Summary

Fixes preserve_thinking in the built-in admin chat UI.

The raw /v1/chat/completions API preserves historical reasoning correctly when prior assistant messages include reasoning_content. However, the admin chat UI was storing streamed reasoning only in the UI-private _thinking field and then dropping it when serializing conversation history for the next request.

This made preserve_thinking work with raw API calls but fail inside the built-in chat.

Changes

• Include reasoning_content when building messagesForApi
• Fall back to _thinking so existing localStorage conversations continue to work
• Store streamed thinking as both:
  • reasoning_content, for API history replay
  • _thinking, for existing UI rendering
• Preserve reasoning on normal assistant messages, aborted streams, and assistant tool-call turns

Manual test

Tested locally with Qwen3.6-35B-A3B-oQ8-mtp.

Prompt 1:

Generate two random 20-digit numbers, validate that both are 20 digits, do not use tools, and only give me one of the two.

Prompt 2:

Now give me the second number only.

Before this fix, the built-in chat could not retrieve the hidden second number because prior reasoning was not sent back. After this fix, the built-in chat returns the second number from preserved reasoning history.