Commit Graph

2 Commits

Author SHA1 Message Date
agent_coder a14174a726 fix(ai): drain-hang в writeToServerResponse — расширить patches/ai@6.0.134.patch (#486)
Серверная ai@6.0.134 в writeToServerResponse при backpressure (write()===false)
ждала ТОЛЬКО once('drain'). Если клиент отвалился мид-запись, сокет не дренится,
await не резолвится: read-цикл паркуется НАВСЕГДА, finally{response.end()}
недостижим, reader и буферы висят до рестарта. В autonomous ран продолжает лить
вывод после дисконнекта → КАЖДЫЙ дисконнект мид-ран оставляет висящий пайп.
Плюс read() — fire-and-forget с throw → unhandledRejection.

Патч расширен (index.js и index.mjs): Promise.race drain/close/error с гигиеной
once-слушателей (все три снимаются на первом settle — не копятся по одному на
stall); при close/error — reader.cancel() и выход (безопасно для detached-ранов:
независимый дренаж делает consumeStream); rejection read() поглощается с логом.
pnpm-lock patch_hash перегенерён (patch-commit). Выравнивание версии ai —
отдельный коммит (#495), не здесь.

Тест: трипвайр-спека в apps/server (только там резолвится патченная копия) по
образцу ai-sdk-partial-output.patch.spec — дисконнект мид-запись без drain
завершает ответ (end() вызван, reader не висит), read()-throw не даёт
unhandledRejection, оба dist-билда несут маркер PATCH(docmost #486).
2026-07-11 02:54:03 +03:00
agent_vscode 9b4b38a611 fix(ai): patch ai@6.0.134 — drop O(n²) partialOutput accumulation causing heap OOM on long agent runs (#184)
Production OOM'd (JS heap 1.85 GB / 2 GB limit) during a ~20-step,
~28k-chunk autonomous agent turn. Heap snapshot analysis (memlab) showed a
single DefaultStreamTextResult retaining ~1.7 GB via the never-consumed
leftover tee() branch of its internal baseStream.

Root cause in ai@6.0.134: streamText substitutes the default text() output
strategy even when the caller passes NO `output` option. Its
createOutputTransformStream then accumulates the ENTIRE turn text and, on
EVERY text-delta, enqueues `{ part, partialOutput }` where partialOutput is
a flat snapshot of all text so far (JSON.stringify flattens the
cons-string) — O(n²) memory across the turn. Every consumer accessor tees
baseStream and keeps the second branch as the new baseStream; the final
leftover branch is never read, so its controller queue holds every chunk
(28,225 x ~164 KB in the OOM'd run) for the life of the turn.

Fix (pnpm patch on both dist/index.js and dist/index.mjs):
- pass the raw, possibly-undefined `output` option into
  createOutputTransformStream instead of defaulting to text()
- when output == null, publish each text-delta immediately without
  accumulating turn text or producing partialOutput snapshots; streaming
  granularity is unchanged, and callers that DO request an output strategy
  keep the original behavior

Our server never uses partialOutputStream / experimental_output / the
output option, so no behavior changes for us beyond memory.

Regression spec ai-sdk-partial-output.patch.spec.ts drives the real
patched SDK with MockLanguageModelV3: asserts per-delta textStream
granularity, an EMPTY experimental_partialOutputStream (tripwire — yields
one cumulative partial per delta when unpatched), and the PATCH(docmost
marker in both installed dist bundles. Also documents the patch in
AGENTS.md (must be re-created when bumping `ai`) and CHANGELOG.md.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-05 02:13:17 +03:00