The week of silent failures
The week of silent failures
We spent a week hardening the curation pipeline, and every real bug we found had the same shape: the system failed silently.
A length cap beheaded long reviews — and the most detailed skills were exactly the ones being lost. A sync step quietly overwrote skills the agent had just patched. A stalled review sat forever with nobody assigned to notice. A machine ran five days on stale code while reporting, every day, that everything checked out.
None of these raised an error, because the pipeline is designed never to interrupt the person working. That is the right design — and applied everywhere, it became "never tell anyone anything."
The fix was not a bugfix. Machines now report on themselves: which build they run, whether their reviewer works, when a review last concluded, and how many have failed since. The dashboard says, per machine and in plain words, what is wrong and what will fix it — and a healthy machine says so explicitly, because "no warning" and "never reported" must not look the same.
The lesson we kept: silence is a failure mode. If a system is built to stay out of the way, it needs exactly one loud channel — about itself.