Distillation

Date: Sep 16, 2026

• 🟡 (13:46) Candidate diff modifies 2 files with 654 insertions and 62 deletions: packages/gateway/src/pipeline.ts (270 changed lines in the diff summary) and packages/gateway/test/openai-responses-recall-aware-stream.test.ts (446 changed lines).
• 🟡 (13:46) In packages/gateway/src/pipeline.ts:9839-9846, terminal-response projection reads terminalResponse?.incomplete_details, accepts it only when it is a non-array object, and extracts reason only when it is a string; this value is then used while formatting the terminal Responses event.
• 🟡 (13:46) Search found 67 matches related to Responses stream state/projection. applyResponsesEvent() is defined in packages/gateway/src/stream/openai-responses.ts:582, with calls shown at lines 1393, 1597, 1767, 2158, and 2250.
• 🟡 (13:46) streamResponsesRecallAware() is exported from packages/gateway/src/pipeline.ts:7794; related stream machinery includes validateResponseLifecycle() at line 9144, shiftedOutputIndex() at line 9721, visibleOutputIndexMap() at line 9953, remapVisibleOutputIndices() at line 9963, enqueuePrincipal() at line 10054, and enqueueProjectedPrincipal() at line 10163.
• 🟡 (13:46) Sparse-principal projection in packages/gateway/src/pipeline.ts tracks forwardedVisibleOutputIndices (Map<number, number>) and remaps source output indices to visible output indices before forwarding chunks.
• 🟡 (13:46) Responses streaming logic maintains deferredEvents with candidate-index handling and transactionalEvents: Uint8Array[]; it validates lifecycle events with validateResponseLifecycle(), accumulates state through applyResponsesEvent(), flushes deferred events through projected enqueueing, and shifts continuation/recall output indices with shiftedOutputIndex().
• 🟡 (13:46) The modified regression tests include sparse-terminal visibility coverage using item id msg_sparse_terminal_visible, continuation-output assertions containing "continuation", recall-count tracking via let recalls = 0, and a case using output_index: 7.