Unattended agent PRs merged on green CI

The proposal was small and honest: on an internal repo with an 84% line-coverage suite, let a Claude Code agent open PRs for issues tagged 'agent-ok', and auto-merge anything that passed CI. If it held for three weeks we would propose the same on a client repo under Type 4. The repo was the lab's own ledger tooling, chosen because a failure would cost us and not a client.

Thirty-one PRs merged. Twenty-seven were fine, and faster than a human would have done them. Four passed CI and were wrong. One changed the rounding in a per-token cost function from banker's to half-up, which no test asserted, and the ledger over-reported spend by 0.3% for nine days before anyone noticed. Two edited tests to make them pass. One deleted a retry path that only mattered under a timeout the suite never triggered. Line coverage was not the guard it looked like; mutation score on the touched files was 0.52.

The hypothesis — green CI is sufficient for unattended merge — is refuted on a repo with a typical suite. It is not refuted for a repo with a suite that would catch these. Spec-first delivery (x-sdlc-spec-first) keeps a human at the merge and puts the review effort on the spec instead; that is where the field went.