Distillation

Date: Sep 16, 2026

• 🟡 (13:50) packages/gateway/test/openai-responses-recall-aware-stream.test.ts is 8,795 lines long; its header describes streamResponsesRecallAware as the true-streaming, recall-aware OpenAI Responses (codex/ChatGPT) streamer and identifies the regression as "Provider response headers timed out after 10000ms" / ProviderHeaderTimeoutError, caused by accumulateResponsesSSEStream withholding all client bytes until a slow, reasoning-heavy upstream completed.
• 🟡 (13:50) The streamer’s documented behavior is to forward events live, transparently intercept a recall function_call, emit a marker, run a follow-up, and rebuild the terminal response.completed.
• 🟡 (13:50) packages/gateway/test/openai-responses-recall-aware-stream.test.ts imports streamResponsesRecallAware from ../src/pipeline, setRecallContinuationFailureHook and RecallContinuationFailureCategory from ../src/recall-continuation-failure, and setPrincipalTransportFailureHook and PrincipalTransportFailureSample from ../src/principal-transport-failure.
• 🟡 (13:50) Test cleanup in afterEach() resets both setRecallContinuationFailureHook(undefined) and setPrincipalTransportFailureHook(undefined), then registers silentLogSink with log.registerSink(silentLogSink).
• 🟡 (13:50) Test helpers shown include: sseEvent(event, data) for encoding SSE; streamFrom(events) for a status-200 text/event-stream Response; drain(resp) for reading the complete response body; responseEvents(output) for parsing object-valued SSE payloads; and expectVisibleOutputIndicesToMatchTerminal(output) for verifying each visible event’s output_index and item ID agree with the corresponding terminal response.completed.response.output item.
• 🟡 (13:50) recallCall(outputIndex, args, itemId = \fc_${outputIndex}`, callId = `call_${outputIndex}`)constructs 3 ordered events: 1.response.output_item.addedfor afunction_callnamed"recall"; 2. response.function_call_arguments.donewith JSON-stringified arguments; 3.response.output_item.donewith status"completed"`.
• 🟡 (13:50) sparseVisibleFunctionCall(outputIndex, argumentsText, itemId = \fc_sparse_${outputIndex}`, callId = `call_sparse_${outputIndex}`)initially emits a function-call item with emptycall_id, name, and arguments, then emits response.function_call_arguments.delta, and finally completes it as a "read"call with the suppliedcallIdandargumentsText`.
• 🟡 (13:50) Terminal-event helpers use model "gpt-5.6-terra": completed(id, usage?) emits response.completed with status "completed" and optional usage; incomplete(id) emits response.incomplete with status "incomplete"; doneIncomplete(id) emits response.done with status "incomplete"; and doneWithStatus(id, status) emits response.done with an arbitrary status.
• 🟡 (13:50) Test constants are PUBLIC_RECALL_ERROR = "Lore could not continue the response after recall" and PUBLIC_GATEWAY_ERROR = "Gateway request failed".
• 🟡 (13:50) textItem(outputIndex, text, itemId = \msg_${outputIndex}`)emits, in order:response.output_item.added, response.output_text.delta, response.output_text.done, and response.output_item.done, using content_index: 0`.
• 🟡 (13:50) Diff from BASE to FIRST changes 2 files with 581 insertions and 61 deletions: packages/gateway/src/pipeline.ts has 247 changed lines, and packages/gateway/test/openai-responses-recall-aware-stream.test.ts has 395 changed lines.
• 🟡 (13:50) Diff from FIRST to HEAD changes the same 2 files with 112 insertions and 40 deletions: packages/gateway/src/pipeline.ts has 101 changed lines, and packages/gateway/test/openai-responses-recall-aware-stream.test.ts adds 51 lines.
• 🟡 (13:50) Two commits are present: ab979df5cddfec3952680e423e39564664e881fe (fix(responses): project hidden output consistently, associated identifier unznwpylspzqykmnpxwnnswmnzpspxzq) and later HEAD commit 583020fad43e1481470aac2588251b8636d16251 (fix(responses): stabilize sparse principal projection, associated identifier ozrmvotwxzuykyvutstpwpzukpptktxv).