Claude Code job on GitHub Actions stayed green for 2.5 days

The stall ran from the night of 9/18 to the morning of 9/21. The author had already been burned by this pattern: on 9/10 a job returned "Not logged in" and stayed green, and on 9/11 a "session limit" message did the same. The response on 9/11 was to split failures into two types — authentication problems that need a human (red) and usage limits that resolve on their own (kept green). Usage limits were kept green to avoid alert fatigue, and the reason was written to a file that a separate watchdog was supposed to read.

The flaw was structural: the watchdog itself called Claude and ran inside the same weekly quota as the job it was watching, so when the limit hit, the role that was supposed to announce the stop stopped along with it. That watchdog also ran every two hours, 84 times a week, mostly concluding there was nothing waiting. The stop was discovered only by accident, while the author was doing something else and happened to open the history file.

The lesson the author draws is that the watchdog's job was placed on the same dependency as the thing it was meant to monitor. What the fix hinges on is separating the path that reports a stop from the path that stops — notifications now go out via curl without calling Claude, and a Python-only daily aggregation no longer depends on the quota. For anyone running similar scheduled agents, the benefit will depend on whether these paths are truly independent in practice.