SQLite for Local Dev — Postgres-Portable, Bun + Drizzle

🎯 Why: Design for Change

The goal of writing software is to be able to change it safely — including changing the database. Every rule here (snake_case, plural tables, NOT NULL FKs, JSON document store) exists so the schema ports to Postgres with a dialect swap, not a rewrite. Portability is change-safety: when SQLite stops fitting, you move with a Drizzle config edit instead of a month of forensics.

SQLite here is the development database for a TypeScript/Bun app whose production system of record will almost certainly be Postgres. That single fact drives every rule: write the SQLite schema so that switching to Postgres is a Drizzle dialect change and a data copy, not a redesign. Use the same names, the same modeling discipline, and the same integrity rules as postgres-dba — SQLite is just a smaller engine running them.

This skill is about the SQLite schema, the Drizzle layer, and operating SQLite. For schema-design philosophy (why NOT NULL FKs, why a surrogate key) the authority is postgres-dba; this skill does not re-argue it, it ports it. For app-layer TypeScript use the language skills.

How to use this skill

1. Match the task to a rule below or in the decision guide.
2. Open the matching references/*.md for the rationale, the wrong way, the right way, and the explicit-override escape hatch.
3. Copy the closest file from templates/ and adapt it — the templates already encode every convention, so you start compliant and portable.
4. Apply the rule unless you can state, in a code comment, the specific reason it does not apply. "SQLite let me" is not a reason — SQLite lets you do almost anything; Postgres will not.

The hard rules (non-negotiable defaults)

1. Naming is identical to postgres-dba. snake_case, lowercase, unquoted. Tables plural (orders), columns singular (shipped_at), primary key literally id, FKs <table_singular>_id, booleans is_/has_, timestamps past-tense _at. In Drizzle the TypeScript property may be camelCase, but the column name argument is always the snake_case name — createdAt: integer('created_at', …). The database never sees a quoted mixed-case identifier. See references/naming.md.
2. bun:sqlite + drizzle-orm/bun-sqlite, one Database per process. That is the dev driver. Every connection runs the pragma block (Rule 3) at open. SQLite has one writer at a time; a single shared Database instance plus serialized writes is the model — do not invent a pool of writers. See references/drizzle.md.
3. Pragmas are mandatory and per-connection. Every connection sets, in order: journal_mode = WAL, foreign_keys = ON, busy_timeout = 5000, synchronous = NORMAL. foreign_keys is OFF by default in SQLite — forgetting it silently disables every ON DELETE you wrote. This is the single most common SQLite data-integrity bug. See references/drizzle.md.
4. Tables are STRICT, types come from the Postgres-portable set only. Declare STRICT (SQLite ≥ 3.37) so a column actually rejects wrong types instead of silently coercing. Use only the type vocabulary in references/types.md: INTEGER, TEXT, REAL, BLOB mapped to a concrete Postgres target. No VARCHAR(n), no DATETIME, no BOOLEAN, no NUMERIC — those are affinity theater in SQLite and lie about the Postgres column you will eventually create.
5. FKs are NOT NULL by default with explicit ON DELETE. Same rule as postgres-dba Rule 4. A nullable FK carries a // nullable-fk: <reason> comment or it is a defect. Many-to-many is a compound primary key, no surrogate id on a pure junction. See references/portability.md.
6. Enums are text + a CHECK (col IN (...)), mirroring the Postgres enum. SQLite has no enum type. Use Drizzle text({ enum: [...] }) and an explicit check constraint so the values match the CREATE TYPE you will write in Postgres. Booleans are integer({ mode: 'boolean' }) + CHECK (col IN (0,1)). See references/portability.md.
7. Timestamps are one chosen representation, app-set, always UTC. SQLite has no timestamptz. Default: integer epoch-milliseconds via Drizzle { mode: 'timestamp_ms' }, set by the app, never CURRENT_TIMESTAMP (which emits a non-ISO, tz-ambiguous string). This ports to Postgres timestamptz mechanically. See references/types.md.
8. JSON is text({ mode: 'json' }) with a relational spine, mirroring the JSONB rule. Keys/FKs/hot fields are real columns; open-ended data is a JSON document, hot fields lifted out via a GENERATED ... STORED column and indexed. This is postgres-dba Rule 8 with jsonb → text JSON. See references/portability.md.
9. Set-based / multi-row business logic lives in an explicit transaction in a named module — not scattered across callers. This is the one place SQLite cannot match postgres-dba Rule 7 (no plpgsql). Quarantine that logic in one repository/service function wrapped in a db.transaction(...) so the Postgres cutover has exactly one place to consider promoting to a function. See references/portability.md.

Decision guide

Templates

• templates/schema.ts — a Drizzle SQLite schema demonstrating every rule: snake_case columns under camelCase keys, id surrogate key, a text+CHECK enum, integer boolean, epoch-ms timestamps, NOT NULL FKs with explicit onDelete, one annotated nullable-FK override, a compound-key junction, a STORED generated column, and a JSON body with a lifted+indexed hot field.
• templates/db.ts — the dev connection: bun:sqlite + Drizzle, the mandatory pragma block applied at open, and startup migration.
• templates/db.production.ts — the production connection: pragmas tuned for a server (mmap, cache), single-writer discipline, IMMEDIATE transaction helper, and a VACUUM INTO backup hook.
• templates/drizzle.config.ts — drizzle-kit config for the SQLite dialect, with the one-line change that points it at Postgres later.

What this skill will not do

• Bless a SQLite-only construct that has no Postgres equivalent in the schema (e.g. relying on rowid aliasing, type affinity coercion, or WITHOUT ROWID for a normal table). If Postgres can't express it, it doesn't belong in a portable schema.
• Bless foreign_keys left at the default. An app that doesn't set the pragma has no referential integrity, full stop.
• Bless CURRENT_TIMESTAMP / DATETIME('now') for stored timestamps — the format is not ISO-8601 and not tz-aware (Rule 7).
• Bless a multi-writer connection pool against one SQLite file.
• Bless SQLite in production by default. It can be production — under the explicit conditions in references/production.md, and only there.
• Re-derive schema-design philosophy. That is postgres-dba's job; this skill ports its conclusions to a smaller engine.