Thinking models: raw <think> tag content displayed instead of being collapsed

[erwinh22]

Description

When using thinking models (e.g. moonshotai/kimi-k2-thinking via NVIDIA NIM), the reasoning blocks wrapped in think tags are rendered as raw plaintext in the response output instead of being collapsed or hidden.

Steps to Reproduce

1. Configure a thinking model (e.g. moonshotai/kimi-k2-thinking via NVIDIA NIM provider)
2. Send any prompt (e.g. 'what is today s date?')
3. Observe the response includes the raw think tag content as plaintext

Expected Behavior

The content between think tags should be:

• Collapsed/hidden by default, OR
• Rendered in a toggleable reasoning section (similar to how other AI coding clients handle thinking model output)

Actual Behavior

The raw think tags and all reasoning content are displayed as plaintext in the response, making the output noisy and hard to read. The internal reasoning of the model (e.g. 'The user is asking a simple non-code question about today s date...') is shown directly in the chat.

Environment

• OpenCode version: latest
• Provider: NVIDIA NIM
• Model: moonshotai/kimi-k2-thinking
• OS: Windows 11

Context

Many thinking/reasoning models use think blocks (DeepSeek R1, Kimi K2-Thinking, QwQ, etc.). Most AI coding clients (Cursor, VS Code Copilot, etc.) handle these tags by collapsing or hiding the reasoning content. Adding this support would improve the experience for all thinking models.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Support parsing <think> and <thinking> tags as reasoning blocks #11439: Same problem — <think> and <thinking> tags rendered as plain text instead of as reasoning blocks. That issue also identifies the root cause (removal of extractReasoningMiddleware in v1.1.46+) and proposes a fix.
• MiniMax M2 <think> tags #4033: MiniMax M2 <think> tags — same raw tag display issue with a different thinking model.
• GLM 4.7 thinking process not properly formatted - missing leading tags #7779: GLM 4.7 thinking process not properly formatted — related rendering issue with thinking model output.

If your issue is the same as #11439, please add a thumbs-up reaction there rather than keeping this open as a separate report.

[erwinh22]

Closing as duplicate of #11439 — our fix PR is at #15382 which addresses the rendering-layer approach described in #11439.