Inside Claude Code: How a Local AI Agent OS Was Reverse‑Engineered

Recently a GitHub project called shareAI‑lab published a thorough reverse‑engineering of Anthropic’s Claude Code, including all research data and analysis notes.

Claude Code is Anthropic’s flagship AI coding product, but its core logic is hidden in a heavily obfuscated 50 k+ line JavaScript bundle packaged with the CLI. The obfuscation deliberately renames and encrypts code to hide algorithms and prompts.

Because the JavaScript must run locally, the obfuscation can be undone, providing an entry point for reverse‑engineers. The shareAI‑lab team used Claude Code itself to analyze its own obfuscated output, slicing the 50 k lines into 15 chunks, then manually debugging and filling gaps to reconstruct a 95 % accurate inferred architecture .

https://github.com/shareAI-lab/analysis_claude_code

The reverse‑engineered architecture is organized into several layers:

User Interaction Layer

All entry points—CLI, VSCode plugin, or web UI—feed commands into a unified scheduler that encodes requests for the Claude model brain.

Agent Core Scheduling Layer

The central Agent Loop (nO) manages every agent’s behavior. It decides whether a new task is needed, which tools to invoke, which agents to awaken, how to compress history, and whether error recovery is required. Decisions are dispatched via the h2A message queue (asynchronous streaming) and the wu session flow generator , with the wU2 compression engine optimizing context.

The model itself is only a tool called by this scheduling engine; the real coordination happens in the engine and runtime logic.

Tool Execution & Management Layer

This “middle platform” schedules sub‑agents for specific functions (e.g., shell execution, file access). Core components include:

MH1 Tool Engine : discovers tools, validates parameters, assigns tasks.

UH1 Concurrency Scheduler : limits concurrency and prevents resource contention.

SubAgent Manager : isolates sub‑tasks into independent agents.

Permission Gateway : checks whether an agent may run a command, access a file, or use network resources.

Each request generates a dedicated sub‑agent, which operates under strict permission and state controls.

Tool Ecosystem

Claude Code ships with hundreds of modular tools (file I/O, command execution, web search, task management, performance diagnostics, etc.). Tools are defined as YAML files that can be hot‑loaded, audited, and granted permissions dynamically.

Storage & Persistence System

The bottom layer provides memory through a three‑tier architecture:

Current session stored in Messages for real‑time interaction.

Mid‑term summaries stored in Compressed via the wU2 compressor.

Permanent preferences saved in CLAUDE.md (languages, project structure, tool preferences).

System state cached in StateCache (tool usage counts, error flags, permission status).

This local state replaces cloud‑based memory, enabling a “human‑like” recall without external servers.

Real‑Time Steering

Unlike typical trigger‑based AI tools, Claude Code’s h2A queue uses a “double‑buffer + conditional consumption” mechanism, allowing processing to start as soon as input begins. The pseudo‑code below illustrates the logic:

class h2AAsyncMessageQueue{  enqueue(message){  // Zero‑delay path  if(this.readResolve){    this.readResolve({done:false,value:message});    this.readResolve=null;    return;  }  // Buffer path  this.primaryBuffer.push(message);  this.processBackpressure(); }}

If a consumer is waiting, the message is delivered immediately; otherwise it is buffered with back‑pressure control, enabling Claude to generate text, adjust tasks, and respond to new input simultaneously.

Intelligent Context Compression

Claude Code’s wU2 compressor activates when token usage exceeds 92 % of the context limit, preserving only the most important 30 % of messages and summarizing the rest:

// Compression trigger  if(tokenUsage > CONTEXT_THRESHOLD * 0.92){    const compressedContext = await wU2Compressor.compress({      messages: currentContext,      preserveRatio: 0.3,      importanceScoring: true    });}

This importance‑weighted approach keeps critical information while discarding redundant data, maintaining memory fidelity over long interactions.

Overall, the reverse‑engineered documentation shows that Claude Code is not merely a smart autocomplete tool but a locally‑run, distributed agent operating system with real‑time steering, sophisticated context management, and a rich, auditable tool ecosystem.

Future AI coding assistants are expected to evolve from simple extensions into stable, secure, and organized intelligent‑agent platforms.