Staff Review
A local, staff-engineer-grade code review — for your working tree, before anyone else sees it.
Open a GitHub-style review of any diff in your browser and let any AI coding agent leave a thorough, inline review on it — then fix, document, or skip each comment. No PR, no cloud, no waiting.
# set up the repo (writes the /staff-* skills)
staff install
# open “main vs. working tree” in your browser
staff main..WT
# then, in your agent:
/staff-review main..WT → /staff-resolveWhy you’ll want it
Review any diff, locally, in seconds
Compare a branch, a commit, a range, your staged changes, or your uncommitted working tree — main..WT, <sha>..<sha>, release..main. A clean split/unified diff opens in your browser with inline comments, replies, and resolutions. Catch issues before you push — no GitHub PR required.
A thorough review from any harness or model
Staff Review ships editable skills that drive a staff-engineer-level review: trace every changed hunk through its edge cases, read the callers and tests, and leave concrete, actionable comments. Use Claude Code out of the box, or any agent that can read a SKILL.md and run a shell command.
Capture project-specific concerns once
Flag a comment with Document, then run /staff-resolve — the agent writes it up as a library entry. Every future review cross-checks the diff against that library, so the gotcha your team keeps re-learning gets caught automatically. Commit the library and the whole team benefits.
Loop review → resolve for higher quality
/staff-loop runs the review and the fixes in isolated subagents, round after round, until a fresh review finds nothing new (or a cap you set). Each round’s fixes get re-reviewed, so regressions and missed issues surface on their own.
Staff Review itself is 100% local — a small web server reading your local git history. Your agent still talks to whatever model you choose; pick a local one if you want zero data to leave the machine.
Do your human reviewers a favor
Install with Homebrew and open your first review in about a minute.