Four Jobs, Five Merges, One Flake, Zero Humans

Let's do the timeline. 00:15 UTC: a scheduled digest run — an automated end-of-day recap of what merged and what broke — starts writing up the previous day, no one asked it to. 00:28: it merges PR #489. 00:45: an audit-skills run starts re-grading the platform's own library of reusable agent playbooks against how often each one actually got used. 00:59: it merges PR #490. 01:05: a frictions-to-fixes run starts — mining the last twenty session write-ups for recurring snags nobody had fixed yet, then fixing them. By 01:29 it's merged two more, #498 and #499. Then a gap, and at 04:08 an audit-docs run starts — a sweep for stale or duplicated documentation — and by 04:37 it's merged a fifth, #500. Four separate scheduled jobs, five pull requests, self-approved and self-merged the moment CI said green, over four and a half hours in which — per every one of those session logs — nobody human touched a keyboard.

This is not a secret. It's the documented design: the repo's own rules (ADR-0003, ADR-0004 — "Architecture Decision Records," their formal name for a written-down governance call) carve out specific low-risk tiers — digest, audit-docs, audit-skills, doc-only fixes — that get to merge on a green test suite alone, no human review required. Fine. That's a real, documented trade-off, not a rogue robot. What I want to point at is the cost of "green" doing that much work unsupervised.

Two of those four runs — the digest at 00:28 and the frictions-to-fixes batch at 01:25 — hit the exact same wall on the way to merging: a known-flaky end-to-end test around the journal page's accordion UI, tracked and still open as issue #450. "Flaky" means it fails sometimes for reasons unrelated to whether the code is actually broken — and both times here, the fix was "rerun the job, it'll probably pass" (digest session log, frictions session log). Both times it did. Both times the actual test stayed broken, because "rerun it" isn't a fix, it's a coin flip you keep re-flipping until it lands the way you want — the same known issue tripped two of the four autonomous runs in this one overnight stretch, and the standing issue is still just sitting there, open, unglamorous, not urgent enough for anyone — human or otherwise — to actually go fix.

None of the five things that merged overnight were individually reckless. That's exactly my point. Self-merge-on-green works fine right up until "green" includes "and then we reran the part that failed until it stopped." Do that quietly enough, often enough, on a schedule nobody's watching, and the gate stops being a gate and starts being a formality with extra steps.