feat(provider): add preserveReasoningInContent config option to fix Qwen preserve_thinking interoperability

[jgrcic]

Issue for this PR

Closes # (No existing issue)

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

This adds a preserveReasoningInContent config flag to solve an interoperability issue with the preserve_thinking feature on Qwen models (specifically tested with Qwen3.6-35B-A3B-FP8 running on vLLM).

Currently, OpenCode automatically extracts historical reasoning into a separate reasoning_content property. However, for Qwen to successfully access its previous reasoning turns in vLLM, the historical reasoning must be embedded directly within the assistant's content string, wrapped in explicit <thinking>...</thinking> tags. When OpenCode strips these out, the model loses its chain of thought in subsequent prompts.

To fix this:

1. I added preserveReasoningInContent to the Model schema in src/config/provider.ts.
2. Updated src/provider/transform.ts so that when this flag is enabled, it correctly transforms the reasoning history into <thinking> tags within the content body and unsets reasoning_content.

How did you verify your code works?

I tested this locally against my own vLLM instance running Qwen3.6-35B-A3B-FP8.
I prompted the model to mentally generate two random numbers, but only output one. In the next turn, I asked it for the second number.
Without this flag, the model fails to recall the second number because the reasoning is stripped. With the flag enabled in my opencode.json, I verified via mitmproxy that the payload correctly contains the <thinking> tags, and the model successfully recalled the second number.

Screenshots / recordings

No UI changes.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍