Agent-Skills.md

Agent Skills: PR Hygiene

Surface open PRs that have genuinely unaddressed review feedback, filtering out CodeRabbit auto-summaries and CI-bot noise. Use before ending a session, when asked "which of my PRs need attention", or to gate a session close on unresolved review threads.

UncategorizedID: idvorkin/chop-conventions/pr-hygiene

Repository

idvorkin/chop-conventions
idvorkin
11

Install this agent skill to your local

pnpm dlx add-skill https://github.com/idvorkin/chop-conventions/tree/HEAD/skills/pr-hygiene

Skill Files

Browse the full folder contents for pr-hygiene.

Download Skill

Loading file tree…

skills/pr-hygiene/SKILL.md

Skill Metadata

Name
pr-hygiene
Description
Surface open PRs that have genuinely unaddressed review feedback, filtering out CodeRabbit auto-summaries and CI-bot noise. Use before ending a session, when asked "which of my PRs need attention", or to gate a session close on unresolved review threads.

PR Hygiene

git-safe-push guards one PR footgun (pushing to a merged branch). This guards its sibling: open PRs quietly piling up with review comments you never resolved or replied to. GitHub gives no nudge, and a naive "the last comment isn't mine" filter over-flags ~3× — it can't tell a CodeRabbit auto-summary or a CI-bot ping from a real ask.

pr-hygiene enumerates your open PRs across GitHub identities and sorts them into three tiers so you act only on the ones that need it.

When to use

  • Before ending a work session — run it, resolve/reply to anything 🔴.
  • "Which of my PRs need attention?" / "Do I have stale review comments?"
  • As a close-out gate: it exits nonzero when any 🔴 remains.

Install

It is a single self-bootstrapping uv run --script file — no packaging needed. Symlink it onto $PATH the same way as git-safe-push:

ln -sf ~/gits/chop-conventions/skills/pr-hygiene/pr_hygiene.py ~/.local/bin/pr-hygiene

The PEP 723 shebang resolves typer on first run. Requires gh (authenticated).

Usage

pr-hygiene                        # ranked markdown table (🔴 first)
pr-hygiene --json                 # machine-readable JSON (+ counts, errors)
pr-hygiene --repo idvorkin/chop-conventions   # filter to one repo
pr-hygiene --author octocat       # override author (repeatable)
pr-hygiene --no-fail              # always exit 0 (don't gate on 🔴)

Default authors are idvorkin and idvorkin-ai-tools. Bitbucket repos (e.g. igor2) are invisible to gh and simply aren't listed. Per-PR GraphQL failures are reported inline (tier error), never crash the run.

How it classifies — the whole point

For each PR one GraphQL round-trip pulls review threads (isResolved), reviewDecision, reviews, and issue comments. Tiers:

  • 🔴 needs action — ≥1 unresolved review thread that is a real ask (a human reviewer, or a bot finding that isn't an auto-summary), OR reviewDecision == CHANGES_REQUESTED, OR a human reviewer has the last word and the author hasn't responded since (push, comment, or review).
  • 🟡 awaiting merge — flagged only because a CodeRabbit auto-summary or a github-actions/CI comment is the last event, with no open threads. Merge-ready, not stale.
  • 🟢 clean — no review activity at all.

Two noise rules make this work:

  1. CodeRabbit walkthrough/summary comments and CI-bot (github-actions) comments are NOISE, never asks — screened by bot login at the issue level.
  2. A bot finding inside a review thread is a real ask — CodeRabbit and Copilot post genuine inline findings under the same login, so thread comments are screened on the body (auto-summary markers) rather than the login. Get this backwards and either every summary flags red, or every real finding gets dropped.

Cross-identity nuance: a comment counts as a reviewer ask only when its author differs from the PR author — so a self-note on your own PR is ignored, while a thread left by your other identity (e.g. idvorkin reviewing an idvorkin-ai-tools PR) correctly counts.

The classification logic lives in pure, unit-tested functions in pr_hygiene.py (see tests/); gh shell-outs are injected so tests run offline.