<think> tags in plain string content are displayed as raw text instead of being stripped or collapsed #21515

Description

When using models that return thinking content as inline tags in a plain string content field ( via OpenAI-compatible API), opencode renders the raw tags and thinking content directly to the user:
{ "role": "assistant", "content": "<think>用户用中文说\"你好\"，这是一个简单的问候。我应该用中文友好地回复。\n</think>\n\n你好！有什么我可以帮助你的吗？" }
Expected: Only show 你好！有什么我可以帮助你的吗？ — the thinking content should be stripped or collapsed into a reasoning block.

Actual: The entire string including ... is displayed as plain text.

Context
This is related to #11439 and #18748. Many OpenAI-compatible providers use inline tags rather than structured reasoning_content fields. OpenCode should handle both formats.

Suggested Approach
At the rendering layer, parse , , tags and either:
Strip thinking content entirely, or
Display it as a collapsible reasoning block (consistent with native reasoning_content handling)
Preserve raw tags in storage so multi-turn LLM context is not broken (as noted in #11439)
Handle edge cases: unclosed tags during streaming, tags inside code blocks should be left alone

Plugins

No response

OpenCode version

1.4.0

Steps to reproduce

No response

Screenshot and/or share link

Operating System

Windows 10

Terminal

No response