@deepseek-ai/dsh-tool-vision-read
DSH 插件：vision_read — 将图像读取路由到专用视觉模型（例如 Kimi K3），让仅支持文本的智能体能够查看图像
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Mappedinfo/dsh-tool-vision-read#921d4af1a2333d58d1020703c20b0c2ab93833ab
Tool Vision Read 以 @deepseek-ai/dsh-tool-vision-read 发布，当前版本为 0.1.0-rc.6。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Unofficial community plugin. Independently developed and maintained; not part of the official DeepSeek Harness distribution.
A lightweight DeepSeek Harness plugin that registers a vision_read tool: it reads an image file through a dedicated vision model route and returns a text description — so a text-only agent (a model whose route does not declare image input, e.g. a chat model without vision) can still "see" images.
It is the same idea as routing different roles to different models (e.g. oh-my-pi's modelRoles), applied to one narrow capability: image description. No third-party CLI (modlens etc.), no manual delegation — the plugin does it in one call.
| Mode | Behavior | Cost |
|---|---|---|
direct (default) | The plugin reads the file, commits it through the attachment service, and makes one llm.stream call to the configured vision provider/model with a text+image message. | One round trip, no agent loop. |
subagent | The plugin starts an in-process subagent pinned to the vision route (agentOptions), which calls read_image itself and can iterate (zoom, OCR, follow-ups). | Full agent loop, more flexible. |
The tool always routes to the configured vision route, regardless of the calling model. Configuration parsing fails when either route id is omitted; a call resolves the route before image I/O and fails with guidance when the resolved route does not declare image input (declare it in the provider settings, e.g. defaultInput: [text, image] for pi-ai routes).
k3-256k, which accepts image input) — any provider that supports image content blocks works.Copy this package under packages/vision/tool-vision-read (or install from git), then:
pnpm install
Register the package in tsconfig.base.json (add ./packages/vision/*/src to the @deepseek-ai/dsh-* wildcard and the @deepseek-ai/dsh-*/invariant wildcard) and in tsconfig.host.json references, then mount it — see the official adding-a-package cookbook.
Install the Git repository into a profile through the DSH plugin manager:
dsh plugin --profile web add github:Mappedinfo/dsh-tool-vision-read
For local development, link the checkout instead:
dsh plugin --profile web add link:/absolute/path/to/dsh-tool-vision-read
The package declares a DSH Profile Bundle, so a successful install adds it to dsh.profile.bundles. Restart dsh web; its bundled cordis.patch.yml mounts vision_read automatically. The default route is kimi-coding / k3-256k. Override it for one launch without editing the package:
DSH_VISION_PROVIDER=my-provider DSH_VISION_MODEL=my-vision-model dsh web
The profile's own cordis.patch.yml is applied after the Bundle. Use an id-targeted override when the route or other options must be durable:
- id: tool-vision-read
config:
provider: my-provider
model: my-vision-model
mode: direct
Do not insert a second tool-vision-read row in the profile. Remove the Bundle with dsh plugin --profile web remove @deepseek-ai/dsh-tool-vision-read. The @deepseek-ai/* peer packages are satisfied by the DSH installation's module closure ($DSH_HOME/profiles/node_modules flat fallback) — autoInstallPeers: false keeps pnpm from pulling older registry copies.
| Key | Type | Default | Meaning |
|---|---|---|---|
provider | string | — (required) | Registered provider route owning the vision model. |
model | string | — (required) | Vision model id on that route. |
toolName | string | vision_read | Model-facing tool name. |
mode | 'direct' | 'subagent' | 'direct' | Execution mode. |
maxImageBytes | number | attachment limits | Cap on image bytes sent to the vision route. |
maxOutputTokens | number | 1024 | Cap on the vision route's output tokens. |
prompt | string | see source | Instruction sent beside the image; {{path}} and {{focus}} placeholders. |
vision_read(file_path: string, focus?: string)
Returns { path, provider, model, description } — the vision model's text description of the image. Accepts PNG/JPEG/WebP/GIF paths only; paths resolve against the calling session's workspace cwd.
A text-only agent (deepseek-v4-flash) calling vision_read on a campus-gate photo, with the description produced by Kimi K3-256K through the kimi-coding route:
vision_read demo in the DeepSeek Harness GUI
user: 请用 vision_read 看一下 /Users/shiqi/Downloads/微信图片_20260816082109_883_131.jpg 并描述内容
agent: (vision_read) → "这是一张横构图、白天拍摄的现代城市/园区街景照片……天空与云约占画面上方 2/3……
左侧一栋多层建筑转角呈弧形……中右一座较低的建筑带弧形屋顶边缘和竖向格栅外立面……"
src/index.ts # plugin (name/inject/apply/Config) + vision_read tool
src/invariant.ts # package invariant companion (no runtime invariant)
cordis.patch.yml # auto-mounted DSH Profile Bundle layer
lib/ # reference build emitted from the deepseek-harness monorepo
tests/ # vitest spec (runs in the monorepo context)
docs/dsh-discussion-draft.md # DeepSeek Harness "Show Your Plugins!" draft
lib/ is the reference build generated from the package inside the deepseek-harness monorepo (packages/vision/tool-vision-read); the tests run against the monorepo toolchain. The git dependency installs the committed lib/ directly.vision_read on a JPEG received a correct description from Kimi K3-256K.MIT