Pull request scheduling AI skill

October 5, 2026

William Martin of GitHub has written an agent skill that pays down technical debt in GitHub’s command line interface cli/cli by opening a managed stream of pull requests. The skill and William’s commentary are in pull request 14095, merged in August 2026. The interesting parts are the set of constraints that keep the loop from outrunning the human who has to read the diff.

Martin aims to clear a backlog of roughly 1900 findings. A single cleanup pull request would be unreviewable. The skill is the alternative: one verified gap per run, opened ready for review, and then the run stops. The goal is a change so small that reviewing it takes about two minutes.

A stalled review halts production rather than letting work pile up behind it, so tick frequency does not matter: merging one a day yields one a day. An unattended run may try up to three targets. The first attempt that validates becomes the pull request. Failure notes from the others ride along so the next run does not rediscover them. A run that lands nothing is silent - the absence of a pull request is the signal. A human handles review comments, and the one-open-pull-request cap means the loop waits.

Steering goes through a committed file .experiments/tech-debt-burndown/memory.md. A human owns the Current focus section. The run may append rejected targets and failed attempts below it with a 150-line budget, and must re-check any factual claim it finds there. The pull request is opened ready for human review.