pr-review

Files (3)

• SKILL.md
• pr-context.py · scripts/
• test_pr_context.py · scripts/

SKILL.md (first lines)

Review the **published PR**, not local `HEAD` or uncommitted changes. The review is **read-only**: leave the checkout, index, working-tree files, and external state (PR comments, pushes, remote refs) as you found them. Fetching Git objects and refs is the one permitted write. Fixes require a separate request.

## Scope

From the target worktree, run the helper by its absolute path under this skill's base directory (`--pr <URL>` or `--pr <number> --repo <owner/name>` selects an explicit PR):

```bash
python3 <skill-dir>/scripts/pr-context.py
```

Handle the helper's `status`:

- **`found`**: proceed with the discovered PR and repository when the target is clear; otherwise ask which PR to review. Then run `diff_command`.
- **`ambiguous`**: ask which listed PR to review and rerun with `--pr`.
- **`needs_pr`**: ask for a PR; unresolved discovery does not prove no PR exists.
- **`failed`**: report the error and stop until it is resolved.

## Review

Use `gh pr view <PR-URL> --json files` as the file checklist.

Read the full PR diff. Read surrounding code at `pr.headRefOid` (`git show <sha>:<path>`, `git grep <pattern> <sha>`), not the working tree. If the commit is missing, fetch `refs/pull/<number>/head` from the target repository without changing the checkout; if unavailable, disclose the missing context.

For each behavioral change, trace affected callers and dependencies, examine relevant tests, and account for compatibility. Confirm or rule out each suspected issue, judging against the repository's conventions and agent instructions.

Prioritize correctness, regressions, security, data loss, concurrency, and compatibility. A missing test is a finding when it leaves a concrete changed behavior unprotected. Run checks only when they are read-only; record the rest as unverified.

The review is done when every changed file is either reviewed or listed as excluded with a reason.

Before reporting, re-read the PR's published head and base SHAs. If either changed, refresh the metadata, diff, and file checklist, then revisit affected conclusions.

## Report

Open with a scope header:

- **Scope**: PR URL, target branch, and published head SHA; the comparison is GitHub's PR diff.
- **Exclusions**: skipped files with reasons; local-only changes are outside this review.
- **Limits**: missing context, unverified behavior, and unrun checks. Tests run from a different revision are not verification of the PR head.

Then list **actionable** findings introduced by the changes, ordered by severity:

- `critical`: data loss, security exposure, or a broken primary path in normal use.
- `high`: wrong behavior or a regression on a common path.
- `medium`: wrong behavior under an edge case or uncommon configuration.
- `low`: a minor defect with limited impact or an easy workaround.
- `nit`: a non-blocking stylistic inconsistency or readability/presentation issue with no behavioral impact.

For each finding give:

- **Severity and short title.**
- **File and line**, pointing to the smallest relevant changed range.
- **Trigger and impact**: the conditions under which it occurs and what breaks; nits excluded.
- **Suggested fix**, briefly, when useful.

List uncertain concerns after the findings, each with the evidence that would settle it. With no actionable findings, say so; the header's limits keep that from claiming the code is proven correct.

From bswan0002/skills, MIT.

More from bswan0002/skills