OpenHands Bug Exposes Gap Between Service Health and Session Liveness in AI Agents

A real incident logged in OpenHands PR #4548 revealed that a browser-tool conversation remained stuck for over eight hours while health-check endpoints continued returning HTTP 200 responses. The root cause was an unbounded shutdown wait in AsyncExecutor.close(), where remaining tasks were never cancelled, leaving the session unable to make progress despite appearing healthy at the service layer. The fix introduces task cancellation and a 10-second timeout on shutdown, explicitly framed as a best-effort safety net rather than a guarantee of clean resource cleanup. A separate multi-agent system called CodeFlowMu surfaced a related but distinct problem: a recovery path could reactivate a worker before the task had completed the lifecycle transition needed for legitimate execution. Both cases illustrate that reliable agent recovery requires verifying at least five separate layers — service health, session liveness, task eligibility, causal prerequisites, and formal acceptance — rather than treating a single healthy status as sufficient to resume work.