<think> tag is displayed as plain text instead of hidden reasoning in OpenAI Compatible provider

[jwadow]

App Version

3.26.4

API Provider

OpenAI Compatible

Model Used

gemini-2.5-pro

Roo Code Task Links (Optional)

No response

🔁 Steps to Reproduce

1. Run Roo Code in OpenAI-compatible mode.
2. Use a model such as gemini-2.5-flash or gemini-2.5-pro (though the issue happens regardless of model).
3. Generate output that contains <thinking> ... </thinking>.
4. Observe that the <think> block is shown as normal text (without <think> tag).

💥 Outcome Summary

It seems the handling of <think> is missing in OpenAI-compatible mode.
When using Roo Code in OpenAI-compatible mode, the <thinking> tag from model output is rendered as plain text in the response instead of being collapsed into a gray "Thinking..." spoiler (as expected).
The <think> tag should be rendered as a hidden reasoning block (collapsed "Thinking…" spoiler), not exposed directly to the user.

📄 Relevant Logs or Errors (Optional)

[roomote]

I'm exploring a potential fix for the <think> tag rendering issue in OpenAI-compatible mode and will open a PR to share progress.

[roomote]

Opened PR #7617 with a fix for this issue: #7617

The fix adds XmlMatcher to the BaseOpenAiCompatibleProvider to properly parse <think> tags and convert them to reasoning chunks, which will render as collapsed "Thinking..." spoilers in the UI as expected.

All CI checks have passed successfully. The fix will benefit all OpenAI-compatible providers. Feedback welcome!

[jwadow]

Opened PR #7617 with a fix for this issue: #7617

The fix adds XmlMatcher to the BaseOpenAiCompatibleProvider to properly parse <think> tags and convert them to reasoning chunks, which will render as collapsed "Thinking..." spoilers in the UI as expected.

All CI checks have passed successfully. The fix will benefit all OpenAI-compatible providers. Feedback welcome!

It would be reasonable to check <thinking> as well as <think> and all that stuff. The LLM creators do all sorts of things.
I don't know how the Roo Code works from the inside, maybe even give the user a choice to type the custom text into the input field in settings - which exactly is the "thinking" tag, so that OpenAI Compatible provider becomes more universal.