Distillation

Date: Sep 16, 2026

• 🔴 (12:40) Immutable accumulation-candidate review returned PASS for identity/integrity verification but concluded DO-NOT-MERGE because of four must-fix correctness blockers.
• 🔴 (12:40) Reviewed immutable candidate tklkqxytmpkzzryqlolnpvxouropxznp at commit ce88c614c0ae55d229cb605aa7c7369d918d3650, parent kvpmnlwnznqxomuzywlvmotqtplywzvm / 683dff02101f15198b18e3088a0b57dade290056, base 2cbf40060c67df229731cf1b9e34d5bf22ec2145, with plain range-diff SHA-256 02d06b2f84d7d2d8cfa5f22443384d12a2a9bc922ade879449de9b0b9d971d1c.
• 🔴 (12:40) Immutable accumulation candidate changed exactly 3 files: packages/gateway/src/pipeline.ts, packages/gateway/src/stream/openai-responses.ts, and packages/gateway/test/openai-responses-recall-aware-stream.test.ts.
• 🔴 (12:40) Candidate range was conflict-free; the conflict warning concerned only the unrelated fix/responses-private-references bookmark. Review started and ended on clean empty child d890b184d786a4348e30bcdac7412baa7ae01ae9.
• 🔴 (12:40) Immutable review touched no files, VCS state, services, network, or databases. Required repo-setup skill was unavailable, so the reviewer loaded jj-guide and read AGENTS.md, quality/REVIEW.md, both package manifests, and gateway build/bundle scripts.
• 🔴 (12:40) Accumulation blocker 1: omitted Codex continuation indices can bypass the shifted <100_000 ceiling. assertSparseCoordinates() checks an output index only when the original payload owns output_index (pipeline.ts:8386-8400), and continuation validation runs before Codex normalization (pipeline.ts:11001-11013). Codex normalization can then assign an omitted index and mutate state (openai-responses.ts:1580-1623, openai-responses.ts:1757-1768), while shiftedOutputIndex() checks safe-integer arithmetic but not maxSparseIndex (pipeline.ts:9849-9855).
• 🔴 (12:40) Concrete omitted-index counterexample: 1. principal recall accepts output_index: 99_999; 2. contIndex becomes 100_000; 3. a done-only continuation item omits output_index; 4. pre-normalization sees no coordinate; 5. Codex normalization assigns local index 0; 6. forwarding/merging accepts shifted index 100_000, violating the <100_000 contract.
• 🔴 (12:40) Existing continuation-boundary tests at openai-responses-recall-aware-stream.test.ts:7406-7500 use explicit coordinates and miss normalized omitted coordinates. Required fix: give the normalizer the remaining shifted coordinate space or reject the normalized coordinate before authoritative-map mutation; add principal-boundary tests for omitted done-only and data-only continuation indices.
• 🔴 (12:40) Accumulation blocker 2: broad prefix/suffix lifecycle parsing at pipeline.ts:8432-8435, pipeline.ts:8702-8842, and pipeline.ts:8844-8966 accepts forged reasoning lifecycle names. Codex validation handles only exact known names and has no rejecting default (openai-responses.ts:1709-1743).
• 🔴 (12:40) Forged names such as response.reasoning_summary_private.done and response.reasoning_text_private.done can seed authoritative lifecycle state; a later output_item.done or terminal snapshot with matching text then passes tracked-part checks, defeating terminal-only reasoning rejection before recall execution.
• 🔴 (12:40) User requires an exact lifecycle-event allowlist before mutation and principal/continuation tests proving forged names fail before recall or nested-recall callbacks and “never expose their text.”
• 🔴 (12:40) Accumulation blocker 3: item_reference events return immediately after auxiliary handling in the principal path (pipeline.ts:10389-10391) and continuation path (pipeline.ts:11014-11022), bypassing applyResponsesEvent() (pipeline.ts:10408-10416, pipeline.ts:11029-11037) and retained-state checks (pipeline.ts:10465-10474, pipeline.ts:11085-11094).
• 🔴 (12:40) User stated the existing item_reference test at openai-responses-recall-aware-stream.test.ts:2497-2539 proves only public suppression and “never verifies private accumulation or limit-crossing reference state.”
• 🔴 (12:40) Required reference fix: accumulate references canonically before excluding them from client projection, and add principal and continuation limit-crossing tests so every accepted provider event enters private authoritative state before filtering or limits.
• 🔴 (12:40) Accumulation blocker 4: materializeCompletedCodexMessages() handles only messages and only done parts (pipeline.ts:7971-7996), while missing output_item.done is accepted for nearly every Codex item without checking nested lifecycle completion (pipeline.ts:9349-9373), including when terminal output is absent (pipeline.ts:9461-9467).
• 🔴 (12:40) User stated that a fully finalized reasoning summary with no item-done and absent/empty terminal output remains only in the lifecycle map and “never reaches rawOutputItems,” causing recall continuation and post-response state to lose accepted reasoning.
• 🔴 (12:40) A message containing one finalized part plus one unfinished delta receives a partial content array; finalizeResponsesAcc() prefers that array over accumulated text (openai-responses.ts:824-847), silently dropping the unfinished accepted delta while reporting success.
• 🔴 (12:40) Existing missing-done tests at openai-responses-recall-aware-stream.test.ts:3240-3362 cover only message text. Required tests: absent/empty terminal output for completed reasoning summary/content and mixed complete/incomplete message parts. Required behavior: reject unfinished lifecycles or consistently preserve every accepted semantic part.
• 🔴 (12:40) Positive accumulation finding: non-reference principal and continuation events call applyResponsesEvent() before evaluating retainedStateBytes (pipeline.ts:10408-10474, pipeline.ts:11029-11094), preserving done-only identity completion and limit-crossing function-call state.
• 🔴 (12:40) Positive accumulation finding: preserveLifecycleFailureState() reconstructs only allowlisted type and text/refusal values (pipeline.ts:10097-10138); principal and continuation lifecycle-limit matrices assert sparse holes and absence of negative leakage.
• 🔴 (12:40) Positive accumulation finding: sparse message holes survive finalization because finalizedMessageContent() writes by original content index (pipeline.ts:7971-7984) and finalizeResponsesAcc() skips sparse holes rather than converting them into opaque blocks (openai-responses.ts:824-843).
• 🔴 (12:40) Positive accumulation finding: reasoning snapshots receive bidirectional checks when present. The helper rejects untracked nested parts, wrong types, extra fields, contradictions, and incomplete tracked lifecycles (pipeline.ts:7998-8042, pipeline.ts:9094-9163, pipeline.ts:9398-9445); tests verify principal and nested-continuation rejection before additional recall side effects.
• 🔴 (12:40) Positive accumulation finding: resource-limited terminal usage has an idempotent double-accounting guard at pipeline.ts:10929-10939; resource-limit catch accounting occurs at pipeline.ts:11307-11318; normal completion uses the same helper at pipeline.ts:11446; regression coverage at openai-responses-recall-aware-stream.test.ts:7794-7894 asserts exact combined usage.
• 🔴 (12:40) Positive accumulation finding: unsuccessful terminal delivery rolls back before finish() (pipeline.ts:11689-11696), and the outer failure path rolls back at pipeline.ts:11852 before callback delivery at pipeline.ts:11979; new tests inspect rollback state inside onComplete.
• 🔴 (12:40) Positive accumulation finding: failed callbacks receive empty content/raw output (pipeline.ts:9742-9746), resource-limit terminals expose output: [] (pipeline.ts:11957-11959), and internal failure projections strip recall/reference identities.
• 🔴 (12:40) Performance concern: iterative maximum-index scanning and append loops eliminate argument-spread failures, but production still has quadratic duplicate-identity scans at pipeline.ts:8560-8577 and repeated findIndex terminal matching at pipeline.ts:9498-9513.
• 🔴 (12:40) The 63,000-item regression monkeypatches away the first quadratic scan at openai-responses-recall-aware-stream.test.ts:7709-7732; this proves the spread fix but not acceptable production complexity near the 100,000-frame ceiling.
• 🔴 (12:40) Fail-first evidence was judged strong but incomplete: it targeted exact parent kvpmnlwn / 683dff02101f15198b18e3088a0b57dade290056 and produced 10 failures with 6 controls plus independent done, terminal, materializer, and reasoning-text mutations. Reported fixed, adjacent, matrix, static, and bundle runs were coherent.
• 🔴 (12:40) Existing mutations do not cover omitted indices assigned during shifted Codex normalization, forged lifecycle event names, reference early-return branches, missing-done reasoning materialization, or mixed finalized/unfinished parts. Tests were not rerun during immutable review because focused runs already existed and the blockers were static counterexamples caused by missing tests.
• 🟡 (12:40) Assistant accepted the accumulation review as DO-NOT-MERGE, kept ce88c614 frozen, and planned to repair all 4 blockers in a later child with new tests that fail on the parent.
• 🟡 (12:40) Assistant reported that unified provider handling passed all 664 integration tests and planned to finish the repeated guard battery plus static/bundle gates before freezing it for exhaustive validation and fresh reviews.
• 🟡 (12:41) Work queue contained 8 ordered high-priority tasks: 1. Prove expanded accumulation regressions fail on frozen parent — completed; 2. Freeze and exhaustively validate accumulation correction — pending; 3. Repair accumulation review blockers with failing-first tests — pending; 4. Resolve projection security blockers in additive child — pending; 5. Fix provider metadata, accounting parity, and categorical logs — in progress; 6. Resolve recovery review blockers and revalidate — pending; 7. Resume Codex and private-reference repair lanes — pending; 8. Push reviewed candidates and update or create PRs — pending.
• 🟡 (12:41) Assistant stated the provider guard battery covers exactly 12 accepted-failure cases spanning plain routing, principal routing, diagnostics, continuation, and both in-progress projections; it was being run 10 times before final source-exact gates.