dsh-vision-bridge

<h1 align="center">dsh-vision-bridge</h1> <p align="center">_{A vision layer for text-first models in DeepSeek Harness}</p>

<a href="./README.zh-CN.md">简体中文</a> · <a href="#install">Install</a> · <a href="#configure-vision-providers">Configure</a> · <a href="#security">Security</a>

dsh-vision-bridge is an installable DeepSeek Harness bundle that adds image understanding to text-first model routes. It preserves the Harness model list, adds a small glasses control to eligible models, and delegates image analysis to a separately configured vision provider.

The plugin supports Gemini-native, OpenAI-compatible Chat Completions/Responses, and Anthropic-compatible Messages APIs. The selected vision provider returns bounded text analysis; image blocks are never forwarded directly to the active upstream model.

Highlights

• One model list: Vision Bridge no longer appears as a duplicated provider group.
• Per-model glasses toggle: Gray means the bridge preference is off; blue means it is on.
• Selection stays explicit: Clicking the glasses only changes the preference. Clicking the model name selects the model and applies that preference.
• Persistent without model churn: Glasses preferences are remembered in the Harness web client without triggering a model switch or model-directory refresh.
• Clear service provenance: Hover the glasses to see the active vision provider and model.
• Multiple vision providers: Add, switch, and delete isolated provider profiles from Settings.
• Native-vision aware: Models that already advertise image input do not receive a bridge toggle.
• Reasoning UI preserved: The original Harness reasoning-effort menu remains unchanged.

Model selector behavior

Control	Result
Gray glasses	Vision Bridge is disabled for that model.
Blue glasses	Vision Bridge is enabled for that model.
Click glasses	Toggle and remember the preference only; the selected model does not change.
Click model name/row	Select the model. Blue routes through Vision Bridge; gray uses the normal upstream route.
Hover glasses	Show the vision provider and model that will perform image understanding.

The glasses control appears only when a matching bridge route exists and the upstream model is text-only or has unknown image capability. The preference is stored per upstream model in the local Harness client.

How it works

1. Enable the glasses for an eligible text-first model, then click the model name to select it.
2. Attach an image and ask a visual question normally.
3. Harness validates and stores the attachment in the session.
4. The bridge route replaces image blocks only in the provider-bound request copy with controlled attachment markers. The original session and transcript keep the images.
5. The active upstream model calls vision_bridge; the tool reads the latest session image through Harness attachment services.
6. The plugin sends a bounded request to the configured vision provider and returns only its text analysis to the active agent.

Explicit workspace paths remain supported through image_paths; they are resolved with Harness filesystem policy.

Requirements

• DeepSeek Harness 0.1.0-rc.5 or a compatible 0.1.x release
• Node.js ^22.19 or >=24
• An upstream model route that supports Harness tool calls
• An API key for at least one image-capable endpoint

…