Our software is open source so you can solve your own problems without needing help from others. And if you solve a problem and are so kind, you can upstream it for the rest of the world to use. Check out our [post about externalization](https://blog.comma.ai/a-2020-theme-externalization/).
Development is coordinated through [Discord](https://discord.comma.ai) and GitHub.
* **New features**: We believe openpilot is mostly feature-complete, and the rest is a matter of refinement and fixing bugs. As a result of this, most feature PRs will be immediately closed, however the beauty of open source is that forks can and do offer features that upstream openpilot doesn't.
* **Negative expected value**: This a class of PRs that makes an improvement, but the risk or validation costs more than the improvement. The risk can be mitigated by first getting a failing test merged.
[Projects / openpilot bounties](https://github.com/orgs/commaai/projects/26/views/1?pane=info) is the best place to get started and goes in-depth on what's expected when working on a bounty.
There's lot of bounties that don't require a comma 3/3X or a car.
In order for your fork's data to be eligible for the training set:
* **Your cereal messaging structs must be [compatible](../cereal#custom-forks)**
* **The definitions of all the stock messaging structs must not change**: Do not change how any of the fields are set, including everything from `selfdriveState.enabled` to `carState.steeringAngleDeg`. Instead, create your own structs and set them however you'd like.
* **Do not include cars that are not supported in upstream platforms**: Instead, create new opendbc platforms for cars that you'd like to support outside of upstream, even if it's just a trim-level difference.