Distillation

Date: Sep 16, 2026

• 🔴 (21:59) User required the recall follow-up’s stream flag to never diverge from how the continuation is consumed; streaming input may be accumulated into a non-streaming continuation before the recall loop proceeds.
• 🔴 (21:59) User required the lore context-warning read path to “Never throws / never blocks.”
• 🔴 (21:59) User required packages/gateway/src/recall.ts to never import pipeline.ts; pipeline-owned parsing and stream-accumulation functions are passed through RecallFollowUpCtx, including parseJSON, which wraps accumulateNonStreamResponse.
• 🟡 (21:59) Investigation found buildRecallFollowUpRequest() around packages/gateway/src/recall.ts:1243; its stream argument is required with no default and must match how the caller consumes the upstream response.
• 🟡 (21:59) Investigation found appendCodexRateLimitEvent() at packages/gateway/src/codex-rate-limits.ts:93, with repeated-value accumulation at line 118. In packages/gateway/src/pipeline.ts, quota events are appended from failed responses around lines 18434–18436, recovered responses around lines 18496–18498, and ordinary continuation responses around lines 18620–18622; cumulative metadata is assigned to recovered, continuation, and final responses around lines 18521, 18632, and 18654.
• 🟡 (21:59) Investigation confirmed recoveryRequestBase = currentModifiedReq at packages/gateway/src/pipeline.ts:18442; makeJSONRecallCtx(recovery) starts around line 18444, disables cacheTools and systemTTL for recovery around line 18453, and sets requireSuccessfulCompletion: recovery around line 18472.