerated file. A hand edit fails tests/packages/db/src/schema/auth.test.ts.
• Schema files group by concern, not one per table: auth.ts holds the Better Auth tables, console.ts the ones the console owns. A new table joins the file for its concern, or starts a new one when it is a new concern, and that file needs an export from index.ts: export * from "@/schema/<name>". Miss that export and the table never reaches a migration.
• For examples, read auth.ts (tables, relations, indexes) and waitlist.ts (a minimal non-auth table).
• Conventions: text primary keys (.$defaultFn(() => crypto.randomUUID()) on non-auth tables), timestamp("created_at").defaultNow().notNull(), snake_case columns, onDelete: "cascade" on FKs, and an index() on every FK column. Both bend for the same kind of column: use onDelete: "set null" when the row outlives the reference and the column is only provenance, and skip the index when nothing queries by it and the table is small enough that the delete-time scan is free. The console.ts tables are both: their actor_id must not take a rule or a history entry with it when an account is deleted, and although the lists join, search and sort through it, every one of those drives off user's primary key rather than this column, so an index here would serve only the delete-time sweep. Index a FK when a query narrows on the column itself, or when the table is large enough for the sweep to matter. When a set null column is the only record of who acted, store the readable text beside it (activity.actor, allowlist.actor) so a row does not become anonymous, and can name an actor that was never an account at all.

2. Generate and review

bun run db:generate

Read the generated packages/db/drizzle/NNNN_*.sql. Done when that SQL, its meta/NNNN_snapshot.json, and a new meta/_journal.json entry all appear and the SQL matches the schema edit.

3. Point at a throwaway database first

Local work runs against a disposable container, never the POSTGRES_URL in the root .env. That URL is the shared Neon database, and canary and production both read it, so a migration applied there is applied to production, and a column renamed under running code takes that code down. This is not hypothetical: it has happened, and step 0 and this step both exist because of it.

bunx pglaunch -k                                    # disposable Postgres, -k keeps it across restarts

Put the URL it prints in the worktree's own .env as POSTGRES_URL, then migrate and seed into it. The output is colorized, so a script that captures it must strip the escapes first (| sed $'s/\033\\[[0-9;]*m//g'): an escape left on the end of the URL makes drizzle-kit migrate exit 1 with no message at all. One container per worktree, so two branches with different migrations cannot corrupt each other. Two things follow:

• Seed your own fake rows rather than copying real ones. The shared database holds real people, and a screenshot or a response body from it must never reach a PR.
• Applying to the shared database is a deliberate, separately-authorized act. It happens when the change merges, not while it is being built.

4. Apply

bun run db:migrate

This is local or ad-hoc only. On production and canary deploys the API build auto-applies pending migrations (.github/scripts/migrate-on-deploy.ts, gated on VERCEL_ENV/VERCEL_GIT_COMMIT_REF, PR previews skipped), so a migration merged to canary applies itself on the next deploy.

5. Make the running stack see it

bunx turbo run build --filter=@packages/db

The API consumes @packages/db's built dist. If dev is running and the API imports the new table, restart dev entirely: bun --hot does not pick up new files or exports reliably (see the dev skill).

6. Inspect data

bun run db:studio

Notes

• POSTGRES_URL comes from the root .env, which is why step 3 overrides it per worktree: the value that ships in .env is the shared one.
• Never edit an applied migration; generate a new one instead.