OpenCode stumbles on MiniMax think tags

This is a small but very real interoperability bug in the agent stack: model vendors are pushing reasoning formats that help performance, while client tools still assume plain assistant text.

• –MiniMax M2.7 explicitly expects its full thinking trace to be round-tripped, so stripping tags for UI cleanliness can hurt model quality.
• –OpenCode already has UI controls for showing thinking blocks, but the harder problem is preserving reasoning in message history without confusing the parser.
• –SGLang’s `minimax-append-think` parser appears to address transport formatting, not the higher-level chat-memory contract OpenCode needs.
• –The likely fix is provider-aware message handling, not just a frontend parse tweak: keep reasoning intact in stored history, but expose sanitized content to the agent loop.
• –This is exactly the kind of edge case that shows up when “open-weight + self-hosted + agentic coding” starts replacing closed APIs.