gen-pr

A CLI and GitHub Action that automatically generates pull requests using AI (specifically, Large Language Models or LLMs) from GitHub issues and pull requests.

This tool delivers the ultimate "Vibe Coding" experience, allowing humans to focus solely on writing issues while AI handles all implementation details. Our vision is to create a workflow where developers only need to describe what they want, and the AI translates those requirements into working code.

Features

• Multiple AI Coding Tools: Supports various coding tools, including:
  • Aider: An interactive AI pair programming tool
  • Codex CLI: OpenAI's coding agent
  • Claude Code: Anthropic's agentic coding tool
  • Gemini CLI: Google's AI coding assistant
• Planning Capability: Uses LLMs to analyze source code and develop implementation strategies before making any changes.
• Flexible Integration: Works as both a CLI tool and a GitHub Action

Our Use Cases

Currently, we use gen-pr to create scaffolding pull requests. We have five different gen-pr workflows to generate PRs with various combinations of LLMs and coding tools. We review the pull requests, select the best one, and then continue to write code on the PR manually. This approach has successfully reduced the amount of manually written code. Below is an analysis report generated by calc-ai-contrib for June 2025, during the development of Exercode. AI indicates the amount of code written fully automatically by gen-pr. We are very satisfied with the results and consider the AI contribution ratio a key metric for measuring our productivity.

╔══════════════════════════════════════════════════╗
║           CONTRIBUTION ANALYSIS REPORT           ║
╠══════════════════════════════════════════════════╣
║ Date: 2025-06-01 to 2025-06-30 (PRs: 91)         ║
║ Total Edits: 9,925 (+6,729 / -3,196)             ║
╠══════════════════════════════════════════════════╣
║ AI vs Human: [█████░░░░░░░░░░░░░░░░░] 24% / 76%  ║
║ Contributors: 1 AI, 6 Human                      ║
╚══════════════════════════════════════════════════╝

📊 DETAILED BREAKDOWN
────────────────────────────────────────
🤖 AI   : [████░░░░░░░░░░░░]  24% |    2,365 Edits (+1,685 / -680)
👥 Human: [████████████░░░░]  76% |    7,560 Edits (+5,044 / -2,516)

Requirements

• For development:
  • asdf
• For execution:
  • Node.js and npx (for @openai/codex, @anthropic-ai/claude-code, and @google/gemini-cli)
  • Python (for aider)
  • gh

Usage

GitHub Actions

See action.yml and .github/workflows/gen-pr-example.yml.

CLI

Here are some examples for creating PRs for issue #89.

Without Planning

Claude Code:

npx --yes gen-pr@latest --issue-number 89 --coding-tool claude-code

Codex:

npx --yes gen-pr@latest --issue-number 89 --coding-tool codex-cli

Gemini CLI:

npx --yes gen-pr@latest --issue-number 89 --coding-tool gemini-cli

With Planning (Strongly Recommended for Aider)

gen-pr can generate an implementation plan by reading files in the target repository using Repomix. This feature is particularly useful for non-agentic coding tools like Aider.

Gemini 2.5 Pro (gemini/gemini-2.5-pro) for planning and Aider for coding:

npx --yes gen-pr@latest --issue-number 89 --planning-model gemini/gemini-2.5-pro --reasoning-effort high --repomix-extra-args="--compress --remove-empty-lines --include 'src/**/*.ts'" --aider-extra-args="--model gemini/gemini-2.5-pro --edit-format diff-fenced --test-cmd='yarn check-for-ai' --auto-test"

Claude Opus 4 on Bedrock (bedrock/us.anthropic.claude-opus-4-1-20250805-v1:0) for planning and Aider for coding:

npx --yes gen-pr@latest --issue-number 89 --planning-model bedrock/us.anthropic.claude-opus-4-1-20250805-v1:0 --reasoning-effort high --repomix-extra-args="--compress --remove-empty-lines --include 'src/**/*.ts'" --aider-extra-args="--model bedrock/us.anthropic.claude-opus-4-1-20250805-v1:0 --test-cmd='yarn check-for-ai' --auto-test"

Gemini 2.5 Pro (gemini/gemini-2.5-pro) for planning and Claude Code for coding:

npx --yes gen-pr@latest --issue-number 89 --planning-model gemini/gemini-2.5-pro --reasoning-effort high --repomix-extra-args="--compress --remove-empty-lines --include 'src/**/*.ts'" --coding-tool claude-code

o4-mini (openai/o4-mini) for planning and Codex for coding:

npx --yes gen-pr@latest --issue-number 89 --planning-model openai/o4-mini --reasoning-effort high --repomix-extra-args="--compress --remove-empty-lines --include 'src/**/*.ts'" --coding-tool codex-cli

DeepSeek R1 on OpenRouter (openrouter/deepseek/deepseek-r1-0528:free) for planning and Gemini CLI for coding:

npx --yes gen-pr@latest --issue-number 89 --planning-model openrouter/deepseek/deepseek-r1-0528:free --reasoning-effort high --repomix-extra-args="--compress --remove-empty-lines --include 'src/**/*.ts'" --coding-tool gemini-cli

Grok 4 (xai/grok-4) for planning and Aider for coding:

npx --yes gen-pr@latest --issue-number 89 --planning-model xai/grok-4 --reasoning-effort high --repomix-extra-args="--compress --remove-empty-lines --include 'src/**/*.ts'" --aider-extra-args="--model gemini/gemini-2.5-pro --edit-format diff-fenced --test-cmd='yarn check-for-ai' --auto-test"

Local Gemma 3n via Ollama (ollama/gemma3n) for planning and Aider for coding:

npx --yes gen-pr@latest --issue-number 89 --planning-model ollama/gemma3n --repomix-extra-args="--compress --remove-empty-lines --include 'src/**/*.ts'" --aider-extra-args="--model ollama/gemma3n --edit-format diff-fenced --test-cmd='yarn check-for-ai' --auto-test"

For PR (#103)

Codex:

npx --yes gen-pr@latest --issue-number 103 --coding-tool codex-cli

Config File

You can create a YAML configuration file named gen-pr.config.yml or gen-pr.config.yaml in the root of your repository to set default values for options. This config file works for both CLI usage and GitHub Actions. Command-line flags (CLI) or workflow inputs (GitHub Actions) will override values in this file. For example:

repomix-extra-args: "--compress --remove-empty-lines --include 'src/**/*.ts'"
aider-extra-args: '--model gemini/gemini-2.5-pro --edit-format diff-fenced --no-gitignore'
coding-tool: claude-code
test-command: 'yarn check-for-ai'

Supported Model Format

The tool requires model names defined on llmlite in the format provider/model-name:

• OpenAI: openai/gpt-4.1, openai/o4-mini and more
• Azure OpenAI: azure/gpt-4.1, azure/o4-mini and more
• Google Gemini: gemini/gemini-2.5-pro, gemini/gemini-2.5-flash and more
• Anthropic: anthropic/claude-4-sonnet-latest, anthropic/claude-3-5-haiku-latest and more
• AWS Bedrock: bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0, bedrock/us.anthropic.claude-3-5-haiku-20241022-v1:0 and more
• Google Vertex AI: vertex/gemini-2.5-pro, vertex/gemini-2.5-flash and more
• xAI: xai/grok-4, xai/grok-3, xai/grok-3-mini and more
• OpenRouter: openrouter/deepseek/deepseek-r1-0528:free, openrouter/deepseek/deepseek-chat-v3-0324:free and more
• Ollama: ollama/gemma3n, ollama/deepseek-r1, ollama/qwen3 and more

Environment Variables

Each provider uses standard environment variables for authentication:

• Coding Tools
  • Codex CLI: OPENAI_API_KEY
  • Claude Code: ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN
  • Gemini CLI: GEMINI_API_KEY (or GOOGLE_GENERATIVE_AI_API_KEY)
• Planning Models
  • OpenAI: OPENAI_API_KEY
  • Anthropic: ANTHROPIC_API_KEY
  • Google Gemini: GEMINI_API_KEY (or GOOGLE_GENERATIVE_AI_API_KEY)
  • Azure OpenAI: AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_API_VERSION
  • AWS Bedrock: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION (or AWS_REGION_NAME)
  • Google Vertex AI: GOOGLE_APPLICATION_CREDENTIALS or default service account
  • xAI: XAI_API_KEY
  • OpenRouter: OPENROUTER_API_KEY
  • Ollama: OLLAMA_BASE_URL (default: http://localhost:11434), OLLAMA_API_KEY (optional)

License

Apache License 2.0