CodeRun AI Agent is a VS Code extension that uses Qwen 3.7 through the browser API (chat.qwen.ai) as an autonomous coding agent. It operates via a Think → Plan → Act → Verify loop, reading/writing/editing files, running terminal commands, and searching code — all by embedding tool calls as text JSON blocks in the conversation.

🌟 How It Works

Unlike traditional AI extensions that use native function-calling APIs, CodeRun communicates with Qwen through the browser chat API — the same API the chat.qwen.ai website uses. Tool calls are embedded as json {"tool_calls":[...]} blocks in Qwen's response text, then parsed and executed by the extension.

User Prompt → Extension → Qwen Browser API → Qwen responds with text
  → Extension parses ```json {tool_calls}``` blocks
  → Executes tools (read_file, write_file, run_terminal, etc.)
  → Sends results back to Qwen as next user message
  → Loop continues until task is complete

🔑 Features

Feature	Description
Agent Loop	Multi-iteration Think → Plan → Act → Verify loop with up to 20 iterations
19 Tools	File I/O, directory ops, search, terminal, web requests, planning
Terminal Integration	VS Code shell integration with ANSI cleaning, streaming output, auto shell detection
File Checkpoints	SQLite-backed snapshots before every file modification with one-click undo
Diff Review	Inline Accept/Reject buttons for proposed file changes in the chat UI
Permission System	Per-tool Allow/Deny with persistent Always Allow/Always Deny decisions
Thinking Blocks	Collapsible thought process cards with accumulated reasoning across iterations
Tool Cards	Collapsible execution cards with status, args, and results inline in chat
History Restoration	Full conversation history loaded from Qwen's API with tool call/result reconstruction
Custom Context	Full conversation serialized into every request — own context management

🛠️ Available Tools (19)

Category	Tools
File	`read_file`, `write_file`, `edit_file`, `delete_file`, `patch_file`
Directory	`create_folder`, `delete_folder`, `list_directory`
Search	`search_files`, `find_in_files`, `list_symbols`, `get_file_info`
Terminal	`run_terminal`, `terminal_input`, `stop_terminal`
Web	`web_request`
Planning	`create_plan`, `update_plan`
Utility	`get_current_datetime`

All tools accept multiple parameter name variants (e.g., file_path, filepath, path) for robustness.

🔌 Authentication

Uses cookie-based authentication from chat.qwen.ai. Paste your browser cookie string (containing the token=eyJ... JWT) in the settings panel. The cookie is stored in VS Code's encrypted secrets storage.

📁 Project Structure

src/
├── extension.js           ← VS Code activation, IPC bridge, auth, 30+ message types
├── agentLoop.js           ← Core agent loop: stream → parse → execute → loop
├── promptBuilder.js       ← System prompt assembly with workspace context
├── providerQwen.js        ← Qwen Browser API client + context serialization
├── providerManager.js     ← Returns Qwen provider
│
├── tools.js               ← 18 tool async generators (file, dir, terminal, search, etc.)
├── toolDefinitions.js     ← JSON schemas for all tools
├── toolRegistry.js        ← Name→impl mapping with alias resolution (bash, read, etc.)
├── toolExecutor.js        ← Result formatting for LLM context
│
├── terminalManager.js     ← VS Code Terminal API, shell integration, child_process fallback
├── permissions.js         ← Permission system with persistent decisions
├── diffManager.js         ← Diff patch storage, apply/reject
├── checkpointManager.js   ← SQLite file snapshots for undo
│
├── projectKnowledge.js    ← SQLite knowledge base (sql.js)
├── searchManager.js       ← Search abstraction (SQLite → filesystem fallback)
├── contextManager.js      ← Intent classification, editor context
├── planningManager.js     ← Plan creation and step tracking
├── executionManager.js    ← Plan execution engine
├── verificationManager.js ← Post-execution verification
├── learningManager.js     ← Framework/convention detection
├── timelineManager.js     ← Chronological event log
│
├── ChatSpace.js/.css      ← Chat UI: tool cards, diff cards, terminal cards, thinking blocks
├── Dashboard.js/.css      ← Sidebar, settings, conversation list
├── MarkdownRenderer.js    ← Client-side markdown → HTML
│
├── constants.js           ← System prompt, event types, storage keys
├── config.js              ← VS Code settings reader, API key management
├── utils.js               ← Shared helpers
├── events.js              ← Simple pub/sub event bus
├── symbolParser.js        ← Multi-language symbol extractor (JS, TS, Python, Go, Rust, Java, etc.)
├── memoryManager.js       ← Simple in-memory store
├── mcpManager.js          ← MCP placeholder stub
├── skillsManager.js       ← Skill prompt fragments
├── workspaceContext.js    ← Workspace folder detection
├── conversationStore.js   ← LocalStorage conversation CRUD
├── settingsStore.js       ← LocalStorage settings persistence
├── agent.js               ← Public agent API wrapper
├── index.html             ← Standalone browser entry point

🚀 Quick Start

Prerequisites

Node.js >= 18.x
VS Code >= 1.80.0

Debug

Open the project in VS Code
Press F5 — the Extension Development Host window opens
Paste your Qwen browser cookie in the login panel

Package

npm install -g @vscode/vsce
vsce package

💻 Tech Stack

Layer	Technology
Extension Host	VS Code Extension API (Node.js)
Frontend	Vanilla JS + CSS (no framework)
Database	SQLite via `sql.js` (WASM)
LLM	Qwen 3.7 Max (browser API)
Auth	Cookie-based (chat.qwen.ai JWT)
Build	None — raw ES modules

This creates a .vsix file that can be installed via:

code --install-extension ai-agent-<version>.vsix

📦 Repository

Find the source code on GitHub:

Repo: github.com/nbsgr/qwen-coderun-agent
Issues: github.com/nbsgr/qwen-coderun-agent/issues
Clone: git clone https://github.com/nbsgr/qwen-coderun-agent.git

License

MIT

Qwen-CodeRun-Agent