Draftpilot

Draftpilot writes code in your codebase based on your instructions.

This project is the standalone open-source "brains" behind draftpilot.com. You can use this freely on your projects with your OpenAI API key.

Goal & Philosophy

Draftpilot does not aim to write complete code without intervention in every case (though it would be nice to do so in simple cases). It aims to partner with the user to translate intent into code changes in a reliable way, and learn when it gets things wrong.

When Draftpilot is working well, users should be able to start most of the changes using natural language, continuing in the IDE for the difficult bits, like a junior engineer pair programming with a senior engineer. Draftpilot should also be able to use unix tools, git, and the web where appropriate.

How to use

You'll need to set the environment variable OPENAI_API_KEY to your OpenAI key. We recommend having GPT-4 API access, though you can use --gpt4 never to stick with 3.5.

Draftpilot is currently focused on Javascript & Typescript projects, though it will work for other types of codebases. You may want to create a custom extractor (see pyExtractor.ts for an example) for best results.

You can run draftpilot without installing in your codebase with npx:

npx -y draftpilot@latest

Or add the following alias to your .bashrc/.zshrc for convenience:

alias dpt="npx -y draftpilot@latest"

You can see all commands with --help.

After usage, configuration and temporary files will be generated in the .draftpilot folder. This folder can be inspected for partial output in case things go wrong, but should not be checked into git.

Tips for use / Limitations

Due to token limits, Draftpilot works on codebases with smaller files. If you have very large files, i.e. > 1000+ lines, Draftpilot will not be able to load as much context into the prompt. In the future we may support chunking file edits for large files.

Draftpilot is not the best tool for large refactors.

While you can give a vague request and hope it gets figured out, it's best to provide as much context as possible - which files to read & edit, and how you want the changes made.

If openAI embedding fails, try running with batch-size 1 to see what went wrong:

node cli.js index --verbose --batchSize 1

Development instructions

Draftpilot uses npm - run npm i to install dependencies, npm run watch to run the server and frontend in watch mode, and npm run test to run tests.

All prompts are in the src/prompts folder in embedded-typescript format. If you make changes to the .ets files, run npm run ets to regenerate the .ets.ts files, which provide type-safety

How it works

Context

The first step is onboarding the assistant with context about the codebase, which includes the main libraries used and the purpose of key folders and files. This is done with the AI in partnership with the user.

Planning

In the planning phase, the assistant tries to determine how to best fulfill the request. If needed, it can read individual files, run a command, search the codebase, or search the web for context. The plan is presented to the user for approval or modification.

Execution

In the execution phase, the plan is put into action - files are created, edited, and deleted. After execution, the user can inspect the results and ask for modifications or a redo.

Validation

In the validation phase, the assistant tries to get the code into a reasonable shape.

Commit

In the commit phase, the changes are summarized into a git commit. Commits by draftpilot are prefixed so that it's clear from git blame that these changes were written by an AI. I recommend separating AI commits from user changes so future AI learns from humans and not generated code.

Contributing

Code contributions, issue reports, and test cases are always welcome.

In cases where the agent could be more intelligent, it is massively helpful to provide enough context to reproduce and debug the issue. Draftpilot records its API requests to the /tmp folder, so you can send those as well, stripping out any sensitive information.

License

Draftpilot is available under the AGPL 3.0 license. You can use generated code freely commercially (subject to any copyright concerns that exist generally with AI-generated code), but if you distribute any derived works based on this project, you must make modifications available to the public.