Distillation

Date: Sep 16, 2026

• 🟡 (15:42) packages/gateway/test/openai-responses-recall-aware-stream.test.ts adds a created-at validation case (lines 1401-1429): principal stream uses created("resp_created_at_principal", "gpt-5.6-terra"), recallCall(0, { query: "created-at validation" }), and completed("resp_created_at_principal"); with validation: "codex", invalid continuation data must produce PUBLIC_RECALL_ERROR while suppressing "response.completed", "private created-at", and "private recall result".
• 🟡 (15:42) Added projection-performance test "updates the principal projection without sorting once per indexed event" (lines 1431-1491): generates exactly 64 indexed textItem(...) events and matching terminal output messages (msg_projection_0 through msg_projection_63), spies on Array.prototype.sort via vi.spyOn, counts sorts of non-empty all-number arrays as numericProjectionSorts, and requires numericProjectionSorts <= 4; output must contain "answer 63" and no "response.failed", and the spy is restored in finally.
• 🟡 (15:42) Projection-performance test configures both onRecall and runFollowUp to throw new Error("should not run"), confirming the 64-item principal-only stream should require neither recall nor follow-up execution.
• 🟡 (15:42) Added parameterized test "allows only semantic $name after recall" (lines 1493-1574) with 2 ordered cases: 1. "custom continuation events" injects response.custom carrying diagnostic secret "private-custom-continuation-event"; 2. "terminal extension fields" injects terminal field private_extension: "private-terminal-extension".
• 🟡 (15:42) The semantic allowlist test uses principal response ID "resp_allowlisted_principal", follow-up response ID "resp_allowlisted_followup", model "gpt-5.6-terra", recall query "allowlisted continuation", private recall result "private recall result", visible message ID "msg_allowlisted_answer", visible text "allowlisted answer", and terminal usage { input_tokens: 11, output_tokens: 3 }.
• 🟡 (15:42) Semantic allowlist expectations require visible "allowlisted answer" and exactly 1 event: response.completed; event: response.custom, "private-custom-continuation-event", and "private-terminal-extension" must not appear in streamed output or serialized completedResponse, demonstrating that only semantic continuation fields/events are forwarded after recall.