[Enhancement]: More configurations for GitHub Copilot chat (language, model, etc.)

[jdcola]

Before Reporting

• I have checked FAQ, and there is no solution to my issue
• I have searched the existing issues, and there is no existing issue for my issue

What happened?

1. Select "Chinese" option for "Feature -> Chat -> Reply in language"
2. The chat window always replay in English

How to reproduce the bug.

It just happened!

Relevant log output

No response

macOS version

15.1.1

Xcode version

16.1

Copilot for Xcode version

0.35.2

[intitni]

I guess you are using GitHub Copilot Chat? Language, temperature and additional system prompt doesn't work for GitHub Copilot chat.

[intitni]

Looks like they have a parameter for that now, will give it a try next release.

userLanguage: T.Optional(T.String()),

[jdcola]

Looks like they have a parameter for that now, will give it a try next release.

Thanks! (I'm also using Zed's chat feature where GH Copilot automatically answers in the same language of the question.)

By the way, is there an option to choose the underlying model of GitHub Copilot? Because Claude is better than OpenAI models. (Also available in Zed's chat panel.)

[intitni]

I wish I know. There is no documentation so we need to dig into the obfuscated source code for it.

[intitni]

Ok since Zed is open source, we can copy their implementations. zed-industries/zed@6f06558

By glancing at the code, I think they are calling the endpoint directly. I am not sure if it is the right way to go since some other repositries were blocked because of that.

[intitni]

But zed is apparently more popular than this project. If zed survives, I think it would be safe to just call the API directly?

But I don't have time for it right now, if anyone wants to contribute, let me know.

[intitni]

After digging into the language server code, I can't find anything that allows users to specify the model for GitHub Copilot chat.

For anyone interested, the chat feature will call getBestChatModelConfig to get the model configuration. The first arugment is a string array of model family names, the function will use the name to determine the correct model configuration. But the name list is hardcoded as far as I can tell. Other than that, the getFirstMatchingModelMetadata seems to be returning nothing, so only the fallback values are used.

If you really want to use Claude in GitHub Copilot chat right now, you can open the language-server.js in VSCode, format the code and search for something like

case "gpt-4o":
  return {
    modelId: i.id,
    uiName: i.name,
    modelFamily: r,
    maxRequestTokens: await $St(this.ctx, i),
    maxResponseTokens: 4096,
    baseTokensPerMessage: 3,
    baseTokensPerName: 1,
    baseTokensPerCompletion: 3,
    tokenizer: "o200k_base",
    isExperimental: (c = i.isExperimental) != null ? c : !1,
  };

and change the model id to claude-3.5-sonnet. Use at your own risk though. The system prompt will not be updated accordingly so it still says it's using gpt-x.

Update:

I am not very sure but if it doesn't reach the fallback value in your case, you can try editing the model family list by searching for code that looks like:

function Yo(e) {
  switch (e) {
    case "user":
    case "inline":
      return ["gpt-4o", "gpt-4-turbo", "gpt-4"];
    case "meta":
    case "suggestions":
    case "synonyms":
      return ["gpt-4o-mini", "gpt-3.5-turbo"];
  }
}

And add claude-3.5-sonnet to the start of the first array. The function name may be different.