fix: recover agent threads after laptop sleep instead of erroring

[vedranjukic]

Summary

After laptop sleep, running agent threads would show "Error connecting to bridge" or "Agent stopped responding after 90 seconds" even though the agent was still alive in the sandbox. This happened because:

1. Stale timers from before sleep fire immediately on wake
2. The bridge WebSocket died during sleep, so health checks fail
3. Activity timeouts fire as expired, marking the thread as error

Three changes fix this:

• Activity timeout recovery: Before erroring, attempts to reconnect the bridge and request replay. If the bridge is reachable (common — it just needs a new WebSocket), recovers the thread with replay instead of failing.
• Grace timer deferral: On reconnect failure, silently cleans up the handler instead of marking the thread as error. Recovery is deferred to reconcileAndReconnect on the next subscribe_project.
• Stale handler cleanup: reconcileAndReconnect now cleans up all stale active handlers and their timers before reconnecting, preventing ghost timeouts from the pre-sleep session.

Test plan

• Start an agent run, close laptop lid for 2+ minutes, reopen — thread should recover (not error)
• Start an agent run, disconnect WiFi for 60s, reconnect — thread should recover
• Verify normal timeouts still work (kill the bridge process manually — should eventually error after retry)
• Verify reconcileAndReconnect properly restores running threads after sleep

Made with Cursor