Database Schema & Migrations

Plexus uses Drizzle ORM with SQLite (default) or Postgres. Schema files live in packages/backend/drizzle/schema/.

Schema Organization

drizzle/schema/
├── index.ts          ← must export every table; missing exports = "No schema changes" from drizzle-kit
├── postgres/         ← Postgres table definitions (PLEXUS_DB_TYPE=postgres)
└── sqlite/           ← SQLite table definitions (default)

Always edit the correct dialect subdirectory. When adding a new table, update drizzle/schema/index.ts to export it or drizzle-kit generate will silently report "No schema changes."

Migration Rules (non-negotiable)

NEVER edit existing migration files. Migrations represent the historical change sequence. Editing them causes out-of-sync history, data loss risk, and broken deployments.

NEVER manually create migration SQL files or edit meta/_journal.json. Drizzle-kit ignores migrations not in the journal. Manual creation causes conflicting migrations and corrupts the migration system.

NEVER modify a live database directly with SQL commands. Always use migrations.

The Only Correct Workflow

1. Edit schema .ts files in postgres/ or sqlite/.
2. Validate locally (optional): run bun run generate-migrations from the repository root and inspect the generated SQL. Leave generated artifacts in place and uncommitted; do not delete them.
3. Never run drizzle-kit generate directly.
4. When authorized to commit, include only the schema .ts changes, not generated SQL, journals, or snapshots. The pre-commit hook blocks migration artifacts.
5. After the PR merges to main, CI auto-generates and commits the migrations.

Migration Naming

Migrations must have a descriptive, semantic name. The wrapper script enforces this:

bun run generate-migrations                        # auto-derives name from branch
bun run generate-migrations --name add_quota_checkers  # explicit name

If --name is omitted, the script derives a name from the current git branch (e.g., pi/issue-424-1779050379120 → auto_issue_424). On the main branch, --name is required.

This produces files like 0044_add_quota_checkers.sql or 0044_auto_issue_424.sql instead of 0044_rare_skullbuster.sql.

• Use snake_case starting with a verb: add_, create_, drop_, rename_, update_, fix_, remove_, convert_, jsonb_, etc.
• Auto-derived names use the auto_ prefix.
• Do not use random or meaningless names. The lint:migrations script rejects them.

Bypasses (maintainers only):

• Pre-commit hook: ALLOW_MIGRATIONS=1 git commit
• PR check: add the migrations-ok label to the PR

Adding a New Table — Checklist

• Create the table definition in the appropriate dialect directory.
• Export the new table from drizzle/schema/index.ts.
• If Postgres, add any new enum values to postgres/enums.ts (e.g. quotaCheckerTypeEnum).
• If validating locally, run bun run generate-migrations --name <descriptive-name> from the repository root and inspect the SQL.
• Leave generated migration artifacts in place and uncommitted. When authorized to commit, include only the .ts schema changes.

Type Definitions

Infer types from the schema rather than writing them manually:

import { InferSelectModel, InferInsertModel } from 'drizzle-orm';
import { myTable } from '../drizzle/schema';

type MyRow    = InferSelectModel<typeof myTable>;
type NewMyRow = InferInsertModel<typeof myTable>;

Shared inferred types also live in packages/backend/src/db/types.ts.

Dialect-Aware Timestamp Conversions

SQLite and Postgres use different column types for timestamps. Never write inline dialect === 'postgres' ? ... : ... timestamp logic. Always use the shared utilities from src/utils/normalize.ts.

Pattern 1: text (SQLite) vs timestamp (Postgres)

Used by columns like mcp_request_usage.created_at, mcp_debug_logs.created_at.

• SQLite text('col') → Drizzle expects a string (ISO 8601)
• Postgres timestamp('col') → Drizzle expects a Date object

import { toDbTimestamp } from '../../utils/normalize';
import { getCurrentDialect } from '../../db/client';

const createdAt = toDbTimestamp(record.created_at, getCurrentDialect());
// SQLite  → "2026-02-09T17:36:14.297Z"  (string)
// Postgres → Date object

Pattern 2: integer(timestamp_ms) (SQLite) vs bigint(number) (Postgres)

Used by columns like quota_snapshots.checked_at, quota_snapshots.resets_at, quota_snapshots.created_at.

• SQLite integer('col', { mode: 'timestamp_ms' }) → Drizzle expects a Date object
• Postgres bigint('col', { mode: 'number' }) → Drizzle expects a number (epoch ms)

import { toDbTimestampMs } from '../../utils/normalize';
import { getCurrentDialect } from '../../db/client';

const checkedAt = toDbTimestampMs(result.checkedAt, getCurrentDialect());
// SQLite  → Date object
// Postgres → 1739122574297  (number)

Both functions accept Date | number | string | null | undefined and return null for invalid/nullish values.