hisence999/DSH-vison
DSH-vison is a DeepSeek Harness plugin that adds image understanding for text-only models by converting images into descriptions from a configured vision-capable model.
Install
via GitHub · GitHub source
Installing from GitHub runs the project's build script, which pnpm blocks until you allowlist it — run the command once and pnpm prints the exact key to add under `allowBuilds` in ~/.dsh/profiles/web/pnpm-workspace.yaml.
What happened when we ran it
Installed cleanly when we ran it
Every command here is run in a throwaway container against a clean profile, and the result is whatever the harness recorded — not a guess from the source. Last run 10d ago.
Show it in your README
For maintainers: the badge serves this listing's latest sandbox verdict, so a re-run updates it on its own — and it links readers to the full result here.
Due diligence
What it is — DSH-vison provides image understanding for text-only models by converting images into descriptions from a configured vision model.
Who it is for — When you use a text-only model to process user messages containing images, you need image understanding capability. If you use a multimodal model, you can skip installing this plugin.
Watch out — Sandbox test passed: it was installed in a fresh profile and registered by harness. No obvious pitfalls found.
The verdict — I would install it because the sandbox test passed and it enables global image handling for text models.
Generated by grok-4.6, and a starting point rather than a verdict. Where it says a plugin installs or does not, that is from a real run in a clean profile — everything else is read off the repository. Trust the source over this.
DSH-vison is a DeepSeek Harness plugin that adds image understanding for text-only models by converting images into descriptions from a configured vision-capable model. It wraps llm.resolveModelInfo so image admission and tool checks can be allowed for text-only models, then listens to agent/pre-step to process image-bearing user messages. The image remains in session history for UI display, while a session-surface replacement and deriveMessages wrapper present the generated description to the model. tools/post-execute applies the same conversion to read_image results, and system-prompt/assemble identifies the model active for the current step.
The plugin is intended for installations where an agent may receive images but the selected model is not multimodal. Multimodal models are detected and passed through without interception. Configuration is stored in the dsh-image-vision settings namespace; the profile patch uses enabled: true and patchAdmission: true as fallback settings.
DSH-vison requires at least one configured image-capable model and does not cache descriptions, so images are sent for recognition again in later turns. Its terminal surface includes the Windows PowerShell and Unix shell installation and removal scripts, which copy files and modify $DSH_HOME profiles, cordis.patch.yml, and the DSH dsh-host-apiproxy whitelist. It is a poor fit where a multimodal model already handles all images directly, or where modifying those host files is not acceptable.
The plugin stores settings in the dsh-image-vision namespace. The settings page exposes an enable switch, the recognition model, and the prompt. Its profile patch can provide these fallback values:
| Key | Value |
|---|---|
enabled |
true |
patchAdmission |
true |
patchAdmission controls the reversible wrapper around the shared llm service. The plugin also requires dsh-image-vision to be added to the dsh-host-apiproxy setting namespace whitelist; otherwise the settings page cannot find its namespace.
The plugin package consists of index.js, client.js, and package.json. For a profile installation, these files are placed under $DSH_HOME/profiles/<profile>/node_modules/dsh-image-vision/; the README gives web as the usual Web profile and shows the corresponding Windows path under C:\Users\<user>\.dsh\profiles\web\.... The profile’s $DSH_HOME/profiles/<profile>/cordis.patch.yml contains an image-vision entry naming dsh-image-vision.
The setup scripts patch the WEB_SETTINGS_NAMESPACES list in dsh-host-apiproxy. Reinstalling or upgrading DSH can overwrite that host file, requiring the patch to be applied again.
The repository supplies install.ps1 and uninstall.ps1 for Windows PowerShell, plus install.sh and uninstall.sh for Linux and macOS. Uninstallation removes the package, the image-vision patch entry, and the API proxy whitelist change. Configuration remains unless the removal script receives -PurgeConfig on Windows or --purge-config on Unix. Restart DSH after installation or removal.
At least one configured model that supports images is required. Descriptions are not cached; the same image is recognised once per turn, but later sends invoke the vision model again. Failed recognition retries other configured image-capable models at 0.6-second intervals. If all attempts fail, placeholder text is used instead of sending the image to a text-only model.
Written from the project's own documentation and kept in sync with it. Where the two disagree, the source is authoritative — read the README on GitHub
Same category
zhu1090093659
A `describe_image` vision tool for text-only models: images (local path, URL, attachment) go to a configurable OpenAI-compatible vision endpoint and only the returned text enters the session.
Installed cleanly when we ran it
liustack
Vision bridge for text-only models: paste an image, get structured JSON evidence (OCR, layout, semantics).
Installed cleanly when we ran it
Anionex
为纯文本模型"看图“设计更好的视觉工具箱和技能，支持多图理解，图片问答，前端UI还原、GUI 自动化等，并可选无缝接入多个主流agent，直接识别粘贴图片｜ A vision toolkit and skill designed for text-only llms — image Q&A, long-screenshot OCR, frontend UI restoration, and GUI automation, with optional seamless integration for Codex, Claude Code, Pi, Oh My Pi, and OpenCode
Installed cleanly when we ran it
ysr666
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.
Installed cleanly when we ran it
Anionex
Vision for text-only models: paste an image and the model switches to a Vision Toolkit variant for image Q&A, multi-image comparison, long-screenshot OCR, screenshot-to-UI reproduction, element grounding, and pixel diff. No API key by default — images are processed by the author-hosted free service, 100 per machine per day; configurable to your own provider.
Installed cleanly when we ran it
omdsh-dev
Chrome sidebar extension that lets DeepSeek Harness operate your browser directly, no vision capabilities required. 一款 Chrome 侧边栏扩展程序，可让 DeepSeek Harness 直接操控您的浏览器，无需视觉能力。
Installed cleanly when we ran it
