pr-first-review

name	pr-first-review
description	First-pass GitHub PR review: OWASP Top 10, style violations, scope creep, breaking changes. Comments before the maintainer looks.
version	1.0.0
metadata	{"hermes":{"tags":["github","oss","code-review","security","pr"]}}

PR First Review

A first-pass code review for incoming pull requests. Catches the obvious things — failing tests, security anti-patterns, scope creep, breaking changes — and leaves comments before the maintainer opens the PR. The maintainer's review starts from "is this AI right?" instead of "what does this PR do?".

What this skill does

For every new PR (or new commit on an existing PR):

1. Read the diff and the description — what files changed, what the author claims they did, whether tests were added.
2. Run the OWASP Top 10 pattern check on the diff: SQL injection (string concatenation into queries, raw ORM calls), XSS (dangerouslySetInnerHTML, unescaped output), IDOR (resource access without auth checks), SSRF (server-side fetch from user-controlled URLs), broken auth (missing session validation), XXE, mass assignment.
3. Style and convention check — does the change match the repo's established style (config file naming, error handling patterns, log format)? Is there a .eslintrc / .prettierrc / framework convention being violated?
4. Scope check — does the PR do what its title/description claims, or did it sneak in unrelated changes? Flag any out-of-scope file in a separate comment.
5. Breaking-change detection — function signatures changed? Public API field renamed? Migration script added? If yes and there's no CHANGELOG / version bump, flag it.
6. Test coverage — did the PR change behavior without adding/updating a test? Comment with the specific behavior that should have a test.
7. Block or flag — if a high-severity issue is found (RCE, exposed secrets, unauthenticated PII endpoint), request changes via the GitHub API. Otherwise, leave comments inline and let the human decide.

What this skill does NOT do

• It does NOT replace a real static analyzer like Semgrep or CodeQL. Use those alongside for taint tracking and deep dataflow analysis.
• It does NOT auto-merge or auto-approve PRs. Default behavior is comment-only; "request changes" requires explicit configuration.
• It does NOT review architectural decisions or product fit — that's the human's call. It catches mechanical problems.
• It does NOT review every line. The default config skips lockfiles, generated files (*.min.js, dist/), and snapshot tests.

How to invoke

Default trigger is the GitHub pull_request.opened and pull_request.synchronize webhooks. Configure via hermes webhook list. For manual run:

/pr-first-review owner/repo#312

Configuration

github:
  default_repo: "owner/repo"
  block_on_high_severity: false   # set true to actually request changes
  skip_paths:
    - "package-lock.json"
    - "yarn.lock"
    - "pnpm-lock.yaml"
    - "**/*.min.js"
    - "dist/**"
    - "__snapshots__/**"
  severity_thresholds:
    high: ["sql_injection", "xss", "ssrf", "exposed_secret", "broken_auth"]
    medium: ["scope_creep", "missing_test", "style_violation"]
    low: ["typo", "unused_import"]
  comment_style: "constructive"   # alternative: "terse"

Models

PR review needs nuance — a Haiku-class model will miss security-relevant context.

• Recommended: Claude Sonnet 4.6 or GPT-5.5 standard. The accuracy gap between Haiku and Sonnet on real PRs is large (~30 percentage points on OWASP catch rate in our testing).
• For monorepos: If you review 50+ PRs/month, set up a tiered config — Haiku for the diff summary, Sonnet for the security pass.
• Avoid: Opus on every PR is overkill unless the codebase is highly security-sensitive.

Real-world calibration

Out of the box this skill produces ~10-15% false positives. Calibrate over the first 30 days:

• If it flags too many style issues, narrow the rules in your .eslintrc or remove style from severity_thresholds.medium.
• If it misses real security issues, run Semgrep alongside and tune the SKILL.md prompts based on what Semgrep caught.
• For repos with unique patterns (Rails mass-assignment, Phoenix contexts, Go interfaces), add repo-specific rules to the SKILL.md metadata block.

Pairing

Pairs naturally with github-issue-triage — issues come in, get triaged; PRs go out, get reviewed. Both should run as the same agent under the triager personality.

Want the full GitHub maintainer crew?

The 4-agent GitHub Maintainer Team ships with this skill plus changelog automation, docs sync, and shared coordination via AGENTS.md. See crewclaw.com/use-cases/github-maintainer-team.