mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-05-17 03:57:18 +00:00
ff63da2652
* refactor(serve): extract createInMemoryChannel helper from httpAcpBridge.test.ts (#4156 A1) Sub-PR A1 of issue #4156 (Stage 1.5b Mode A daemon). Pure refactor with zero behavior change. Extracts the inline paired NDJSON channel construction (`new TransformStream` × 2 + `ndJsonStream` × 2) that was duplicated across `httpAcpBridge.test.ts` into a production helper `createInMemoryChannel()` at `packages/cli/src/serve/inMemoryChannel.ts`. The helper is added to `packages/cli/src/serve/index.ts`'s barrel export alongside the rest of the serve module's public API. The helper is intentionally bare — it returns only the stream pair, no lifecycle / teardown surface. Two reasons: 1. Consumer behavior diverges widely (stuck channel, crashable child simulation, no-op, real in-process termination); a one-size-fits-all `close()` would either pull test-fixture concerns into a production module or force a single shape on consumers that don't want it. 2. The SDK's `ndJsonStream` outer wrapper does not reliably propagate close on `Stream.writable` to the opposite `Stream.readable`; consumers needing to simulate a child exit hold their own underlying `TransformStream` references and close those directly. 10 of 11 inline call sites in `httpAcpBridge.test.ts` migrate cleanly to the new helper. The 11th (`makeChannel` at line 151) keeps the inline 4-line construction because its `kill()` closure needs the underlying `ab` / `ba` writables to simulate child-process termination — a comment above the function explains the asymmetry. The helper is also a primitive for the future A2 PR's `inProcessAcpBridge.ts`, which will use it to wrap an in-process `QwenAgent` without spawning a `qwen --acp` child (see issue #4156 §3 decision 1 and §8). Test plan: - New `inMemoryChannel.test.ts`: 5 tests covering bidirectional round-trip, ordering preservation, and bidirectional direction isolation - Existing `httpAcpBridge.test.ts`: 70 tests, identical count and behavior before vs after migration - `vitest run packages/cli/src/serve/inMemoryChannel.test.ts packages/cli/src/serve/httpAcpBridge.test.ts` — 75/75 pass - `tsc --noEmit -p packages/cli/tsconfig.json` — clean for changed files 🤖 Generated with [Qwen Code](https://github.com/QwenLM/qwen-code) * refactor(serve): address Copilot review feedback on createInMemoryChannel Two small follow-ups from #4160 review: 1. inMemoryChannel.test.ts:113,137 — handle the pending `reader.read()` that the isolation tests intentionally leave hanging when the timeout wins the race. `reader.releaseLock()` in `finally` rejects that pending read per Web Streams spec; without a rejection handler this could surface as an unhandled rejection / flaky test signal. Added a no-op rejection handler via the two-arg `.then(onResolve, onReject)` form so the cleanup-path rejection settles cleanly. 2. inMemoryChannel.ts:11 — the JSDoc said "two `TransformStream<...>` pairs" which reads ambiguously as "two pairs of TransformStream" (i.e., 4 streams). The implementation creates exactly two TransformStreams (one per direction). Reworded to "two `TransformStream<...>` instances (one per direction)" to disambiguate. Tests still 5/5 pass. 🤖 Generated with [Qwen Code](https://github.com/QwenLM/qwen-code) * refactor(serve): expose abort() teardown primitive on createInMemoryChannel + route test through barrel Two follow-ups from #4160 review: 1. Expose `abort(reason?)` on the helper return value (per @wenshao critical comment). Reasoning: the helper previously returned only the `Stream` pair, leaving consumers no way to tear the channel down. `ndJsonStream`'s outer wrapper does not reliably propagate `close()`, but `abort()` on the underlying byte-level `TransformStream` is forceful-by-spec — pending reads on both sides settle immediately so GC can reclaim. This unblocks the future Stage 1.5b in-process bridge (#4156, sub-PR A2) which needs teardown on daemon shutdown. The settlement shape is documented honestly in JSDoc: at the inner byte-level layer pending reads reject with the supplied reason; at the outer SDK-wrapped `Stream` the wrapper translates that into a clean `{done: true}` signal. Either way, pending operations no longer hang — that's the teardown invariant we care about. 2. Route the test's import through the `serve/index.js` barrel rather than the source file (per @wenshao suggestion). Without a test that exercises the public API path, a typo or missing re-export in the barrel would go undetected in CI. Tests: 8/8 helper tests pass (5 existing + 3 new abort tests covering teardown invariant + idempotency + no-reason variant). 70/70 existing httpAcpBridge tests still pass. 🤖 Generated with [Qwen Code](https://github.com/QwenLM/qwen-code) |
||
|---|---|---|
| .. | ||
| src | ||
| index.ts | ||
| package.json | ||
| test-setup.ts | ||
| tsconfig.json | ||
| vitest.config.ts | ||