Distillation

Date: Sep 16, 2026

• 🔴 (22:39) User/project directive: “Always accumulate into the internal state for postResponse.”
• 🔴 (22:39) User/project privacy invariant: “never exposes non-semantic %s frames after recall interception”.
• 🔴 (22:39) User/project semantic-validation invariant: “never treats forged reasoning lifecycle names as semantic %s frames”.
• 🟡 (22:39) Candidate diff in packages/gateway/src/pipeline.ts introduces projection/privacy tracking including recallIndices, unresolvedToolIndices, visibleOutputIndices, projectionInvalid, newlyPrivateRecallIndex, and discardDeferredCandidate(outputIndex).
• 🟡 (22:39) Candidate tests in packages/gateway/test/openai-responses-recall-aware-stream.test.ts include sparse/lower-index visibility scenarios using fixtures such as textItem(7, "sparse visible text", "msg_sparse_terminal_visible"), "msg_lower_terminal_visible", and textItem(7, "already visible", "msg_sparse_visible_first").
• 🟡 (22:39) Candidate tests check that streamed output does not contain "private result" and inspect response.failed via events.find(({ event }) => event === "response.failed")?.data.response.
• 🟡 (22:39) Candidate tests exercise deferred recall privacy with recallCall(9, { query: "private query" }, "fc_late_deferred_recall").
• 🟡 (22:39) Non-semantic-frame coverage conditionally injects rawSSEEvent("diagnostic", privateDiagnostic) after recall interception.
• 🟡 (22:39) Forged semantic-prefix coverage uses sseEvent("response.reasoning_summary_private.done", ...) to ensure lifecycle names that merely resemble reasoning events are rejected as non-semantic.
• 🟡 (22:39) Full truncated tool output was saved at /home/byk/.local/share/opencode/tool-output/tool_0ac5eaeaf001S2x3uxOXwT1PrY.