Distillation

Date: Sep 17, 2026

• 🟡 (02:01) Independent correctness review task ses_f52ff5084ffe7snDvSQgJjIjEs completed with verdict DO-NOT-MERGE for frozen patch SHA-256 a2bce01664538243cb8ba16c6ca8f0364c6f79463ad6677567044a1b3e1e74fe against HEAD ab9408c81cc9916177641ae8e9f08624fc27b5b9, tree d3323b963b9ef2de41a7ff9884299c74b93886ba, and SEA SHA-256 753b05afb86195246b211237ca5b11e26a7c08da863d70a711da2c93ee453c8a.
• 🟡 (02:01) Correctness MUST-FIX 1: cancellation became durable before runner finalizers joined. SessionRunner published Step.Interrupted and atomically transitioned cancelling -> cancelled in packages/core/src/session/runner/llm.ts:307-374, while the enclosing scope remained active until line 569. Regression packages/core/test/session-runner.test.ts:4055-4106 demonstrated one durable interruption event at lines 4078-4084 while an injected runner finalizer remained blocked until line 4103. Although SessionRunCoordinator joins the owner before cleanup at packages/core/src/session/run-coordinator.ts:107-108, cleanup then encountered cancelled and merely cleared the claim at packages/core/src/session/execution/local.ts:130-131.
• 🟡 (02:01) Required lifecycle correction from the review: while a runner finalizer is blocked, the claim must remain cancelling and no Step.Interrupted may exist; only after the finalizer joins should exactly one atomic interruption settlement occur and the claim be cleared.
• 🟡 (02:01) Correctness MUST-FIX 2: crashes after durable cancellation could permanently wedge a Session. Durable cancelled crash boundaries existed in the completed-assistant path at packages/core/src/session/execution/local.ts:146-148, durable Step.Interrupted path at packages/core/src/session/execution/local.ts:151-164, and runner settlement at packages/core/src/session/runner/llm.ts:359-371. Recovery quarantined both cancelling and cancelled at packages/core/src/session/recovery.ts:121-139; rows with recovery_error were excluded at lines 71-76; the unique Session claim index was defined at packages/core/src/session/sql.ts:237-239; acquisition used onConflictDoNothing() and returned without execution at packages/core/src/session/execution/local.ts:52-70; and no production path cleared execution recovery_error.
• 🟡 (02:01) Required recovery correction from the review: an expired cancelled claim has a known durable outcome and must be fenced-cleared during recovery; only an expired cancelling claim has an uncertain outcome and should be quarantined. Existing regression packages/core/test/session-projector.test.ts:2018-2065 covered only cancelling, with no crashed-cancelled regression for either settlement path.
• 🟡 (02:01) Correctness concern: initial and final Snapshot.capture calls at packages/core/src/session/runner/llm.ts:290 and :532 lacked immediate ownership checks. Snapshot.capture writes a content-addressed Git tree at packages/core/src/snapshot.ts:129-140; the initial capture’s preceding general claim check was at llm.ts:219, with the next explicit check only before provider dispatch at line 376. A stale owner could therefore write orphaned snapshot state after takeover, although later event publication remained fenced.
• 🟡 (02:01) Review confirmed the candidate correctly handled: durable Step.Started notification interruption; completed-assistant cancellation; same-owner cancellation phase guards; durable arming of explicitly resumed admit-only input; recovery identity validation; bounded scans; migration/schema/index synchronization; and generated current-client/OpenAPI/SDK event surfaces. Residual risks remained the disclosed synchronous check-to-external-call window, process-local interruption routing, and bounded wall-clock skew.
• 🟡 (02:01) Review provenance verified exactly 34 paths—29 modified and 5 untracked—with a clean index at the base commit; forward cached applicability passed, reverse applicability against the frozen candidate passed, every patch base blob matched the base commit, and every postimage matched the frozen candidate. The reviewer ran no tests, builds, or services.
• 🟡 (02:01) Assistant rejected v19 and selected four fixes: 1. move all cancelling -> cancelled settlement exclusively into post-join coordinator cleanup, 2. make expired cancelled claims safely clearable on startup, 3. add immediate ownership checks before both snapshot writes, and 4. add cancellation crash and finalizer-order regressions.
• 🟡 (02:01) Revised work queue: 1. Move cancellation settlement after runner finalizers join—in_progress, high priority; 2. Clear expired durably cancelled claims on recovery—pending, high; 3. Fence snapshot writes at ownership boundaries—pending, high; 4. Add cancellation crash and finalizer-order regressions—pending, high; 5. Run full package validation and isolated artifact restart smoke—pending, high; 6. Freeze revised candidate and obtain independent correctness and security approval—pending, high.
• 🟡 (02:02) Cancellation architecture decision: external interruption should make the runner stop, join tool fibers, and propagate the interrupt without publishing Step.Interrupted or changing the execution phase. After the outer scope and runner finalizers complete, coordinator cleanup performs the single durable interruption transaction and clears the claim. User decline remains a distinct deliberate in-runner settlement path.
• 🟡 (02:03) Cancellation-settlement implementation locations identified: renewCancellation, settleCancellation, and settleInterruption in packages/core/src/session/runner/llm.ts; beginCancellationSettlement in packages/core/src/session/runner/publish-llm-event.ts:476; SessionExecutionClaim.settleCancellation in packages/core/src/session/execution/claim.ts:84; and coordinator-cleanup settlement paths in packages/core/src/session/execution/local.ts:129-178.
• 🟡 (02:05) Production edits were applied to packages/core/src/session/runner/llm.ts, packages/core/src/session/runner/publish-llm-event.ts, and packages/core/src/session/recovery.ts: runner-owned explicit-cancellation publication/phase settlement was removed in favor of post-finalizer coordinator cleanup, and recovery handling was split so cancelled is clearable while only cancelling is quarantined.
• 🟡 (02:05) After the production edit, the sole remaining SessionMessageTable match in packages/core/src/session/runner/llm.ts was the import at line 37 (import { SessionExecutionTable, SessionMessageTable, SessionRecoveryTable } from "../sql"); a follow-up edit updated packages/core/src/session/runner/llm.ts to clean up the now-unused import.
• 🟡 (02:05) Planned finalizer-order regression strengthening in packages/core/test/session-runner.test.ts:4055-4106: while releaseFinalizer remains blocked, assert zero session.next.step.interrupted.1 events, execution phase cancelling, and inability to insert replacement claim replacement-during-finalizer; after releasing the finalizer, assert exactly one interruption event and no execution claim.
• 🟡 (02:05) Planned startup recovery regression: seed an expired cancelled execution claim and prove startup clears it without quarantine, recovery continuation, or a lingering blocked Session claim.