---

What Is This?

DeepSeek v4 for Visual Studio embeds the DeepSeek V4 model directly into your editor. Select code, paste screenshots, drag in files — AI is right beside you, always ready to respond.

It's more than a chat window; it's a complete AI workflow system:

• Multi-Agent Collaboration Engine — Four specialized agents with automatic task dispatch and handoff
• Skills Engine — Define reusable AI workflows with Markdown
• MCP Protocol — Connect to any tool ecosystem with automatic Function Calling
• 1M Token Context — Handle large codebases with intelligent compression that preserves information
• Three Editing Methods — Patch / Insert / Create with four-level matching for precise application
• RAG Retrieval-Augmented Generation — Pluggable knowledge base integration
• Three OCR Engines — Read your error screenshots (PaddleOCR-Sharp / Windows / MCP)
• 🌐 Internationalization (i18n) — Auto-switch between Chinese and English, user-customizable translations

---

Feature Overview

Feature	Description
🧠 DeepSeek V4	Streaming chat · Deep Thinking (Reasoning) · Dual model (Pro / Flash)
🤖 Multi-Agent System	Ask / Explore / Plan / Edit agents with automatic Handoff collaboration
🔧 MCP Protocol	Multi-server connections · Function Calling · Tool allowlists · Persistent config
📐 Skills System	Slash commands · Project/User/Built-in tiers · YAML frontmatter metadata
📝 Three Editing Methods	apply_patch / insert_edit_into_file / create_file, four-level matching + Healing repair
📚 1M Context	900K token budget · Context compression · No more file truncation
🔍 RAG Retrieval	Pluggable provider interface · Smart caching · Auto-injected into conversation context · 🚧 Built-in vector DB in development
🌐 Web Search	Baidu Qianfan (1500 free/month) + DuckDuckGo dual engine · Auto fallback on quota exhaustion
📄 File Parsing	50+ formats · Code/Docs/PDF/Word/Excel all supported · Drag & drop parsing
🖼️ Image OCR	PaddleOCR-Sharp local · Windows built-in · MCP remote three engines
🌐 Internationalization	Auto-detect system language · Manual override in Options · User-customizable translations
📊 Code Diff Preview	Red/green diff markers in editor · Accept/Undo per hunk · Apply all at once
💡 Ghost Text Completion	Inline grey predictions · Context-aware · Configurable debounce delay
💬 Chat Window	WebView2 rendering · Markdown/code highlighting · Multi-session persistence · Live plan display
⚙️ Visual Configuration	Tools → Options one-stop settings · Context/Search/OCR categorized management

---

Multi-Agent Collaboration System

This extension includes four specialized agents that collaborate automatically via Handoff — no manual switching needed:

Agent	Role	Capabilities	Can Handoff To
Ask 🤔	Q&A Assistant	Code explanation, read-only analysis, knowledge Q&A	Explore
Explore 🔍	Explorer	Codebase search, file discovery, structure analysis, reference tracking	Ask, Plan, Edit
Plan 📋	Planner	Task decomposition, solution design, plan.md generation	Edit, Explore
Edit ✏️	Executor	Code write/delete, file operations, post-edit diagnostics	Explore, Ask

Typical Collaboration Flow

User Question → Ask (Analyze problem)
                  ↓ needs planning
               Plan (Create plan → generate plan.md)
                  ↓ Handoff
               Edit (Execute changes → notify Explore to investigate files)
                  ↓ completed
               Ask (Summarize and report)

Each agent has its own system prompt, tool allowlist, and permission boundaries, ensuring safety and control.

---

Skills System

🔥 This is the core feature that sets this extension apart from ordinary AI plugins.

What Is a Skill?

A Skill is a Markdown file (SKILL.md) with YAML frontmatter that describes "when to trigger, what to do". Once loaded, the AI gains professional workflow instructions for that domain.

---
name: code-review
description: 'Review code for quality, security, and performance. Use when: code review, PR review, code audit'
argument-hint: '[file path or code snippet]'
user-invocable: true
---

# Code Review

## Review Process
1. Analyze from five dimensions: correctness, security, performance, maintainability, best practices
2. List issues by priority: 🔴 Critical → 🟡 Medium → 🟢 Suggestion
3. Provide fix suggestions and code examples for each issue

Three Skill Source Levels

Level	Path	Use Case
📁 Project	.github/skills/ .agents/skills/ .claude/skills/	Version-controlled, team-shared
👤 User	~/.copilot/skills/ ~/.agents/skills/	Personal preferences, cross-project
🏭 Built-in	BuiltInSkills/ (bundled with extension)	Out-of-the-box, e.g. code-review

Usage

Type / in the chat window to trigger slash command autocompletion. Select a skill and the AI loads the corresponding workflow:

/code-review  UserService.cs
/tdd          Implement user login feature
/triage       #42 How should this bug be handled

---

MCP Protocol Integration

Connect to external tool servers via Model Context Protocol (MCP), infinitely expanding AI capabilities:

• Multi-server simultaneous connections: Each server runs in its own process, no interference
• Automatic Function Calling: AI determines when to call MCP tools automatically
• Tool Allowlists: Each agent declares which tools it is allowed to use, fine-grained permission control
• Persistent Configuration: %LocalAppData%\DeepSeekVS\mcp_servers.json stores server list
• Built-in OCR Server: PP-OCRv5 integrated by default (uvx paddleocr-mcp)
• Internal Tool Filtering: Internal tools like OCR are automatically hidden from the AI's visible list to prevent accidental calls

Configuration entry: Chat window → 🔌 MCP button → Add/Manage servers.

---

1M Context & Compression

Leverage DeepSeek V4's 1M token context window:

Token Budget Management

Parameter	Value	Description
Token Limit	900K	Reserve 100K for output
File Size Limit	No limit	No more file content truncation
Auto Compression Threshold	85%	Trigger compression when usage reaches threshold

Context Compression

When usage exceeds 85%, early conversation turns are automatically compressed:

• Preserve last 3 turns in full
• Earlier turns compressed into concise summaries, injected as system messages
• Supports both LLM summarization and rule-based extraction modes
• Compressed summaries can be re-compressed (progressive)
• Real-time ContextStats for querying token distribution across dimensions

Configure compression parameters under Tools → Options → DeepSeek Chat → Context Management.

---

RAG Retrieval-Augmented Generation

⚠️ Planned: The IRagProvider interface and RagService registration/caching infrastructure are in place. A built-in local vector database provider is under development. You can currently integrate a custom RAG backend by implementing IRagProvider.

Pluggable RAG integration providing AI with project knowledge base support:

• Provider Interface (IRagProvider): Register any RAG backend
• Smart Caching: Reuse results for consecutive queries with Jaccard similarity ≥60%
• Auto Injection: Retrieval results injected into context before each conversation round
• Multi-provider Support: Switch active provider by name
• 🚧 Built-in Local Vector DB: Zero-config solution based on SQLite + local embedding model (planned)

// Register a custom RAG provider
var ragService = new RagService();
ragService.RegisterProvider(new MyCustomRagProvider());
ragService.SetActiveProvider("MyProvider");
ragService.IsEnabled = true;

📋 See Roadmap / TODO for RAG-related plans.

---

Web Search

Search Engine	Features	Quota
Baidu Qianfan	Excellent Chinese search results	1500 free/month, auto switch on exhaustion
DuckDuckGo	Completely free, privacy-protecting	May not be accessible in China

• Intelligently generates search keywords based on conversation context
• Search results automatically injected into chat context
• Seamless fallback to DuckDuckGo when Baidu quota is exhausted

---

File Parsing

Support drag-and-drop or paste of 50+ file formats, automatically extracting text content:

Category	Formats
Code	.cs .py .java .js .ts .go .rs .cpp .c .h .swift .kt .rb .php .sql .html .css .xml .json .yaml .toml .proto etc.
Documents	.txt .md .rst .log .csv
Office	.doc .docx .xls .xlsx
PDF	.pdf (parsed via UglyToad.PdfPig)
Images	.png .jpg .jpeg .bmp .gif .tiff .webp → auto OCR

How to use: Drag files directly from File Explorer into the chat window, or Ctrl+V to paste.

---

Image OCR

Three OCR engines for different scenarios:

Engine	Chinese Accuracy	Setup Difficulty	Best For
Windows Built-in	Average	Zero config	English screenshots, quick lookups
PaddleOCR-Sharp	≥95%	Auto-download model	Chinese error screenshots (recommended)
MCP OCR	Depends on server	Requires server config	When you have a custom OCR service

💡 Simply Ctrl+V paste an error screenshot — AI automatically recognizes the text and analyzes the problem without manually typing error messages.

---

Code Diff Preview

After AI modifies code, changes are marked in the editor with red (deleted) / green (added) markers:

• Live Preview: Preview all changes before applying
• Per-hunk Confirmation: Accept / Undo each diff block
• Apply All: Accept all changes at once when confirmed
• Post-Edit Diagnostics: Automatically check for newly introduced compilation errors

Toggle this feature via DeepSeekOptionsPage.

---

Ghost Text Code Completion

In-editor grey ghost text predictions, similar to GitHub Copilot:

• Context-Aware: Uses current file content and cursor position
• Configurable Debounce: Adjustable trigger interval to avoid excessive requests
• Caching: LRU cache for 10 most recent completion results
• Non-Intrusive: Grey text, Tab to accept, Esc to cancel

Enable and configure under Tools → Options → DeepSeek Chat.

---

Installation

Recommended: Download VSIX Package

1. Releases → Download DeepSeek_v4_for_VisualStudio.vsix
2. Close all Visual Studio instances
3. Double-click the .vsix file → Follow the prompts to install
4. Restart Visual Studio

Advanced: Build from Source

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

Build Environment Requirements:

Component	Version
Visual Studio	2022 (17.14+)
.NET Framework SDK	4.7.2
Visual Studio SDK	Via VS Installer → Modify → Check "Visual Studio extension development"
Windows	10/11 x64

---

🌐 Internationalization (i18n)

DeepSeek v4 for Visual Studio supports Chinese and English interfaces, automatically following your system language, or manually switchable.

Automatic Language Detection

The extension automatically detects your Windows system UI language on startup:

• Chinese system → Displays Chinese UI and prompts
• English / Other systems → Displays English UI and prompts

Manual Language Switching

1. Open Tools → Options → DeepSeek Chat
2. Find the Language / 语言 category
3. Select from the Display Language / 显示语言 dropdown:
   • auto — Auto-detect system language (default)
   • zh-CN — Force Chinese
   • en — Force English
4. Click "OK" — the interface updates immediately

Custom Translations

You can create custom translation files to override default translations:

1. Create zh-CN.user.json or en.user.json in the extension's Resources\Locales\ directory
2. Add the key-value pairs you want to override, for example:

{
  "ui.welcomeMessage": "Hello! This is my custom welcome message!\nStart asking!"
}

3. Restart Visual Studio or toggle the language to apply

Coverage

i18n covers the following:

• UI Text — Tool window titles, button labels, dialogs
• AI Prompts — System prompts, skill prompts, Agent routing prompts
• Output Messages — Welcome messages, error messages, API status messages
• Settings Page — All option display names and descriptions

---

Quick Start

① Get an API Key

Visit platform.deepseek.com/api_keys → Create API Key → Copy.

② Configure the Extension

Tools → Options → DeepSeek Chat → Paste API Key → Select model.

Setting	Recommended	Description
API Key	Your key	Obtain from platform.deepseek.com
Selected Model	deepseek-v4-pro	Pro model has stronger reasoning
Enable Deep Thinking	✅ On	Show model's reasoning process
Reasoning Effort	high	Reasoning depth (high / max)
Search Provider	Baidu Qianfan	Recommended for China, 1500 free/month
OCR Engine	PaddleOCR-Sharp	Best for Chinese text recognition
Show Diff Markers	✅ On	Preview changes before applying
Copilot Enable	✅ On	Inline code completion
Token Budget	900000	1M context upper limit
Auto Compression	✅ On	Auto compress when token limit exceeded

③ Open Chat Window

View → Other Windows → DeepSeek Chat, or click the toolbar icon.

④ Quick Reference

Action	Method
Code Q&A	Type your question; AI automatically reads the currently open file
File Content Parsing	Drag files from File Explorer into the chat window
Screenshot Error Recognition	Ctrl+V paste screenshot, auto OCR recognition
Web Search	Check 🌐 Web Search in the chat window
Invoke Skills	Type / to select slash commands
Manage MCP Servers	Chat window → Click 🔌 MCP button
Switch Agent	Agent selector at the top of the chat window
Multi-session Management	Left sidebar session list → New/Switch/Delete

---

Architecture Overview

┌─────────────────────────────────────────────────────────┐
│                   Visual Studio 2022                     │
│  ┌─────────────┐  ┌──────────────┐  ┌───────────────┐  │
│  │  Chat Window │  │  Diff Viewer  │  │  Ghost Text   │  │
│  │  (WebView2)  │  │  (Adornment)  │  │  (Tagger)     │  │
│  └──────┬───────┘  └──────┬───────┘  └───────┬───────┘  │
│         │                 │                   │          │
│  ┌──────┴─────────────────┴───────────────────┴───────┐  │
│  │                 AgentDispatcher                    │  │
│  │         (Central Router · Handoff Management)       │  │
│  └──────┬──────┬──────┬──────┬────────────────────────┘  │
│         │      │      │      │                           │
│    ┌────┴─┐ ┌─┴───┐ ┌┴───┐ ┌┴────┐                      │
│    │ Ask  │ │Expl.│ │Plan│ │Edit │                      │
│    └──────┘ └─────┘ └────┘ └─────┘                      │
│                                                          │
│  ┌───────────────────────────────────────────────────┐  │
│  │                  Service Layer                      │  │
│  │  DeepSeekApi │ SkillService │ McpManager │ OCR     │  │
│  │  FileParser  │ EditPatch    │ WebSearch  │ RAG     │  │
│  │  ContextMgr  │ Compressor   │ DiffMarker │ ChatPst │  │
│  └───────────────────────────────────────────────────┘  │
│                                                          │
│  ┌───────────────────────────────────────────────────┐  │
│  │                  External Services                  │  │
│  │  api.deepseek.com  │  MCP Servers  │  Search APIs  │  │
│  └───────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────┘

Core Services

Service	Responsibility
AgentDispatcher	Multi-agent central routing, Handoff coordination, workflow orchestration
DeepSeekApiService	DeepSeek API calls, streaming responses, Thinking/Reasoning control
SkillService	Skill discovery, loading, YAML parsing, slash command completion
McpManagerService	MCP server lifecycle management, tool aggregation and invocation
EditPatchService	Three editing method parsing, four-level matching, Healing repair, diagnostics check
ContextCompressorService	Context compression, LLM/rule dual-mode summarization
RagService	RAG provider registration, activation, retrieval result injection
ConversationContextManager	Conversation context construction, token budget management, message trimming
WebSearchService	Dual-engine search, auto switching, intelligent keyword generation
OcrService	Three-engine OCR unified interface (PaddleOCR-Sharp / Windows / MCP)
FileParserService	50+ format file text extraction
ChatHtmlService	WebView2 HTML/CSS/JS generation, Markdown rendering
CodeDiffService	Code difference calculation and editor markers
ChatPersistenceService	Multi-session persistent storage

---

Development Guide

Project Structure

DeepSeek_v4_for_VisualStudio/
├── Models/                  # Data models
│   ├── AgentModels.cs       # Agent task plan models
│   ├── AgentTypes.cs        # Agent type enums and definitions
│   ├── ContextModels.cs     # Context statistics and compression models
│   ├── DeepSeekModels.cs    # DeepSeek API request/response models
│   ├── EditPatchModels.cs   # Edit patch models
│   ├── McpTypes.cs          # MCP protocol types
│   ├── RagModels.cs         # RAG retrieval models
│   ├── SkillDefinition.cs   # Skill definition models
│   └── TreeModels.cs        # File tree models
├── Services/                # Business services
│   ├── Agents/              # Agent implementations
│   │   ├── BaseAgent.cs     # Agent base class
│   │   ├── AskAgent.cs      # Ask agent
│   │   ├── ExploreAgent.cs  # Explore agent
│   │   ├── PlanAgent.cs     # Plan agent
│   │   └── EditAgent.cs     # Edit agent
│   ├── AgentDispatcher.cs   # Agent dispatcher
│   ├── ChatHtmlService.cs   # Chat HTML rendering
│   ├── CodeDiffService.cs   # Code diff service
│   ├── ContextCompressorService.cs  # Context compression
│   ├── ConversationContextManager.cs # Conversation context management
│   ├── DeepSeekApiService.cs # API service
│   ├── EditPatchService.cs  # Edit patch service
│   ├── FileParserService.cs # File parsing
│   ├── McpManagerService.cs # MCP management
│   ├── OcrService.cs        # OCR service
│   ├── RagService.cs        # RAG service
│   ├── SkillService.cs      # Skills service
│   └── WebSearchService.cs  # Search service
├── View/                    # UI views
│   └── DeepSeekChatControl* # Chat window control (WebView2)
├── CodeCompletion/          # Code completion
│   ├── GhostTextTagger.cs   # Ghost text tagger
│   └── InlinePredictionManager.cs  # Inline prediction management
├── Commands/                # VS commands
├── Settings/                # Options pages
├── ToolWindows/             # Tool windows
└── Utils/                   # Utilities

Debugging

1. Open .slnx in VS 2022
2. Set to Debug configuration
3. F5 to launch Experimental Instance
4. In the experimental instance, open/create a project → View → Other Windows → DeepSeek Chat

Testing

This extension includes 86 xUnit tests covering core paths such as model serialization, patch parsing, context management, and API streaming responses.

Running Tests

# Run all tests
dotnet test DeepSeek_v4_for_VisualStudio.Tests\DeepSeek_v4_for_VisualStudio.Tests.csproj

# With coverage report
dotnet test DeepSeek_v4_for_VisualStudio.Tests\DeepSeek_v4_for_VisualStudio.Tests.csproj `
    /p:CollectCoverage=true /p:CoverletOutputFormat=cobertura

Test Tech Stack

Component	Version	Purpose
xUnit	2.9.x	Test framework
Moq	4.20.x	Mock framework
FluentAssertions	6.12.x	Assertion library
coverlet	6.0.x	Code coverage

Test Structure

DeepSeek_v4_for_VisualStudio.Tests/
├── Unit/
│   ├── Models/       # Serialization, enums, tool call parsing
│   ├── Services/     # Patch parsing, 4-level matching, context management
│   └── Utils/        # String extensions
├── Integration/       # API streaming, persistence, agent dispatch
├── TestData/          # Test JSON/skill files
└── Fixtures/          # DI container fixture

---

FAQ

Q: Why is the chat window blank?

Make sure the WebView2 Runtime is installed. VS 2022 usually includes it, but if missing, download it from developer.microsoft.com/microsoft-edge/webview2.

Q: API call fails with 401?

Check that your API Key is correct: Tools → Options → DeepSeek Chat → API Key. Make sure the key comes from platform.deepseek.com and your account has sufficient balance.

Q: OCR Chinese recognition is inaccurate?

Switch the OCR engine to PaddleOCR-Sharp (Tools → Options → DeepSeek Chat → OCR Engine). The ChineseV5 model will be automatically downloaded on first use.

Q: Baidu search is not working?

Baidu Qianfan requires an API Key (configured in the Options page under the Search category). If you don't have a Baidu key, you can switch to DuckDuckGo (completely free), though it may be slower to access from within China.

Q: How do I add a custom Skill?

Create a .github/skills/ folder in your project root, and add a SKILL.md file (see Skills System for format). Restart the chat window for it to be discovered.

Q: Does this extension conflict with GitHub Copilot?

No conflict. This extension's Ghost Text completion is independent of GitHub Copilot and can be used alongside it. To disable this extension's completion, uncheck "Copilot Enable" in the options page.

---

🗺️ Roadmap / TODO

The following features have reserved interfaces or infrastructure in the architecture and are planned or under development:

🔍 RAG Retrieval-Augmented Generation

Current status: IRagProvider interface and RagService registration/caching infrastructure are in place, but no built-in provider implementation yet.

Planned Item	Description	Priority
Built-in Local Vector DB Provider	SQLite + local embedding model (e.g., all-MiniLM-L6-v2) for out-of-the-box project-level code indexing and semantic search	🔴 High
Automatic File Indexing	Auto-scan code files and build vector index on project open, no manual configuration needed	🔴 High
Embedding Model Config UI	Embedding model selection in options page (local / API), with DeepSeek Embedding API support	🟡 Medium
Hybrid Retrieval Strategy	BM25 keyword + vector semantic hybrid retrieval for improved code search accuracy	🟡 Medium
Incremental Index Updates	Auto incremental index update on file changes, avoiding full rebuild	🟢 Low

🧪 Testing & Quality

Planned Item	Description	Priority
Unit Test Generation Skill	Based on the tdd skill, auto-generate xUnit unit tests for selected code	🟡 Medium
Integration Test Expansion	Increase integration test coverage for Agent Handoff flows and MCP tool call chains	🟡 Medium
UI Automation Tests	Automated regression tests for the WebView2 chat window	🟢 Low

🔧 Tools & Integration

Planned Item	Description	Priority
Enhanced Git Integration	PR description generation, auto commit message writing, code review comments	🟡 Medium
Solution-Level Code Indexing	Cross-project symbol index and reference tracking for large solutions	🟡 Medium
Custom MCP Server Templates	One-click deployment templates for common MCP servers (e.g., database query, API docs)	🟢 Low
Local Model Support	Support offline use via Ollama / LM Studio and other local inference backends	🟢 Low

🎨 User Experience

Planned Item	Description	Priority
Code Lens	In-editor AI action entry points (explain code, generate tests, find references)	🟢 Low
More Built-in Skills	Professional workflows like debug-analyzer, api-designer, sql-optimizer	🟡 Medium
Session Export	Export conversations as Markdown / PDF for sharing and archiving	🟢 Low

💡 Feature suggestions and code contributions are welcome via Issues!

---

Acknowledgments

• DeepSeek — Powerful AI model support
• PaddleOCR — Excellent OCR engine
• Markdig — Fast Markdown parser

---

License

This project is open source under the MIT License.

---