What it does This repo compiles publicly available leaks, reverse-engineering notes, and community analysis of Claude Code v2.1.88 into five deep-dive reports—available in English, Chinese, Japanese, and Korean. It maps the codebase structure (~1,884 files, ~512K lines), explains the core agent loop, and documents telemetry, hidden feature flags, and an alleged “undercover mode” for hiding AI authorship in open-source commits.

The interesting bit The reports read like a field guide to a black box. Animal codenames (Capybara, Tengu, Numbat), random-word feature flags (tengu_frond_boric), and a claimed “KAIROS” autonomous mode with heartbeat ticks and push notifications. The authors also flag that rejecting a remote settings update can apparently kill the app entirely.

Key highlights

• Five analysis topics: telemetry, codenames, undercover mode, remote control/killswitches, and future roadmap
• ~40 built-in tools, ~80 slash commands, React/Ink terminal UI
• Core loop: user → messages → Claude API → tool execution → loop back
• Production harness layers: permissions, streaming, concurrency, compaction, sub-agents, MCP
• Quadrilingual docs; explicitly non-commercial (disclaimer prohibits commercial use)

Caveats

• All claims are compiled from “publicly available online references and discussions”—not first-party sources
• The repo contains no actual Claude Code source code; it is analysis and documentation only
• Some findings (e.g., undercover mode, remote killswitches) are unverified and potentially speculative

Verdict Worth a skim if you’re building CLI agents and want to see how a well-funded competitor allegedly structures theirs. Skip if you need actionable code or verified facts; this is reconnaissance, not engineering reference.