About this template
The git diff is ground truth. Claims are suspects.
If you review AI agent PRs and have ever caught a confident "all tests pass" that turned out to be fiction, this template is your second pair of eyes.
Coding agents invent confidence. The PR body says refactored and tested; the diff says three files changed and no test was touched. This fork treats the git diff as ground truth: it reads the PR body or chat summary, reads what actually changed, and lists every claim the diff does not support.
When it cannot verify a claim, it says UNVERIFIED. Fail-closed, not benefit of the doubt.
How this code review gate works
- Fork the template and sign in to Claude Code.
- Optionally export a
GH_TOKENso it can rungh pr diffagainst remote PRs. - Paste:
Compare the PR body to the diff. List hallucinated claims. - Drop the mismatch list straight into your review comment.
You still read the diff. You still decide the merge. The gate just makes sure the summary never substitutes for the code.
What you get when you fork
- The
agent-claim-vs-diffskill, built around diff-as-ground-truth review - Fail-closed defaults in
~/vault/gate.md, tunable to your repo - A demo prompt that works on a local diff or a PR number
- Your own copy in an isolated Firecracker microVM on your subdomain, running even when your laptop is closed
What you connect
Bring your own keys; they live only inside your VM:
- Claude login or
ANTHROPIC_API_KEY - Optional GitHub token, needed only for remote PRs
Who it's for
You, if agents write a lot of your PRs and you refuse to rubber-stamp green checks alone.
Not you, if you already trust agent summaries without reading the diff. This template will keep disagreeing with you.
Fork it
$25+/mo, 30-day money-back, no free trial. Fork it - your copy boots in about 3 minutes, in time to check your next agent PR before you review it.
Part of the Software Factory collection.