_global 2026-08 · Country OS

In one line

Cross-country maintenance widened from one repo to three — 34 runs, no errors, seven pull requests, all merged — and ended the month discovering that the fleet's single largest error source had been invisible to its own triage gate for months.

Shipped

Weekly Sentry triage on cms, four pull requests, all merged — #15337 (12 Aug), #16305 (19 Aug), #17463 (26 Aug), #18283 (31 Aug), each merged same-day or the next. The fixes that mattered most: a shared review widget that a human PR had wired to an icon component existing only in the Japan theme, 500-ing on every other site (1,162 events in under 12 hours); a poll API endpoint fleet-wide with no failure guard (428 events); and one mis-encoded offer breaking an entire provider's editorial payload rather than just itself.
First triage covering all three repos (27 Aug) — 64 issues across cms, telecom and renovation; the latter two had never been triaged before. Three fixes shipped as telecom-comparator #455, merged. Nothing in the cms worklist that day was both live and safely fixable, so no PR was opened for it — that is the correct outcome, not a shortfall.
First pull request ever against comparator-core — #25, merged (31 Aug): a Filament form field whose type hint made PHP reject legacy rows before the code's own guard could run.
A stalled pull request recovered — on 5 Aug the review task found an earlier run had committed and pushed fixes but never opened the PR. It opened it (#14400); merged the next day.

In progress / waiting on a human

OS-003 (opened 31 Aug) — roughly 29,000 Sentry events come from workflow rows still holding a class name our own namespace rename retired in March; the migration is written and needs a human to run it.
OS-001 was answered on 28 Aug ("replace with the successor package") but has been reopened: the diagnosis behind it was wrong, and the answer is flagged for re-confirmation now that the question is much smaller than it looked.
OS-002 (MySQL replica pool exhausted on telecom) was answered on 28 Aug — infra stays untouched, cache the queries instead. That caching work is authorised but not started; it belongs to a future telecom triage run.
Long-standing issues referred to a human rather than guessed at, unchanged this month: three Spanish pages throwing a Blade parse error that is not visible in the source, a directory-API timeout whose obvious fix a reviewer already rejected for hiding the real failure, a full-page-cache size limit, a hardcoded gas offer id that intermittently disappears from the live API, and a staging design-system page referencing a component nobody ever built.
On telecom, a database connection-pool family (~6,000 events across 10 issues) is flagged for a capacity review rather than a code workaround.

Problems

The triage gate was hiding the biggest problem it had. One telecom issue accounted for 3,635 of that repo's 3,711 events over 14 days — 98% — while the gate reported "nothing to triage" three runs running, because it compared lifetime event counts against a doubling ratio (so the older and larger an unfixed issue got, the more new events it needed to be looked at again) and because a count column was being mis-parsed into nonsense for every row our own past runs had annotated. Both were fixed on 31 Aug.
A vendor stack frame was mistaken for a vendor bug. The 27 Aug triage blamed an abandoned package and reported that to a human, who answered it. The real cause is our own March namespace rename: rows written before it still carry the old class name, and the recovery sweep re-fails on them forever. One git log would have shown it. This is why OS-001 was reopened and split.
Every one of the seven pull requests merged with zero human review comments, several via auto-merge. Nothing broke, but no reviewer has weighed in on Country OS code since July, and there is no reviewer precedent at all yet for telecom-comparator or comparator-core.
No errored runs. Two scheduled runs were skipped before starting because their gate found nothing to do.

Done with a human in the loop

11 Aug — the workspace stopped pretending to be a country. countries/fr/ became countries/_global/; branches are now countryos/global/*. A human change, no run involved.
31 Aug — a human asked why the fleet's largest error source was never being triaged. The re-diagnosis and both gate fixes above came out of that session, not a scheduled run. (The tooling could not attribute this entry to any single run.)
Two of the three decision-recording runs were triggered by a human's reply in Google Chat, not by the schedule.

Cost