Distillation

Date: Sep 16, 2026

• 🟡 (13:40) Test changes in packages/gateway/test/openai-responses-stream.test.ts add imports for log, afterEach, and makeResponsesAccState; define silentLogSink with no-op info, warn, error, and captureException; and restore that sink in afterEach().
• 🟡 (13:40) packages/gateway/test/openai-responses-stream.test.ts adds a test named uses created metadata and failure usage in the plain streaming route, covering metadata from response.created and usage from a failure terminal in the plain streaming path.
• 🟡 (13:40) In /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/pipeline.ts, the recall-aware Responses-stream finish(resp, successful) function at lines 9621-9633 is idempotent via completionAttempted; it calls recallDiagnostics.finish("completed" | "failed"), invokes opts.onComplete(resp, successful), sets completed = true only after the callback succeeds, and logs openai-responses recall-aware onComplete error: if the callback throws.
• 🟡 (13:40) The same recall-aware stream uses KEEPALIVE_INACTIVITY_MS = 30_000, emits : keepalive, and sequenceChunk() rewrites parseable SSE frames with monotonically increasing sequence_number values beginning at 0.
• 🟡 (13:40) settleRecall() in /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/pipeline.ts rejects exhausted recall admission with new RecallContinuationFailure("depth_exhausted"); races opts.onRecall(input) against abort; rolls back late or aborted callback results; records diagnostics and budget usage with Buffer.byteLength(result.resultText) and result.coverage; and logs recall final continuation: budget exhausted reason=${stopReason} when the budget yields a stop reason.
• 🟡 (13:41) accumulateNonStreamResponse() in /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/pipeline.ts detects SSE bodies even when stream: false was sent or text/event-stream is absent, then routes them through protocol-specific strict accumulators. For openai-responses, it uses validation: codex ? "codex" : "public", stopAtTerminal: true, and requireCompletedTerminal: true.
• 🟡 (13:41) Buffered OpenAI Responses JSON is parsed by parseResponsesNonStreamEnvelope(). accumulateNonStreamResponse() throws ResponsesTerminalError(response, parsed.status) whenever the status is not completed, independently of requireValidCompletion.
• 🟡 (13:41) parseResponsesNonStreamEnvelope() derives status from json.status, defaulting to "unknown", and calls assertValidNonStreamCompletion(json, "openai-responses") for statuses completed and incomplete.
• 🟡 (13:41) preserveIncompleteResponsesTerminal() converts only a ResponsesTerminalError whose status === "incomplete" back into its contained GatewayResponse; all other errors are rethrown.
• 🔴 (13:41) User asserted that when content parsing fails under required completion validation, usage-only projection may be parsed for validation and retained in NonStreamCompletionError, but invalid usage still throws and the projection is never returned as success.
• 🟡 (13:41) NonStreamCompletionError stores the parsed GatewayResponse, uses message upstream response did not complete, and is thrown after usage-only projection when required non-stream completion validation fails.
• 🟡 (13:41) assertValidNonStreamCompletion() rejects a non-null json.error with upstream response contained an error; for OpenAI Chat Completions, it requires a first choice with a message and string finish_reason, otherwise throwing upstream OpenAI request did not complete.
• 🟡 (13:41) Commit 0dd569b520bad37adbaea230f32100c7f83f463c (Change ID: lpmvqrurxwlmrmwnyllmxlvkmplvwnlz) on bookmarks fix/responses-provider-failures and fix/responses-provider-failures@git is titled fix(responses): unify provider failure handling; author/committer is Burak Yigit Kaya <ben@byk.im>. It changes 6 files with 3,398 insertions and 500 deletions: 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.
• 🟡 (13:41) Commit 24880457763b6f10831856d01c41bd72d62ddc28 (Change ID: wswlsmrnuzlqtxyskqotksrlqtzplrkl) is titled fix(responses): reject non-monotonic provider failures; author/committer is Burak Yigit Kaya <ben@byk.im>. It changes 4 files with 364 insertions and 27 deletions: packages/gateway/src/pipeline.ts, packages/gateway/src/stream/openai-responses.ts, packages/gateway/test/openai-responses-recall-aware-stream.test.ts, and packages/gateway/test/openai-responses-stream.test.ts.