Is it possible to support GitHub Copilot Chat?

[pidgeon777]

As the titles says.

References:

https://docs.github.com/en/copilot/quickstart

https://github.com/CopilotC-Nvim/CopilotChat.nvim

[pidgeon777]

It is important to note that GitHub Copilot Chat supports various models, including GPT-4, GPT-4o, o1, and o1-mini. For more detailed information and updates, you can refer to the GitHub Copilot Chat changelog.

Given the widespread use and active subscriptions to GitHub Copilot Chat among developers, it would be highly beneficial for the lsp-ai plugin to incorporate support for these models. This integration would enhance the development experience by leveraging the advanced capabilities of GitHub Copilot Chat.

[arunoruto]

As much as I remember, Copilot is based on OpenAIs ChatGPT. I remember helix-gpt having a GitHub Copilot integration. I looked a bit more at the code and found out, that they are using the https://api.githubcopilot.com/chat/completions endpoint. I guess you can just use the open_ai examples and replace the endpoint with the one from githubcopilot. Here is some more information about the API:

After doing more analysis, I determined that the API endpoint at https://api.githubcopilot.com/chat/completions was nothing more than an OpenAI API proxy, with a different auth wrapper (which authenticated using the Copilot bearer token instead of an OpenAI key). I could query any model, with nearly identical behavior to the real OpenAI API.

I tried it out with helix, but I am getting the following error:

2024-11-26T21:55:58.707 helix_lsp::transport [ERROR] lsp-ai err <- "ERROR lsp_ai::transformer_worker: generating response: error decoding response body: expected value at line 1 column 1\n"
2024-11-26T21:55:58.707 helix_lsp::transport [ERROR] lsp-ai err <- "\n"
2024-11-26T21:55:58.707 helix_lsp::transport [ERROR] lsp-ai err <- "Caused by:\n"
2024-11-26T21:55:58.707 helix_lsp::transport [ERROR] lsp-ai err <- "    expected value at line 1 column 1\n"
2024-11-26T21:55:58.707 helix_lsp::transport [ERROR] lsp-ai <- InternalError: error decoding response body: expected value at line 1 column 1

EDIT My current lsp-ai config for helix can be found in flake repo. You can kinda read like json, just replace the = with : and remove the ; at the end;

EDIT 2 I looked a bit more into the whole Open AI API compatibility. While api.githubcopilot.com is being used now, it used to be a different endpoint until the beginning of 2024. According to GitHubs changelog, they used to have copilot-proxy.githubusercontent.com as their API endpoint. This is also referenced in this line of helix-gpt.
What does this mean? Chat should work with the https://api.githubcopilot.com/chat/completions endpoint, but completion will not. Doing some manual curl commands, the /chat/completions endpoint expects a messages entry in the data, while for completions a prompt is used. I tried manually submitting a request to https://copilot-proxy.githubusercontent.com/v1/engines/copilot-codex/completions as described by helix-gpt, but my token isn't the right one for such requests... Looking at the code, the copilot API token is used to obtain a session token, which is then used further to make requests.
tldr;

• Copilot chat should work using https://api.githubcopilot.com/chat/completions as an endpoint and the output of gh auth token as the token value. I used set -gx GH_AUTH_TOKEN $(gh auth token) to set the env variable in fish.
• Completions need a different endpoint which requires the prompt property to be set.

[arunoruto]

I think I got chat working using https://api.githubcopilot.com/chat/completions! The key for it to work is to set the model to be an empty string "model": "". I am getting an output, even tho it is not that fast.

As for the code completion, this is gonna be tricky...
Github deviates from the OpenAI API here, and requires a new token every x minutes/hours (not sure, but it changes with time). helix-gpt sends a request to https://api.github.com/copilot_internal/v2/token with the copilot token (different than the one obtained with gh auth token) and in that response, a new token is obtained which can be used to make new requests.

tldr; chat can be used now (please test it and give feedback) and completions cannot be configured with the current state of the code (I guess?)

EDIT: it is possible to use the chat API to make completions as shown in the wiki: https://github.com/SilasMarvin/lsp-ai/wiki/Configuration#instruction
If someone has some nice prompts for this, enlighten us!