LLM Code Review — Multi‑Model

Tiny CLI wrapper around git diff + the llm tool to get fast, actionable PR reviews from multiple models at once.

---

🚀 One‑liner to run

review --models "gpt-4o,gemini-2.0-flash,anthropic/claude-3-7-sonnet-latest" origin/master...HEAD

Swap origin/master for origin/main if your default branch is main.

---

1) Prereqs (once)

• git and bash (any Linux/macOS shell is fine)
• llm CLI (Python‑based)
• Optional: bat for nice Markdown output

Quick install (recommended in a venv):

python3 -m venv ~/.llm-env
source ~/.llm-env/bin/activate
pip install --upgrade pip llm
# (Optional) On Debian/Ubuntu for pretty output
sudo apt install -y bat && sudo ln -s /usr/bin/batcat /usr/local/bin/bat || true

Place the review script somewhere on your PATH, e.g. ~/bin/review, and make it executable:

chmod +x ~/bin/review

---

2) Add API keys (once)

The script can call OpenAI, Gemini, and Claude via llm. Set keys like this:

llm keys set openai       # for OpenAI (e.g., gpt-4o)
llm keys set google       # for Gemini (e.g., gemini-2.0-flash)
llm keys set anthropic    # for Claude (e.g., anthropic/claude-3-7-sonnet-latest)

Quick test:

llm -m gpt-4o "hello"
llm -m gemini-2.0-flash "hello"
llm -m anthropic/claude-3-7-sonnet-latest "hello"

---

3) Usage

Compare current branch to default branch:

review --models "gpt-4o,gemini-2.0-flash,anthropic/claude-3-7-sonnet-latest" origin/main...HEAD

Only staged changes:

review --cached

Add extra context (PR description, ticket, etc.):

git log -1 --pretty=%B | review --context -

Set models once per shell:

export LLM_MODELS="gpt-4o,gemini-2.0-flash,anthropic/claude-3-7-sonnet-latest"
review origin/main...HEAD

---

4) What you get

• A short readiness rating (★☆☆ / ★★☆ / ★★★)
• Blocking items to fix before human review
• Non‑blocking suggestions
• Test plan gaps

---

Troubleshooting

• No output? Probably no diff. Try review --cached or confirm your range (e.g., origin/main...HEAD).
• Auth error? Re‑run the key setup commands above.
• Command not found? Ensure llm and review are on your PATH, or activate your venv.

---

🔧 Codex Setup (Automated)

If you’re running this inside Codex, use the included codex-startup.sh script in this repo (from the project root).

This script will:

1. Create a Python venv and install llm (plus Gemini/Claude plugins) and the review CLI.
2. Persist your API keys for both the llm tool and your shell (adds exports to ~/.bashrc).
3. Generate helper scripts under tools/:
   • tools/pre_pr_review.sh – runs the multi-model review
   • tools/auto_fix_from_review.sh – asks an LLM for a minimal patch based on the review

Required Codex Secrets (exact names)

Set these three secrets in Codex exactly as named:

Secret Name	Value (example)	Used For
OPENAI_API_KEY	sk-...	OpenAI models (e.g., gpt-4o)
ANTHROPIC_API_KEY	sk-ant-...	Claude models (e.g., claude-3-5-sonnet-*)
GOOGLE_API_KEY	AIza...	Gemini models (e.g., gemini-2.0-flash)

The startup script will run llm keys set for each provider and export these in the shell (with fallbacks in ~/.bashrc).

Typical Codex Pre-PR Flow

# Run an initial review
bash tools/pre_pr_review.sh

# (Optional) Try an auto-fix pass based on .ai-review.md
bash tools/auto_fix_from_review.sh .ai-review.md || true

# Run a second review after fixes
bash tools/pre_pr_review.sh

# Open a PR (uses repo’s default branch as base)
gh pr create --fill \
  --body-file .ai-review.md \
  --title "chore: codex multi-model review" \
  --base "$(git remote show origin | sed -n '/HEAD branch/s/.*: //p')"

🗣 Using Codex with This Setup

With the Codex integration and codex-startup.sh in place, you can work entirely in plain English.

Example

You could type to Codex:

"Refactor src/app/example.component.ts to use async/await instead of promises and run a multi-model review before opening a PR."

What Happens Behind the Scenes

When you make a request like this, Codex will:

1. Interpret your instruction and make the requested code change(s) in your repo.
2. Run tools/pre_pr_review.sh – this triggers the multi-model review (review CLI) with OpenAI, Gemini, and Claude using the API keys you set in Codex secrets.
3. Save the AI review output to .ai-review.md.
4. Optionally run tools/auto_fix_from_review.sh – this will feed .ai-review.md to an LLM to propose a minimal patch and apply it.
5. Run a second review so you can see if the fixes resolved the earlier feedback.
6. Open a PR using gh pr create with the .ai-review.md contents as the PR body.

Because codex-startup.sh persists your keys and sets everything up automatically, this works in every Codex session without re-configuring.

You can mix and match tasks — ask Codex to:

• Make a specific change and run the review
• Just run a review on your current branch
• Apply auto-fixes from the last review
• Open a PR after review passes

All using plain language instructions.