[BUG] Support for Qwen3-Coder-Next model

[oleg-kn]

Problem (one or two sentences)

Roo said
I am Roo, a knowledgeable technical assistant focused on answering questions and providing information about software development, technology, and related topics. I can analyze code, explain concepts, and access external resources to help you with technical questions.

<function=attempt_completion>
<parameter=result>I am Roo, a knowledgeable technical assistant focused on answering questions and providing information about software development, technology, and related topics. I can analyze code, explain concepts, and access external resources to help you with technical questions.


Model Response Incomplete
The model failed to use any tools in its response. This typically happens when the model provides only text/reasoning without calling the required tools to complete the task.Details

Context (who is affected and when)

everyone i guess

Reproduction steps

Linux, llama.cpp latest build.

Expected result

answer

Actual result

The model provided text/reasoning but did not call any of the required tools. This usually indicates the model misunderstood the task or is having difficulty determining which tool to use. The model has been automatically prompted to retry with proper tool usage.

Variations tried (optional)

Same model/setup works fine in another Roo fork (Kilocode)

App Version

latest

API Provider (optional)

OpenAI Compatible

Model Used (optional)

unsloth/Qwen3-Coder-Next-GGUF:Q8_K_XL

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

Date/time: 2026-02-05T16:48:13.142Z
Extension version: 3.46.2
Provider: openai (proxy)
Model: unsloth/Qwen3-Coder-Next-GGUF:Q8_K_XL

The model provided text/reasoning but did not call any of the required tools. This usually indicates the model misunderstood the task or is having difficulty determining which tool to use. The model has been automatically prompted to retry with proper tool usage.

[roomote]

I'd like to help with this issue, but I need some clarification to ensure I implement the right solution.

Looking at the error output, it appears the Qwen3-Coder-Next model is outputting XML-style function calls (e.g., <function=attempt_completion>) rather than native OpenAI-style JSON tool calls. Roo Code expects the latter format when using the OpenAI Compatible provider.

Could you please provide more details on the following:

1. llama.cpp configuration: What command-line arguments or settings are you using when running llama.cpp? Specifically, are you using any tool-call format flags like --jinja or setting a specific chat template?

2. Kilocode differences: You mentioned the same model works in Kilocode. Do you know what specific configuration or settings Kilocode uses that might differ from Roo Code? Does Kilocode perhaps parse XML-style tool calls, or does it configure the model differently?

3. Model template: Are you using the model's built-in chat template, or a custom one? Qwen3 models typically support tool calling, but the format may vary depending on the template used.

4. Expected behavior: Would you prefer Roo Code to:

   • Add support for parsing XML-style tool calls (like the <function=...> format), or
   • Provide configuration options to help the model output native JSON tool calls

Please tag @roomote in your response with the answers, and I'll be happy to propose an implementation plan once I have this information.

[oleg-kn]

@roomote

llama-server -hf unsloth/Qwen3-Coder-Next-GGUF:Q8_K_XL
--timeout 2400
--host 0.0.0.0
--mmap
--jinja
--threads -1
--ctx-size 262144
--temp 1.0
--top-p 0.95
--min-p 0.01
--top-k 40
--seed 3407
--fit on

3. probably built-in chat template.
4. both!

[roomote]

Thanks for the additional details! I can see you're using --jinja with the built-in chat template.

Based on your feedback, I'll work on implementing support for parsing XML-style tool calls (like <function=attempt_completion>) that Qwen3-Coder-Next outputs. This approach will allow the model to work out of the box without requiring users to configure specific templates.

I'll explore the codebase and open a PR with my attempt at addressing this. Feedback and guidance are welcome!

[roomote]

Opened PR #11220 with my attempt at addressing this issue.

The PR adds an XmlToolCallParser that detects and converts XML-style tool calls (like <function=attempt_completion><parameter=result>...</parameter></function>) to the standard tool call events that Roo Code expects.

This should allow Qwen3-Coder-Next running via llama.cpp with --jinja to work properly. The parser handles both streaming and complete scenarios, and passes through non-tool-call text unchanged.

Feedback welcome!