miloco-server的vision_understand 工具调用失败

[WayneRoot]

1. 现象：

2. miloco部署compose
   `services:
   backend:
   container_name: miloco-backend
   image: ghcr.io/xiaomi/miloco-backend:latest

   NOTICE: MUST be host for miot device discovery.

   network_mode: host
   expose:

   • ${BACKEND_PORT:-9000}
     environment:
   • BACKEND_HOST=${BACKEND_HOST:-0.0.0.0}
   • BACKEND_PORT=${BACKEND_PORT:-9000}
   • AI_ENGINE_HOST=${AI_ENGINE_HOST:-192.168.1.200}
   • AI_ENGINE_PORT=${AI_ENGINE_PORT:-8080}
   • BACKEND_LOG_LEVEL=${BACKEND_LOG_LEVEL:-info}
   • TZ=${TZ:-Asia/Shanghai}
     volumes:
   • ./data:/app/miloco_server/.temp
   • ./log/backend:/app/miloco_server/.temp/log

   NOTICE: Mount configuration files, if you want to use your own configuration files, please mount them here.

   - ./config/server_config.yaml:/app/config/server_config.yaml

   - ./config/prompt_config.yaml:/app/config/prompt_config.yaml

   restart: unless-stopped
   healthcheck:
   disable: true`

3. 推理模型
   本地模型qwen3-4b-instruct，通过flask提供API服务。文本模型正常，视觉模型向miloco发起工具调用时，miloco返回失败。tool_call的调用参数如下，也可见上图。
   工具调用格式正确： {"name":"local_default__vision_understand","arguments":{"request_id":"analyze_living_room","query":"客厅里有没有人","location":"living room"}}

请确认一下是否是参数问题还是其它问题？