Production-grade agentic AI coding assistant powered by Qwen 3.6-27B, DeepSeek V4-Pro, and GLM 5.1.
| Model | Strength | Cost | Context |
|---|---|---|---|
| Qwen 3.6-27B | Fast all-round coder | $0.20/$0.50 per 1M | 128K |
| DeepSeek V4-Pro | Deep reasoning, FIM autocomplete | $0.60/$1.50 per 1M | 1M |
| GLM 5.1 | #1 SWE-bench, hardest tasks | $1.05/$3.50 per 1M | 203K |
- Sidebar Chat with streaming + markdown rendering + syntax highlighting
- Agent Mode — AI autonomously reads, edits, tests, and commits code
- Smart Routing — auto-picks the best model by task complexity
- Image Support — paste screenshots or attach images for visual context
- Context Injection — auto-includes current file, errors, and git diff
- Copilot-style ghost text completions as you type
- Powered by DeepSeek V4 Fill-in-Middle (FIM) for precise code completion
- Falls back to Qwen 3.6 for chat-based completion
| Category | Tools |
|---|---|
| Files | read_file, edit_file (with diff preview), search_files, list_directory |
| Terminal | run_command |
| Diagnostics | get_diagnostics |
| Git | git_status, git_diff, git_log, git_commit, git_push, git_pull, git_branch, git_stash, generate_commit_msg |
| GitHub | github_create_pr, github_create_issue, github_list_prs |
| Testing | run_tests (auto-detect framework), generate_tests |
| Docs | generate_docs |
| HTTP | http_request |
| Packages | install_package (npm/yarn/pip/go) |
- Diff Preview — shows VS Code diff editor before applying agent edits
- Confirmation Dialogs — asks before destructive actions (edit, commit, push)
- "Allow All" Option — skip confirmations for the current session
- Token Budget — auto-trims conversation to fit model context window
- Cost Tracking — real-time cost display in status bar
| Provider | Model | Sign Up |
|---|---|---|
| Together.ai | Qwen 3.6 | https://api.together.xyz |
| DeepSeek | DeepSeek V4-Pro | https://platform.deepseek.com |
| Zhipu AI | GLM 5.1 | https://open.bigmodel.cn |
git clone <your-repo>
cd ai-coder-extension
npm install
npm run compileIn VS Code Settings (Ctrl+,), search "DeepZen AI" and set:
aiCoder.qwenApiKey— Together.ai keyaiCoder.deepseekApiKey— DeepSeek keyaiCoder.glmApiKey— Zhipu AI key (format:id.secret)
Press F5 to launch in development mode.
To install permanently:
npm install -g @vscode/vsce
vsce package
code --install-extension deepzen-ai-1.0.0.vsix| Shortcut | Action |
|---|---|
Enter |
Send chat message |
Ctrl+Enter |
Send as agent task |
Ctrl+V |
Paste screenshot into chat |
Ctrl+Shift+A |
Run Agent Task (command palette) |
Ctrl+Shift+M |
Switch Model |
Ctrl+Shift+N |
New Chat |
Ctrl+Shift+G |
Smart Commit (AI generates message) |
Ctrl+Shift+T |
Generate Tests for current file |
Ctrl+Shift+I |
Toggle inline autocomplete |
Type at the start of your message to force a specific model:
| Command | Routes To | Best For |
|---|---|---|
/qwen or /quick |
Qwen 3.6 | Quick tasks, general coding |
/deepseek or /medium |
DeepSeek V4 | Debugging, refactoring |
/glm or /hard |
GLM 5.1 | Complex tasks, architecture |
- Paste: Ctrl+V to paste screenshots from clipboard
- Attach: Click 📎 to attach image files
- Use Cases: UI bugs, error screenshots, design mockups → code
| Usage | Monthly Cost |
|---|---|
| Light (2-3 hrs/day) | ~$10-12 |
| Medium (4-6 hrs/day) | ~$22-25 |
| Heavy (8+ hrs/day) | ~$40-55 |
src/
├── extension.ts # Entry point, registers all commands
├── api/
│ ├── CloudClient.ts # API client (JWT auth, retry, caching, FIM, vision)
│ └── SmartRouter.ts # Complexity-based model routing
├── chat/
│ └── ChatViewProvider.ts # Sidebar UI (markdown, images, streaming)
├── agent/
│ ├── AgentEngine.ts # Agentic loop (23 tools, context injection)
│ └── tools/ # All tool implementations
│ ├── readFile.ts
│ ├── editFile.ts # With VS Code diff preview
│ ├── terminalExec.ts
│ ├── searchFiles.ts
│ ├── listDirectory.ts
│ ├── diagnostics.ts
│ ├── git.ts # 9 git tools + AI commit messages
│ ├── github.ts # PR & issue creation via gh CLI
│ ├── testing.ts # Auto-detect test framework
│ └── docsHttpPkg.ts # Docs, HTTP, package management
├── completion/
│ └── InlineCompletionProvider.ts # Copilot-style ghost text
└── context/
└── ContextGatherer.ts # Auto context (file, errors, git, structure)
MIT