Feature Description

Add GLM (ZhipuAI) as a first-class built-in provider in the auxiliary client's PROVIDER_REGISTRY, alongside OpenRouter, Nous, xAI, etc.

Current Behavior

Users must configure GLM through the custom endpoint path:

auxiliary:
  vision:
    provider: glm
    model: glm-4v-flash
    base_url: "https://open.bigmodel.cn/api/paas/v4/"
    api_key: "<key>"

This triggers the "custom" provider branch in _resolve_task_provider_model (line 3985-3987), which:

1. Strips the provider identity, breaking all provider-specific logic downstream
2. Requires manual base_url and api_key configuration
3. Bypasses the ZAI/GLM-specific vision handling in resolve_vision_provider_client

Proposed Behavior

auxiliary:
  vision:
    provider: glm    # or "zhipu" or "zhipuai"
    model: glm-4v-flash
    # base_url and api_key auto-resolved from ZHIPU_API_KEY env var

This would require:

1. Add "glm" / "zhipu" / "zhipuai" to PROVIDER_REGISTRY with:

   • api_key_env_vars: ["ZHIPU_API_KEY", "ZHIPUAI_API_KEY", "GLM_API_KEY"]
   • inference_base_url: "https://open.bigmodel.cn/api/paas/v4"
   • default_auxiliary_model: "glm-4v-flash" (for vision)

2. Extend ZAI vision logic to also cover GLM provider:

   # In resolve_vision_provider_client, line 3566:
   if requested in ("zai", "glm") and not resolved_base_url:

3. Extend _build_call_kwargs ZAI vision max_tokens skip to also match GLM:

   _skip_max_tokens = (
       provider in ("zai", "glm")
       and ("4v" in _model_lower or "5v" in _model_lower or "-v" in _model_lower)
   )

Motivation

• ZhipuAI is a major Chinese LLM provider with multimodal models (GLM-4V, GLM-5V)
• The open.bigmodel.cn endpoint is OpenAI-compatible but has GLM-specific quirks (rejects max_tokens on vision calls, different error codes)
• Users currently hit error 1210 (see vision_analyze tool fails with GLM (glm-4v-flash) provider: error 1210 request format mismatch #26827, [Bug] auxiliary task provider identity lost when base_url + api_key are both set #26879) because the custom provider path doesn't handle these quirks

Related

• vision_analyze tool fails with GLM (glm-4v-flash) provider: error 1210 request format mismatch #26827 (vision_analyze fails with GLM)
• [Bug] auxiliary task provider identity lost when base_url + api_key are both set #26879 (provider identity lost in auxiliary task config)