Qwen 3.5 and Gemma 4 can't call tools

Before submitting your bug report

• I've tried using the "Ask AI" feature on the Continue docs site to see if the docs have an answer
• I'm not able to find a related conversation on GitHub discussions that reports the same bug
• I'm not able to find an open issue that reports the same bug
• I've seen the troubleshooting guide on the Continue Docs

Relevant environment info

- OS: macOS
- Continue version: 1.2.22
- IDE version: 1.115.0
- Model: Qwen 3.5 9B, Gemma 4 26B-A4B
- config:
  
name: Local Config
version: 1.0.0
schema: v1
models:
  - name: Autodetect (1)
    provider: lmstudio
    model: AUTODETECT
    apiBase: http://SERVER_IP:1234/v1

Description

When using the model I loaded on my server from LM Studio on my macbook and VS Code, they are trying to run a tool like that:

Now I need to find the user.js file. Let me search for it.

<tool_call> <function=file_glob_search> <parameter=pattern> **/user.js </tool_call>

But the tool_call is not being registered by Continue and ends up in the thinking part instead which acts up like a message ending.

same for:

read_file failed with the message: File "src/test.vue" does not exist or is not accessible. You might want to check the path and try again.

Please try something else or request further instructions.

that end up in the thinking part with:

<tool_call> <function=edit_existing_file> <parameter=filepath> src/components/Navbar.vue <parameter=changes>
... changes

To reproduce

Load any of these models with reasoning and try using them to modify files with it.

Log output

I can only share the following:

Type: Chat
Result: Success
Prompt Tokens: 5095
Generated Tokens: 0
ThinkingTokens: 4528
Total Time: 61.03s
To First Token: 0.53s
Tokens/s: 74.8

Reproduce the issue in VS Code using the provided Continue configuration, LM Studio endpoint, and Qwen 3.5 or Gemma 4 model with reasoning enabled. Trace how the shown <tool_call> output is handled and verify that tool calls are registered and executed instead of being placed in the thinking content.