Breaking

GitHub repo dissects Claude Code internals from leaked source maps

July 2, 2026 at 08:19 EDT

AI Agents
Software Dev & Coding
Open Source

A GitHub repository that reverse-engineers the internals of Anthropic's AI coding agent "Claude Code" from leaked source maps has drawn attention among developers. It offers a detailed look at the design of a tool that autonomously understands, edits and executes code from the terminal.

June 29, 2026 · Claude Code, rebuilt from scratch

Inside the minimal loop that lets an AI write its own code

A from-scratch re-implementation strips Claude Code down to the "boring but important parts" — tool calls, permissions, memory and a single-threaded master loop — after an accidental source leak made its internals public.

80%+

of Anthropic's own internal code written by Claude

~500k

lines of TypeScript exposed via accidental source maps

~2,300

source files leaked in npm package v2.1.88

~$13

rough cost per developer, per active day

The master loop ("nO")

A single-threaded core cycle — one permitted action at a time, then review.

OBSERVE

read state

→

CHOOSE

one permitted action

→

EXECUTE

run + summarize

→

REVIEW

loop back ↺

Parallelism is controlled through limited spawning of sub-agents; permissions, memory and real-time steering are integrated into the same loop.

What works

Called by some "the best coding agent"
~70% of usage runs through /feature sub-agent workflows
Strong context recovery across sessions
Tool, permission & memory parts prized over the UI

The gaps

Sessions/context lost on crashes, drops or restarts
Reliability gap — human review still needed often
Practical value of multi-agent Teams questioned
New "agentjacking" security risk via fake bug reports

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.