Cannot remove think tag for an agent model

[OlegAndreych]

Hello

How to get rid from <think> tag in an agent model output?
I've tried to use qwen3-235b-a22b:free as an agent model using custom OpenAI provider.
Unfortunately, I see <think> tag in model output like <think>Okay, let's tackle this problem. The user is getting.... It hinders tool usage and leads to exceeding the token limit.
I've tried to create a .aider.model.settings.yml in home directory and at project root, but no luck, <think> tag is still there.
I've noticed that if I use qwen3-235b-a22b: without :free suffix there's no <think> tag in model output. Maybe it somehow matches internal aider desk settings or something.

.aider.model.settings.yml content is:

- name: openai-compatible/qwen3-235b-a22b:free
  edit_format: diff
  weak_model_name: openai/qwen-2.5-coder-32b-instruct:free
  use_repo_map: true
  extra_params:
    max_tokens: 128000
    top_p: 0.95
  use_temperature: 0.6
  editor_model_name: openai/qwen-2.5-coder-32b-instruct:free
  editor_edit_format: editor-diff
  reasoning_tag: think
  remove_reasoning: think

- name: openai/qwen3-235b-a22b:free
  edit_format: diff
  weak_model_name: openai/qwen-2.5-coder-32b-instruct:free
  use_repo_map: true
  extra_params:
    max_tokens: 128000
    top_p: 0.95
  use_temperature: 0.6
  editor_model_name: openai/qwen-2.5-coder-32b-instruct:free
  editor_edit_format: editor-diff
  reasoning_tag: think
  remove_reasoning: think

[OlegAndreych]

I’ve tried “Ask” mode in aider desk with qwen3-235b-a22b:free as the main model. It detects thinking just fine.
So, the aforementioned issue should be of the agent model and aider desk.

[wladimiiir]

Thanks for the feedback.

I am aware that currently, when the model uses <think> tags for thinking, it is not properly parsed in Agent mode as distinct Thinking and Answer sections. This issue will be addressed in the next update.

However, I appreciate your idea of stripping the thinking for the next iteration to free up context. This approach will require testing to determine if it hinders model performance by preventing the model from seeing its previous thoughts. It will definitely need to be a configurable option, allowing users to choose whether thinking should be included in further messages to the LLM.