Distillation

Date: Sep 16, 2026

• 🔴 (13:02) User asserted the stream must “Always feed the accumulator,” as documented near line 3136 in packages/gateway/src/stream/openai-responses.ts.
• 🔴 (13:02) User asserted that an injected recall tool_use never leaks to the client; the optimization near lines 2702–2704 is safe only when no recall tool_use can appear in the stream.
• 🔴 (13:02) User asserted that failed usage/accounting never depends on client demand, as documented near line 2864 in packages/gateway/src/stream/openai-responses.ts.
• 🟡 (13:03) Repository search found 100 matches, with more available, for accumulateResponsesSSEStream. Its definition is at packages/gateway/src/stream/openai-responses.ts:2092, with an internal call at line 2844.
• 🟡 (13:03) Production call sites for accumulateResponsesSSEStream include packages/gateway/src/llm-adapter.ts lines 2985 and 2996; and packages/gateway/src/pipeline.ts lines 12230, 15746, 15845, 18567, and 19117.
• 🟡 (13:03) Test coverage for accumulateResponsesSSEStream is extensive in packages/gateway/test/openai-responses-stream.test.ts, with listed matches spanning lines 125 through 4538, and in packages/gateway/test/codex-usage-metadata.test.ts at lines 65, 80, 98, 115, 122, and 132.