DeepSeek Harness Plugin Hub

Publish and manage complete Harness Profiles. Discover Plugins for your next setup.

Explore

PluginsPresetsDocsNews

Community

Publish a pluginContactReport an issue

Resources

Plugin Hub on GitHubDeepSeek HarnessSystem statusPrivacy notice
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

Independent and unofficial. Not affiliated with, authorized by, or endorsed by DeepSeek.

Vision Assist — DSH Plugin for DeepSeek Harness
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in
← Plugins
V

dsh-vision-assist

Vision Assist

DeepSeek Harness vision assistant plugin: equips models without vision capabilities with a switchable multimodal recognition model. Images in the input field are automatically saved to disk and rewritten as text prompts; the main model can view images by calling the vision_recognize tool. The recogn

The plugin will be installed here. Keep web if you are unsure.

npx -y @deepseek-ai/dsh plugin --profile web add github:wjling/dsh-vision-assist#072d8f77f12c5037a4f8e81adb9346f6b4e4a1d8
READMECompatibilityVersions

Description

DeepSeek Harness vision assistant plugin: equips models without vision capabilities with a switchable multimodal recognition model. Images in the input field are automatically saved to disk and rewritten as text prompts; the main model can view images by calling the vision_recognize tool. The recognition model can be hot-switched in the vision-assist namespace under settings.

Compatibility and provenance

Vision Assist is published as dsh-vision-assist and currently resolves to version 1.0.0. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
web
Release source
github
Registry updated
9/2/2026

Versions

1.0.0stable
9/2/2026

Related plugins

Loading related plugins…

Latest
1.0.0
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
web
License
MIT
Source
github
GitHub
★ 0
Weekly downloads
0
Last push
9/20/2026
View source ↗
README badge

Click the badge to copy Markdown for your README.

Do you maintain this Plugin?Claim benefit · Priority security scan

Verify the GitHub repository declared in package.json to manage this listing. After you claim it, Hub will prioritize a security scan of the current version and publish the result when it passes.

Claim this Plugin →
Report an issue

Related plugins

More verified plugins in vision-media.

Tool Describe Image@linxin666/dsh-tool-describe-imageModel-facing describe_image tool for the dsh web GUI: gives a text-only model image understanding by asking a vision-language model at an OpenAI-compatible endpoint to describe one image (local path, http(s) URL, or attachment reference). Hot-pluggable — Modlens@liustack/modlensPlug-in vision for text-only LLMs, powered by the free Antigravity CLIDeepseek Ivideodeepseek-ivideoiPolloWork HyperFrames Video Studio and 27 editable video templates as a native DeepSeek Harness conversation view.Image Gendsh-image-genBring ChatGPT-like image generation to DeepSeek Harness — Gemini, OpenAI, Seedream, DashScope, local ComfyUI & more.

README

dsh-vision-assist

Vision for text-only models in DeepSeek Harness (DSH). Give a model that cannot see images a pluggable multimodal "vision sidecar": paste an image and recognition is handed to a configurable VLM, so the conversation never trips the model does not support image input rejection.

🇨🇳 中文:给没有视觉能力的主模型配一个可随时切换的多模态识别模型,输入框图片开箱即用。

When the routed model cannot accept images, this plugin intercepts the request at the LLM call boundary and rewrites the image blocks into a "local attachment path + a call to vision_recognize" text hint. The message body and the UI history keep the original image; only the request sent to the vision-less model is rewritten. The actual recognition is done by the multimodal model you pick in settings.

# install from this repo
dsh plugin --profile web add github:wjling/dsh-vision-assist
# restart DSH to activate
# local development
git clone https://github.com/wjling/dsh-vision-assist
cd dsh-vision-assist
pnpm install
pnpm run build:client   # rebuild lib/client.js after editing src/client.js

功能

  1. vision_recognize 工具(host 级注册,任意会话可用):用选定多模态模型识别本地图片, 支持三档精细度 coarse / default / fine,可附加自定义 prompt。
  2. 模型调用边界图片降级:当会话路由模型无视觉输入时,发往该模型的请求在 llm.streamWithRegistration 边界被克隆改写(图片块 → 附件仓路径 + vision_recognize 指引);能看图的模型不受任何干预。GUI 气泡与会话历史始终保留原图。
  3. 热更新配置 + 设置页 UI:识别模型可在 DSH 设置 → 插件(可配置标签页)的 "视觉助手 vision-assist" 卡片里直接改,或改 settings 的 vision-assist 命名空间,均即时生效。

安装

dsh plugin --profile web add github:wjling/dsh-vision-assist
# 重启 DSH 生效

宿主共享包 @deepseek-ai/dsh-llm / dsh-settings / dsh-tools 按 peerDependencies 声明(运行时由 DSH 宿主提供,不遮蔽宿主版本); 开发环境通过 devDependencies 安装同名版本用于本地解析。schemastery 是唯一常规依赖。client 段构建产物为 lib/client.js(esbuild CJS 工厂, react 由宿主模块系统外部提供)。

配置

方式一:DSH 设置页(推荐)——设置 → 插件 → 可配置标签页 → "视觉助手 vision-assist" 卡片:

  • 启用识别接管(总开关)
  • 识别模型 provider(如 codemaker)
  • 识别模型 model(如 gemini-3.7-flash)
  • 识别超时(毫秒,1000–600000)

方式二:settings.yaml(热更新,两种方式写的是同一份配置):

vision-assist:
  enabled: true          # 总开关,false 时工具与图片改写都停用
  provider: codemaker    # 识别模型所在 provider(需在 llm-pi-ai 中配置)
  model: gemini-3.7-flash  # 识别模型 id
  timeoutMs: 120000      # 单次识别超时(毫秒)

要求:选定的模型在 llm-pi-ai.providers.<provider>.models 中声明了 input: [text, image](未声明的模型会被 DSH 视为纯文本而拒绝图片)。

工作原理

输入框贴图发送
   │  api-proxy 模态预检(resolveModelInfo 声明图片能力)→ 消息正常进入会话
   ▼
GUI 气泡显示原图,会话历史保留图片
   │  主模型回合发起 llm 调用(llm.stream / preparedCall.stream)
   ▼
包裹 streamWithRegistration:路由到无视觉 pi-ai 模型?
   ├─ 否 → 原样放行
   └─ 是 → 克隆请求,图片块(含 tool-result 嵌套)替换为
          "附件仓路径 + 调用 vision_recognize" 的文本指引
                      ▼
               主模型正常接单 → 调用 vision_recognize(path=..., mediaType=...)
                      ▼
          ctx.llm.stream 直连选定多模态模型 → 返回识别文本 → 继续任务

兼容性

  • 开发/测试环境:DSH 0.1.0-rc.8,Node ≥ 20
  • 插件全部逻辑在用户空间(~/.dsh + 插件包本体),更新 DSH 后无需任何重装或补丁
  • host + client 双段:client 段仅注册设置卡片(settings.plugin.item 插槽), 改动 src/client.js 后需执行 npm run build:client 重建

FAQ

  • 图片会不会丢? 不会。原图始终保存在 ~/.dsh/attachments/v1/objects/(内容寻址), 界面与会话历史保留图片,只有"发往无视觉模型的那次请求"里被替换为识别指引。
  • 怎么换识别模型? 设置 → 插件 → "视觉助手 vision-assist" 卡片里改 provider/model, 或改 settings 的 vision-assist.provider/model,立即生效。
  • 识别模型报 "does not support image input"? 该模型在 llm-pi-ai 里没声明 input: [text, image],补上声明后重试。
  • 会话切成视觉模型后图片没有改写,正常吗? 正常——插件只干预无视觉的模型。

Roadmap

  • client 设置 UI(在 Web 设置页直接选识别模型)
  • 多图并行识别
  • 识别结果缓存(同 attachmentId 不重复调模型)

License

MIT