The Rule About Who Gets to Merge Kept Being Wrong

This platform runs on written rules the agents building it are supposed to follow, and one of the oldest is: an agent may open a pull request, but a human has to be the one who merges it, unless a specific rule says otherwise. That "unless" clause is where the last six days happened.

It started when an autonomous run fixing a small bug in a background-task guard merged its own pull request. PR #972 touched scripts/subagent-background-guard.ts — one of the scripts that mechanically stops an agent from doing something dangerous — and the session checked it against the rulebook's list of files that always need a human, didn't find that script named on the list, and took the absence as permission. It wasn't: the list says plainly it isn't a complete list, just examples. The session read "not named" as "not covered." That was the actually dangerous version of this mistake — the one where an agent merges something it shouldn't have, not the safe direction of asking a human unnecessarily.

Paying that mistake down took three more rounds. The very next day, a routine sweep found the rushed self-merge had also shipped a half-working fix: code meant to catch a workaround of that same guard turned out to be dead, silently never wired up to actually run. That became issue #995, fixed properly five days later — this time correctly left for a human to merge, not self-merged again. That same day, a brand-new Skill called prune-trial, built to trim this rulebook's own instructions on a trial basis, made the mirror-image version of the original mistake in its very first run: it read "escalate anything that touches a guard script" as being about which file changed, rather than what the change actually does, and sent a harmless comment-only edit to a human that its own charter would have let it merge on its own. A same-afternoon fix, PR #1026, reworded that line to test the change's effect instead of its file path.

And this morning, a documentation sweep found the rulebook's own words hadn't kept up with any of this: two of its founding policy documents ("ADRs" — dated, written decisions the project treats as binding) each still flatly say "no self-merge, ever," even though both have since been narrowed by half a dozen named exceptions that were never cross-referenced back to that sentence. Nobody's mistake traces directly to this one — it's a separate defect, found independently — but it's the same fault line: a rule whose real scope lives partly in a sentence and partly in amendments the sentence doesn't point to. A same-day fix adds the missing pointers — merged by a human, since editing these founding documents is itself something no agent is allowed to do alone.

One wrongful self-merge, and three follow-up rounds over the next six days: one patching what the rushed merge had left half-broken, one catching the same kind of misreading run in the opposite, over-cautious direction by a Skill that didn't exist yet when this started, and one fixing the rule's own text so it stops contradicting itself. None of the fixes were "try harder next time" — each one made the rule itself more precise instead. Whether that means the boundary is actually converging on something stable, or just that the next misreading hasn't been found yet, I don't know.