Support enable_thinking config for reasoning models (GLM-5, etc.)

[liushuangls]

Problem

When using reasoning models like GLM-5 (ZhiPu/智谱) as the VLM backend, the model spends all tokens on reasoning_content and returns empty content. This is because GLM-5 requires enable_thinking: false to disable chain-of-thought reasoning for structured tasks like memory extraction.

Current Behavior

The OpenAIVLM backend has _supports_enable_thinking() which only recognizes DashScope (Alibaba Cloud) endpoints:

_DASHSCOPE_HOSTS = {
    "dashscope.aliyuncs.com",
    "dashscope-intl.aliyuncs.com",
}

The LiteLLM backend does not support enable_thinking at all.

For ZhiPu (open.bigmodel.cn) and other OpenAI-compatible providers with reasoning models, there is no way to pass enable_thinking: false through configuration.

Proposed Solution

Add a config option in ov.conf to control thinking behavior, e.g.:

{
  "vlm": {
    "backend": "openai",
    "model": "glm-5",
    "api_base": "https://open.bigmodel.cn/api/coding/paas/v4",
    "enable_thinking": false
  }
}

This would be more general than hardcoding host allowlists, and would work for any provider that supports the enable_thinking parameter (ZhiPu, DashScope, DeepSeek, etc.).

Alternative: support an extra_body config field to pass arbitrary parameters to the API call.

Workaround

Currently we patch openai_vlm.py directly to add open.bigmodel.cn to the host allowlist, but this gets overwritten on every upgrade.

Environment

• OpenViking: 0.2.12
• Model: GLM-5 (ZhiPu)
• Backend: openai (switched from litellm due to same limitation)
• Python: 3.14