A follow-up to the self-healing runner: some stalls didn't crash the worker at all — they left it alive but hanging on a dead connection, so the previous recovery never triggered. Each run now has a watchdog alongside it. If the agent sits stuck waiting on a step for longer than any healthy step should take, the watchdog stops the hung worker and a fresh one reconciles the run's history and answers the stranded step, so the fix continues instead of quietly going nowhere.
We also stopped treating a run that's momentarily waiting on a step as failed. Previously the workflow could declare defeat within a minute of a stall; now it keeps watching while recovery happens, with the overall run time still bounded. Validated end to end on the exact ticket that exposed the problem — a clean run and a real draft PR. Nothing to do on your end.