🤖 Auto-generate PR Description Action: Supercharge Your Pull Requests!

This GitHub Action uses OpenAI, Claude or any OpenRouter model to automatically craft detailed, insightful pull request descriptions. Say goodbye to vague PR summaries and hello to clear, concise, and context-rich descriptions that enhance your team's collaboration and code review process.

🚀 Features

• Automatically generates detailed PR descriptions
• Three providers — OpenAI (default), Claude and OpenRouter — each in its own module
• Provider and model selectable by input or environment variable
• Supports GitHub Actions workflow
• Context-aware: the model sees the diff plus the PR's current description and its full comment thread (issue comments, review summaries and inline code comments), so ticket links, decisions and reviewer concerns survive regeneration

What the model receives

Section	Content	Role
<diff>	the PR diff (base...head) from the GitHub API	source of truth for what changed
<current_description>	the PR body, labelled as human-written or previously auto-generated	why — intent, ticket links, notes to preserve
<comments>	issue comments, review bodies and inline review comments (with file:line) from trusted authors	discussion context, resolved concerns, the archived original description

Budgets keep the request bounded: 100k chars of diff, 5k of description, 20k of comments (2k per comment). Anything cut is marked as truncated. Comment reads are best-effort — if the token lacks the scope, generation still proceeds on the diff alone.

🔒 Safety

• No shell sees PR data. The diff comes from the GitHub API; branch names, which the PR author controls, never reach a command line. Only diffs too large for the API fall back to local git, addressed by commit SHA.
• Only trusted comments reach the model. On a public repository anyone can comment, and the comment thread is part of the prompt. Comments are used only from the repository's owners, members and collaborators, the PR author, and the action's own archived copy of the original description.
• Issue-closing keywords are guarded. A generated Closes #N / Fixes #N / Resolves #N is turned into Refs #N unless the human-written description or a trusted comment already contained that reference — so the model cannot be talked into closing unrelated issues on merge.
• Secrets are not sent to the provider. The hunks of .env*, *.pem, *.key, *.p12/*.pfx, SSH keys, master.key, credentials.*/secrets.* config files and similar are replaced by a placeholder before the diff leaves the runner.
• Bounded runtime. Provider requests time out after 3 minutes and are retried on rate limits, 5xx and network errors with exponential backoff.
• Fork PRs. Use the pull_request trigger. Do not switch to pull_request_target to get secrets for fork PRs: the action would then run with a write token for untrusted code changes.

📝 ToDo

• Handles rate limiting and retries API calls
• Configurable prompt templates for description generation
• Supports multiple languages for generated descriptions

📋 Requirements

• GitHub repository
• An API key for the provider you use: OpenAI, Anthropic (Claude) or OpenRouter

🛠️ Installation

1. Create a .github/workflows/auto-pr-description.yml file in your repository.
2. Add the following content to the file:

name: Auto-generate PR Description
on:
  pull_request:
    types: [opened, reopened, synchronize]

jobs:
  generate-description:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      pull-requests: write
      issues: write
    steps:
      # Optional: only needed for diffs too large for the GitHub API.
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
          persist-credentials: false
      - name: Auto-generate PR Description
        uses: yuri-val/auto-pr-description-action@v1
        with:
          openai_api_key: ${{ secrets.OPENAI_API_KEY }}
          github_token: ${{ secrets.GITHUB_TOKEN }}

3. Add your provider's API key to your repository secrets (OPENAI_API_KEY, ANTHROPIC_API_KEY or OPENROUTER_API_KEY).

🔀 Providers

provider	Default model	Key (input, or environment variable)
openai (default)	gpt-6-luna	openai_api_key / OPENAI_API_KEY
claude	claude-haiku-5-5	anthropic_api_key / ANTHROPIC_API_KEY (or CLAUDE_API_KEY)
open-router	deepseek/deepseek-v4.1-flash	openrouter_api_key / OPENROUTER_API_KEY

Inputs win over environment variables, so the provider can be set once for the whole job:

    env:
      AI_PROVIDER: claude
      ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
    steps:
      - uses: yuri-val/auto-pr-description-action@v1

Each provider is its own module in providers/ (openai.js, claude.js, open-router.js).

Which model writes best

Measured 2026-10-10 on three real PRs and one 5-commit release (all eight candidates followed the output rules; times and costs are per description):

Provider / model	Character	Avg time	Cost
openai / gpt-6-luna	shortest, accurate, fewest tokens	3.5 s	$0.0007
openai / gpt-5.6-luna	accurate, somewhat generic	4.7 s	$0.0016
claude / claude-haiku-5-5	most specific; the only model to catch every non-obvious change (e.g. a changed default with an upgrade note)	4.7 s	$0.0014
open-router / deepseek/deepseek-v4.1-flash	detailed and accurate	4.4 s	$0.0028
open-router / z-ai/glm-5.3-flash	detailed, occasional overstatement	13.9 s	$0.0012
open-router / xiaomi/mimo-v2.6-flash	detailed, slower	15.8 s	$0.0011
open-router / google/gemini-3.8-flash	concise, odd grouping, most expensive	4.6 s	$0.0067
open-router / qwen/qwen3.8-flash	long, very slow	57.9 s	$0.0025

⚙️ Configuration

You can customize the action by providing the following inputs:

Input	Description	Required	Default
provider	openai, claude or open-router (env AI_PROVIDER)	No	openai
model	Model for the provider (env AI_MODEL)	No	per provider, see above
openai_api_key	OpenAI API key (env OPENAI_API_KEY)	For openai	N/A
anthropic_api_key	Anthropic API key (env ANTHROPIC_API_KEY / CLAUDE_API_KEY)	For claude	N/A
anthropic_workspace_id	Only for Anthropic keys not scoped to a workspace (env ANTHROPIC_WORKSPACE_ID)	No	N/A
openrouter_api_key	OpenRouter API key (env OPENROUTER_API_KEY)	For open-router	N/A
openai_model	Deprecated alias of model for openai	No	N/A
github_token	GitHub token with repo permissions	Yes	${{ github.token }}
temperature	Sampling temperature (0.0 to 1.0) for non-reasoning OpenAI models only	No	0.7

📤 Outputs

The action provides the following outputs:

• pr_number: The number of the pull request updated
• description: The generated pull request description

🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to check the issues page.

📝 License

This project is MIT licensed.

👤 Author

Yuri V

• GitHub: @yuri-val

🙏 Acknowledgements

• OpenAI, Anthropic and OpenRouter for the models
• GitHub Actions for the seamless integration

---

If you find this action helpful, please consider giving it a ⭐️!