Second-order finding from the #218 review (see also #219, #221).
#218 correctly makes a one-off timeout advisory: a client-side timeout cannot distinguish slow-but-healthy from stuck-broken, so it is not proof the thing is broken — only proof we did not get to check it. That reasoning holds for a single run.
It stops holding for a streak. An advisory that recurs across N consecutive runs is no longer "couldn't verify right now" — it is "unverifiable for a week," which is a real signal about the system.
The risk is that this is more ignorable than the failure mode it replaces, not less. A red gate blocks and eventually gets attention, even if only as "known-red." A green run carrying a loud advisory line gets no attention at all, because nobody reads the logs of a passing run. The nine-day outage happened with a red gate; the same neglect against a green one would last longer.
Suggested: track advisory recurrence per section (label is already stable) and escalate to fatal — or alert — after N consecutive runs. Pick N deliberately; 3 is probably right for a per-push gate.
Credit: raised unprompted by the agent that authored #218 when asked to challenge its own fix.
Second-order finding from the #218 review (see also #219, #221).
#218 correctly makes a one-off timeout advisory: a client-side timeout cannot distinguish slow-but-healthy from stuck-broken, so it is not proof the thing is broken — only proof we did not get to check it. That reasoning holds for a single run.
It stops holding for a streak. An advisory that recurs across N consecutive runs is no longer "couldn't verify right now" — it is "unverifiable for a week," which is a real signal about the system.
The risk is that this is more ignorable than the failure mode it replaces, not less. A red gate blocks and eventually gets attention, even if only as "known-red." A green run carrying a loud advisory line gets no attention at all, because nobody reads the logs of a passing run. The nine-day outage happened with a red gate; the same neglect against a green one would last longer.
Suggested: track advisory recurrence per section (label is already stable) and escalate to fatal — or alert — after N consecutive runs. Pick N deliberately; 3 is probably right for a per-push gate.
Credit: raised unprompted by the agent that authored #218 when asked to challenge its own fix.