dsh-vision-router: Pixel-Level Vision Tools for DeepSeek Harness
ysr666/dsh-vision-router
Eyes for text-only DeepSeek Harness agents: free built-in vision chain and pixel-level image tools, one-command install.
Give text-only DeepSeek Harness agents 'eyes' and turn image viewing into ordinary tool calls.
dsh plugin --profile web add github:ysr666/dsh-vision-router- Category
- Tools & Capabilities
- Platform
- DSH-Plugin
- Author
- ysr666
- Distribution
- Plugin
dsh-vision-router Key Features
dsh-vision-router Repository
dsh-vision-router Repo Summary
dsh-vision-router is a vision plugin (DSH plugin) for DeepSeek Harness, providing "eyes" for text-only DeepSeek Harness agents. It solves the core problem that traditional vision plugins convert images to text descriptions—lossy, one-shot, and blind to pixels. This plugin keeps the original pixels on the vision model's side and lets DeepSeek focus on reasoning, making "looking at an image" an ordinary tool call. It includes a built-in free vision chain (no key required) and eleven pixel-level vision tools for Q&A, grounding, cropping, pixel diff, color recognition, OCR, SVG tracing, cutout, and screenshots. Installation is one command, no Python needed, and image interactions behave like ordinary tool-calling turns—grounded, measurable, and repeatable. The project is maintained by ysr666, licensed under MIT, last updated in 2026-08, and verified with 149 tests.
What core features does the dsh-vision-router DSH plugin support?
The plugin works out of the box and is free by default: the vision tool chain ends with a five-model OVHcloud anonymous fallback—no account, no key, 2 requests/minute per IP per model, roughly 10 RPM across independent buckets. Users can also provide their own vision models. Version v1.4.0 adds automatic recognition and direct-channel bridging of undeclared vision models, doctor repair for stale version-pinned exemptions, and a spotlight-guided onboarding walkthrough, plus hardened settings-save verification and vision-backend compatibility.
How do I install dsh-vision-router?
Installation is a single command. The package ships its own composition patch (dsh.bundle.patch); dsh plugin add wires the row, the admission wrapper, and the attachment limits automatically—zero manual file edits. Taking over the official DeepSeek route is an optional setting (stealth mode, off by default). The install command is:
dsh plugin --profile web add github:ysr666/dsh-vision-router
What vision tools does dsh-vision-router provide?
The plugin includes eleven pixel-level tools covering visual Q&A (vision_qa), grounding (vision_ground), cropping (vision_crop), pixel diff (vision_pixel_diff), color recognition, OCR, SVG tracing, cutout, and screenshots. These tools work as ordinary tool calls: DeepSeek keeps reasoning while the vision tools do the seeing, with results that are measurable and repeatable.
Which DeepSeek Harness versions and platforms does dsh-vision-router support?
The plugin targets the DSH Web profile, requires Node.js >= 22, and needs no Python runtime. It depends on @deepseek-ai/dsh-llm-deepseek and @deepseek-ai/dsh-anonymous-user-id, and ships with the official dsh.bundle manifest for deep integration with DeepSeek Harness.
How do I configure and manage the lifecycle of dsh-vision-router?
Configuration options include vision model selection, anonymous fallback toggle, and stealth mode. The plugin supports full installation and lifecycle management, and provides a troubleshooting guide. See the Configuration and Troubleshooting sections of the repository README for more details.