Configuration

Prefer the WebUI panel (recommended), or configure in profile's cordis.patch.yml under the plugin id:

- id: analyze-image-tool
  config:
    baseURL: https://api.siliconflow.cn/v1     # Any OpenAI-compatible endpoint
    apiKey: ***                                  # Leave empty to use below chain
    model: Qwen/Qwen3-VL-32B-Instruct           # Vision/multimodal model id on the endpoint
    maxTokens: 2048
    timeoutMs: 60000                            # Increase for large images / slow endpoints, e.g. 120000
    maxImageBytes: 10485760
    defaultQuestion: Describe this image thoroughly. Include any visible text verbatim, the overall layout, and notable details.
    composerNoteTemplate: 用户在这条消息里粘贴了一张图片（附件 ID: {attachment_id}）。要查看图片内容，请调用 analyze_image 并传入该附件 ID（attachment_id 参数）。

composerNoteTemplate supports {attachment_id} and {image_index} placeholders.

Endpoint Examples (same config, pick one)

Use case	baseURL	Model example	Notes
SiliconFlow (default)	`https://api.siliconflow.cn/v1`	`Qwen/Qwen3-VL-32B-Instruct`	Requires key, strong vision
Alibaba DashScope	`https://dashscope.aliyuncs.com/compatible-mode/v1`	`qwen3-vl-flash`	Compatible mode, cost-effective
Zhipu	`https://open.bigmodel.cn/api/paas/v4`	`glm-4.6v-flash`	Free tier available
OpenRouter	`https://openrouter.ai/api/v1`	`google/gemini-2.5-flash`	Multi-model aggregator
Local Ollama	`http://localhost:11434/v1`	`qwen3-vl:4b`	No key needed, fully offline

API Key Resolution Chain (in order)

apiKey saved in WebUI panel（ctx.settings user layer）
Plugin config apiKey（cordis.patch.yml）
Environment variable VISION_API_KEY, then SILICONFLOW_API_KEY（can be in ~/.dsh/.env or exported）
dsh credential channel（VISION_API_KEY, then SILICONFLOW_API_KEY）
Local endpoint (localhost) requires no key

Leaving apiKey empty in the panel = no override, continues down the chain.

Tool Usage

Model calls analyze_image for image-related tasks:

path（mutually exclusive with attachment_id）: absolute local path, http(s) URL, or data: URL（支持 PNG/JPEG/WebP/GIF/BMP/TIFF/HEIC，默认上限 10MB）。
attachment_id（mutually exclusive with path）: attachment ID of image pasted into composer（形如 sha256:…）, see "Paste image recognition" below.
prompt（optional）: specific question, e.g. "extract all visible text verbatim", "count buttons", "describe layout". Default is detailed description including all visible text.

analyze-image-tool

AI 분석

설치

문서

Configuration

Endpoint Examples (same config, pick one)

API Key Resolution Chain (in order)

Tool Usage