Distillation

Date: Sep 16, 2026

• 🔴 (19:57) User stated failed usage/accounting never depends on client demand, reflected by // Failed usage/accounting never depends on client demand. in packages/gateway/src/stream/openai-responses.ts around lines 2863-2865; on response.failed, finish(false, finalizeUnsuccessfulResponsesAcc(state)) runs before await waitForDemand().
• 🟡 (19:57) packages/gateway/src/stream/openai-responses.ts changed finish from finish(successful: boolean) to finish(successful: boolean, response?: GatewayResponse), with onComplete(response ?? finalizeResponsesAcc(state), successful).
• 🟡 (19:57) Gateway stream failure handling emits a fixed response.failed payload before completing local accounting: id: state.id || "resp_error", object: "response", created_at: state.createdAt ?? 0, model: state.model, status: "failed", output: [], usage: null, and error: { type: "server_error", message: PUBLIC_GATEWAY_ERROR }; it then calls finish(false, finalizeUnsuccessfulResponsesAcc(state)).
• 🟡 (19:57) Responses terminal-event handling declares the first semantic terminal authoritative; bytes delivered afterward are treated as transport tail and discarded on cancellation. It calls await opts.onValidatedEvent?.(event, data) and breaks when opts.stopAtTerminal is true.
• 🟡 (19:57) packages/gateway/src/pipeline.ts was modified to import/use cloneResponsesAccState and RESPONSES_PROVIDER_FAILURE_LOG, and added provider-failure/terminal rollback state including continuationProviderFailureSeen, normalizationStateBeforeTerminal, terminalRecallSnapshot, and terminalOutputSnapshot.
• 🔴 (19:58) User specified that accumulated text must be replaced with the final version because the final version is more reliable, reflected by // Replace accumulated text with the final version (more reliable) in packages/gateway/src/stream/openai-responses.ts around line 814.
• 🟡 (19:58) Search reported exactly 12 codexRateLimits matches across the gateway: packages/gateway/src/pipeline.ts lines 18679-18682 merges currentResp.codexRateLimits with continuationResp.codexRateLimits; packages/gateway/src/translate/types.ts line 349 declares codexRateLimits?: Array<Record<string, unknown>>; packages/gateway/src/translate/openai-responses.ts line 1281 iterates resp.codexRateLimits ?? []; and packages/gateway/src/stream/openai-responses.ts line 63 declares the field, line 729 appends quota records, lines 1048-1049 include it in finalized state, and lines 1065-1066 include a structuredClone(state.codexRateLimits) copy.