wanshichenguang/dsh-tool-vision

DSH plugin: image_describe (识图) tool over the DashScope OpenAI-compatible vision API (qwen3.7-flash). Bring your own DASHSCOPE_API_KEY.

Bundle verified MIT JavaScript Unknown
Bundle verified

Listed

1

Vision

Bundle verified

★ 1 View on GitHub
VersionUnknown
LanguageJavaScript
LicenseMIT
View on GitHub

Preview

Preview 1 of 2: wanshichenguang/dsh-tool-vision
Preview 2 of 2: wanshichenguang/dsh-tool-vision

What it does

Adds an image_describe tool backed by the DashScope OpenAI-compatible vision API; on text-only sessions, pasted images are stored as local paths for the model and rendered inline in the chat transcript.

Best for

  • Text-only DSH sessions that need to inspect pasted or dropped images through a tool.
  • Agents that need descriptions or focused answers about local image files.
  • DashScope users who want inline transcript rendering while the model receives a readable local image path.

Not ideal for

  • Deployments without a DashScope API key; the tool registers but fails when invoked.
  • Teams requiring a vision provider other than the DashScope OpenAI-compatible endpoint.
  • Images above the configured 20 MiB default per-image limit.
  • Deployments requiring per-call approval unless they add a separate `tools/pre-execute` policy.

README

dsh-tool-vision

English 中文

A DeepSeek Harness plugin: the model-facing image_describe（识图） tool. It reads a local image file and asks the DashScope OpenAI-compatible API (qwen3.7-flash) to describe it or answer a focused question.

Works in every agent preset: the bundle patch inserts the tool at the host plane, so every session sees it in the tool catalog.

Install

dsh plugin --profile web add dsh-tool-vision

Restart dsh web (or install it from the Plugin Market with one click). The tool appears as image_describe in the model’s tool catalog.

Paste bridge

On the web surface, pasting or dropping an image into the composer while the session’s model is text-only (positively confirmed from provider metadata) uploads each draft image to the plugin’s host route, releases the draft, and appends [image: <path>] text to the prompt before sending — so the request never trips the host’s image admission, and the model can read the file with image_describe. Vision-capable or unknown models keep the native intake (the image attaches inline, unchanged).

The uploaded copies live in the harness data home ($DSH_HOME/media, default ~/.dsh/media) under the same naming convention as the harness’s own image-to-path conversion.

The sent message also renders the images back inside its user row in the transcript: a conversation node matches the [image: <path>] markers in the user message and displays the copies served by the host route while hiding those transport markers from the visual bubble. The stored path text remains unchanged for the text-only model.

API key

The plugin ships no key. Set the key in the environment that starts DSH:

DASHSCOPE_API_KEY=sk-xxxx

You can also pass it per-row in cordis.patch.yml:

- id: tool-vision
  name: 'dsh-tool-vision'
  config:
    apiKey: sk-xxxx

Without a key the tool registers normally and fails with VISION_NO_API_KEY on use, naming the fix.

Tool

Tool Args Behavior
image_describe path (string), prompt (string, optional) Resolves the image (absolute path, or relative to the workspace), sends it as a base64 data URL plus the prompt to POST {baseUrl}/chat/completions, and returns the assistant text. Omitting prompt uses defaultPrompt (a complete content description).

Config

Key Default Meaning
baseUrl https://dashscope.aliyuncs.com/compatible-mode/v1 DashScope OpenAI-compatible endpoint.
apiKey DASHSCOPE_API_KEY environment variable DashScope API key; an explicit config value wins over the environment.
model qwen3.7-flash Vision model id.
defaultPrompt 请详细描述这张图片的内容：画面主体、关键细节、文字信息（如有）、布局与结构。 Instruction used when a call omits prompt.
timeoutMs 120000 Cooperative tool-call timeout budget (ms).
maxImageBytes 20971520 Cap on one image’s byte size.
maxOutputChars 65536 Cap on the returned description text.

Security

  • No credentials ship with the package; the key stays in your environment or profile config.
  • The HTTP client sets redirect: 'error': the configured endpoint receives the credential, and a redirect fails the call instead of forwarding the key or the image to another origin.
  • The image bytes go only to the configured endpoint.

Known Limitations and Deferred Work

  • Single provider, single request shape — the tool targets the DashScope OpenAI-compatible endpoint only; a provider-selection seam is deferred until a second vision backend needs the same schema.
  • No vision-specific permission policy — the tool executes without requesting ctx.approval; a deployment that needs confirmation must add a tools/pre-execute policy.

License

MIT

Frequently Asked QuestionsFAQ

Use the verified command dsh plugin --profile default add github:wanshichenguang/dsh-tool-vision in a DSH-enabled shell. The command resolves the public package metadata and keeps the plugin attached to the catalog identity shown on this page.

tt-a1i/archify#integrations/deepseek-harness★ 14,012

Generate validated, self-contained interactive architecture, workflow, sequence, data-flow, and lifecycle diagrams from repositories or system descriptions.

MIT
liustack/modlens★ 2,925

Vision bridge for text-only models: paste an image, get structured JSON evidence (OCR, layout, semantics).

MIT
GanyuanRan/Aegis★ 1,048

Software-engineering method pack for coding agents, with skills for baseline-first planning, systematic debugging, prompt hygiene, verification before completion, and repair/retirement tracking.

MIT
ysr666/dsh-vision-router★ 680

Free vision for text-only agents: built-in keyless vision chain plus pixel tools (Q&A, grounding, crop, pixel diff, colors, OCR, SVG trace, cutout, screenshots); paste an image to use it.

MIT