SKILL.md
MX Migration Author
Migrations run as a release-phase step (see docs/superpowers/specs/2026-05-05-database-migration-release-phase-design.md). Schema changes must be backwards-compatible with the previous release's code, because old replicas keep serving traffic until rolling deploy finishes.
§A Invariants
- mx-core ships rolling deploy. New and old pods run simultaneously
during cutover.
- Therefore: release
N's migration must work with releaseN-1's
code without errors. The contract-side change (drop / rename / type change) must wait one release after the expand-side change.
mx-migrateruns to completion before anymx-corepod starts. The
app refuses to start if schema is behind. This means migration failures block the deploy rather than crash-loop replicas.
- Drizzle migrations run inside a transaction by default. CONCURRENTLY
index creation requires --> statement-breakpoint to break out.
§B Decision tree
Is the change additive (table / nullable column / index / FK)?
├── yes → use §C.1 (one release)
└── no → expand first, contract later — see §C.2 (multiple releases)§C.1 Expand templates (single release)
| Operation | Safe approach |
|---|---|
| Add table | CREATE TABLE (drizzle generates) — safe; old code has no reference. |
| Add nullable column | ADD COLUMN col TYPE — safe. |
| Add NOT NULL column | Three releases. Phase 1: add nullable + dual-write in code. Phase 2: backfill (separate SQL migration). Phase 3: ALTER COLUMN col SET NOT NULL. |
| Add NOT NULL with DEFAULT | ADD COLUMN col TYPE NOT NULL DEFAULT <expr> — single release, but consider table size: a wide DEFAULT rewrites the table. |
| Add index (small table) | CREATE INDEX is fine. |
| Add index (large table, est. >1M rows) | CREATE INDEX CONCURRENTLY. Must be its own statement and not inside a drizzle transaction. Annotate the breakpoint and the lint allow. |
| Add FK (large table) | Two releases: ADD CONSTRAINT ... NOT VALID, then VALIDATE CONSTRAINT. |
§C.2 Mutate / contract templates (multiple releases)
| Operation | Stages |
|---|---|
| Drop column | R1: stop reading/writing the column in code (deploy code only). R2: ALTER TABLE ... DROP COLUMN. |
| Rename column | R1: add new column. R2: dual-write (read old preferentially, write both). R3: backfill, switch reads to new. R4: drop old column. |
| Change column type | R1: add new column. R2: backfill + dual-write. R3: cut over readers/writers. R4: drop old. |
| Drop NOT NULL | One release — relaxing constraint is safe. |
| Add NOT NULL | See §C.1 row. |
| Drop table | R1: stop all references in code. R2: DROP TABLE IF EXISTS. |
| Drop index | One release — only loses performance; not a correctness break. |
§D Working with the lint guard
pnpm -C apps/core run lint:migrations flags dangerous patterns (no-drop-column, no-drop-table, no-bare-not-null-add, no-rename-column, no-alter-type, no-bare-create-index). The lint runs in CI on every PR.
If the lint flags a statement and the change is genuinely safe in this context (e.g. baseline cutover with no existing consumers, large-batch rebuild on a known-cold table), add a per-statement annotation above or on the same line as the statement:
-- migration-lint:allow=no-drop-column reason=baseline cutover, no consumers yet
ALTER TABLE users DROP COLUMN email;reason= is mandatory. Bare migration-lint:allow=* without reason= is itself a lint error. Reviewers must read the reason; do not slip allow annotations past review without justification.
§E Self-checklist (run after generating each migration)
[ ] Is this migration backwards-compatible with the previous release's code?
[ ] Does any DDL acquire AccessExclusive on a hot table? (DROP/RENAME/TYPE)
[ ] Indexes on large tables are CONCURRENTLY (and broken out of drizzle's tx)?
[ ] If a backfill is needed, is it staged so it does not block?
[ ] Is a follow-up contract migration tracked (issue / next-release plan)?
[ ] `pnpm -C apps/core run lint:migrations` passes locally?
[ ] Migration applied cleanly against an ephemeral testcontainer?§F Authoring workflow
- Edit
apps/core/src/database/schema/*.ts. - Run
pnpm -C apps/core exec drizzle-kit generateto scaffold the SQL. - Read the generated SQL. Restructure if the change implies a
contract operation — split into expand-now / contract-later migrations across releases.
pnpm -C apps/core run lint:migrations. Fix or annotate.- Apply locally:
pnpm -C apps/core run migrate. - Add tests covering the schema change (repository spec, e2e where
user-facing). For NOT NULL adds with default, also write a test that reads existing rows.
- Commit. PR title prefix:
migration:ordb:. PR body must include
the expand-contract phase plan.
§G What this skill must not do
- Decide unilaterally to add
migration-lint:allow=— that's a human
judgment call. Suggest the annotation and the reason, surface the trade-off, but require explicit user confirmation.
- Combine expand and contract into one migration, except for the
baseline (idx ≤ 0003).
- Suggest schema changes without first reading
apps/core/src/database/schema/ to understand current shape.
- Use
applyMigrationsor any boot-time mutation. Migration is a
release-phase step, end of story.
§H Cross-references
release-coreskill — release pipeline, version bump, image tag.mx-review— broader review conventions.api-conventions— controller / DTO conventions, often coupled to
schema.
zod-patterns— DTO shape often follows schema; update both.- Spec doc:
docs/superpowers/specs/2026-05-05-database-migration-release-phase-design.md.