Claude Code tasks sit idle 3 days as prior runs stay stuck

The trigger for the discovery was a habit, not an alert: the tasks write a row to a database at start and another at finish, so on 09-27 the user could see that the start rows for 25, 26 and 27 were missing entirely. That distinction mattered. A missing finish row would have meant a run that stopped midway; a missing start row meant the run never began at all. Without those two rows, a day that ran and did nothing would look identical to a day that never ran.

The user first suspected the Mac had been asleep, since scheduled tasks run locally. Two checks ruled it out: uptime showed the machine had been up four days, and a log file timestamped 09:56 on 2026-09-25 showed activity around the 9:10 and 10:01 scheduled times. The real cause was in the execution history — both tasks had started, stalled seconds later, and sat marked "running" for 46 hours, with no new runs created and the run counter frozen since the 24th.

The shape of this failure is familiar outside Claude Code, in batch jobs that exit abnormally without clearing a lock file, queue workers blocked by a leftover job, or connection pools holding a connection that no longer answers. What ties them together is that no failure notice goes out, because the stuck run is "still in progress" rather than failed. The takeaway the user drew is that detection has to be written by the side that did run — the side that didn't will never report itself — and that the start and finish rows logged from the beginning are what made the diagnosis possible at all.