Restart continuation resumes every in-flight run at once, and its 5-minute loop guard misses a crash loop that repeats every 22 minutes #24059

What happened

User: "cpu load goes to 100% at login" and "the orchestrator is trying to run multiple sims at once with different agents".
The Mac was force-restarted at 20:05 and load averages were 339 / 623 / 898 (1 / 5 / 15 min) at 20:32.
Each login starts t3 serve (launchd, RunAtLoad), which immediately re-launches the same agent tasks that overloaded the machine.

Diagnosis

• A delegating thread started 4-5 delegated iOS tasks within 4-11 s, three times in one afternoon. Each task drives a simulator through agent-device.
• After the 20:05 restart, effect-worker.start -> ProviderTurnStartService.start -> resumeThread/startTurn created 4 continuation runs within 1 s. A second restart at 20:32 continued 2 of them again.
• continueRestartedRun dispatches with start_immediately and has no limit on how many run at once.
• The loop guard counts only links within CONTINUATION_LOOP_WINDOW_MS (5 min) of the source run's completion. With ~22 min between restarts it counted 1 (< 3), so it never engaged. PR fix(server): limit restart loops within a five-minute window #20306 notes that slower loops "can continue indefinitely".
• Contributing: LocalDeviceHost.ts has no cap on concurrent simulator sessions (see [Bug]: Thread device claims are never released — device_close and thread settle leave simulators "in use" indefinitely #17470).
• Suggested direction: stagger or cap simultaneous continuations, and detect repeated restarts over a longer horizon.

Steps to reproduce

Not run deterministically, because it crashes the machine.

1. On a 16 GB Mac, have one thread delegate 4+ tasks that each boot an iOS simulator.
2. Let the machine overload until it restarts.
3. Let the launch agent start t3 serve at login.
4. All in-flight runs continue at once. Repeat.

Version

0.1.3-nightly.20261023.3553 for the 20:32 restart. The build at the 20:09 start is unknown. The t3 triage CLI was also 0.1.3-nightly.20261023.3553.

Environment

macOS 27.2.0 arm64, 16 GB RAM, Node v24.22.0 (triage CLI), Nightly desktop app plus launchd t3 serve, Claude provider.

Evidence

task  ord status       requested(UTC)  completed(UTC)  kind
A     2   cancelled    09:09:35        09:32:07        continuation
A     3   cancelled    09:32:15        09:36:27        continuation
D     2   cancelled    09:09:36        09:32:06        continuation
D     3   cancelled    09:32:15        09:36:28        continuation
B,C   2   interrupted  09:09:35-36     09:10:26-27     continuation
task start waves (UTC): 05:09:30-41 (4 tasks), 06:26:36-44 (4), 07:02:34-38 (4)
provider-runtime.continue effects since 2026-10-16: 44; up to 8 within one hour
swap at 20:40: 5830 of 6144 MB used

Related issues

PR #18461 and PR #20306 (the continuation cap and its window). #17470 and #16614 (simulator claims and hub CPU). None covers this problem.

Fix applied or workaround

Stopped the service, ran xcrun simctl shutdown all, disabled the launch agent, and set continueThreadsAfterServerUpdate to false in settings.json.

Filed by

claude (claude-sonnet-5-5) via Claude Code, running the t3 triage playbook