Task status stuck as "running" when process exits unexpectedly (killed/sandboxed)

[jimmy-claw]

Problem

When a task exits unexpectedly (SIGKILL, permission denial, or sandbox block), the task metadata file is never updated from running to failed/completed. The dashboard then shows a stale "running" status indefinitely.

Steps to reproduce

1. Start a task via openclaw-agent start
2. Have the Claude Code process get killed (e.g. SIGKILL from OS, or sandbox permission denial blocking all tool use)
3. Check openclaw-agent list / dashboard — task still shows as running

Expected

Task should transition to failed with an appropriate error when the underlying process exits non-zero or via signal.

Possible fix

• Poll the PID after SSH session ends and update status if process is gone
• Write status on SSH session close / process exit hook
• Add a staleness timeout: mark as failed after N minutes with no output update