VSCode LM API in RooCode Exhausts GitHub Copilot Chat Quota Faster Than Native Github Copilot with Chat Agent Auto Fix Usage

[mifhud]

App Version

v3.23.6

API Provider

VS Code Language Model API

Model Used

claude 3.7 sonnet, claude sonnet 4, etc instead of gpt-4.1

Roo Code Task Links (Optional)

Hi roocode team,

I've encountered an issue where using the VSCode LM API in RooCode seems to exhaust the GitHub Copilot chat request quota significantly faster than using GitHub Copilot directly in VSCode.

Observations
Generating longer chat responses (e.g., explanations, refactor suggestions) in RooCode leads to rapid quota exhaustion.

The same tasks and prompts in native Github Copilot with Chat Agent Auto Fix are more quota-efficient.

This suggests that RooCode may be using more tokens or chat invocations per request.

🔁 Steps to Reproduce

1. Connect GitHub Copilot to RooCode using the VSCode LM API.
2. Perform one of the following coding scenarios:

• Case A: Writing a Unit Test

  • Create a new Go/JavaScript/Python file.

  • Write a complex task case file or you can use existing project that have complex task case.

  • Ask Copilot (via RooCode) to generate a unit test for that function.

  • Repeat the same prompt in native GitHub Copilot (VSCode chat sidebar).

• Case B: Starting an App from Scratch

  • Open a new file and type:
    "Create a basic Express.js server" or
    "Generate a REST API in Go with routing and middleware".

  • Observe the response length and token usage from RooCode.

  • Try the same request using GitHub Copilot directly.

3. Monitor how quickly the quota is consumed over several similar requests.
4. Repeat the same process in GitHub Copilot
5. Compare quota consumption.

💥 Outcome Summary

Expected chat quota usage similar to GitHub Copilot in VSCode, but got much faster quota exhaustion when using RooCode via the VSCode LM API.

📄 Relevant Logs or Errors (Optional)

Here’s an example of a long chat response in GitHub Copilot that consumes less quota. The quota is only deducted again when a ‘continue’ or follow-up iteration is triggered

[daniel-lxs]

Thanks for the detailed report and reproduction steps.

There are a few reasons you might be seeing higher token usage when using GitHub Copilot via Roo Code:

1. System prompts and tool context:
   Roo Code includes a more comprehensive system prompt than GitHub Copilot, along with extra tool definitions, instructions, and task metadata. This leads to longer requests and higher token consumption per interaction, especially for larger tasks or when more context is available.

2. Cross-provider compatibility:
   We design our prompt format to be provider-agnostic so that it works consistently across models like Claude, Gemini, OpenAI, and others. We can't easily optimize just for GitHub Copilot's LM API without compromising compatibility or introducing special cases.

3. Potential misuse of the API:
   It's possible there’s some subtle mismatch in how we’re using the VS Code LM API compared to GitHub's first-party integrations. If that’s the case, we’d need more concrete evidence (e.g. logs, examples, or API behavior differences) to investigate. For now, this seems more like a feature request than a bug, but we’re open to revisiting this if you can pinpoint specific differences in how the APIs behave or respond.

Let us know if you can share more diagnostic details if you believe this to be a bug.