We traced a class of Starko AI runs that ended with "agent stopped without reporting a result" partway through a fix. The cause was subtle: a single slow network response inside the sandbox could make the agent's worker conclude its work claim had lapsed and shut down quietly — at a random moment, mid-task.
The runner now supervises itself. If the worker exits while the job is clearly unfinished, it restarts, reconciles what already happened, and picks up exactly where it left off — including answering the step it was stranded on. We validated the fix end to end on a previously failing ticket, which now completes with a real draft PR. Nothing to do on your end.