What Is dsh-vision-router? Pixel Vision for Text-Only Agents

dsh-vision-router is a vision routing plugin for DeepSeek Harness (DSH) that gives text-only agents eyes: image input becomes ordinary tool calls, backed by a free built-in vision chain and 14 pixel-level vision tools—one-command install, no Python, with DeepSeek always the reasoning brain. This guide explains what dsh-vision-router is, its core features, install/update/uninstall commands, typical usage and troubleshooting, so you can let your agent see images and work on them step by step in the DSH Plugin ecosystem.

What Is dsh-vision-router?

dsh-vision-router solves the problem of DeepSeek text-only models not being able to see images: it routes the image turn directly to a vision model, pixel-faithful, while DeepSeek stays the reasoning brain. The positioning and facts below all come from the official README (source):

dsh-vision-router (GitHub repository ysr666/dsh-vision-router) is maintained by ysr666 and open-sourced under the MIT license. Most DSH vision plugins convert images into text descriptions for DeepSeek—a lossy description bridge; dsh-vision-router uses a routing bridge instead, handing the image turn directly to a vision model at pixel fidelity. It ships a built-in free vision chain (a five-model OVHcloud anonymous fallback, no account or key) and 14 pixel-level vision tools. Its tool layer draws on community projects such as Anionex's dsh-vision-toolkit, but all code is independently implemented. It is part of the DSH Plugin ecosystem, in the tools category.

What Are the Core Features of dsh-vision-router?

dsh-vision-router's core capabilities revolve around routing-bridge image turns, a free vision chain, 14 pixel-level tools and continuous multi-step image work: DeepSeek is always the brain, and the vision model is only the eyes called on demand. These capabilities all come from the official README (source):

• Routing bridge, not description bridge: image turns go straight to a vision model at pixel fidelity; DeepSeek reasons, and text turns stay untouched in model, cost and context.
• Free built-in vision chain: a five-model OVHcloud anonymous fallback by default, no account or key (2 requests/minute per IP per model, about 10 RPM in theory); user-configured vision models run first.
• 14 pixel-level vision tools: Q&A, grounding, element detection, cropping, pixel diff, color extraction, OCR, SVG tracing, cutout, HTML screenshot and long-screenshot OCR, all running on sharp, potrace, tesseract and system Chrome—no Python.
• Continuous multi-step image work: an image turn is a text turn that calls tools, e.g. vision_ground → vision_crop → vision_describe → vision_pixel_diff; the agent iterates until the task is done.
• Capability-aware auto routing: optionally enable Auto mode to prioritize already-configured models using measured capability evidence, never inferring capability from model names.
• A verifiable pixel loop: reference → vision_html_screenshot → vision_pixel_diff (diff ratio plus worst 8×8-grid regions) → fix → re-check, making UI restoration measurable (the README example ends at a 2.54% final diff).

How to Install and Enable dsh-vision-router?

Installing dsh-vision-router is a single command; the package wires its own composition patch, and after enabling the "👁 Vision" control in the composer you can use image turns like ordinary tool-calling turns. Requirements: a DeepSeek Harness Web profile and Node ≥ 22 (source):

1. Install dsh-vision-router: run the install command in your DSH terminal to add the plugin to the current profile:

npx @deepseek-ai/dsh plugin --profile web add dsh-vision-router

Wait for the command to finish; from a Harness source checkout use pnpm dsh plugin --profile web add dsh-vision-router.

2. Verify the installation: check that a vision-router line appears in the config:

npx @deepseek-ai/dsh --profile web --dump-config | grep vision-router

If vision-router shows up, the install succeeded; after first adding it to a long-lived Web process, let that process reload once so the plugin bundle is discovered.

3. Enable the "👁 Vision" control: when you need image input, click "👁 Vision" next to the composer until it shows "👁 Vision ✓", then paste or upload the image—pasting or uploading does not enable Vision automatically (source).

4. Update dsh-vision-router: a plain npx @deepseek-ai/dsh plugin --profile web update dsh-vision-router also updates, but pnpm v11 silently holds back releases younger than 24 hours; specify the version explicitly to be safe:

npx @deepseek-ai/dsh plugin --profile web add dsh-vision-router@<version>

You can also use the one-click update in Settings → Vision Router; settings live in the profile and survive upgrades.

5. Uninstall dsh-vision-router: run the remove command to drop the dependency and the bundle layer:

npx @deepseek-ai/dsh plugin --profile web remove dsh-vision-router

If you manually disabled the official DeepSeek row, re-enable it in your profile patch.

Typical dsh-vision-router Usage

dsh-vision-router is enabled by default after installation; daily use is four steps: turn on the Vision control, send an image, let the agent call vision tools, and iterate until you are satisfied. The following four steps cover typical usage from image Q&A to pixel-level restoration.

1. Enable "👁 Vision" and send an image: click "👁 Vision" in the composer until a checkmark appears, then paste or upload the image and send. The mode persists across sends and does not auto-reset; choosing a different ordinary model turns Vision off.

2. Ask the agent to look at the image or ground elements: just ask, and the agent calls tools such as vision_describe; use vision_ground when you need to locate something:

vision_ground image="ref.png" target="the send button"
vision_describe paths=["ref.png","impl.png"] question="list the differences" json=true

3. Verify with a pixel diff: after restoring a UI, compare against the reference with vision_pixel_diff:

vision_pixel_diff original="ref.png" rebuilt="screenshot.png"

The output reports a diff ratio plus the worst 8×8-grid regions; keep fixing until it converges.

4. Take local HTML screenshots and transcribe long screenshots: use vision_html_screenshot for local HTML pages and vision_long_screenshot_ocr for long captures:

vision_html_screenshot source="page.html" width=1200 height=720 fullPage=true
vision_long_screenshot_ocr image="chat-log.png" chunkHeight=1200 overlap=120

dsh-vision-router Troubleshooting

The four issues below come from the official README's Troubleshooting section: coexistence with dsh-web-ui, a UTF-8 BOM startup failure, a fresh release not taking effect, and a stale sharp conflict—the commands are copy-paste ready (source):

1. Images get rewritten when using dsh-web-ui / dsh-web-ui-all together: symptom: with both plugins installed, image uploads are converted into describe-image references and the vision plugin never receives the original image. cause: dsh-web-ui's send hook rewrites image uploads. Fix: open Settings → Plugin config → Image understanding and turn off "Rewrite images to describe-image references on send" (or set interceptImageSend: false); the switch is read on every send, so no hook reinstall or DSH restart is needed.

2. Startup fails with Unexpected token ... is not valid JSON: symptom: dsh web exits immediately at startup with a JSON.parse error. cause: the profile's package.json was saved as UTF-8 with BOM by an editor, and the invisible character makes JSON.parse fail. Fix: run the repair command, which removes the BOM and validates the JSON again:

npx dsh-vision-router repair --profile web

Use npx dsh-vision-router doctor --profile web to diagnose without modifying the file.

3. A fresh release does not take effect (downloaded 0 / added 0): symptom: updating reports 0 downloaded and 0 added, and nothing changes. cause: pnpm v11 silently holds back versions younger than 24 hours. Fix: install the target version explicitly (dsh plugin --profile web add dsh-vision-router@<version>) or run npx dsh-vision-router repair to lift the version pin.

4. Pixel tools fail with colourspace: parameter space not set after an upgrade: symptom: vision_* pixel tools report a colorspace error after upgrading from an early version. cause: a stale sharp 0.34.0 still sits in the profile and conflicts with the host's sharp 0.35.3 in the same process. Fix: delete node_modules/sharp and node_modules/@img under the profile and restart, or run pnpm install inside the profile.

Use Cases and Notes

dsh-vision-router fits every scenario where DeepSeek needs to see an image: UI restoration and verification, screenshot Q&A, icon or element grounding, OCR extraction, color picking and SVG tracing. Typical use cases include front-end restoration (quantified with the pixel loop), visual checks after UI automation, text extraction from long screenshots, and design-file Q&A with element grounding (source). Notes: the free chain is rate-limited by default (2 requests/minute per IP per model). For heavier use, configure a free key channel or a local Ollama backend; the vision_screenshot desktop capture is off by default and requires an explicit opt-in; vision_html_screenshot needs system Chrome while every other tool works without a browser; image text is treated as untrusted evidence, and the agent never executes instructions found inside images (source).

Project Links

dsh-vision-router is an MIT open-source project maintained by ysr666. Plugin details page: dsh-vision-router plugin details.

This page is an independent guide rewritten from the plugin's official README — for the authoritative documentation and the latest changes, defer to the source: ysr666/dsh-vision-router. A plugin is third-party code that runs on your machine once installed; inclusion is not an endorsement — review the source before installing.