Fix Qwen(3.5) thinking format for Qwen3.5 models

[Airead]

What happened?

Summary

Fix the Qwen thinking format handling in the OpenAI completions provider to correctly toggle thinking mode for Qwen3.5 models.

Problem

When using mlx-community/Qwen3.5-9B-4bit with the pi coding-agent, toggling the thinking/reasoning mode does not work correctly. The previous implementation shared the same code path for Z.ai and Qwen, only sending a top-level enable_thinking parameter:

if ((compat.thinkingFormat === "zai" || compat.thinkingFormat === "qwen") && model.reasoning) {
    (params as any).enable_thinking = !!options?.reasoningEffort;
}

Qwen3.5 models require enable_thinking to also be set inside chat_template_kwargs for the thinking toggle to take effect.

Fix

Separate Z.ai and Qwen into distinct code paths. Qwen now sends enable_thinking both at the top level and inside chat_template_kwargs:

if (compat.thinkingFormat === "zai" && model.reasoning) {
    // Z.ai uses top-level enable_thinking: boolean
    (params as any).enable_thinking = !!options?.reasoningEffort;
} else if (compat.thinkingFormat === "qwen" && model.reasoning) {
    // Qwen uses both top-level and chat_template_kwargs for compatibility
    const enableThinking = !!options?.reasoningEffort;
    (params as any).enable_thinking = enableThinking;
    (params as any).chat_template_kwargs = { enable_thinking: enableThinking };
}

Steps to reproduce

1. Configure a custom model using mlx-community/Qwen3.5-9B-4bit with thinkingFormat: "qwen"
2. Start the pi coding-agent with this model
3. Enable thinking/reasoning mode
4. Send a message — thinking works as expected
5. Toggle thinking mode off
6. Send another message — the model still outputs thinking content, the toggle has no effect

"llama.local": {
    "baseUrl": "http://127.0.0.1:8000/v1",
    "apiKey": "sk-xxxxxxxxx",
    "api": "openai-completions",
    "models": [
    {
        "id": "Qwen3.5-9B-4bit",
        "name": "Qwen3.5-9B-4bit",
        "reasoning": true,
        "input": ["text", "image"],
        "compat": {
        "supportsDeveloperRole": false,
        "thinkingFormat": "qwen"
        },
        "contextWindow": 65536,
        "maxTokens": 32768,
        "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }
    }
    ]
}

Expected behavior

No response

Version

0.57.1

[badlogic]

Implemented in 03ad7ec.

For local Qwen-compatible servers, set:

"compat": {
  "thinkingFormat": "qwen-chat-template"
}

Keep qwen for DashScope-style endpoints that expect top-level enable_thinking.