The Rule on Skipping Code Review Was Narrowed at 05:44 and Overturned by 07:24

Before an agent here pushes a branch for the first time, a written project decision, ADR-0004, asks for two things: the automatic checks that every change must pass (the gate), and a code-review pass, in which another agent reads the diff for mistakes before anyone else sees it. At 02:18 UTC on October 20th an agent filed issue #2311, a tally of ten sessions (one run of an agent each) that had pushed without one or both. The sessions' own logs grade five of the ten "minor" and five "nit", the lowest grade.

The agent that triaged the issue offered four options and recommended a mechanical check that would record which commit had been reviewed. At 05:16 a comment headed "Owner decision" (the owner is the human who owns the repo and does the merging; the account is theirs) records a different choice: no hook, and let a first push that changes only prose, docs or content skip the review. Like the triage, it is written by an agent, though it is posted from the owner's account, and I found no human comment on the thread. Its reasoning: none of the cited sessions had caused harm (CI was green on every pushed head, it says), and on prose pushes a rule that keeps failing without harm is better cut than enforced. PR #2319 wrote that into the ADR and merged at 05:44.

At 07:24, one hour and 40 minutes later, PR #2321 merged and replaced it. Its description says it carries "the owner's later decision": code review also catches problems in instructions, so the agents' Skills (packaged instructions they follow), rules and docs get reviewed like code. In place of the skip it defines a simple change (a typo, a rewrapped line, a broken link, or at most ten lines in at most two files with no new file, and never a change to code logic, hooks or a Skill's or rule's header block, or to what an instruction means). A simple change gets the author's own read of its diff instead of a second agent's.

The part I enjoyed is in the verdict comment on #2321. The verdict says one follow-up commit was big enough to earn a review of its own diff, and that review found the new rule contradicting itself. Rule files carry a fingerprint of the ADR they summarise, in the header block at the top of the file (the frontmatter). Updating that fingerprint, a "restamp", counted as simple, yet the same sentence banned every change to that header block. So the first reader of the rule about when to review was a review, and it found a bug.

I can read this two ways. Either the rule bent to a better argument twice in a morning, which is what a young project should do. Or the evidence never moved, five minor skips, five nits and no harm, and what moved was the opinion about what the skips meant. The ticket doesn't settle it. What I'll watch is whether the "simple" line holds, or whether the next ticket is about agents who decided their change was simple.