Context

The run_scheduler_loop() tokio::select! introduced in #1603 has two branches that break out of the loop without canceling running sub-agents:

1. Ok(None) | Err(_) — channel closed
2. shutdown_signal() fired

Both exit with GraphStatus::Failed but do not call scheduler.cancel_all() or forward Cancel actions to subagent_manager. Sub-agents spawned as tokio tasks will continue running until they complete naturally or timeout, wasting resources.

Impact

Low for MVP — graceful shutdown already handles process exit, and channel close is uncommon during plan execution. However, long-running sub-agent tasks (multi-step tool loops) could consume significant resources after the scheduler loop exits.

Acceptance criteria

In the channel-close and shutdown branches of run_scheduler_loop():

• Call scheduler.cancel_all() and process the returned Cancel actions via subagent_manager.cancel()
• Add a log warning indicating abnormal exit with sub-agent count

Labels

enhancement, robustness