Distillation

Date: Sep 16, 2026

• 🟡 (14:15) User provided /home/byk/Code/opencode-lore-responses-projection-security/packages/gateway/test/openai-responses-recall-aware-stream.test.ts, an 8,795-line Vitest suite for streamResponsesRecallAware, the true-streaming, recall-aware OpenAI Responses streamer.
• 🟡 (14:15) The documented regression is "Provider response headers timed out after 10000ms": buffered accumulateResponsesSSEStream withheld all client bytes until a slow, reasoning-heavy upstream completed, causing opencode’s 10-second ProviderHeaderTimeoutError on ChatGPT sessions. streamResponsesRecallAware instead forwards events live, intercepts a recall function call, emits a marker, runs a follow-up, and rebuilds terminal response.completed.
• 🟡 (14:15) The test suite imports streamResponsesRecallAware from ../src/pipeline, setRecallContinuationFailureHook and RecallContinuationFailureCategory from ../src/recall-continuation-failure, and setPrincipalTransportFailureHook and PrincipalTransportFailureSample from ../src/principal-transport-failure; afterEach clears both hooks and registers silentLogSink.
• 🟡 (14:15) Test helpers include sseEvent() for constructing SSE payloads, streamFrom() for creating a text/event-stream Response, drain() for reading the complete client-facing stream, responseEvents() for parsing emitted events, and expectVisibleOutputIndicesToMatchTerminal() for verifying each visible event’s output_index and item ID match the rebuilt terminal response’s output.
• 🟡 (14:15) Test fixtures include created(id, model), recallCall(outputIndex, args, itemId, callId), sparseVisibleFunctionCall(outputIndex, argumentsText, itemId, callId), and completed(id, usage?). recallCall emits response.output_item.added, response.function_call_arguments.done, and response.output_item.done for a completed recall; sparseVisibleFunctionCall begins with blank call_id, name, and arguments, streams an arguments delta, then completes as a read call.
• 🟡 (14:16) User provided an implementation excerpt around lines 11,392–11,616 showing continuation recall processing, synthetic IDs of the form msg_${state.id || "lore"}_${shiftedRecallIndex}, updates to contState.items, propagation of contState.terminalEvent, construction of continuationFollowUpInput and continuationRetryBaseline, reset of recallContinuationTransportRetries to 0, and recall-continuation failure categorization/logging.
• 🟡 (14:16) The implementation rebuilds the terminal response.completed to reflect visible output only, computes visibleIndices via visibleOutputIndexMap(), remaps continuation chunks with remapVisibleOutputIndices(chunk, visibleIndices), clears keepalives, handles client disconnects during continuation and terminal delivery, and exposes transactional commit/rollback behavior for pendingCommits.
• 🟡 (14:16) The implementation continues processing deferred chunks after terminal-delivery logic, increments deferredBytes by chunk.byteLength, and tracks source output indices where available.