currently roo code does not have good tool-calling support of openai's new open mode gpt-oss

[LuyiTian]

App Version

3.25.8

API Provider

OpenRouter

Model Used

openai/gpt-oss-120b

Roo Code Task Links (Optional)

I think the openai GPT-oss model have different tool-calling format, which make it fail frequently when used in Roo code, gpt-oss also support tool-calling inside thinking process, which I think is not supported in roo code and caused error too. we should improve roo code to support the gpt-oss model

🔁 Steps to Reproduce

tool calling and mcp server calling often fail

💥 Outcome Summary

tool calling and mcp server calling often fail

📄 Relevant Logs or Errors (Optional)

[roomote]

Hi! I'm here to help out the maintainers and am going to investigate this issue with GPT-OSS model tool-calling support. I'll work on improving the compatibility to ensure tool-calling and MCP server calling work properly with the new OpenAI GPT-OSS models. Thanks for reporting this!

[LuyiTian]

[roomote]

I've successfully implemented support for GPT-OSS tool-calling! 🎉

Summary of Changes

I've created PR #6816 that addresses the tool-calling issues with OpenAI's GPT-OSS models (gpt-oss-20b and gpt-oss-120b) when used through OpenRouter.

What was fixed:

• ✅ Added proper handling for tool calls in OpenRouter streaming responses
• ✅ Support for tool calls that appear within reasoning/thinking blocks (specific to GPT-OSS)
• ✅ Added new ApiStreamToolCallChunk type to properly handle tool call data in streams
• ✅ Comprehensive test coverage for various GPT-OSS tool-calling scenarios

Technical approach:

• The solution detects GPT-OSS models and uses an XmlMatcher to parse tool calls embedded in reasoning blocks
• It handles both standard OpenAI-style tool calls and the GPT-OSS specific format
• Maintains backward compatibility with existing models

The PR is currently going through CI checks. Once merged, GPT-OSS models should work properly with tool-calling and MCP server interactions in Roo Code.

PR Link: #6816

[jayteaftw]

Hey I am also having this issue however I am running GPT-OSS-120B with VLLM