[Bug] Kimi K2.6 reasoning content leaks into main response instead of being rendered as thinking block #32800

Description:

When using Kimi K2.6 through the direct Moonshot AI provider, the model's reasoning content is not separated into a thinking/rationale block. Instead, it is streamed as part of the main assistant response, making it look like the model is "talking to itself" in the answer.

This does not happen with Kimi K2.7 under the same setup — reasoning is correctly handled there.

Environment:

• OpenCode version: 1.17.7
• Provider: Moonshot AI (direct, via /connect)
• Model: Kimi K2.6
• OS: Windows 11 + WSL
• Terminal: WezTerm

Steps to reproduce:

1. Run opencode.
2. Connect Moonshot AI provider via /connect.
3. Select Kimi K2.6 model via /models.
4. Send any prompt.
5. Observe that reasoning text appears inline in the assistant response rather than in a separate thinking block.

Expected behavior:

Reasoning should be rendered as a collapsible Thought: block (when /thinking is enabled) or hidden from the response entirely (when /thinking is disabled).

Actual behavior:

Reasoning text is mixed into the normal response text, as if it were part of the answer. The /thinking command has no effect on this behavior — the reasoning content is never recognized as a separate block.

Additional context:

• /thinking command state does not affect this behavior.
• The issue appears specific to Kimi K2.6; Kimi K2.7 works correctly.
• Possibly related to open issues around reasoning_content handling for Kimi K2.6: Kimi K2.6: thinking is enabled but reasoning_content is missing in assistant tool call message #29690, [Kimi K2.6] reasoning_content is missing in assistant tool call messages #29619, Bug: 400 Error (Extra inputs not permitted) when sending 'reasoning' field in chat history for kimi-k2.6 #27852, opencode 上下文大了以后 不显示thinking #32545.