I think I found a cause of crashed conversations on the cloud — long-running commands would cause a lock to be held, and if another lock-using operation (e.g. event load when re-connecting the websocket) also tried to acquire the lock, it would also acquire the lock and cause the fastapi thread to stall, causing liveness probes to fail: https://github.com/OpenHands/software-agent-sdk/pull/2622 (cc @straight-sundown-84319)

taking a look 👀 - code LGTM! Just sending the agent to verify the artifact / reproduce the issue