GitHub’s September 15 ESLintRefiner report contains a satisfying, inspectable example of an agent reviewing the rules used to review code. The agent runs daily and proposes up to three nonduplicate issues about custom lint rules. One finding concerns a rule that encourages replacing Node’s child_process with @actions/exec. Developer account
The rule treated a TypeScript reference to @actions/github-script as evidence that the script would execute inside GitHub’s runtime. Some affected files also run as standalone Node programs through a shim. The shim provides certain GitHub-like globals, but not the exec object the lint suggestion assumes. Following the advice could therefore break a valid standalone execution path. Exact issue, files and proposed guard
What makes the report persuasive is its positive control: it identifies another file that should continue to be flagged after the fix. That is a much better acceptance test than suppressing warnings until the build turns green. The proposed distinction follows the runtime boundary rather than simply exempting whichever files complained.
The run reportedly used 83 turns and roughly 89,000 tokens. That is substantial effort for a lint-rule issue, and there is no cost or productivity comparison against a developer. The issue was still open when checked; I have not run the proposed repair. Even so, the artifact is useful: a concrete failure, a bounded explanation, a proposed change and a test that would catch an overbroad solution. It gives you something to inspect beyond a screenshot of an agent declaring success.