Vision Proxy for Copilot

A VS Code extension that gives text-only models (like GLM 5.2) the ability to "see" images by proxying them through a vision model first.

Works with Volcengine Ark (火山引擎方舟) and any OpenAI Responses API compatible endpoint.

How It Works

flowchart LR
    A[User sends image] --> B[Vision Model<br/>doubao-seed-2.0-pro]
    B --> C[Text Description]
    C --> D[Text Model<br/>GLM 5.2]
    D --> E[Response streamed<br/>to Copilot Chat]

When an image appears in the conversation — whether pasted by the user or returned by a tool (e.g. screenshots) — the extension:

1. Extracts all images from user messages and tool results
2. Sends them to a vision model (doubao-seed-2.0-pro) to get a detailed text description
3. Appends the image description to the conversation as plain text
4. Forwards everything to the text-only model (glm-5.2) via the OpenAI Responses API
5. Streams the response back to Copilot Chat

This way, text-only models can handle image inputs without crashing with Model only support text input errors.

Features

• 🖼️ Vision Proxy — Automatically describes images from user messages and tool results (e.g. screenshots from MCP tools) before passing them to text-only models
• 🧠 Adjustable Reasoning Effort — Switch between minimal / low / medium / high directly from the Copilot Chat input toolbar (native configurationSchema integration)
• 🔧 Full Tool Calling — Supports Copilot agent mode with function calls and tool outputs, including images returned by tools
• 📡 Responses API — Uses the modern OpenAI Responses API (/responses endpoint) with reasoning support
• 📊 Context Window Usage — Reports token usage back to Copilot Chat so the context usage indicator works correctly
• 🔐 Secure API Key Storage — API key stored in VS Code SecretStorage, never written to settings files
• 🌊 Streaming — Real-time streaming responses via SSE with proper event handling

Configuration

Reasoning Effort

The reasoning effort can be adjusted directly in the Copilot Chat input toolbar — no need to open settings. A dropdown button appears next to the model selector showing the current effort level. Click it to switch between:

• Minimal — Fastest, least reasoning
• Low — Light reasoning, faster responses
• Medium — Balanced reasoning depth and speed
• High — Most thorough reasoning, highest quality

API Key

Use the command Vision Proxy: Set API Key (Ctrl+Shift+P) to store your API key securely in SecretStorage. You can also check the status with Vision Proxy: Show Status.

Usage

1. Install the extension
2. Run Vision Proxy: Set API Key and enter your Volcengine Ark API key
3. Open Copilot Chat
4. Select GLM 5.2 from the model picker
5. Start chatting — paste images and they'll be automatically described
6. Use agent mode — tools that return images (e.g. screenshots) are also handled automatically

Supported Models

The extension is designed for Volcengine Ark (火山引擎方舟) models but works with any OpenAI Responses API compatible endpoint:

Commands

Building

npm install
npm run compile
npx vsce package --no-dependencies --allow-missing-repository

Install the generated .vsix with:

code --install-extension vsion-proxy-copilot-0.0.8.vsix

Requirements

• VS Code 1.116.0 or later
• GitHub Copilot Chat extension
• A Volcengine Ark API key (or compatible OpenAI Responses API endpoint)

License

MIT