dsh-vision 👁️

给 DeepSeek Harness 里没有原生识图能力的模型(如 deepseek-v4-flash)加上识图工具。
Give image-recognition ability to DeepSeek Harness models without native vision (e.g. deepseek-v4-flash).
把本地图片或网络图片 URL 交给视觉大模型(默认阿里云百炼 qwen3.7-flash),返回中文文字描述。
Delegates local image paths / URLs to a vision LLM (default: Alibaba Cloud Bailian qwen3.7-flash) and returns a Chinese description.
English | 中文

中文
特性
- 🖼️ 支持本地图片路径、网络图片 URL
- 🔑 OpenAI 兼容格式,不绑定特定厂商(默认阿里云百炼)
- ⚙️ 配置优先级:插件 config > 环境变量 > 默认值
- 📦 零额外运行时依赖(只用 Node 内置
fetch)
安装
pnpm install && pnpm run build
# 在插件父目录执行:
dsh plugin --profile web add ./dsh-vision
# 重启 dsh,然后新建会话即可使用 vision 工具
配置
方式一:环境变量
export DASHSCOPE_API_KEY=sk-xxx
export VISION_MODEL=qwen3.7-flash-2026-07-15
export DASHSCOPE_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
方式二:profile 补丁层 ~/.dsh/profiles/<name>/cordis.patch.yml
- id: dsh-vision
config:
apiKey: sk-xxx
model: qwen3.7-flash-2026-07-15
baseURL: https://dashscope.aliyuncs.com/compatible-mode/v1
使用
模型会自动调用 vision 工具,参数:
| 参数 |
必填 |
说明 |
image |
✅ |
本地图片绝对路径(如 C:\a.png)或 http(s) URL |
prompt |
❌ |
识别要求,默认"请详细描述这张图片的内容" |
提示:配合图片上传
DeepSeek Harness 默认的 DeepSeek adapter 声明模型纯文本,上传图片会被 MODEL_DOES_NOT_SUPPORT_IMAGES 拦截。
要让用户能直接粘贴图片(图片块渲染为 [图片附件: sha256:...] 标记),需要对
dsh-llm-deepseek 打一个小补丁(见 PATCHES.md)。
English
Features
- 🖼️ Local image paths and remote http(s) URLs
- 🔑 OpenAI-compatible API — vendor-agnostic (Bailian by default)
- ⚙️ Config precedence: plugin config > environment > defaults
- 📦 Zero extra runtime deps (built-in
fetch only)
Install
pnpm install && pnpm run build
# from the PARENT directory:
dsh plugin --profile web add ./dsh-vision
# restart dsh, then open a NEW session — the `vision` tool appears in the model's toolset
Configuration
Option A — environment variables:
export DASHSCOPE_API_KEY=sk-xxx
export VISION_MODEL=qwen3.7-flash-2026-07-15
export DASHSCOPE_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
Option B — profile patch layer ~/.dsh/profiles/<name>/cordis.patch.yml:
- id: dsh-vision
config:
apiKey: sk-xxx
model: qwen3.7-flash-2026-07-15
baseURL: https://dashscope.aliyuncs.com/compatible-mode/v1
Usage
The model calls the vision tool automatically:
| Param |
Required |
Description |
image |
✅ |
Absolute local path (e.g. C:\a.png) or http(s) URL |
prompt |
❌ |
Instruction; defaults to "describe this image in detail" |
Note: enabling image upload
The stock DeepSeek adapter declares models text-only and rejects uploads with
MODEL_DOES_NOT_SUPPORT_IMAGES. To let users paste images directly (rendered as
[图片附件: sha256:...] markers), apply the small patch described in PATCHES.md.
Articles / 相关文章
License
MIT