Blackbox Blast
BB Preflight
← Incident log

CASE · Major · monitoring-gap

Unverified reportCollected from public reports

A scheduled Claude Code job did nothing for 2.5 days while GitHub Actions stayed green (unverified report)

A GitHub Actions job that ran Claude Code every four hours to revise blog articles produced nothing for 2.5 days, from the evening of September 18 to the morning of September 21. Every run in that period showed green, and the eight runs on record for September 20 to 21 all ended at once on the usage limit.

3 min read

OBSERVED
CATEGORY
monitoring-gap
BLAST RADIUS
uptime
DETECTED BY
manual-observation
TAGS
#claude-code #github-actions #usage-limit #exit-code #silent-failure #watchdog

CH 1

Cause

The weekly usage limit ran out, and Claude printed a one-line limit message. The step ended in || true, pipefail was off so the pipeline's status came from tee, and alerts only fired on red. The watchdog meant to report blocked runs also ran on Claude and the same quota, so the work and its monitoring stopped together.

CH 2

Consequence

No article was revised for 2.5 days (September 18 to 21, 2026). On September 21, while looking into something else, the author opened a status file and found eight consecutive usage-limit entries. On September 20, two other Claude-based jobs had also hit the limit and failed.

CH 3

Fix

The author actually changed three things: the monitor now starts every six hours instead of every two, cutting 77 unneeded starts a week; the reason a run stopped also goes into a daily summary that does not use Claude; and leaving a marker is now treated separately from the marker reaching a person. The article also recommends moving the pass/fail decision out of shell into tested Python, counting what each run produced, keeping 'zero results' apart from 'could not check', and sending alerts with plain curl, though the author says their own job still greps inside the YAML.

What happened

A freelance engineer ran Claude Code on a schedule: every four hours, a GitHub Actions workflow on a self-hosted runner called claude -p to rewrite blog articles for better search rankings. From the evening of September 18 to the morning of September 21, 2026, the job changed nothing. Every run was green and took about 25 seconds, for work that normally takes tens of minutes.

The chaos on the ground

Nobody was told: failure alerts fire only when a run turns red. On September 21, while looking into something else, the author opened the JSON file that records why runs stopped, and found that the eight runs on September 20 and 21 had all ended on the usage limit. On September 20, two other jobs that relied on Claude, a weekly editorial meeting and a business review, had also hit the limit and failed.

Root cause

The weekly usage limit had run out, and Claude answered each call with a single line saying so. Three habits turned that into a green check. The command ended in || true, which throws away the exit code. The shell did not set pipefail, so the pipeline reported the status of tee, which succeeds. And notifications only went out when a run turned red.

There was a watchdog meant to catch exactly this, reading the status file and reporting problems. But it also used Claude, from the same weekly pool, so it went quiet at the same moment as the job it was watching.

The fix

The author actually changed three things. The monitor now starts every six hours instead of every two, part of cutting 77 unneeded starts a week. The reason a run stopped now also appears in a daily summary that runs on plain Python, without Claude. And leaving a marker is now treated as a separate question from whether the marker reached a person.

The article also recommends fixes anyone can add to their own workflow: move the success-or-failure decision out of shell into Python, where it can be tested against real error messages; count what each run produced, and report "zero results" separately from "could not check"; and send alerts with plain curl, so a quota problem cannot silence the alarm about the quota problem. The author is candid that their own job still greps inside the YAML.

The same lesson shows up in an earlier entry: a green check means the process exited cleanly, not that the work happened. And a monitor that shares a failure mode with the thing it watches is not a monitor.

Sources

PREFLIGHTIs your own Mac prepared?BB Preflight is a free Mac tool that checks 21 things, from AI agent permissions to where your keys live, without changing anything.Run the check →