Integrate DeepSeek V3.2 Thinking Mode

[tanzhenxin]

Summary
Qwen Code already has built-in support for Thinking Mode, but we currently don’t have a model wired up that uses this capability. DeepSeek V3.2 provides a clear Thinking Mode API, making it a good first model to integrate so we can validate and exercise our thinking-mode pipeline.

Why This Matters

• Thinking Mode is increasingly common across modern coding agents (Claude Sonnet 4.5, Gemini 3.0 Pro, etc.).
• Integrating DeepSeek V3.2 allows us to test real-world reasoning performance using an actual model that exposes the feature.
• This helps improve tool-calling accuracy, multi-step reasoning, and overall agent reliability.

Scope of Work

• Implement DeepSeek V3.2 Thinking Mode based on their spec:
  https://api-docs.deepseek.com/guides/thinking_mode
• Add model configuration and routing so DeepSeek V3.2 can run in Thinking Mode within Qwen Code.
• Verify compatibility with:
  • Tool calling
  • Reasoning/trace streaming
  • Conversation/session storage
  • Plan mode
  • Telemetry
• Add configuration options or presets for using DeepSeek V3.2 in Thinking Mode.
• Update documentation with usage examples.

Expected Outcome
DeepSeek V3.2 can run end-to-end in Qwen Code using Thinking Mode, allowing us to properly exercise and validate the existing thinking-mode pipeline.