dsh-visual-plugin

Vision bridge plugin for DeepSeek Harness: when the main model has no vision, forward user images to a configurable OpenAI-compatible vision model and show the results in a Web UI right panel.

• Host half — agent/pre-step image interception (auto-describes every attached image before it reaches the text-only model), the vision_describe tool for follow-up questions, /vision-bridge/{test,balance,recent} routes, and a deepseek-vision wrapper adapter that lets the web gateway admit image uploads.
• Browser half — a right-side floating panel (shell.overlay) to configure url / model / api_key, test the connection, and watch recent image descriptions with thumbnails; a sidebar toggle; and a vision_describe tool card in the conversation.

Distributed as a publishable dual-half dsh bundle: "dsh": { "bundle": { "patch": "./cordis.patch.yml" }, "client": { "platform": "web" } }.

Install

# from an npm tarball / local checkout
dsh plugin --profile web add ./dsh-visual-plugin-0.1.0.tgz

# or from a git host (prebuilt lib/ is committed, so no prepare build is needed;
# pnpm >= 10 still requires an allowBuilds entry only when a prepare script runs)
dsh plugin --profile web add github:jyh20030112/dsh-visual-plugin

Then restart dsh web (client-package metadata is cached per boot), open the Web UI, and click 视觉桥接 / Vision Bridge in the sidebar footer.

Configure

In the panel:

1. 接口地址 / Endpoint URL — an OpenAI-compatible /chat/completions base, e.g. https://api.deepseek.com (or a GLM-4V/SiliconFlow/Moonshot endpoint).
2. 模型名称 / Model name — a vision-capable model, e.g. glm-4v-flash.
3. API Key — entered once, stored through the credentials seam (never in settings, never echoed back).
4. Click 保存配置 then 测试连接; a successful test shows latency.