Distillation

Date: Sep 16, 2026

• 🟡 (22:22) Fail-first regression run with Vitest v4.1.8 produced exactly 3 failures and 225 passes across 228 tests; 1 of 2 test files failed. All failures were in packages/gateway/test/openai-responses-recall-aware-stream.test.ts.
• 🟡 (22:22) The fail-first regression suppresses malformed non-JSON principal Codex quota events failed at packages/gateway/test/openai-responses-recall-aware-stream.test.ts:294: output leaked both private-principal-quota-sentinel and event: codex.rate_limits, although "safe answer" remained present.
• 🟡 (22:22) The fail-first regression suppresses malformed non-JSON continuation Codex quota events failed at packages/gateway/test/openai-responses-recall-aware-stream.test.ts:329: output leaked both private-continuation-quota-sentinel and event: codex.rate_limits, although "safe continuation" remained present.
• 🟡 (22:22) The fail-first regression bounds Codex quota metadata by bytes across the principal and continuation failed at packages/gateway/test/openai-responses-recall-aware-stream.test.ts:394: emitted canonical quota metadata measured exactly 16,535 bytes, exceeding MAX_CODEX_RATE_LIMIT_BYTES of 16,384 bytes.
• 🟡 (22:22) Investigation located Codex quota handling in packages/gateway/src/stream/openai-responses.ts: import of appendCodexRateLimitEvent at line 37, ResponsesAccState.codexRateLimits at line 61, accumulation through applyResponsesEvent() at lines 618–619, finalized propagation at lines 913–914, and validated forwarding at lines 2380–2385.
• 🟡 (22:23) applyResponsesEvent() in packages/gateway/src/stream/openai-responses.ts:612-620 handles event === "codex.rate_limits" by calling appendCodexRateLimitEvent((state.codexRateLimits ??= []), parsed) and returns the accepted normalized event or undefined.
• 🟡 (22:23) The validated forwarding path in packages/gateway/src/stream/openai-responses.ts:2379-2387 suppresses a parsed codex.rate_limits event unless acceptedCodexRateLimit is truthy; accepted quota events are forwarded as JSON.stringify(acceptedCodexRateLimit), while other event types retain their original data.
• 🔴 (22:24) User stated sparse numeric identities can never need more address space than the number of frames the parser is willing to process.
• 🔴 (22:24) User required the sparse-index ceiling to ensure that arithmetic remains practical and that index + 1 always changes the value.
• 🟡 (22:24) accumulateResponsesSSEStream() in packages/gateway/src/stream/openai-responses.ts:1931-1934 computes maxSparseIndex as Math.min(opts.maxFrames ?? DEFAULT_MAX_SSE_FRAMES, DEFAULT_MAX_SSE_FRAMES), capping caller-supplied frame overrides at the production ceiling.
• 🟡 (22:25) Investigation showed the malformed non-JSON leak path: accumulateResponsesSSEStream() attempts JSON.parse(data) at packages/gateway/src/stream/openai-responses.ts:1957-1965; parse failures throw "malformed Responses stream event" only when opts.validation is set, while the tolerant no-validation path merely continues.
• 🟡 (22:25) Strict event processing validates parsed records and public/Codex events before mutation: isRecord(parsed) at lines 1967–1969, validatePublicResponsesEvent(...) at lines 1984–1986, and validateCodexResponsesEvent(...) at lines 1986–1988.
• 🟡 (22:26) packages/gateway/src/stream/openai-responses.ts:2187 stores the return value of applyResponsesEvent(state, event, parsed) as acceptedCodexRateLimit; this value later controls whether a codex.rate_limits event is emitted.
• 🟡 (22:26) Investigation indicates quota limits are enforced against the codexRateLimits array belonging to the current ResponsesAccState; the failing combined principal-plus-continuation test demonstrates separately accumulated stream segments can collectively exceed the global 16 KiB output budget even when each segment is locally accepted.