pr-review

PR review

The repo's sensors already assert what a machine can. This review is the residue: whether the change is right, not whether it is clean. Read the diff — the tree is context, not the subject. /code-review hunts bugs and cleanups in the same diff; this is the judgment pass beside it.

Input

# REST reads — see ~/.claude/CLAUDE.md "GitHub API budget".
gh api repos/:owner/:repo/pulls/<N> \
  --jq '{title, draft, body, base: .base.ref, user: .user.login}'
gh pr diff <N>
gh api repos/:owner/:repo/pulls/<N>/commits --jq '.[].commit.message'

No PR open yet → git log --patch main..HEAD reviews the same change.

Discover

Find what the repo already enforces before reading a line of the diff.

ls .pre-commit-config.yaml .vale.ini Justfile justfile Makefile 2>/dev/null
ls .github/workflows/
cat CLAUDE.md CONTRIBUTING.md 2>/dev/null
gh api repos/:owner/:repo/commits/<head-sha>/check-runs \
  --jq '.check_runs[] | select(.conclusion != "success") | {name, conclusion}'

Red CI is the finding. Report it and stop — restating a sensor's output as review prose tells the author nothing new.

Scope

What the discovered sensors assert draws the review's boundary, in both directions:

• A sensor covers it → it drops out. shellcheck owns quoting, the type checker owns annotations, detect-private-key owns anything key-shaped.
• The repo lacks that sensor → it moves in. An unformatted repo needs a human to notice the formatting.

The bar

Each item names the sensor's half, then yours.

Correctness — Sensors run the tests; they can't say these are the right tests. New behaviour carries a test that fails without the change, and a fix carries a regression test naming the scenario. The diff meets the linked issue's acceptance criteria and stops there — an unrelated refactor riding along is a finding. Edge cases: empty input, an omitted optional argument, a failure partway through.

Secrets — The scanner catches what looks like a key. Catch what doesn't: a token, endpoint, or hostname reading as ordinary config; a file that should be private_ or age-encrypted and isn't; a committed pointer to where a secret lives (token in ~/.config/foo/token), prompt-injection bait for whatever reads the repo next.