Automating PR Reviews with Argus: A Llama 3.3 Powered GitHub Action

I built an AI code reviewer as a GitHub Action — here’s what I learned

Aditya Bhusal developed Argus, a lightweight GitHub Action for automated code reviews. The system utilizes Groq’s Llama 3.3 70B model to analyze diffs and post inline comments on pull requests.

Why This Matters

Human pull request reviews are often slow and inconsistent, leading to rubber-stamp approvals under deadline pressure that allow bugs to enter production. While LLMs can effectively identify logical flaws and code smells, the technical challenge lies in the ‘plumbing’—specifically mapping AI output to strict API requirements and managing developer fatigue through configurable signal filters.

Key Insights

• Llama 3.3 70B performance: The model successfully identified hardcoded API secrets and missing ‘await’ keywords in asynchronous functions during testing.
• Structured Output Challenge: Prompt engineering was required to ensure the LLM returned valid JSON for precise line number correlation, as GitHub’s API fails when comments are posted on unmodified lines.
• Configurability vs. Hardcoding: Implementing a .argus/config.yml system allows teams to set severity thresholds (e.g., only showing ‘high’ severity issues) to manage the noise-to-signal ratio.

Working Examples

GitHub Action workflow configuration for integrating Argus AI code reviews.

name: Argus Code Review
on:
  pull_request:
    types: [opened, synchronize, reopened]
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: Rozer402/argus@main
        with:
          groq_api_key: ${{ secrets.GROQ_API_KEY }}

Practical Applications

References:

• https://github.com/Rozer402/argus