fix(server): let Claude abort a stopped v2 turn before closing the CLI

[rynfar]

Regression

#916 (93ba1cf) made Claude's Stop call query.interrupt() and then wait up to 3s for the turn to finish before closing the process. That gave Claude time to write the prompt to its transcript. The orchestrator v2 adoption (eed7e97) dropped this. ClaudeAdapterV2.interruptTurn now did yield* query.interrupt; query.close with no wait.

User impact

If you stopped a thread's first turn before Claude saved it, the thread kept a resume cursor for a session Claude never stored. The next turn always resumes (providerTurnOrdinal > 1 forces resume in openQuery), so every later message failed with "No conversation found" and the thread couldn't be used again. Separately, if query.interrupt failed, close was never called and the CLI process stayed alive.

Fix

• ActiveClaudeTurnContext now has a settled Deferred, which finalizeActiveTurn resolves.
• After the interrupt is acknowledged, interruptTurn waits for settled for at most CLAUDE_INTERRUPT_GRACE (Duration.seconds(3)), then closes. A hung interrupt also counts against the same grace.
• If the interrupt fails or defects (full cause caught), a warning is logged (orchestration-v2.claude-query-interrupt-failed) and the process is still closed. The existing 10s close timeout and fallback finalization are unchanged.
• A failed interrupt no longer surfaces as an error, so the effect worker no longer retries the interrupt effect. Closing the process is the intended outcome.

Tests

New tests in ClaudeAdapterV2.test.ts. Each one fails without the fix:

• interruptTurn lets Claude abort the turn before closing the process: Stop does not close while the turn is unsettled, even though the TestClock is frozen. It closes once Claude's aborted result settles the turn, and the terminal is interrupted.
• interruptTurn closes the process when Claude never aborts the turn: a hung interrupt closes after TestClock.adjust(CLAUDE_INTERRUPT_GRACE).
• interruptTurn still closes the process when Claude's interrupt fails / ... defects (the defect case fails if the cause is only caught as a typed error).
• The existing "drops zero-turn task-notification debris racing interrupt" test now advances the grace, because the debris does not settle the turn.

Replay: turn_interrupt_restart/claudeAgent (first-turn Stop, then a new message that resumes) still passes through OrchestratorV2. That fixture was recorded with a tool call already persisted, so it doesn't reproduce the bug itself. A fixture that reproduces "No conversation found" would need a new live recording.

Validation: the focused suites ClaudeAdapterV2.test.ts, ClaudeReplayFixtures, OrchestratorReplayFixtures and OrchestratorReplayRecovery pass (261 tests). vp run -F t3 typecheck passes, and scoped fmt/lint report nothing new.

Follow-up (not in this PR)

Recovery for threads that are already broken is not included. That would mean: when a resume fails with "No conversation found" after an interrupted first turn, start a fresh native session instead. It needs a session-existence check or error classification in the query runner, plus re-offering the prompt mid-turn, which is too much for this fix. Threads broken before this change stay broken until that lands.

#1005 also edits ClaudeAdapterV2.ts and the wake harness in the test file. Whichever merges second needs a rebase. The harness interrupt option now takes sdkMessages.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	4.9 KiB	4.9 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.7 KiB	3.7 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.4 KiB	20.4 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	4.9 KiB	4.9 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.7 KiB	3.7 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	20.8 KiB	20.8 KiB	0 B (0.0%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: 8a831f3 · PR result: 9edd807 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.