Distillation

Date: Sep 16, 2026

• 🔴 (20:24) User stated recall follow-up invariant: the follow-up request’s stream flag must never diverge from how the continuation is consumed.
• 🟡 (20:24) In packages/gateway/src/pipeline.ts, non-stream recall couples build, forward, content-type assertion, and parsing: Codex uses runRecallFollowUpStreamAccumulated() because /backend-api/codex/responses mandates streaming and rejects stream: false with 400 {"detail":"Stream must be set to true"}; other backends use runRecallFollowUpJSON() with stream:false.
• 🟡 (20:24) makeJSONRecallCtx() parses JSON through accumulateNonStreamResponse() and parses SSE through accumulateResponsesSSEStream() with stopAtTerminal: true, requireCompletedTerminal: true, validation set to "codex" or "public", and requireSuccessfulCompletion: recovery.
• 🟡 (20:24) Non-stream recall continuation recovery in recoverRecallContinuation() calls runRecallRecovery(), merges usage and ordered codexRateLimits, rejects recovered output containing another recall tool or lacking a usable continuation, and returns failRecall(category, false) for non-abort recovery failures. AbortError and foregroundAbort.signal.aborted are rethrown.
• 🟡 (20:24) mergeFailedContinuationMetadata() preserves usage and codexRateLimits from ResponsesTerminalError and NonStreamCompletionError; failed normal follow-ups attempt recovery under "follow_up_failed", while recall on the final round or after a budget stop recovers under "depth_exhausted".
• 🟡 (20:24) Recall quota metadata is accumulated in order through appendCodexRateLimitEvent(), then attached to continuationResp, recovered continuations, and the final currentResp, allowing newer quota-window updates to follow older ones.
• 🔴 (20:24) User stated empty-completion telemetry must never throw and must never block the read path.
• 🟡 (20:24) Before adding any lore context-warning banner, packages/gateway/src/pipeline.ts checks isEmptyCompletion(currentResp) and records protocol, model, sessionID, stopReason, outputTokens, and recallDepth through captureEmptyCompletion(); it also logs an empty completion → client warning.
• 🔴 (20:24) User stated recall source coverage is private and must never be emitted to the client.
• 🟡 (20:24) The streaming recall callback result type in packages/gateway/src/pipeline.ts includes private coverage?: readonly import("@loreai/core").RecallCoverage[], alongside anchorText, resultText, and optional commit()/rollback() transaction hooks.
• 🔴 (20:25) User stated every Responses event must always be accumulated into internal state for postResponse.
• 🟡 (20:25) In packages/gateway/src/pipeline.ts, streaming Responses events are normalized with normalizeCodexEvent(), lifecycle-validated by validateResponseLifecycle(), seeded through seedImplicitCodexItem(), and then accumulated via applyResponsesEvent() even when recall-related events will be hidden from the client.
• 🟡 (20:25) Streaming recall interception tracks suppressed recall calls in parsedRecallInputs, pendingRecalls, and recallIndices; unresolved sparse Codex function-call identities use unresolvedToolIndices and unresolvedToolBytes, while non-recall function calls set otherToolSeen.
• 🟡 (20:25) Recall and unresolved-tool events are deferred or suppressed client-side but byte-counted. Limits raise SSEStreamLimitError with exact messages including Responses stream exceeded byte limit, Responses retained state exceeded byte limit, and recall stream exceeded deferred event limit.
• 🟡 (20:25) applyResponsesEvent() may return an accepted codex.rate_limits record; appendCodexRateLimitEvent(publicCodexRateLimits, acceptedCodexRateLimit) determines the public quota event, while internal accumulation occurs regardless. The public event is reserialized through JSON.stringify(publicCodexRateLimit).
• 🟡 (20:25) Streaming Responses parsing ignores empty data and [DONE]; malformed JSON in an event whose name starts with response. throws malformed JSON in Responses event ${event}, mismatched payload types throw Responses payload type does not match ${event}, and non-JSON non-message events are forwarded or deferred depending on active recall/unresolved-tool state.