Distillation

Date: Sep 16, 2026

• 🔴 (16:13) User asserted the required privacy invariant/test case: never exposes non-semantic %s frames after recall interception.
• 🟡 (16:13) Assistant identified a terminal-reconciliation bug: when a sparse ordinary tool resolves at the terminal, it must be registered in the visible-index registry before deferred frames are flushed. Expected discrimination: resolving at a lower index must fail with output: []; resolving at a higher index must succeed.
• 🟡 (16:14) Assistant defined the correction architecture: 1. register every terminal-resolved ordinary tool before deferred delivery; 2. rebuild public SSE from exact event schemas so unknown frames never cross the wire; 3. keep principal output transactional until the terminal establishes the final visible map, preventing sparse live/terminal contradictions and pre-recall extension leaks; 4. replace repeated middle insertion with sets plus terminal-time sorting; 5. merge and roll back continuation quota metadata with the rest of the accumulator.
• 🟡 (16:14) Assistant stated discriminating tests would be added around terminal ordinary-tool resolution, exact public event projection, transactional principal delivery, visibility-map construction, and continuation quota rollback before replacing the registry.
• 🟡 (16:14) Inspection of packages/gateway/src/pipeline.ts:11555-11784 showed deferred continuation frames are flushed at flushHeldContinuation() before mergeContinuation(), terminal output is built through buildOutputItems() and visibleOutputIndexMap(), transactional frames are remapped during delivery, and non-terminal/non-recall principal events are still described and encoded as verbatim at :11729-11749.
• 🟡 (16:15) Assistant chose wire-boundary projection rather than weaker validation: unsupported events may still be parsed and internally accounted where reviewed, but only exact recognized lifecycle event names may produce client frames, and each recognized frame will be reconstructed from its schema.
• 🟡 (16:15) Assistant stated the terminal will always be rebuilt from validated accumulated state so sparse live indices cannot disagree with terminal output slots.
• 🟡 (16:15) Test helpers in packages/gateway/test/openai-responses-recall-aware-stream.test.ts include responseEvents() at :72-88, expectVisibleOutputIndicesToMatchTerminal() at :90-112, sparseVisibleFunctionCall() at :149-180, completed() at :182-190, and textItem() at :222-252.
• 🟡 (16:16) Inspection located buildTerminal(res: GatewayResponse) at packages/gateway/src/pipeline.ts:9840; related visible-output construction and sorted-index logic begins around :9904.
• 🟡 (16:17) Inspection confirmed isSupportedResponsesOutputItemType() in packages/gateway/src/stream/openai-responses.ts:238-242 accepts only string types present in RESPONSES_OUTPUT_ITEM_TYPES; isValidResponsesOutputItemStatus() at :244-252 validates optional statuses against type- and phase-specific sets.
• 🟡 (16:17) Existing regression forwards a principal data-only non-recall tool after its name resolves at packages/gateway/test/openai-responses-recall-aware-stream.test.ts:2518-2557 verifies a terminal-resolved function_call named read with arguments {"path":"file"} is publicly delivered, does not invoke recall, and does not emit PUBLIC_RECALL_ERROR.