Distillation

Date: Sep 16, 2026

• 🔴 (13:47) User stated continuation lifecycle events must “NEVER reach the client — the principal stream already emitted,” establishing that lifecycle frames from a recall follow-up must not be exposed after the principal stream has emitted its lifecycle.
• 🟡 (13:47) packages/gateway/test/openai-responses-recall-aware-stream.test.ts includes a recall-depth exhaustion case configured with maxRecallDepth: 1 and a follow-up response ID resp_depth_exhausted_followup.
• 🟡 (13:47) packages/gateway/test/openai-responses-recall-aware-stream.test.ts verifies merged follow-up behavior: continuation text remains visible, the replay anchor stays hidden, raw recall function_call data does not leak, and no response.in_progress event reaches the client.
• 🟡 (13:47) packages/gateway/test/openai-responses-recall-aware-stream.test.ts includes sparse-index ordering cases: a higher-index late recall after already-visible output is handled separately from a lower-index recall after visible output, with the latter test explicitly named "fails closed when a lower-index recall follows visible output" and requiring "private query" to remain hidden.
• 🟡 (13:48) packages/gateway/test/openai-responses-recall-aware-stream.test.ts includes "stops reading upstream while the client applies backpressure"; the test coordinates a blocked third pull with Promise.withResolvers<void>() and expects pulls to equal exactly 3.
• 🟡 (13:48) packages/gateway/test/openai-responses-recall-aware-stream.test.ts lines 8100–8134 test "holds no-index continuation events after chained recall detection": a malformed continuation emits a recall function_call at output_index: 0, then a no-index response.custom event containing secret "must-not-leak". The client output must omit the secret and contain PUBLIC_RECALL_ERROR.
• 🟡 (13:48) packages/gateway/test/openai-responses-recall-aware-stream.test.ts lines 8136–8217 use test.each for "allows only semantic $name after recall" with 2 ordered cases: 1. "custom continuation events" carrying event secret "private-custom-continuation-event"; 2. "terminal extension fields" carrying terminal secret "private-terminal-extension".
• 🟡 (13:48) In the semantic-allowlist test, the Codex-validated follow-up resp_allowlisted_followup emits visible message msg_allowlisted_answer with text "allowlisted answer" and usage { input_tokens: 11, output_tokens: 3 }. The merged output must contain the answer and exactly one response.completed, while response.custom, the custom-event secret, and terminal private_extension secret must be absent from both client output and serialized completedResponse.
• 🟡 (13:48) The semantic-allowlist test invokes principal recall query "allowlisted continuation"; onRecall returns anchorText: buildAnchor(query) and private resultText: "private recall result", and runs with validation: "codex".
• 🟡 (13:48) packages/gateway/test/openai-responses-recall-aware-stream.test.ts lines 8219–8259 test "recall-only: fails the response when the continuation fails": runFollowUp throws Error("follow-up unavailable"), but client output exposes only response.failed and PUBLIC_RECALL_ERROR; it must omit "follow-up unavailable", "lore-recall", "Searching", and response.completed.
• 🟡 (13:48) The recall-only continuation setup-failure test requires RecallContinuationFailureCategory[] to equal ["follow_up_setup"] and logs exactly 2 sanitized errors in order: 1. "recall follow-up stream failed category=follow_up_setup"; 2. "openai-responses recall-aware stream failed category=follow_up_setup". Joined logs must not contain the internal exception text "follow-up unavailable".