AI jobs stop silently: 11日間 queue empty, 25 notifications unread

The author runs several recurring AI jobs for a content site — daily research, weekly article planning, and analytics intake. The four incidents share one shape: none of them crashed loudly. One weekly blog-lessons job had been dead for three weeks before anyone asked about it, because there was no path to notify anyone on failure. The author's conclusion is that an automation that stops silently is worse than none, since it creates the illusion that it is working.

After that first miss, the author built a notification for a research queue that pulls one item at a time; when fewer than three remain, it should alert. The queue still went empty from Oct 3 for 11 days, and the notification file shows daily re-detections, the last being the 25th, 11 days after the first. The author notes that even they initially missed the existing notice and opened a duplicate, which buried it further in a growing list.

A separate weekly planning job that writes seven article plans stalled after two articles when an AI usage limit hit; because the plan record was meant to be written as one sheet later, the two finished articles linked to a record that never existed. In another case, a background agent returned exit code 0 four times without finishing any job. The author's stated fixes — escalating repeated notifications and writing records first — were not yet implemented at the time of writing.