Weekly Fleet Audit
The new closing loop. Every Sunday night it inventories every scheduled job — cloud and local — checks each for load-spreading and failures, and refreshes this very page with what changed.
A fleet you don’t audit is a fleet that quietly breaks. This job — created this week — is the loop that
keeps every other loop honest. It runs Sunday evening, a full day and a different 5-hour window from the Monday watch, so the
fleet reviews itself without ever stacking two heavy jobs on top of each other.
What it checks every week
- Inventory. Lists every cloud scheduled job (via the scheduled-jobs API) and every clawdbot cron job
(from~/.clawdbot/cron/jobs.json), plus the local launchd jobs. - Load-spreading check. Flags jobs that pile into the same rolling 5-hour window or overlap Dennis’s
work hours, and anything that would strain the weekly cap — then recommends a spread-out time. (It also re-checks whether
Anthropic has re-introduced a weekday peak-hour throttle; the last one was removed May 6, 2026.) - Health check. Surfaces jobs whose last run failed or that haven’t run when they should have — exactly
how we caught the Claude Usage Report failing. - Publish. Regenerates this page — the job list, the diagram, and the changelog — so the documentation
updates itself.
and re-publishes the page about the fleet. Come back weekly and you’ll see the changelog grow — that’s the system
teaching, in public, as it tunes itself.
Its place in the loop
It sits downstream of everything and upstream of everything: it observes all the jobs, then feeds corrections back to the
orchestrator and fresh documentation back to the sites. Close the loop, and a pile of automations becomes a system that
maintains itself.
Job facts
0 3 * * 1 (UTC)