Distillation

Date: Sep 16, 2026

• 🔴 (12:40) Changes in packages/gateway/src/pipeline.ts added imports for projectResponsesInProgressEvent and RESPONSES_PROVIDER_FAILURE_LOG and introduced additional Responses lifecycle/failure handling, including preservation of sequence_number and a sanitizedFailure structure.
• 🔴 (12:43) Repository history showed commit 24880457763b6f10831856d01c41bd72d62ddc28 (fix(responses): reject non-monotonic provider failures) with parent a1020a45be1382893a183e2ff39506e78a08fb76.
• 🔴 (12:43) Repository history showed commit a1020a45be1382893a183e2ff39506e78a08fb76 (fix(responses): preserve trusted failure metadata) with parent 01a6e06b08ba1b54f42c5b4c757ec449eb8e8490.
• 🔴 (12:43) Worktree modifications included 6 files: packages/gateway/src/pipeline.ts, packages/gateway/src/stream/openai-responses.ts, packages/gateway/test/openai-responses-recall-aware-stream.test.ts, packages/gateway/test/openai-responses-stream.test.ts, packages/gateway/test/pipeline-streaming.test.ts, and packages/gateway/test/recall-codex-stream.test.ts.
• 🟡 (12:43) A repository/status artifact had SHA-256 c79cc2ec010e22f01961d8eec7bce3a60b8dfcb02c12b9ce4c0de9c5e173cb4d.
• 🔴 (12:45) Responses streaming tests cover sanitizing principal provider failures received in multiple forms/statuses, rejecting principal terminal failures with non-monotonic source sequences before accepting failure usage, keeping response.created metadata authoritative through hidden recall and provider failure, and finalizing validated failure usage when cancellation wins terminal backpressure.
• 🔴 (12:45) Responses streaming tests cover rejecting response.in_progress when it changes the identity established by response.created, rejecting contradictory statuses or hidden output in response.created/response.in_progress snapshots, and rejecting a terminal carrying status in_progress.
• 🔴 (12:45) Continuation tests cover rejecting non-monotonic source sequences before accepting failure usage, usage overflow while merging a continuation, per-response usage overflow before recall side effects, and merging continuation cache usage into onComplete.
• 🔴 (12:45) A privacy regression test throws Error("private provider failure") and asserts the client output does not contain private provider failure.
• 🔴 (12:47) User-provided directive: a follow-up stream’s own lifecycle and terminal events “NEVER reach the client — the principal stream already emitted”; those follow-up events are suppressed.
• 🔴 (12:47) A malformed named Responses event in a continuation is rejected; failure output is expected to contain exactly one event: response.failed and must not expose Upstream response has no body.
• 🟡 (12:49) Tool reported three SHA-256 values: e02913cfe3f7507cae8a358dab68d27a5960b1fe6eedc5386345cbcc053da488, f596a3263bf661df8884702b00e550d9187f3b369a662ddd989e7ac28540178d, and 5b6b8dc43aae88fc56daad97d7ab55da8ed57ce873d8d0cf347f3eb4e4bed21c.
• 🔴 (12:51) Test coverage includes Anthropic translator emits inclusive Responses cache usage, rejection of terminal events with non-monotonic source sequences before accepting usage, finalization of validated failure usage when cancellation wins terminal backpressure, rejection of impossible cache usage before recall side effects, and rejection of cache usage without input_tokens before recall.
• 🟡 (12:55) Tool reported six SHA-256 values, in order: 1. 5b6b8dc43aae88fc56daad97d7ab55da8ed57ce873d8d0cf347f3eb4e4bed21c, 2. f596a3263bf661df8884702b00e550d9187f3b369a662ddd989e7ac28540178d, 3. e02913cfe3f7507cae8a358dab68d27a5960b1fe6eedc5386345cbcc053da488, 4. 2f13d034476b26316abff935a26a74b8e1c516a152e85952daa80aecc50405f3, 5. d9f6777bd96c6fe0b9f1cf028b5f82843284bcc08c492130c96d9bd82c894ba8, 6. fc65ee28f054d358a744ab8a62ac22001fa7fb03f8971816cb83998995541d6d.
• 🔴 (12:59) Streaming response accumulation in packages/gateway/src/pipeline.ts dispatches by forwarded.effectiveProtocol: openai-responses uses accumulateResponsesSSEStream() with validation: req.codex ? "codex" : "public", stopAtTerminal: true, and requireCompletedTerminal: true; openai uses accumulateOpenAISSEStream() with strict: true, stopAtTerminal: true, and consumeUntilDone: true; gemini uses accumulateGeminiSSEStream() with strict: true and stopAtTerminal: true; other protocols use accumulateSSEResponse() with strict: true and stopAtTerminal: true.
• 🔴 (12:59) Non-stream accumulation calls accumulateNonStreamResponse(upstreamResponse, forwarded.effectiveProtocol, req.codex === true, abortScope.signal, true).
• 🔴 (12:59) When accumulation throws ResponsesTerminalError, packages/gateway/src/pipeline.ts schedules post-response accounting only after downstreamSettled and one setImmediate; it abandons stale generations, drops the owned provisional key on a confident project conflict, and otherwise calls accountUnsuccessfulResponse() while marking an existing session state _dirty = true.
• 🔴 (12:59) An incomplete ResponsesTerminalError without recall tool use is converted through nonStreamHttpResponse(), with upstream usage-limit headers copied to the response; other terminal errors return errorResponse(502, "Gateway request failed").
• 🔴 (12:59) Gemini accumulation requires stopReason to be one of end_turn, max_tokens, or tool_use; otherwise it throws Error("upstream Gemini request did not complete").
• 🔴 (13:02) packages/gateway/test/codex-usage-provisional.test.ts contains a parameterized test for incomplete values [false, true] named preserves current Codex quota during session header migration (incomplete=${incomplete}).
• 🔴 (13:02) The Codex quota migration test makes 2 upstream calls. Each emits codex.rate_limits; primary.used_percent is calls * 10 with window_minutes: 300 and reset_at: 2000000000, while secondary.used_percent is calls * 20 with window_minutes: 10080 and reset_at: 2000100000.
• 🔴 (13:02) The test response uses id: "resp_quota", model gpt-5, text Hello quota, usage { inputTokens: 600000, outputTokens: 3 }, and uses stop reason max_tokens only on call 2 when incomplete=true; otherwise it uses end_turn.
• 🔴 (13:02) The Codex request posts to /v1/codex/responses with stream: true, authorization Bearer account-one, provider openai-codex, project path from dirname(harness.dbPath), initial x-session-id: codex-before-plugin, and tools read, write, and shell.
• 🔴 (13:02) Session-header migration is exercised by a second request adding x-lore-session-id: codex-after-plugin; after migration, the test expects HTTP 200, exactly 2 upstream calls, exactly 1 active session, header x-codex-primary-used-percent: 20, and header x-codex-secondary-reset-at: 2000100000.
• 🔴 (13:02) The migrated response must contain only the second captured codex.rate_limits event, contain event: response.incomplete when incomplete=true or event: response.completed otherwise, and must not expose "input_tokens":600000.
• 🔴 (13:04) Foreground response ownership in packages/gateway/src/pipeline.ts is transferred once by finishForeground(), which copies usage-limit headers and wraps the response body with cleanup using releaseForeground, foregroundAbort.signal, and rollbackRecallPersistence; awaitForeground() releases foreground ownership if an operation throws before transfer.
• 🔴 (13:04) For a non-OK foreground upstream response, the body is read with AbortSignal.any([foregroundAbort.signal, AbortSignal.timeout(foregroundErrorBodyTimeoutMs)]); a timeout logs upstream error body read timed out, and the status logs as upstream error: ${upstreamResponse.status}.
• 🔴 (13:04) Non-OK foreground handling calls captureToolPairing400() with exact fields status, errorBody, messages: modifiedReq.messages, layer: result.layer, model: req.model, and sessionID; it then sets the GenAI span to code 2 with message HTTP ${upstreamResponse.status}, ends the span, and returns finishForeground(sanitizedUpstreamErrorResponse(upstreamResponse)).
• 🔴 (13:06) captureToolPairing400() is exported from packages/gateway/src/pipeline.ts at line 2575.