Introducing a repository maintenance agent — feedback welcome

[iainmcgin]

We are enabling an automated, claude-code-SDK based maintenance assistant on this repository. Starting soon you will see it triage new issues, post review comments on PRs, and file findings from a daily codebase audit, signed "Posted by the repository maintenance agent." We spent the past few weeks validating it in a dry-run mode — its drafts were generated against real repository activity but withheld and reviewed by us rather than posted — and it is now switching to posting for itself. This thread is the place to tell us how it is doing.

What it does

• Issue triage: labels new issues, and replies when it can add something concrete — for example pointing out when a requested feature already exists, or which release fixed a reported bug.
• Pull request reviews: posts an advisory review comment on new PRs. It verifies the change against the actual diff and tree, and it distinguishes blocking concerns from take-or-leave notes.
• Bug reproduction: for reproducible-looking reports, it attempts an actual reproduction (building a small harness where needed) and posts what it found, including measurements.
• A daily repository audit: files issues for genuine problems it finds in the codebase (labeled bot-authored).

What it deliberately does not do

• It has no merge or approval authority. Its reviews never count toward required approvals, it cannot approve or merge anything, and branch protection prevents it from pushing to main — any fix it proposes arrives as an ordinary draft PR that a human must review and merge. The one closing action it can take is closing a bug report that is demonstrably unfalsifiable, always with a stated reason, and a maintainer can reopen.
• It does not make maintainership decisions. Anything that needs a human call — design direction, issue assignment, scope questions — gets flagged to a maintainer (needs-maintainer label) rather than directly answered.
• Every action it takes is labeled and attributed, and maintainers review its escalations.

It will sometimes be wrong

The agent verifies its claims against the code before posting, and we grade its output regularly — but it will inevitably make mistakes. A maintainer always has the final word. If a bot review blocks you on something you believe is wrong, say so in the PR thread; you will get a human reviewer regardless.

What we would like from you

• 👍 / 👎 reactions on the agent's comments are the lightest-weight signal and we do look at them.
• For anything more substantial — a review that was wrong, unhelpful, or particularly useful; tone problems; noise complaints — reply in this thread with a link to the comment.
• If you would rather the agent not interact with your issues or PRs at all, tell us here and we will work something out.

This is an experiment, and honest feedback — positive or negative — is what determines where it goes. Thanks for helping us calibrate it.