pr-review

# PR Review Skill

This skill provides a structured approach to reviewing pull requests using **parallel specialized agents** for thorough coverage, with confidence scoring to minimize false positives.

**Important**: Any text comment you post must be prefixed with:
```
[AUTOMATED]
```
This is important because you are using the Github CLI with the account of your beloved human, and you want to make it clear that the comment is not coming from the human.

---

## Review Workflow

### Step 1: Pre-Review Validation

Before starting, launch a **haiku agent** to check if review should proceed:

```
Check if any of the following are true for the PR:
- The PR is closed
- The PR is a draft
- The PR is trivial (automated, obviously correct single-line change)

If any condition is true, stop and explain why.
```

If validation fails, do not proceed with the review.

### Step 2: Gather Context

Launch **two haiku agents in parallel**:

**Agent A - PR Summary**:
```
Fetch PR details using: uv run scripts/gh_pr.py files owner/repo <number>
If there's a linked issue, fetch it: uv run scripts/gh_pr.py issue owner/repo <issue>
Return a brief summary of:
- What the PR does
- Files changed
- Author's stated intent — based on code and eventual linked issue
```

**Agent B - Guidelines Discovery**:
```
Find all relevant guideline files:
- Root CLAUDE.md (if exists)
- CLAUDE.md files in directories containing modified files
- Any project-specific style guides

Return the file paths and key rules that apply to changed files.
```

### Step 3: Parallel Review Agents

Launch **4 specialized agents in parallel** to review the changes. Each agent should receive:
- PR title and description
- Relevant guideline summaries (from Step 2)
- The diff (via `uv run scripts/gh_pr.py files owner/repo <number> --raw`)

**Agent 1: Bug Hunter (opus)**
See `agents/bug-hunter.md`. Focus on actual bugs and security issues.

**Agent 2: Bug Hunter (opus)**
Same as Agent 1, run twice for coverage. Different agen