Distillation

Date: Sep 16, 2026

• 🔴 (11:04) User supplied /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/pipeline.ts lines 9080–9329. Reasoning output completion rejects changed content type/value, incomplete delta or part lifecycles, and disagreement among authoritativeValue, finalValue, partFinalValue, and the final output item; successful validation sets lifecycle.outputDone = true.
• 🔴 (11:04) seedImplicitCodexItem only operates with a normalizationState, skips response.output_item.added and response.output_item.done, requires a nonnegative safe-integer output_index, and seeds a missing acc.rawItems entry from normalizationState.rawItems by applying a synthetic response.output_item.added.
• 🔴 (11:04) validateResponseLifecycle rejects any Responses event after terminal, duplicate response.created, missing response identity, invalid creation status, nonempty creation output, and any response.* event before response.created. Under Codex validation, response.created.status === "queued" is allowed.
• 🔴 (11:04) response.in_progress validation requires stable response identity, optional status "in_progress", and empty output.
• 🔴 (11:04) Responses terminal events recognized by lifecycle validation are response.completed, response.done, response.incomplete, and response.failed; allowed terminal statuses are "completed", "incomplete", "failed", and "cancelled". Codex validation uniquely permits response.completed with status "incomplete", while response.failed may carry "failed" or "cancelled".
• 🔴 (11:04) For status "incomplete", incomplete_details must be null/undefined or a non-array object, and its optional reason must be "max_output_tokens" or "content_filter"; malformed forms throw "malformed Responses terminal event".
• 🔴 (11:04) assertOutputLifecyclesComplete accepts optional allowedIncompleteIndices. Outside Codex validation it throws when a raw output lacks output_item.done; in sparse Codex streams, omission is retained for non-reasoning items and reasoning without a string ciphertext envelope, but reasoning with string encrypted_content throws as provisional.
• 🔴 (11:04) preserveStreamedReasoning reconstructs missing reasoning summary entries from lifecycle authoritativeValue as { type: "summary_text", text: summaryState.authoritativeValue }.
• 🔴 (11:04) assertTerminalReasoningMatchesLifecycle validates both actual.summary against lifecycle.reasoning and actual.content against lifecycle.content; terminal collections must be arrays, indexed parts must be non-array objects, and part types/values must match lifecycle state.
• 🟡 (11:05) Search found Codex rate-limit handling across tests and implementation: packages/gateway/test/codex-usage-provisional.test.ts, packages/gateway/test/codex-usage-metadata.test.ts, packages/gateway/test/recall-buffered-transaction.test.ts, packages/gateway/src/pipeline.ts, packages/gateway/src/translate/types.ts, packages/gateway/src/translate/openai-responses.ts, and packages/gateway/src/stream/openai-responses.ts.
• 🟡 (11:05) Codex quota metadata uses event type "codex.rate_limits" and optional field metered_limit_name; GatewayResponse/stream state expose codexRateLimits?: Array<Record<string, unknown>>, stream accumulation pushes each quota into state.codexRateLimits, translation emits every stored quota, and recall continuation merging appends currentResp.codexRateLimits before continuationResp.codexRateLimits.
• 🟡 (11:05) Search located streamResponsesRecallAware at /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/pipeline.ts:7799, call sites around lines 15453 and 18871, and streamResponsesPassthrough at /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/stream/openai-responses.ts:2649 with a pipeline call around line 19079.
• 🔴 (11:05) User established the pipeline invariant expressed in the source comment as "Never throws / never" for the relevant post-response/context-warning pathway.
• 🟡 (11:05) Search found accountUnsuccessfulResponse declared at /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/pipeline.ts:13458 and called around lines 15879, 16142, 18746, and 18812.
• 🔴 (11:06) User supplied /home/byk/Code/opencode-lore-provider-error/packages/gateway/src/pipeline.ts lines 13380–13579. Successful post-response processing updates shadow context using actual input, output tokens, worker model, requested model, resolved conversation TTL, and long-context status; marks sessionState._dirty = true; persists session tracking; learns candidate headers when headerSessionId is absent; and schedules background work unless noStore.
• 🔴 (11:06) Commit-triggered curation requires loreConfig().knowledge.enabled, loreConfig().curator.onIdle, and containsGitCommit(req). Its multiplier is 3 when model input cost is at least 5, 2 when at least 1, otherwise 1; effectiveAfterTurns = loreConfig().curator.afterTurns * curationMultiplier, and turnsSinceCuration is raised to that threshold when lower.
• 🔴 (11:06) Post-response processing catches failures, logs "post-response processing failed:", returns false, and always invokes endSpan?.() in finally; success returns true.
• 🔴 (11:06) accountUnsuccessfulResponse(resp, sessionID, resolvedConversationTTL, genAiSpan, endSpan, markDirty?) records only validated provider usage without publishing successful-turn state. It treats usage as present only when at least one token count is numeric and greater than 0, optionally calls markDirty, accounts usage through accountConversationUsage, adds Sentry usage attributes when a span exists, then always sets span status { code: 2, message: "upstream response did not complete" } and calls endSpan().
• 🔴 (11:06) conversationTTLForAccounting first checks the live session’s resolvedConversationTTL, then persisted tracking, and returns only "5m", "1h", or undefined.
• 🔴 (11:06) postResponse delegates through withTenant(sessionState.storageTenantId ?? "", ...) to postResponseForTenant.
• 🔴 (11:06) Full background chains are tracked in inFlightBackground: Set<Promise<unknown>>; trackBackground increments state.backgroundWorkCount, adds the promise, and on either fulfillment or rejection removes it and decrements the count. Reset awaits these chains alongside limiter drain before swapping the database, and ownership includes global-queue waiting time.
• 🔴 (11:06) scheduleBackgroundWorkForTenant combines pipelineGenerationAbort.signal and sessionLifecycleSignal(sessionID) with AbortSignal.any. It skips background work when session auth is stale and no fresh fallback resolves, preventing worker calls that would return 401.
• 🔴 (11:06) Background scheduling resolves the worker provider from getWorkerModel(sessionState.lastUpstream)?.providerID; the provider-aware auth guard exists because a session may hold credentials for provider A while lastUpstream targets provider B, and it skips work rather than repeatedly issuing unauthenticated calls and degrading worker health.