Distillation

Date: Sep 16, 2026

• 🔴 (13:42) User asserted that recall continuation diagnostics must “never alter the fixed public failure,” captured by test recall continuation diagnostics never alter the fixed public failure in packages/gateway/test/openai-responses-recall-aware-stream.test.ts:9172.
• 🔴 (13:42) User asserted that recall must “never execute recall from an incomplete principal,” captured by test never executes recall from an incomplete principal in packages/gateway/test/openai-responses-recall-aware-stream.test.ts:9320.
• 🔴 (13:42) User asserted that recall must “never execute a chained recall from an incomplete continuation,” captured by test never executes a chained recall from an incomplete continuation in packages/gateway/test/openai-responses-recall-aware-stream.test.ts:9346.
• 🔴 [enforced-workflow] (13:42) User’s regression-test discipline in quality/REVIEW.md requires every adversarial-review defect to land a deterministic regression test in the same PR; the test must fail on the base branch, pass on the fix, and drive the real precondition rather than an artificially constructed state.
• 🔴 [enforced-workflow] (13:42) User requires guard tests to fail when the guard is deleted; a test that still passes after guard removal is not considered to test the guard.
• 🔴 [enforced-workflow] (13:42) User’s adversarial state-setup rules require: 1. state already present before the handling operation, 2. delete-then-recreate sequences across lifecycle boundaries, and 3. modify-while-off sequences where capture triggers cannot observe the change.
• 🔴 [requested-tests] (13:42) User requires fan-out registry additions to be covered by corresponding contract batteries: SYNCED_TABLES via sync-registry-contract.test.ts; MIGRATIONS via the db.test.ts schema-version assertion plus migration-specific tests; and AGENTS via agents.test.ts, including name, displayName, binary, and detect() checks and a constant placeholder binary.
• 🔴 [requested-tests] (13:42) User’s standing batteries are: registry contracts in packages/core/test/sync-registry-contract.test.ts; per-operation sync invariants through assertSyncInvariants in packages/core/src/sync-data.ts; and sequence properties in packages/gateway/test/sync.property.test.ts, covering invariants after every operation, prune-floor liveness, eventual convergence, no ping-pong, and currentTier correctness.
• 🔴 (13:42) User identified skip/early-return branches as the highest-risk correctness surface; guards involving NOT EXISTS, latest <> 'upsert', and content matching previously masked data-loss bugs, so every skip branch needs a test that makes it skip something incorrectly and proves it does not.
• 🔴 (13:42) User stated that property-based tests over faithful mocks catch sequence-dependent defects missed by example tests; the #833 property battery found four engine bugs spanning enable/disable/push/pull boundaries.
• 🔴 [enforced-workflow] (13:42) User stated: “Never trust the fix without seeing it fail.” Every bug fix must be driven by a regression test written first, confirmed failing on the base branch, and then confirmed passing with the fix.
• 🔴 (13:42) User prefers freezing memoized arrays shared by multiple callers, such as the syncedColumns PRAGMA cache, so accidental mutation throws a TypeError in strict mode instead of silently corrupting the cache.
• 🔴 (13:42) User stated that db().query monkeypatches are shadowed by the tracing Proxy; SQL-counting tests must use the log.registerSink({ withDbSpan }) seam because assigning db().query = ... is silently ignored.
• 🔴 [enforced-workflow] (13:42) User requires two separate review passes for non-trivial changes: 1. adversarial/correctness review covering break-inducing states, enumerated cases, apply-then-revert mutants, property-test stability ×10, hash-equivalence proofs, frozen-array mutation safety, and stale-cache analysis; 2. security/pentest review covering injection, credential leakage, protocol attacks, and privilege escalation. Adversarial review is mandatory before merge; security review depends on risk profile.
• 🔴 [enforced-workflow] (13:42) User requires reviewers to run pnpm test, pnpm run typecheck, pnpm run lint, and pnpm run format:check; verify property tests across 10 runs; confirm the working tree remains unmodified using hashes; and report per-point PASS / FAIL / CONCERN / MUST-FIX verdicts with file:line evidence plus an overall MERGE / DO-NOT-MERGE decision.
• 🔴 (13:42) User requires adversarial review for correctness-critical paths including the sync engine, hashing, lifecycle, and migrations regardless of change size; mechanical or test-only PRs may skip it at reviewer discretion.
• 🔴 (13:43) User asserted that provisioned-throughput and custom-model-import IDs must “always carry a :vN” version suffix, documented in packages/gateway/test/bedrock-runtime.test.ts:619.
• 🔴 (13:43) Current packages/gateway/src/pipeline.ts behavior: emitTextItem() emits the synthetic Responses SSE sequence response.output_item.added → response.content_part.added → response.output_text.delta → response.output_text.done → response.content_part.done → response.output_item.done.
• 🔴 (13:43) Current packages/gateway/src/pipeline.ts behavior: buildTerminal() reconstructs the terminal event using state.terminalEvent ?? "response.completed", computes inclusive input tokens as inputTokens + cacheReadInputTokens + cacheCreationInputTokens, and includes input_tokens_details with cached_tokens and cache_write_tokens when cache usage exists.
• 🔴 (13:43) Current packages/gateway/src/pipeline.ts behavior: buildOutputItems(hiddenIndices) merges sorted indexes from state.rawItems and state.items, omits hidden indexes and item_reference raw items, preserves raw message content when present, emits refusal content when applicable, and reconstructs text messages and function_call items.