sanbuphy/learn-coding-agent

This repo is a study resource compiled from publicly available discussions and references about Claude Code, Anthropic's command-line AI coding agent. It does not contain the Claude Code source code. Instead, the author gathered publicly shared technical observations about how the tool works internally and organized them into analysis documents available in four languages: English, Japanese, Korean, and Chinese. The main content is a set of five reports covering telemetry collection, internal codenames, a mode that hides AI authorship in open-source repositories, remote configuration controls, and details about upcoming features. The telemetry report explains what usage data the tool sends back, noting there is no accessible opt-out for first-party logging. The codename report documents internal project names such as Capybara, Tengu, and Numbat. The remote-control report describes how the application polls a server for settings updates every hour, and how rejecting certain changes causes the application to exit. The repo also includes an architecture overview of the agent loop at the core of the tool: the user sends a message, the AI returns a response, and if the model requests a tool action the system executes it and loops back. The README notes this loop sits under additional production layers handling permissions, streaming, concurrency, sub-agents, and persistent state. A statistics section covers the analyzed codebase: roughly 1,884 TypeScript files, over 512,000 lines of code, about 40 built-in tools, and over 80 slash commands. A directory reference traces the internal source structure by filename and function. The repository carries a non-commercial research disclaimer throughout, prohibiting commercial or unauthorized use. The full README is longer than what was shown.