[kitlangton]

Bun and Node now use Effect's own SQLite clients instead of ours

Before: packages/core carried its own SQLite client (database/sqlite.ts) with thin Bun and Node drivers on top. On Node every SQLite error came back as UnknownError, so a duplicate-key insert looked the same as a corrupt database. On Bun, opening with readonly: true threw, because we always passed create: true.

After: #sqlite still resolves per runtime, but the Bun and Node entries are now short shims over @effect/sql-sqlite-bun and @effect/sql-sqlite-node 4.1.0. Errors are classified on both (a duplicate key is a ConstraintError), read-only opens work, and Node caches prepared statements. sqlite.ts is deleted. The workerd client stays ours for now; it no longer imports sqlite.ts and keeps its scoped-transaction fix.

  // database/sqlite.bun.ts (sqlite.node.ts is the same shape)
- 100 lines: our own connection, semaphore, transaction and error mapping on top of sqlite.ts
+ import { SqliteClient } from "@effect/sql-sqlite-bun"
+ export const sqliteLayer = (config) => Layer.orDie(SqliteClient.layer(config))

  // database/sqlite.workerd.ts
- import { Sqlite } from "./sqlite.js"   // shared Native service + connection helpers
+ inline Connection over options.storage; die with an Error when given a path instead of storage

What changes in behavior

• sql.withTransaction now opens with BEGIN IMMEDIATE instead of BEGIN. Direct callers: none in Bun or Node production code. The only one in src is drizzle/effect-sqlite/session.ts:141, on the workerd native path, which still uses our client. The rest are tests (sqlite-interrupt, sqlite-workerd).
• Drizzle db.transaction is unchanged: it still reserves a connection and issues its own begin deferred and savepoints. Covered by the drizzle and new Node tests.
• The upstream clients set busy_timeout before the WAL pragma and retry WAL on a lock. The pragmas in database.ts (WAL, synchronous, busy_timeout, cache_size, wal_checkpoint, foreign_keys) still run after open.
• @effect/sql-sqlite-bun moves from devDependencies to dependencies, and @effect/sql-sqlite-node is new (root catalog 4.1.0).

Bundles: sqlite-bundle.test.ts now checks that Node and default builds include @effect/sql-sqlite-node but never bun:sqlite or @effect/sql-sqlite-bun, and that workerd includes neither upstream client. The Node CLI bundle (build-node.ts --bundle-only) builds, starts, and has no bun:sqlite.

Tests: new sqlite-node.test.ts builds a Node bundle and runs it under node: WAL is on, a Drizzle transaction commits two rows, a duplicate insert fails with ConstraintError, Drizzle's typed insert().returning() and select() work (including a constraint failure mid-returning() followed by a working query), and the queued-transaction and queued-reserve interrupt sweep from sqlite-interrupt.test.ts leaks no connection. A Bun case checks the same error and reopens the file read-only. Both failed on dev with the pinned Bun 1.4.4 (UnknownError, and Bun's "read-only database cannot have create: true"); older Bun versions accept that read-only open, so the Bun case only catches it on 1.4.4 and later. sqlite-interrupt.test.ts passes unchanged against the upstream Bun client.

median of 5 interleaved runs, warm 100k-row DB	dev p50	this PR p50	dev ops/s	this PR ops/s
Node point read	13.6 µs	7.9 µs	23.6k	28.0k
Node Drizzle 5-row tx	89.4 µs	72.5 µs	8.4k	9.6k
Node withTransaction 5-row tx	100.0 µs	85.0 µs	7.2k	8.4k
Bun point read	10.3 µs	10.4 µs	26.0k	25.1k
Bun Drizzle 5-row tx	81.2 µs	81.7 µs	7.2k	7.7k
Bun withTransaction 5-row tx	100.4 µs	94.8 µs	5.9k	6.7k

Bun is flat within noise; Node is faster, mostly from the statement cache. Not measured: cold reads on a large database, and lock contention with another process holding the file.

👀 Read the BEGIN IMMEDIATE note above and sqlite.workerd.ts; everything else is deletion or tests.

---

Added by the opencode bot (AI-generated).

Related: #53879

[opencodejr]

The change looks sound. On dev, the new Node test fails with UnknownError. With this PR it gets ConstraintError, WAL is on and the Drizzle transaction commits. I checked this by running the PR's test against both checkouts. The Drizzle, workerd, interrupt, bundle, database, migration and session tests pass (825 tests), and typecheck:tsc in packages/core is clean. @effect/sql-sqlite-node 4.1.0 is built on node:sqlite with no native addon or extra dependencies, so the Node build doesn't gain a native module. BEGIN IMMEDIATE applies only to writable clients, and as you note, no Bun or Node production code calls withTransaction directly.

Public API is unchanged. Only internal #sqlite modules in packages/core are affected, plus the dependency changes.

Two small notes, neither blocking:

• The Bun case in sqlite-node.test.ts also passes on dev under Bun 1.4.2, including the read-only reopen. On this Bun version it doesn't catch the regression described in the PR.
• Nothing passes timeout to sqliteLayer (database.ts only passes filename), so the timeout → busyTimeout alias in sqlite.node.ts is unused. The mapping itself is correct.

Related: #53879

[opencodejr]

Both earlier notes are resolved. The new commit removes the unused timeout alias from sqlite.node.ts, which now passes its config straight to SqliteClient.layer. The description now says the Bun read-only case only catches the regression on Bun 1.4.4 and later, which matches what I saw on 1.4.2. The Node, bundle and database tests pass on the new head, and typecheck:tsc in packages/core is clean. Public API is unchanged.

[opencodejr]

The new commit adds tests and fixes a comment. The added Drizzle typed insert().returning() and select() checks pass on Node, including a working query after a constraint failure. The new tests passed in 5 consecutive runs, and typecheck:tsc in packages/core is clean. Public API is unchanged.

One suggestion (inline): the copied interrupt sweep in the Node fixture could be dropped, to keep that test small.

[opencodejr]

This copied sweep is the most intricate part of a test that otherwise checks simple, deterministic results. It depends on MaxOpsBeforeYield and counted yieldNows, and runs 62 fresh client layers inside a subprocess. Your note in sqlite-interrupt.test.ts says the Node client shares makeSqliteAcquirers with the Bun client, which is already swept there, so it adds little coverage. I'd drop it (and leakedTransaction/leakedReserve from the expected output) to keep the Node test focused on WAL, errors and Drizzle queries.