Scorp1o117/dsh-tool-vision
40Last commit Aug 16, 2026
dsh-tool-vision DSH plugin
dsh-tool-vision bridges the gap between text-only models and visual input. It provides an inspect_image tool and an image bridge that converts pasted images into text hints for the agent loop. Works with any OpenAI-compatible vision API (GPT-4o, Qwen-VL, GLM-4V, etc.).
How to install the dsh-tool-vision DSH plugin
dsh plugin --profile web add dsh-tool-visionCopying does not run this command. Review the repository and version before installing the dsh-tool-vision DSH plugin.
dsh-tool-vision DSH plugin data source
dsh-tool-vision DSH plugin snapshot date: Aug 16, 2026
discovered
What the dsh-tool-vision DSH plugin can do
- Inspect image tool: send local files or URLs to any OpenAI-compatible /chat/completions endpoint and return text answers.
- Image bridge: convert pasted images into text hints before they enter the durable log, preserving the agent-loop invariant.
- Global tool layer: the inspect_image tool is registered globally and available to all agents in the process.
- Web UI settings section (v0.3.0): configure API endpoint, key, model, and bridge options with hot-apply.
- Zero dependencies beyond the dsh SDK; works with OpenAI, Alibaba DashScope, Zhipu, Moonshot, Ollama, etc.
- Supports multimodal models directly: models listed in multimodalModels receive image blocks untouched.
Where the dsh-tool-vision DSH plugin fits
- Let a text-only DSH agent describe the contents of an image and answer questions about it.
- Bridge pasted images into text hints so that the conversation log remains consistent with the durable derivation.
- Automate image analysis workflows (e.g., screenshot analysis, document scanning) using the inspect_image tool.
- Integrate with local vision models via Ollama-compatible endpoints for offline image processing.
- Use the Web UI settings to quickly switch between different vision providers without restarting.
Who the dsh-tool-vision DSH plugin is for
- DSH users who need to add visual understanding to their agents.
- Developers building AI agents that process images in a chat or automation context.
- Users who want to leverage external vision APIs (e.g., GPT-4o, Qwen-VL) without modifying the DSH core.
dsh-tool-vision DSH plugin limitations
- Bridged images become text hints (not pixels) in the conversation, so pixel-precise in-context reasoning is unavailable for text-only models.
- Images are base64-transferred: mind privacy and size limits (default max 10 MB).
- The plugin is independent of the dsh-llm routing/retry system; failures return clear errors to the agent.
dsh-tool-vision DSH plugin: from the repository README
Quoted from the Scorp1o117/dsh-tool-vision README, the upstream source of the dsh-tool-vision DSH plugin. Copyright remains with the original authors.
**GitHub**: [Scorp1o117/dsh-tool-vision](https://github.com/Scorp1o117/dsh-tool-vision) · **npm**: [dsh-tool-vision](https://www.npmjs.com/package/dsh-tool-vision) [](https://github.com/Scorp1o117/dsh-enhancement-suite) [](https://www.npmjs.com/package/dsh-enhancement-suite) Part of the [DeepSeek Harness Enhancement Suite](https://github.com/Scorp1o117/dsh-enhancement-suite) — Vision · Soul/Persona · Long-term Memory · Plugin Marketplace. External vision model for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness). DeepSeek's own models are text-only, and the harness derives every model request strictly from the session log (`llm/stream` requests must equal the durable derivation — the agent-loop invariant). This plugin bridges the gap in two ways: 1. **`inspect_image` tool** — sends an image (local file, or http(s) URL) to **any OpenAI-compatible** `/chat/completions` endpoint that supports `image_url` content parts, and returns the vision model's textual answer into the agent loop. 2. **Image bridge (v0.2.1
Read the full READMERepository license: MIT
dsh-tool-vision DSH plugin questions
How do I install dsh-tool-vision in my DSH profile?
You can install it via npm: `pnpm add dsh-tool-vision` in your profile directory. Then add a mount entry in your profile's `cordis.patch.yml` or use the one-step command: `dsh plugin --profile web add dsh-tool-vision`. The plugin will be automatically discovered and mounted.
Which vision models are supported?
The plugin works with any OpenAI-compatible API that supports `image_url` content parts. Examples include OpenAI GPT-4o/GPT-4o-mini, Alibaba DashScope Qwen-VL, Zhipu GLM-4V, Moonshot, and local Ollama with vision models like llama3.2-vision. You configure the `baseURL` and `model` in the plugin settings.
How does the image bridge work?
When you paste an image into a text-only model, the bridge intercepts the message on the `agent/pre-step` waterfall and replaces the image content with a text hint, such as '[User sent an image, exported to: <path>. Inspect it with the inspect_image tool...]'. The agent can then call `inspect_image` to analyze the image via the configured vision endpoint. This preserves the session log invariant.
Can I use the plugin without a web UI?
Yes. The plugin works entirely in the backend. You can configure it via the profile patch file (`cordis.patch.yml`) or environment variables. The Web UI settings section is an optional convenience for visual configuration; all changes can be made directly in `settings.yaml`.
What should I do if the settings section doesn't appear in the Web UI?
The plugin automatically patches the `dsh-host-apiproxy` namespace allowlist on first start. You need to restart `dsh web` once after the initial patch. After that, the settings section should appear under Settings → 视觉模型. If it still doesn't appear, check that the plugin is correctly mounted and that you have the latest version.