PRE - PR Engine

PRE (PR Engine) is a GitHub Action that automatically reviews pull requests using Large Language Models (LLMs) through the Bellman library.

Features

• Automated code review for pull requests when they are opened
• On-demand reviews triggered by commenting /bellman on a pull request
• Detailed feedback on potential bugs, security issues, and code quality
• Inline comments for specific issues
• Summary and scoring of pull requests
• Customizable system prompts

Usage

See examples in .github/workflows directory

Automatic PR Review

Add this to your GitHub workflow file to automatically review PRs when they are opened:

name: Bellman PR Evaluator

on:
  pull_request:
    types: [opened]

jobs:
  pre-llm-review:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      pull-requests: write

    steps:
      - name: Checkout code
        uses: actions/checkout@v4
        with:
          fetch-depth: 0

      - name: Review PR with LLM
        uses: modfin/pre@master
        with:
          github-token: ${{ secrets.GITHUB_TOKEN }}
          bellman-key: ${{ secrets.BELLMAN_KEY }}
          bellman-url: ${{ secrets.BELLMAN_URL }}
          bellman-model: 'VertexAI/gemini-2.0-flash'

On-Demand PR Review

Add this to your GitHub workflow file to enable on-demand reviews triggered by commenting /bellman on a pull request:

name: Bellman PR Bot
on:
  issue_comment:
    types:
      - created

jobs:
  pr-bot-on-demand:
    # Only run on pull request comments that start with '/bellman'
    if: github.event.issue.pull_request && startsWith(github.event.comment.body, '/bellman')
    runs-on: ubuntu-latest
    permissions:
      contents: read
      pull-requests: write
    steps:
      - name: Checkout code
        uses: actions/checkout@v4
        with:
          # Checkout the PR branch, not the base branch
          ref: ${{ github.event.pull_request.head.sha }}
          fetch-depth: 0

      - name: Extract params
        id: params
        env:
          COMMENT_BODY: ${{ github.event.comment.body }}
        run: |
          if [ -z "$COMMENT_BODY" ]; then
            exit 0 
          fi
          MODEL_OVERRIDE=$(printf '%s' "$COMMENT_BODY" | perl -ne '
            if (/bellman\s+use\s+([a-zA-Z0-9._:\/-]+)/) {
              print $1;
              exit;
            }
          ')
          echo "model_override=$MODEL_OVERRIDE" >> $GITHUB_OUTPUT

      - name: Review PR with LLM
        uses: modfin/pre@master
        with:
          github-token: ${{ secrets.GITHUB_TOKEN }}
          bellman-key: ${{ secrets.BELLMAN_KEY }}
          bellman-url: ${{ secrets.BELLMAN_URL }}
          bellman-model: 'VertexAI/gemini-2.0-flash'
          bellman-model-override: ${{ steps.params.outputs.model_override }}
          system-prompt-addition: ${{ github.event.comment.body }}

Configuration

Input	Description	Required	Default
github-token	GitHub token for API access	Yes	${{ github.token }}
bellman-key	API key for LLM service	Yes	-
bellman-url	URL to Bellman service	Yes	-
bellman-model	LLM model to use for review	No	VertexAI/gemini-2.0-flash
bellman-model-override	Override LLM model to use for review	No	-
system-prompt	Sets a specific system prompt	No	see DefaultSystemPrompt
system-prompt-addition	Additional instructions for the LLM	No	-

Requirements

• A Bellman API key and URL
• GitHub token with permissions to read contents and write to pull requests

Further Development Suggestions

• Add Agent mode allowing the LLM to eg read the entire files, and not just the diffs, if they want to.
• Refactor, cleanup and make testable
• Add built binaries instead of building from source each action
  eg https://full-stack.blend.com/how-we-write-github-actions-in-go.html

License

This project is licensed under the MIT License - see the LICENSE file for details.

Developed by

Modular Finance