Distillation

Date: Sep 16, 2026

• 🔴 (14:31) User asserted sparse numeric identities “never need more address space than the number of frames we are willing to process” in packages/gateway/src/stream/openai-responses.ts:1908-1911.
• 🔴 (14:31) User asserted the sparse-index cap must ensure that “index + 1 always changes the value” in packages/gateway/src/stream/openai-responses.ts:1911.
• 🟡 (14:32) Diff in packages/gateway/src/pipeline.ts changed terminal-event projection to inspect terminalResponse?.incomplete_details, accept only a non-array object whose reason is a string, and assign that value to incompleteReason; the projected response no longer blindly spreads ...terminalResponse.
• 🟡 (14:33) packages/gateway/test/openai-responses-recall-aware-stream.test.ts defines PUBLIC_RECALL_ERROR = "Lore could not continue the response after recall" and PUBLIC_GATEWAY_ERROR = "Gateway request failed".
• 🟡 (14:33) textItem(outputIndex, text, itemId = \msg_${outputIndex}`)inpackages/gateway/test/openai-responses-recall-aware-stream.test.ts:218-248emits, in order: 1.response.output_item.addedfor an assistant message, 2.response.output_text.deltawithcontent_index: 0, 3. response.output_text.done, 4. response.output_item.donewithstatus: "completed"andcontent: [{ type: "output_text", text }]`.
• 🟡 (14:33) Test "forwards events live with NO recall — no header hold-back" verifies a no-recall stream forwards "response.completed" and "hello world" live, without "recall", "lore_marker", or "response.failed"; runFollowUp must not be called.
• 🟡 (14:33) Test "finalizes when the client cancels immediately after a no-recall terminal" verifies onComplete runs exactly once and upstream cancellation occurs after the reader sees event: response.completed and calls reader.cancel().
• 🟡 (14:36) Test "forwards a principal response.failed exactly once when no recall occurs" verifies exactly one response.failed, preserves "provider failed", and does not emit "ended without a terminal event".
• 🟡 (14:36) Test "accepts an empty Codex terminal output after streamed items" verifies streamed text "answer" remains valid when the terminal response.completed has output: []; exactly one completion is emitted and no response.failed appears.
• 🟡 (14:36) Test "accepts a partial Codex terminal output after streamed items" streams items at output_index 0 ("omitted from terminal", ID "msg_omitted") and 1 ("answer", ID "msg_answer"), while terminal output contains only "msg_answer"; it expects "answer", exactly one response.completed, and no response.failed.
• 🟡 (14:36) Parameterized test test.each(["failed", "cancelled"])("hides recall when the principal response.done status is %s", ...) verifies a recall at output_index: 0 is not executed or exposed when the principal response.done status is "failed" or "cancelled": exactly one response.failed, no "name":"recall" or "call_0" in streamed output, and no recall or "call_0" in the completed response.
• 🟡 (14:36) Test "suppresses a recall function_call and emits a marker (mixed tools)" begins with recall call output_index: 0 using { query: "what is lore" }, followed by a non-recall "read" function call at output_index: 1 with item ID "fc_1", call ID "call_1", and arguments {"file":"x"}.