Fix DeepSeek reasoning_content failures in tool-call loops.

Some OpenAI-compatible DeepSeek, MiMo and similar thinking-model routes work on the first request, then fail after a tool call because the adapter drops reasoning_content from conversation history.

1. Why this happens

OpenAI-compatible does not always mean OpenAI-identical. DeepSeek-style thinking models may stream reasoning in chunks before the final tool-call delta. If your framework only stores the final assistant message, the tool call is preserved but the reasoning text is lost.

2. Preserve reasoning on assistant tool-call messages

When the assistant emits tool calls, store any accumulated thinking content on the same assistant history item. For DeepSeek-compatible routes, replay it as reasoning_content on the next request.

{
  "role": "assistant",
  "content": "",
  "tool_calls": [
    {
      "id": "call_123",
      "type": "function",
      "function": {
        "name": "get_time",
        "arguments": "{\"city\":\"London\"}"
      }
    }
  ],
  "reasoning_content": "The user asked for the current time, so I need to call the time tool."
}

3. Test the second tool turn

A single tool call can hide the bug. Use a two-turn regression because the failure usually appears when the provider receives the next request.

1. User asks for a tool call.
2. Assistant streams reasoning chunks, then emits tool_calls.
3. Tool returns.
4. User asks a follow-up that forces another tool call.
5. Assert that the previous assistant message still includes reasoning_content.

4. Keep provider-specific fields behind an adapter

Do not send reasoning_content to every OpenAI-compatible provider. Keep it behind provider capability checks so normal OpenAI routes stay clean, while DeepSeek/MiMo-compatible routes preserve the state they require.

• Normalize incoming streaming chunks into an internal assistant-turn object.
• Attach accumulated thinking before writing conversation history.
• Replay provider-specific fields only for routes that require them.
• Log the final outbound host, model id and message shape when debugging.

5. When a hosted relay helps

If your app is failing against several Chinese model routes, isolate whether the bug is in the upstream provider, your framework adapter, or your local proxy. A small external OpenAI-compatible endpoint can help compare request and response behavior without changing the whole app.