Distillation

Date: Sep 16, 2026

• 🟡 (17:54) packages/gateway/test/openai-responses-recall-aware-stream.test.ts was updated so pre-recall reference identity violations expect PUBLIC_GATEWAY_ERROR and reject PUBLIC_RECALL_ERROR; affected scenarios included reuse of a reference identity by a later output item, collision with synthetic marker msg_resp_reference_marker_1, and continuation references colliding with principal identities.
• 🟡 (17:56) Full execution of packages/gateway/test/openai-responses-recall-aware-stream.test.ts passed all 218 tests in 18.63s.
• 🟡 (17:56) Gateway typecheck ran tsc --noEmit for @loreai/gateway@0.40.0 with no reported errors.
• 🟡 (17:56) Responses-related test files identified were: packages/gateway/test/worker-copilot-responses.test.ts, packages/gateway/test/openai-responses-recall-aware-stream.test.ts, packages/gateway/test/responses-upstream-error-relay.test.ts, packages/gateway/test/responses-compact-routing.test.ts, packages/gateway/test/openai-responses.test.ts, and packages/gateway/test/openai-responses-stream.test.ts.
• 🟡 (17:58) Broader Responses test execution passed 5 test files and all 253 tests in 37.81s.
• 🟡 (17:58) Formatting check found issues in exactly 2 files: packages/gateway/src/pipeline.ts and packages/gateway/test/openai-responses-recall-aware-stream.test.ts.
• 🟡 (17:59) Formatter changes in packages/gateway/src/pipeline.ts collapsed multiline expressions around indexedGet(identityIndexFor(referenceIdentityIndices, acc), identity) and item?.type === "item_reference", and corrected indentation around mergeUsage(state.usage, contState.usage). Formatter changes in packages/gateway/test/openai-responses-recall-aware-stream.test.ts collapsed the "public continuation" content array and terminal item_reference output array to single lines.
• 🟡 (18:01) After additional packages/gateway/src/pipeline.ts and test edits, targeted execution regressed 2 tests: "privately accumulates a done-only sparse Codex item_reference" exposed response.output_item.done containing private identity msg_done_only_private, and "keeps done-only continuation references private" returned response.failed instead of "public continuation". Both failures were classified as protocol errors (principal_protocol and follow_up_protocol).
• 🟡 (18:03) Follow-up fixes in packages/gateway/src/pipeline.ts made a targeted set of 4 tests pass, with 214 of 218 skipped, in 12.72s.
• 🟡 (18:04) A subsequent targeted run exposed 4 failures: terminal item_reference replacement of completed visible message, function_call, and reasoning items did not produce a parseable response.failed; continuation terminal reference replacement yielded a failure payload without response.output: [].
• 🟡 (18:06) The "preflights a reference-driven continuation offset before recall side effects" test still classified the error as follow_up_setup and returned PUBLIC_RECALL_ERROR ("Lore could not continue the response after recall") rather than required PUBLIC_GATEWAY_ERROR ("Gateway request failed"), while the contract required an empty output and no leaked private identity or side effects.
• 🟡 (18:08) Pipeline changes made "preflights a reference-driven continuation offset before recall side effects" pass alone: 1 passed and 217 skipped out of 218.
• 🟡 (18:09) "bounds reference indices before recall and continuation side effects" failed because the emitted failure payload lacked response.output: []; additional changes were applied to packages/gateway/src/pipeline.ts.
• 🟡 (18:10) Terminal replacement failures were correctly emitted as principal_protocol, but their public response.failed.output still retained the previously visible item: a message containing "visible message", a read function call with {"path":"visible"}, or reasoning item rs_visible_reference_replacement, instead of being cleared to [].
• 🟡 (18:12) After clearing public failure output, completion metadata still leaked prior visible state: message completion contained { type: "text", text: "visible message" }, function-call completion contained { type: "tool_use", id: "call_visible_reference_replacement", name: "read", input: { path: "visible" } }, and reasoning completion retained rs_visible_reference_replacement; the contract required both completion.response.content and completion.response.rawOutputItems to be empty.
• 🟡 (18:13) "does not classify a later-resolved sparse ordinary tool as private beside a reference" failed with principal_protocol: output omitted the visible ordinary tool name "read" and identity, returning only response.created followed by a gateway response.failed.
• 🟡 (18:13) "counts private references against retained-state limits before suppression" failed because the generated failure payload had no response.output, rather than required response.output: [].
• 🔴 (18:14) User established the contract: "never forwards response-side item_reference lifecycle events"; the corresponding test took 95ms.
• 🟡 (18:14) Targeted execution had 3 failures and 215 skipped out of 218: 1. "never forwards response-side item_reference lifecycle events" emitted an unexpected gateway response.failed, 2. "privately accumulates a done-only sparse Codex item_reference" also emitted an unexpected response.failed, and 3. "preflights a reference-driven continuation offset before recall side effects" again returned PUBLIC_RECALL_ERROR instead of PUBLIC_GATEWAY_ERROR.
• 🟡 (18:16) "indexes high-cardinality reference matching with linear operations" kept output private and avoided response.failed, but its operation-count assertion failed: actual operations.count was 768 while the test required greater than cardinality * 4 (1024) and at most cardinality * 8 + 16; the implied cardinality was 256.
• 🟡 (18:16-18:17) packages/gateway/src/pipeline.ts received 2 further updates after the high-cardinality operation-count failure; no subsequent test result was shown.