coderabbit

name	coderabbit
description	Use when reviewing uncommitted changes, preparing PRs, requesting or receiving code review, handling CodeRabbit/Greptile/Bugbot/GitHub PR comments, checking security/secrets/a11y/code quality, or deciding whether to accept or reject reviewer feedback. Runs AI review via CLI and covers review triage, false-positive pushback, red/blue team profiles, PR-ready gates. NOT for: runtime debugging or test execution.
execute	scripts/review.sh

CodeRabbit - AI Code Review

Fast AI code reviews via CodeRabbit CLI. Free for open source.

Repositories

Works in any git repo. Free tier covers open source repos. For private repos, ensure CodeRabbit is configured in the repo settings.

Quick Commands

cr review --plain           # Human-readable review
cr review --prompt-only     # For AI agents (minimal tokens)
cr review --type uncommitted # Only unstaged changes
cr review --base main       # Compare against main branch

Workflows

Workflow	Use Case
review	Standard code review
verify	Quick verification for Ralph V-* stories
security	Security-focused review
accessibility	A11y audit for UI changes
secrets	Scan for hardcoded secrets/keys
pr-ready	Pre-PR comprehensive check
red-team	Adversarial security/reliability review profile
blue-team	Architecture/quality/maintainability review profile

Output Modes

Flag	Best For	Token Usage
--plain	Humans reading in terminal	High
--prompt-only	AI agents (Ralph, Claude)	Low
(default)	Interactive TUI	N/A

Integration with Ralph

For V-* verification stories, CodeRabbit runs FIRST as a fast pre-check:

1. cr review --prompt-only --type committed - Quick scan
2. If issues found → Fix before Claude verification
3. If clean → Proceed to full Claude verification

This reduces Claude API costs and catches obvious issues fast.

Receiving Review Feedback

Technical evaluation, not emotional performance. Verify before implementing. Push back when a suggestion is wrong for this codebase.

Response pattern:

1. Read the complete feedback before reacting.
2. Understand the requested change; restate it or ask if unclear.
3. Verify the finding against codebase reality.
4. Evaluate whether it is technically sound for this stack and user decision.
5. Respond with a technical acknowledgment or reasoned pushback.
6. Implement one item at a time and test each fix.

Forbidden responses:

• "You're absolutely right!"
• "Great point!"
• "Thanks for catching that!"

Instead, state the technical requirement, ask a clarifying question, or fix it silently.

Comment type	Action
Real bug / Security	Fix immediately
Important improvement	Fix before proceeding
Style preference	Fix if genuinely better; skip if bikeshed
Over-engineering	Skip with reasoning
False positive	Skip with reasoning

Implementation order for multi-item feedback:

1. Clarify anything unclear first.
2. Fix blocking breakage or security issues.
3. Apply simple fixes.
4. Apply complex refactors or logic changes.
5. Test each fix individually and verify no regressions.

Max 3 review-fix rounds for nitpicks. Push back when the suggestion breaks existing behavior, lacks context, violates YAGNI, conflicts with user-stated architecture, or is technically incorrect. If you were wrong, say "Checked X and you're correct. Fixing." Then fix it.

Evaluator Agent (Weighted Quality Gate)

For high-stakes changes, pair CodeRabbit with the evaluator agent (claude --agent evaluator) for deeper qualitative scoring:

1. CodeRabbit catches structural issues (bugs, security, style)
2. Evaluator scores on 4 weighted criteria: Functionality (20%), Craft (20%), Design (30%), Originality (30%)
3. Score >= 7.0 required to proceed to merge

The evaluator is deliberately adversarial -- it compensates for LLM optimism bias in code review. See ~/Gits/orchestrator/standards/evaluator-grading.md for the full grading rubric.

When to add the evaluator gate:

• Architecture changes or new module introductions
• Agent-generated code (autonomous work products)
• Changes touching >5 files or crossing module boundaries

Configuration

Optional .coderabbit.yaml in repo root for custom rules:

reviews:
  language: en
  path_filters:
    - "!**/*.test.ts"
    - "!**/node_modules/**"

Requirements

• CodeRabbit CLI installed: curl -fsSL https://cli.coderabbit.ai/install.sh | sh
• Authenticated: cr auth login
• Must run from git repository root