Distillation

Date: Sep 16, 2026

• 🔴 (15:10) User directive: streamResponsesRecallAware must never forward response-side item_reference lifecycle events.
• 🔴 (15:10) User directive: deferred recall persistence must never commit when an abort-ignoring recall callback resolves late.
• 🔴 (15:10) User directive: cancellation must never wait for a non-settling recall callback.
• 🔴 (15:10) User directive: cancellation must never wait for non-settling follow-up setup; if an abort-ignoring setup returns a reader late, that reader must be cancelled.
• 🔴 (15:10) User directive: a failed recall-only continuation must remain failed and must never be converted into completed.
• 🟡 (15:10) Additional streamResponsesRecallAware test contracts shown in packages/gateway/test/openai-responses-recall-aware-stream.test.ts: no-recall events are forwarded live without header hold-back; immediate client cancellation after a no-recall terminal still finalizes; a principal response.failed is forwarded exactly once; parallel recalls bind to their own call IDs; repeated call IDs are rejected before recall execution; malformed ordering with arguments before item declaration fails closed; strict recall arguments reject {"query":"x","extra":true}; request-wide frame limits include failed continuation frames and cap combined principal/chained-continuation frames; a continuation’s response.incomplete is preserved.
• 🟡 (15:10) isUsableRecallContinuation() in packages/gateway/src/recall.ts:1189 rejects stop reasons "max_tokens", "pause_turn", and "model_context_window_exceeded", and rejects any tool_use block whose name is absent, non-string, or blank.
• 🟡 (15:10) isUsableRecallContinuation() accepts a continuation only if it contains at least one nonblank text block, a named non-recall tool_use, or a nonblank Responses refusal represented either by an opaque responsesItem === true block or rawOutputItems; recall tool calls alone are not usable final answers.
• 🟡 (15:10) Streaming post-response finalization in packages/gateway/src/pipeline.ts is serialized per sessionID, yields once with setImmediate, drops operations from stale streamingPostResponseGeneration values, logs queue overflow at most once per 30_000 ms, and invokes pumpPendingSessionClaims() after cleanup.
• 🟡 (15:10) awaitStreamingPostResponse() limits waiters to MAX_STREAMING_POST_RESPONSE_WAITERS_PER_SESSION = 16, throws StreamingPostResponseWaitCapacityError("streaming post-response wait queue full") at capacity, and waits via promiseAgainstAbort(() => tail, signal).
• 🟡 (15:10) Remote session binding is enabled when config.remoteGateway || config.hostedMode; requestStorageTenant() derives authenticated ownership with credentialTenantFingerprint(credential) and assigns unauthenticated requests an owner of the form unauthenticated:${crypto.randomUUID()}. Provisional header mappings are capped at 1024 with PROVISIONAL_HEADER_MAPPING_TTL_MS = 5 * 60_000.