[Feature Request] Add Thinking Mode Toggle and Filter `<think>`

[NGC13009]

Validations

• I believe this is a way to improve. I'll try to join the Continue Discord for questions
• I'm not able to find an open issue that requests the same enhancement

Problem

Qwen3, like Claude, supports enabling or disabling the thinking mode at any time, and it is a model that can be deployed locally via Ollama. However, continue still has not provided a convenient option to enable or disable thinking at any time. Users still need to manually input /nothink to control the model, which is very inconvenient. It is hoped that continue will provide a convenient switch to control the model's thinking mode when calling Ollama for local execution.

In addition, I want to configure qwen3 as an autocomplete model, but the current continue's autocomplete does not filter the <think> tag. Even when the thinking mode is turned off, ollama will return an empty <think> tag, and continue will include them in the response, filling them into the code, making autocomplete unavailable.

This issue also appears in the problem of automatically generating titles.

Solution

In summary, three issues are reflected here:

1. A quick toggle switch for switching models provided by ollama is needed.
2. The switchable model is currently not properly configured as an autocomplete model.
3. The current model cannot generate the title and abstract correctly.

[tomasz-stefaniak]

@NGC13009 does the reasoning setting at least partially fix the issues you're having?

https://docs.continue.dev/json-reference#completionoptions

[NGC13009]

@tomasz-stefaniak I have tried these settings, but they cannot resolve the problems I encountered.

I would like to add a thinking toggle button, but for now I have to default to disabling thinking by adding /nothink to the system prompt, and enable thinking by using /think. This method works, but it's not convenient.

Moreover, it completely fails to address the issue of outputting the </think> tag when using autocomplete. This can only be resolved by hoping that the continue plugin will filter out this suggestion (i.e., the initial pair of <think> tags that the model produces at the beginning), while not filtering out subsequent tags, as those might be part of the normal output.

The current version (v1.0.7) still cannot correctly retrieve the title from the history using thinking-type models.