"I told you at the beginning." One maintainer, one correction, three agent pull requests in six days.

We read every human comment on 39 agent pull requests across five public repositories, looking for the same correction written more than once to the same repository's agents. Here is the one we found, and what it costs.

Automatic matching found nothing: maintainers' corrections are one sentence long and worded differently each time. Reading all sixteen found the repeat.

The repeat

In oxc-project/oxc, one maintainer told the Copilot coding agent to remove a stray file and not to add it to .gitignore, on three unrelated pull requests within six days:

Four comments, one correction, three PRs, one person's patience, in public. On two of the same PRs the maintainer also repeated a procedural nudge, "@copilot check the suggestions", a weaker signal of the same thing: nothing carried what they said from one PR to the next.

The rule, as the file it becomes

Reply on the first PR with antelier: forbid /./ in path/to/the-stray-file and the next run answers with this, ready to commit to .github/antelier/rules.yml on the base branch:

version: 1
rules:
  - id: no-stray-file
    kind: forbid
    scope: "path/to/the-stray-file"
    match:
      regex: "."
      in: added
    text: "Remove this file; do not add it to .gitignore. (#19983)"

Rules load from the base branch by name, so a PR cannot disarm the rule it violates. antelier rules export-agents --write AGENTS.md puts the same sentence in front of the agent before it starts.

What we read, by repository

curl had one agent PR to read at all, which fits its stated policy on AI-generated contributions. The other fifteen corrections were single instances in this sample: "Do not open unattended PR's", "You borked it. Did you forget to run pnpm install?", "what is this, why did you add it? Remove it."