Tengu Decoded — A deep dive into the internals of…

Pitch

Description

Comments

About this project

Tengu Decoded explores the internals of Anthropic's Claude Code, revealing how feature flags, telemetry, and device fingerprinting work behind the scenes. Analyzing binaries version by version, this repository serves as a comprehensive research archive, ensuring insights are reproducible and grounded in real code.

Tengu Decoded is a comprehensive exploration of Anthropic's Claude Code, focused on its internal mechanisms such as feature flags, telemetry, device fingerprinting, and the underlying infrastructure, analyzed version by version. The repository serves as a detailed documentation of the functionalities contained within the publicly distributed Claude Code binaries.

The nickname "Tengu" signifies Anthropic's internal label for Claude Code. This repository reveals what each version of Claude Code does, employing various techniques for reverse-engineering.

Key Features

In-Depth Analysis: Investigates various aspects of the Claude Code, including runtime changes, telemetry logging, and the management of feature flags.
Version History: Covers detailed examinations of notable versions:
- Version 2.1.169 (Bun-compiled standalone executable) includes enhanced methodologies for extracting JavaScript source code as cleartext. This results in around 662,000 lines of readable JavaScript, derived directly from real function bodies, allowing for a complete and reproducible research archive.
- Version 2.1.32 analyzed using traditional Unix tools, providing insight into the initial states of the software before the enhancements brought by version 2.1.169.

Example Findings for Version 2.1.169

Runtime Changes: Transitioned to Bun v1.3.14, allowing JS source code to be recovered as plaintext.
Feature Exploration: Notably increased to 218 feature flags, with significant growth in telemetry events to 1086, and around 46 built-in tools.
Security Enhancements: Removed the previous regex injection pipeline in favor of a more sophisticated LLM prefix-classifier while implementing multiple layers of sandboxing for improved security.
New Functionalities: Introduced a background-agent daemon for task management and a managed-agents API for improved session handling.

The repository also houses structured datasets, documentation on methodologies, and a variety of reports that detail the inner workings of Claude Code for each version analyzed. This archive is a vital resource for those interested in understanding the complexities of Claude Code and Anthropic's coding tools.

For more insights and detailed reports, visit relevant directories such as versions/2.1.169/ and versions/2.1.32/ within the repository.