Distillation

Date: Sep 16, 2026

• 🟡 (17:26) packages/gateway/src/stream/openai-responses.ts:1590-1623 showed Codex output-item normalization behavior: a response.output_item.done without an existing item creates seedItem = { ...item }, deletes seedItem.content, and applies a synthetic response.output_item.added; then reconcileCodexDoneItem(...) runs and any mismatch with the existing added item triggers malformedResponsesEvent(). For non-done events, item.id is bound through state.itemIndexById, the output index is removed from unboundTextItems and unboundToolItems, and a nonempty function-call call_id is bound through state.callIndexById; finally parsed.output_index is set to the resolved index.
• 🟡 (17:27) buildTerminal(res: GatewayResponse) at packages/gateway/src/pipeline.ts:9815-9856 rebuilds the suppressed terminal Responses event after recall. It obtains output from buildOutputItems(), maps res.stopReason to status, defaults usage to ZERO_USAGE, calculates inclusive input tokens as inputTokens + cacheReadInputTokens + cacheCreationInputTokens, calculates total tokens as inclusive input plus output, conditionally emits input_tokens_details with cached_tokens and cache_write_tokens, preserves state.terminalEvent/state.terminalResponse, and emits response fields including id, object: "response", created_at, model, status, output, and usage.
• 🟡 (17:27) buildOutputItems(hiddenIndices = new Set()) at packages/gateway/src/pipeline.ts:9858-9920 walks the sorted union of state.rawItems and state.items, skips hidden indices, preserves non-item_reference raw-only items, reconstructs text items as assistant messages, preserves raw content arrays when present, reconstructs refusals as { type: "refusal", refusal }, and otherwise creates output_text content with empty annotations. Tool items are emitted as type: "function_call" with id, call_id, name, arguments, and raw string status or default "completed".
• 🟡 (17:27) Stream abort handling at packages/gateway/src/pipeline.ts:9922-9936 removes the abort listener through cleanupAbort; onStreamAbort() finishes recall diagnostics as "aborted", resumes and clears pending demand, clears keepaliveTimer, and either cancels/releases activeReader or cancels currentPrincipalResponse.body using signal.reason. The abort listener is registered once, and an already-aborted signal invokes the handler immediately.
• 🟡 (17:27) Backpressure setup at packages/gateway/src/pipeline.ts:9937-9959 defines waitForDemand(), which waits while the stream is neither cancelled nor aborted and controller.desiredSize <= 0, then calls signal.throwIfAborted(). The stream tracks principalEventEmitted and ordinaryToolEmitted; safeEnqueue(...) first rejects cancelled work and awaits demand.
• 🟡 (17:27) Search in packages/gateway/src/pipeline.ts found three finish declarations: finish(resp: GatewayResponse, successful: boolean) at line 9613, finish(operation: () => void) at line 14920, and parameterless finish() at line 15000.
• 🟡 (17:29) Recall-aware finish(resp, successful) at packages/gateway/src/pipeline.ts:9613-9625 is guarded by completionAttempted; on first invocation it marks completion attempted, finishes diagnostics as "completed" or "failed", invokes opts.onComplete(resp, successful), sets completed = true, and returns true. If opts.onComplete throws, it logs "openai-responses recall-aware onComplete error:" and returns false.
• 🟡 (17:29) settleRecall(input) at packages/gateway/src/pipeline.ts:9626-9671 first calls recallBudget.admit(recallItemReservation(input)); any admission result throws new RecallContinuationFailure("depth_exhausted"). It starts opts.onRecall(input), defines late-result cleanup that awaits the operation and calls late.rollback?.(), races the callback against an abort promise, removes the abort listener in finally, and schedules late cleanup when abortion wins.
• 🟡 (17:29) After a recall callback resolves, settleRecall(...) checks abortion again; if aborted, it calls result.rollback?.(), logs "late recall rollback failed:" if rollback throws, and rethrows signal.reason. Otherwise it records diagnostics with input, result.resultText, and result.coverage, then calls recallBudget.record(...) with exact result bytes from Buffer.byteLength(result.resultText) and the returned coverage.