This repository provides a comprehensive reverse-engineering study of Claude Code v1.0.33, aimed at researchers and developers seeking to understand modern AI agent system design. It details the core architecture, mechanisms, and implementation logic of a sophisticated AI programming assistant, offering a valuable technical reference for building and analyzing similar systems.

How It Works

The project meticulously analyzes over 50,000 lines of obfuscated code, revealing a layered architecture. Key innovations include a "h2A" asynchronous message queue for zero-latency, high-throughput communication, a multi-agent system with distinct roles (main, sub, task agents) for modularity and isolation, intelligent context management with aggressive compression (92% threshold) and tiered storage, and a robust 6-layer security framework encompassing UI to tool execution.

Quick Start & Requirements

• Analysis Workspace: The primary analysis workspace is located in claude_code_v_1.0.33/stage1_analysis_workspace/.
• Prerequisites: Requires Node.js for analysis scripts (e.g., beautify.js, split.js). Analysis results are presented in Markdown and JSON formats.
• Setup: Primarily involves navigating and reading the provided documentation and code analysis files.

Highlighted Details

• Real-time Steering: Achieved via a "h2A" dual-buffer asynchronous message queue with >10,000 msgs/sec throughput and Promise-based async iterators.
• Multi-Agent Architecture: Features a main "nO" loop engine, isolated "I2A" sub-agents, and concurrent "Task Agents" with permission isolation.
• Context Management: Employs a "wU2" compressor with a 92% auto-trigger threshold for context compression, retaining key information.
• Security: Implements a 6-layer permission validation chain from UI to tool execution, with sandbox isolation for tool environments.

Maintenance & Community

The project is maintained by ShareAI-Lab. Further details and related projects, including an open-source reconstruction, can be found at https://github.com/shareAI-lab/AgentKode. Updates and analysis articles are also published on the ShareAI Lab official WeChat public account.

Licensing & Compatibility

The project is licensed under the Apache License Version 2.0. This license is generally permissive and compatible with commercial use and closed-source linking.

Limitations & Caveats

The analysis is explicitly stated as "not 100% accurate" due to the difficulty of reverse-engineering obfuscated code, with potential for "hallucinations." The findings are intended for reference and learning purposes only.