llm-review

name	llm-review
description	Run an LLM code review (OpenAI Codex + Anthropic Claude in parallel) on uncommitted changes or against a base ref. Use when user asks to "review my changes", "review uncommitted", "review against <ref>", "code review", "LLM review", "review branch vs <ref>", "second opinion on this diff".
allowed-tools	Bash, Read

LLM Code Review

Runs two independent LLM reviewers in parallel (OpenAI Codex CLI and Anthropic Claude CLI) on the current diff and presents both reviews. Cross-vendor coverage catches issues a single model misses.

Activation Triggers

• "review my changes", "review uncommitted"
• "review against ", "review vs ", "review branch vs main"
• "code review", "LLM review", "second opinion on this diff"
• "run the reviewers"

Usage

~/.claude/skills/llm-review/scripts/llm-review.sh           # uncommitted
~/.claude/skills/llm-review/scripts/llm-review.sh master    # current branch vs master
~/.claude/skills/llm-review/scripts/llm-review.sh HEAD~3    # last 3 commits

Workflow

1. Run the script with the appropriate arg.
2. Read both reviews end-to-end — they're printed under Codex (OpenAI) and Claude (Anthropic) headers. They will overlap on critical issues and diverge on judgment calls.
3. Synthesize for the user: present a merged summary with these sections:
   • Critical (both flagged) — high confidence, fix first
   • Critical (one flagged) — investigate, may be a false positive but worth checking
   • Important — should fix
   • Minor — optional
   • Disagreements — where the two reviewers conflict, with your read on which is right
4. Wait for user approval before making any code changes. Never start editing based on the review output without an explicit go-ahead.
5. After fixes, re-run the script to verify the issues were addressed and no regressions were introduced.

Project Customization

If the repo has a REVIEW_GUIDELINES.md at the git root, the script appends its contents to the review prompt automatically. Use this for project-specific patterns (sqlc usage, error wrapping, test conventions, etc.).

Tuning the Review Prompt

The prompt in scripts/llm-review.sh (between PROMPT_EOF markers) is sent only to Claude. Codex's review subcommand does not accept custom prompts when used with --uncommitted or --base (CLI restriction), so codex uses its built-in review prompt. This means the two reviewers may emphasize different things — that's actually useful for cross-coverage.

Failure Handling

• If one reviewer fails (network, rate limit, missing key), the script still prints the other and exits 0.
• If both fail, the script exits 1.
• Common causes: OPENAI_API_KEY or ANTHROPIC_API_KEY not set, codex/claude not on PATH, ref doesn't exist.

Requirements

• codex CLI on PATH (OpenAI Codex CLI)
• claude CLI on PATH
• OPENAI_API_KEY env var
• ANTHROPIC_API_KEY env var
• Git repo with the requested diff scope

Notes

• Claude is invoked with --bare -p to skip hooks, plugins, and memory loading — the subprocess won't recursively trigger your parent session's hooks.
• Codex uses its built-in codex review --uncommitted / codex review --base <ref> modes, which compute the diff themselves; we don't pipe it.
• The script does not modify any files. It only reads the diff and prints to stdout.