⚠️ Archived — superseded by deepseek-v4-for-copilot

This Python proxy was a prototype to work around two DeepSeek V4 + VS Code BYOK bugs. It now lives on as a proper VS Code extension with the same fixes, plus vision, UI integration, and one-click Marketplace install.

DeepSeek V4 for Copilot Chat on the VS Code Marketplace.

• ✅ No proxy process to run
• ✅ No Python / pip dependencies
• ✅ Works in stable VS Code (not Insiders-only)
• ✅ Vision support (image attachments proxied through other Copilot models)
• ✅ Thinking mode, reasoning effort, prompt cache stats
• ✅ API key in OS keychain
• ✅ Inherits all Copilot features: agent mode, tools, skills, MCP

Repo: https://github.com/Vizards/deepseek-v4-for-copilot

A tiny proxy to bring DeepSeek V4 with thinking to VS Code Copilot Chat (BYOK).

DeepSeek V4's thinking mode breaks in Copilot Chat BYOK after the first turn:

1. Multi-turn 400s — DeepSeek requires reasoning_content from previous turns to be included in the next request. VS Code doesn't cache or re-inject it, so the second turn always fails.
2. Missing thinking UI — VS Code renders collapsible thinking blocks when it sees reasoning_text, but DeepSeek sends reasoning_content. The field name doesn't match.

This proxy sat between VS Code and the DeepSeek API to fix both. Both fixes are now built into the extension.

pip install -r requirements.txt
DEEPSEEK_API_KEY=sk-xxx ./start.sh

See git history for the original BYOK configuration steps.