DeepSeek Chat for Visual Studio

📖 Overview

DeepSeek v4 for Visual Studiois a Visual Studio 2022+ extension that deeply embeds the DeepSeek V4 large language model into the IDE. It provides a native-grade chat experience through WebView2, supporting multi-agent collaboration, code editing, inline completions, terminal command execution, web search, OCR image recognition, multimodal image understanding, file parsing, and MCP external tool integration.

Core architecture is built on .NET Framework 4.7.2 + WPF + WebView2, using Visual Studio SDK 17.14, compatible with Visual Studio 2022 17.14 and above.

✨ Core Features

📦 Installation

Option 1: Download VSIX (Recommended)

Download the latest .vsix from Releases, close Visual Studio, then double-click to install.

Option 2: Build from Source

git clone https://github.com/zmy15/DeepSeek-v4-for-VisualStudio.git
# Open .slnx with VS 2022 → Build → F5 to launch experimental instance

🚀 Quick Start

1. Get API Key: platform.deepseek.com/api_keys
2. Open Settings: Tools → Options → DeepSeek Chat → Paste API Key → Select model
3. Open Chat Window: View → Other Windows → DeepSeek Chat (or click toolbar button)

💡 For "seeing" images (image understanding / screenshot analysis / PDF reading), switch the model to deepseek-v4-flash-vision-exp — see "Multimodal Image Understanding" below.

🤖 Multi-Agent Collaboration

Five agents automatically dispatch tasks via the Handoff protocol — no manual switching required:

User Question → Ask (analyze)
                  ↓
              Plan (plan, multi-file/complex tasks)
                  ↓
              Edit (execute code changes, step-by-step confirmation)
                  ↓
              Build (build verification + auto-repair, max 3 rounds)
                  ↓
              Ask (summarize & report)

Routing Mechanism:

• AI automatically classifies user intent (Plan→Edit→Build→Ask)
• Users can explicitly specify: @ask question, @plan task, @edit change, @build compile
• Explore is not independently routed; it is invoked internally by Plan/Ask agents via runSubagent

Handoff Protocol (AgentTypes.cs — HandoffRequest):

• Source agent declares handoff intent (target agent, reason, task description)
• Supports ChainBack — target chains back to source after completion
• Supports ForwardedMessages — reuses message list on handoff to maximize Prefix Cache hits

📐 Skills System

Define reusable AI workflows with Markdown (SKILL.md). Skill file format:

---
name: code-review
description: 'Review code quality, security, performance. Triggers: code review, check code quality, find bugs, security audit, PR review'
argument-hint: '[file or code]'
user-invocable: true
disable-model-invocation: false
---
# Code Review

## When to Use
- User requests code review or inspection
- Self-review before PR submission
- Finding potential bugs, security vulnerabilities, or performance issues

## Execution Steps
1. Read user-provided code or currently open file
2. Analyze: correctness, security, performance, maintainability
3. Sort findings by severity, provide fix recommendations

Skill Sources (priority high to low):

1. Project-level: .github/skills/, .agents/skills/, .claude/skills/
2. User-level: %USERPROFILE%\.copilot\skills\, %USERPROFILE%\.agents\skills\, %USERPROFILE%\.claude\skills\
3. Built-in: BuiltInSkills/ shipped with the extension

Three Trigger Methods:

• User explicit invocation: type /skillname
• AI semantic auto-match: automatically loaded when question semantics match skill description
• Context inference: proactively suggested when conversation context accumulates enough to need a specific skill

🧠 AI Memory System

AI manages three-tier persistent memory via the memory tool:

Supported Operations: view, create, str_replace, insert, delete, rename

Auto-Injection: At the start of each new conversation, user memory and repo memory are automatically injected into the System Prompt. Available to all agents.

🔧 MCP Protocol Integration

Connect external tool servers via Model Context Protocol:

• Transport Modes: HTTP + stdio dual mode
• Tool Classification: Auto-classify read/write by prefix, inject into corresponding agents
• Management UI: McpConfigDialog.xaml — visual configuration

📝 Code Editing Capabilities

Five Edit Tools

💡 insert_edit has been merged into apply_patch, using @@ context markers for insertion positioning.

Four-Tier Matching + Healing Auto-Repair

1. Exact Match → 2. Line-Level Match → 3. Levenshtein Fuzzy Match → 4. Healing Repair

Diff Preview

Before modification, red/green diff markers appear in-editor. Confirm per-hunk through DiffViewerWindow.

💡 Ghost Text Inline Completion

Powered by DeepSeek FIM API (api.deepseek.com/beta/completions):

• Gray prediction text in-editor (GhostTextTagger)
• Tab to accept, Esc to cancel
• InlinePredictionManager manages prediction lifecycle

🌐 Web Search & OCR

Search Engines

OCR Dual Engines

💡 OCR only extracts text; to understand the semantic content of images / screenshots / PDFs (layout, charts, context), switch the model to deepseek-v4-flash-vision-exp (see "Multimodal Image Understanding" below).

Multimodal Image Understanding (deepseek-v4-flash-vision-exp)

After switching the model to deepseek-v4-flash-vision-exp, the AI can directly "see" images instead of only extracting text:

⚠️ The vision model does not support inline FIM completion; it automatically falls back to deepseek-v4-flash for completions.

🛡️ Terminal Approval Security

Three approval modes: BlockAll / AllowAll / SmartBlock (recommended)

🌍 Internationalization

• Auto Chinese/English switching, based on LocalizationService + I18nMarkupExtension
• Custom translations: copy zh-CN.user.json.example to zh-CN.user.json

🧪 Testing

✅ 473+ tests all passing, 5 Agents 100% covered, 26 test files covering Models / Services / Integration

🗺️ Roadmap

💡 Suggestions and code contributions welcome via Issues!

👥 Contributors

📈 Star History

📄 License

MIT License © 2024 zmy15