A collection of ChatGPT assistants, e.g., code viewer, labeler, assigner, etc.

• Overview
  • Why reinvent the wheel?
  • Features
• Usage
  • Configuration
    • Environment variables
    • Inputs
  • Prompt templates
    • Variables available in prompt templates
• Developing
• FAQs
  • Review pull request from forks
  • Choose the ChatGPT API implementation
  • Inspect the messages between ChatGPT server
• License

There're already many chatgpt-actions, why we need to reinvent the wheel?

• Code review is not pull request changeset summarization

  Some of existing actions send the whole pull request diff to ChatGPT, and ask ChatGPT to summarize the pull request, or summarize each hunks. In practice, the summarize has already done by the proposer and the code reviewer is responsible for reviewing the changes, evaluate the quality, and decide whether to merge the pull request.

  This action has a "score" action that asks ChatGPT to score the pull request, and give a short comment as a explanation. That should be a "good hint" when starting a code review.

• Leave too much comments on Github pull requests

  Some of existing actions send each hunks in a pull request diff to ChatGPT, and yields a comment (or comment snippet) for each hunk. That makes them in practically unusable, as receiving 20+ email notification after submitting a pull request is not a good experience. For the reviewer's perspective, we usually just focus on the hunks that are "too good" or "too bad".

  This actions take such experience into consideration, and only leave comments for the hunks that are not "LGTM". (We do have an option to leave comments for good hunks as well).

• Lack of customization of the prompt

  All of existing actions hard-coded the prompt template inside the source code and there's no chance to customize the prompt. When the built-in prompt is not good enough, ChatGPT responses with useless comments.

  At the same time, someone may have good prompt ideas, he/she want to try using ChatGPT as an AI assistant for code quality evaluation and review, but he/she needs to reinvent the wheel for a new action to play on Github.

  This action makes the prompt template configurable. With a set of variables like $title, $description, $filename, $patch, you can customize the message sent to ChatGPT for better experience.

  Prompt engineering is an art to make ChatGPT success, we encourage the community to try discovering new, novel, and insightful prompts to make ChatGPT do code review job better, without reinventing the wheel for another chatgpt-action.

  ANY SUGGESTIONS AND IMPROVEMENTS ARE HIGHLY APPRECIATED!!!

• Scoring your pull requests

      - uses: unsafecoerce/chatgpt-action@main
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          CHATGPT_ACCESS_TOKEN: ${{ secrets.CHATGPT_ACCESS_TOKEN }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          debug: false
          action: score

• Code review your pull requests

      - uses: unsafecoerce/chatgpt-action@main
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          CHATGPT_ACCESS_TOKEN: ${{ secrets.CHATGPT_ACCESS_TOKEN }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          debug: false
          action: review
          review_comment_lgtm: true

• Customizable prompt templates

      - uses: unsafecoerce/chatgpt-action@main
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          CHATGPT_ACCESS_TOKEN: ${{ secrets.CHATGPT_ACCESS_TOKEN }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          debug: false
          action: score
          scoring_beginning: |
            Hi ChatGPT, I have a pull request with title "$title" and the description is as follows,
  
            > $description
  
            I would like to give you the whole diff of the pull request and you need to given a score after
            reviewing the pull request. The score should be a number between 0 and 100 and 85 means the
            the pull request is basically acceptable. You just need to reply a number between 0 and 100,
            e.g., "85", and a comment within 30 words. Reply "OK" to confirm.

name: Code Review

permissions:
  contents: read
  pull-requests: write

on:
  pull_request:

jobs:
  test:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3
        with:
          repository: ${{github.event.pull_request.head.repo.full_name}}
          ref: ${{github.event.pull_request.head.ref}}
          submodules: false

      - uses: unsafecoerce/chatgpt-action@main
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          CHATGPT_ACCESS_TOKEN: ${{ secrets.CHATGPT_ACCESS_TOKEN }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          debug: false
          action: score

      - uses: unsafecoerce/chatgpt-action@main
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          CHATGPT_ACCESS_TOKEN: ${{ secrets.CHATGPT_ACCESS_TOKEN }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          debug: false
          action: review
          review_comment_lgtm: true

See also: ./action.yml

• GITHUB_TOKEN

• CHATGPT_ACCESS_TOKEN: ChatGPT access token, see also: https://github.com/acheong08/ChatGPT.

  The access token can be easily obtained from https://chat.openai.com/api/auth/session after logging into ChatGPT.

• OPENAI_API_KEY: use this to authenticate with OpenAI API, official ChatGPT's behavior using text-davinci-003, see also: https://github.com/transitive-bullshit/chatgpt-api

Note that CHATGPT_ACCESS_TOKEN and OPENAI_API_KEY are not both required. Inside this action, unofficial ChatGPT is preferred if CHATGPT_ACCESS_TOKEN exists. Note that the CHATGPT_ACCESS_TOKEN can expire frequently, so OPENAI_API_KEY should be more convenient if its cost is affordable to you.

• action: The action to run, currently can be review, score
• debug: Enable debug mode, will show messages and responses between ChatGPT server in CI logs.
• chatgpt_reverse_proxy: The URL of the ChatGPT reverse proxy
• review_comment_lgtm: Leave comments even the patch is LGTM
• path_filters: Rules to filter files to be reviewed.

See also: ./action.yml

• review_beginning: The beginning prompt of a code review dialog
• review_patch: The prompt for each chunks/patches
• scoring_beginning: The beginning prompt for scoring a pull request
• scoring: The prompt for the whole pull request

• pull request score (action: score):

  • $title: Title of the pull requests.
  • $description: The description of the pull request.
  • $diff: The whole diff of the pull request.

• code review (action: review):

  • $title: Title of the pull requests.
  • $description: The description of the pull request.
  • $filename: Filename of the file being viewed.
  • $patch: The diff contents of the patch being viewed.

Any suggestions or pull requests for improving the prompts are highly appreciated.

First, you'll need to have a reasonably modern version of node handy, tested with node 16.

Install the dependencies

$ npm install

Build the typescript and package it for distribution

$ npm run build && npm run package

Github Actions limits the access of secrets from forked repositories. To enable this feature, you need to use the pull_request_target event instead of pull_request in your workflow file. Note that with pull_request_target, you need extra configuration to ensure checking out the right commit:

name: Code Review

permissions:
  contents: read
  pull-requests: write

on:
  pull_request_target:

jobs:
  test:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3
        with:
          repository: ${{github.event.pull_request.head.repo.full_name}}
          ref: ${{github.event.pull_request.head.ref}}
          submodules: false

      - uses: unsafecoerce/chatgpt-action@main
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          debug: false
          action: score

      - uses: unsafecoerce/chatgpt-action@main
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        with:
          debug: false
          action: review
          review_comment_lgtm: true

See also: https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#pull_request_target

The javascript's chatgpt package provides two implementations of the ChatGPT API:

• ChatGPTAPI: official ChatGPT using the OpenAI's text-davinci-003.
  • not free
  • requires OPENAI_API_KEY
• ChatGPTUnofficialProxyAPI: unofficial ChatGPT models, rely on third-party server and is rate limited.
  • free
  • requires CHATGPT_ACCESS_TOKEN
  • the proxy server is configurable using chatgpt_reverse_proxy

If both environment variables OPENAI_API_KEY and CHATGPT_ACCESS_TOKEN exists, we prefer the ChatGPTUnofficialProxyAPI implementation.

Set debug: true in the workflow file to enable debug mode, which will show the messages

The MIT License (MIT), Copyright (c) 2023 Tao He.