Claude Code from Source

Claude Code from Source is a technical deep dive that reverse-engineers a widely used AI coding agent by reading its npm source maps and distilling architecture and design choices into 18 focused chapters. It explains the end-to-end agent architecture (six core abstractions), the async-generator-driven agent loop for streaming and error recovery, a 14-step tool-execution pipeline, and multi-agent orchestration techniques like fork agents and prompt-cache sharing that cut input-token costs by ~95%. It also covers practical systems concerns: file-based memory with an LLM-powered “Sonnet” recall that outperforms keyword/embedding search, four-layer context compression, two-phase skill loading with lifecycle hooks and frozen config snapshots for injection resistance, and UI, connectivity, and startup optimizations (240ms cold start, slot reservation behavior, bitmap prefilters).

The work distills ten concrete architectural patterns - e.g., speculative tool execution, concurrent-safe batching, slot reservation escalation, and prompt-prefix sharing - and provides actionable “Apply This” guidance for engineers and technical leaders. The material was produced by extracting original TypeScript from shipped source maps, then synthesized into narrative pseudocode by an automated multi-agent process; no verbatim proprietary code is reproduced. The presentation targets engineers building agentic systems, architects evaluating designs, and anyone curious about how production AI tools are constructed.