Distillation

Date: Sep 16, 2026

• 🟡 (14:19) runActivePipelineRequest() in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/pipeline.ts rejects admission with HTTP 503 and "Gateway is busy" when detached + active + pending session claims reach maxDetachedPipelineRequests, active + streamingPostResponsePending reach maxActivePipelineRequests, or per-admission-key active + pending reach MAX_ACTIVE_PIPELINE_REQUESTS_PER_ADMISSION_KEY.
• 🟡 (14:19) runActivePipelineRequest() combines callerSignal with an internal lifecycle signal via AbortSignal.any(), tracks pending operations in a Set<Promise<void>>, and does not remove the request from activePipelineRequests/detachedPipelineRequests or pump pending session claims until all tracked operations settle.
• 🟡 (14:19) runActivePipelineRequest() distinguishes response-body cancellation from ordinary settlement: after a response is returned, caller abort or wrapped-body cancellation calls markResponseCancelled() once and then settleResponse(); settleResponse() invokes onResponseBodySettled() once and starts asynchronous cleanup.
• 🟡 (14:19) claimPipelineSession() removes the active request while queuing a pending session claim, rejects with PipelineCapacityError("session request queue full") on duplicate/full claims, and removes/rejects an aborted claim before pumping the queue.
• 🔴 (14:19) User asserted that request-local recall diagnostic fingerprints and recall content “never leave this closure.”
• 🟡 (14:19) /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/recall-diagnostics.ts sets MAX_RECALL_DIAGNOSTIC_ROUNDS = 64; createRecallDiagnostics(enabled = true) keeps request-local SHA-256 sets for inputs, results, input/result pairs, and delivered coverage.
• 🟡 (14:19) Recall diagnostics normalize input fingerprints as [scope ?? "all", id || ids || null, id/id list present ? null : query.trim()]; therefore ID lookups ignore query while scope remains part of the comparison.
• 🟡 (14:19) Each recorded recall diagnostic tracks round count, detail-call count, empty trimmed result bodies, total result bytes, repeated input/result/pair status, and positive-length new coverage keyed by identity, revision, kind ?? "detail", offset, and length.
• 🟡 (14:19) Recall diagnostics emit recall-round logs containing round, "detail"/"search" kind, repetition booleans, coverageProgress, and per-result bytes. finish() accepts "completed", "failed", or "aborted" and emits one recall-chain log with rounds, detail calls, repeated counts, empty bodies, total result bytes, coverage-item count, and elapsed time clamped to 0..300_000 ms before clearing all sets.
• 🟡 (14:20) Displayed commit identifiers and subjects, newest first: tklkqxytmpkzzryqlolnpvxouropxz — fix(responses): complete lifecycle reconciliation; kvpmnlwnznqxomuzywlvmotqtplywzvm — fix(responses): bound sparse lifecycle coordinates; lplymwvllmznozzrzoroxvnnuymnusxs — fix(responses): preserve limit-crossing state.
• 🟡 (14:20) Targeted Vitest run in /home/byk/Code/opencode-lore-responses-accumulation with Vitest v4.1.8 passed: 1 test file passed, 48 tests passed, 207 skipped, 255 total; duration 7.76s with transform 3.85s, setup 521ms, import 6.11s, tests 923ms, and environment 0ms.
• 🟡 (14:21) /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts parameterizes rejection of principal reasoning introduced only by output_item.done or the terminal response, for both summary/summary_text and content/reasoning_text; validation occurs before recall execution, private terminal reasoning is not exposed, and [recallCalls, followUps] remains [0, 0].
• 🟡 (14:21) Nested-recall tests reject four terminal reasoning inconsistencies: 1. summary gap, 2. summary contradiction, 3. content gap, 4. content contradiction. Output contains PUBLIC_RECALL_ERROR, excludes "private gap" and "private contradiction", and records [recallCalls, followUps] as [1, 1].
• 🟡 (14:21) Recall-aware stream validation rejects a response.reasoning_summary_text.delta emitted after the corresponding text-completion event; the tested late delta is "evil" after completed text "safe", and the client output must contain "response.failed".
• 🟡 (14:21) The continuation-offset regression test uses 63_000 visible principal items and verifies computation without spreading 126000 retained item keys. It caps constructed SSE frames at 100_000, allows maxRetainedStateBytes: 64 * 1024 * 1024, and temporarily spies on Map.prototype[Symbol.iterator] only to isolate an unrelated quadratic duplicate-identity scan while retaining/validating all 63,000 items.
• 🟡 (14:21) Other located retained-state/resource-limit tests in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/test/openai-responses-recall-aware-stream.test.ts cover: principal event accumulation when crossing the retained-state limit; retention of principal state at the crossing; bounding sparse principal coordinates before mutation; continuation event accumulation at the crossing; iterative reconstruction of a resource-limited continuation with 125500 content blocks; and one-time accounting of terminal Codex continuation usage when candidate flush exceeds the transaction limit.
• 🔴 (14:22) User asserted that an LLM-completion SSE stream “always begins with a data: or event: field.”
• 🟡 (14:22) looksLikeSSE() in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/translate/types.ts returns true when contentType includes "text/event-stream" or, after removing a leading BOM and trimming start whitespace, the body starts with a legal SSE data:, event:, id:, retry:, or comment : field. This handles providers such as ChatGPT/Copilot/Codex and DeepSeek returning SSE for non-streaming requests without the SSE content type, avoiding JSON.parse failures such as Unexpected token 'e', "event: res"... (LOREAI-GATEWAY-38 / -1P).
• 🔴 (14:22) User asserted that response accumulators “always try to populate” token usage.
• 🟡 (14:22) GatewayResponse in /home/byk/Code/opencode-lore-responses-accumulation/packages/gateway/src/translate/types.ts contains required id, model, content, and stopReason; optional rawOutputItems, codexRateLimits, and usage. usage?: GatewayUsage remains optional because providers such as vLLM and partial responses may omit usage at runtime.