review-harness

A tiny add-on that teaches an existing agent to review PRs.

It judges a pull request as YES or NO — does this PR improve the codebase? — against a persistent understanding of that codebase that the harness builds once and updates incrementally.

Inspired by browser-harness: one capability, minimal surface, installable into any host agent with a claude CLI.

Install

git clone https://github.com/burner-agent/review-harness
cd review-harness
./install.sh

Requires claude and gh on PATH. gh must be authenticated with push access to the repos you want to review.

Use

Adopt a repo once:

review-harness adopt owner/repo

Clones, reads, writes MAP.md to an orphan review-harness/memory branch.

Review a PR:

review-harness review https://github.com/owner/repo/pull/42

Output: YES or NO on the first line, reasoning below. Exit code 0 or 1.

How it works

1. Adopt — one-time onboarding pass. Claude reads CLAUDE.md / AGENTS.md / README / top-level dirs and writes a 100–300 line map: purpose, architecture, conventions, hot files, glossary.
2. Catch up — every review invocation compares MAP.md's last-seen-sha to origin/<base>. If main has moved, Claude does a surgical update to the map before judging the PR.
3. Review — Claude reads MAP.md, reads the PR diff + touched files + neighbors, judges.

Memory lives on a dedicated review-harness/memory branch on the target repo, so any machine with the harness installed inherits the accumulated understanding for free. No central server, no vector DB, no embeddings.

Why a markdown map and not embeddings

Code is structured, not fuzzy text. rg + symbol navigation (what claude -p uses natively) is more precise than vector similarity for "find the caller of this function." A curated markdown map captures what a PR reviewer actually needs — purpose, conventions, hot files — in a form that's human-readable, diff-able, and cheap to keep current.

Adding this to an agent

The host agent needs to know this tool exists. The simplest integration: drop a skill file.

For arbos-style skill systems:

---
name: pr-review
trigger: keyword
keywords: pr review, review pr, pull request
---
Run `review-harness review <pr-url>` to get a YES/NO verdict.
Run `review-harness adopt <owner>/<repo>` once before reviewing a new repo.

That's it. The host decides when to invoke and what to do with the verdict.

License

MIT