Review PR

Spawn parallel subagents to critique the git diff of the current branch (vs main) from independent perspectives. Synthesize findings into a prioritized remediation plan.

When to Use

• Before merging a feature branch
• After implementation to catch issues before PR
• When you want a thorough multi-perspective review of code changes
• When the user says "review my changes" or "critique this branch"

Critique Perspectives

One subagent per perspective. Each critic is independent.

Workflow

1. Gather the Diff

Run these commands to collect context for the critics:

# Get the base branch (usually main)
git merge-base main HEAD

# Full diff against main
git diff main...HEAD

# List of changed files
git diff main...HEAD --name-status

# Recent commit messages on this branch
git log main..HEAD --oneline

If the diff is very large, also get a --stat summary and split the diff by file groups when dispatching critics.

2. Dispatch Critics in Parallel

For each perspective, spawn a subagent with the diff content and a focused prompt. Use the Task tool with subagent_type: "generalPurpose" and readonly: true.

Template per critic:

Critique the following code changes from the [{PERSPECTIVE}] perspective.

**Branch:** {BRANCH_NAME}
**Changed files:** {FILE_LIST}
**Diff:** {DIFF_CONTENT}
**Commit messages:** {COMMIT_LOG}

**Your role:** You are a critical code reviewer focused ONLY on [{PERSPECTIVE}]. Be skeptical. Find real problems, not style nitpicks.

**Check:**
[PERSPECTIVE-SPECIFIC CHECKS - see references/critic-prompts.md]

**Output format:**
- **Issues found:** (Blocker / Important / Minor) — cite file and line from the diff
- **Missing considerations:** What the changes don't address
- **Recommendations:** Specific improvements with code suggestions where possible
- **Verdict:** Ready / Needs changes / High risk

For full prompts per perspective, see references/critic-prompts.md.

Important: Each subagent needs the actual diff content. If the diff is too large for a single prompt, split by domain (e.g., backend files vs frontend files) and send relevant portions to each critic.

3. Synthesize Results

When all critics return:

1. Merge by severity — Collect Blockers from all critics first
2. Deduplicate — Same issue from multiple critics = higher priority
3. Locate — Every issue must reference specific file(s) and line(s)
4. Prioritize — Blockers > Important > Minor

4. Critique Summary

# Review PR Summary

**Branch:** {branch_name}
**Files changed:** {count}
**Commits:** {count}

## Blockers
[Must fix before merge — file:line references]

## Important
[Should fix — file:line references]

## Minor
[Nice to have — file:line references]

## Overall Verdict
[Ready / Needs changes / High risk]

5. Create Remediation Plan

Transform findings into an actionable plan. Group by file or concern, not by critic.

# Remediation Plan

## Tasks

### 1. [Task title — addresses Blocker/Important]
**Files:** list of files to change
**What:** Specific change description
**Why:** Which critique finding this addresses

### 2. [Next task...]
...

## Out of Scope
[Issues deliberately deferred with reasoning]

Plan guidelines:

• One task per logical fix (may span multiple files)
• Blockers become mandatory tasks; Important become recommended; Minor are optional
• Reference the critique finding each task addresses
• Keep tasks small and independently mergeable when possible
• If a finding is invalid (false positive), note it in Out of Scope with reasoning

Feedback as Guidance

Core principle: Critics are advisors. You are the decision-maker.

Act on when:

• The issue is clearly a bug or security risk
• Multiple critics flag the same concern
• The fix is low-cost relative to the risk

Defer when:

• The suggestion would expand scope significantly
• It contradicts project conventions (check CLAUDE.md / AGENTS.md)
• The concern is speculative with low probability

Reject when:

• It's a false positive (critic misread the diff)
• The existing pattern is intentional and documented
• The "fix" would introduce more complexity than it solves

Red Flags

• Don't run critics sequentially (perspectives are independent)
• Don't send critics a summary instead of the actual diff (they need real code)
• Don't include style/formatting issues (the formatter handles those)
• Don't create remediation tasks for false positives
• Don't let critics override project conventions
• Don't skip the remediation plan (findings without a plan are just noise)

Integration

• review-plan — Similar pattern but for plans; use review-pr for code
• superpowers:verification-before-completion — Run after remediation to confirm fixes
• superpowers:finishing-a-development-branch — Use review-pr before finishing