Someone Shipped Claude Code's Source Map to npm. All 500,000 Lines of It.

On March 31, a 59.8 MB source map file accidentally shipped inside @anthropic-ai/claude-code v2.1.88 on npm. It contained roughly 500,000 lines of TypeScript across 1,900 files. The entire internal architecture of Claude Code, exposed. Not through a hack. Through a build artifact nobody caught.

I use Claude Code every day. I've built mental models of how it works based on behavior. Almost all of them were wrong.

Alex Kim's analysis and paddo.dev's architecture breakdown reveal the real system. There's a self-healing memory architecture that actively manages context window constraints, deciding what to keep, what to compress, and what to drop. The tool orchestration system has permission layers and hook events I didn't know existed. There's a query engine that handles multi-provider LLM API routing. Sub-agent spawning for parallel task execution. A bidirectional communication layer between IDE extensions and the CLI.

The part that caught me off guard was the scale. 1,900 files. This isn't a wrapper around an API. It's a full production harness with its own state management, error recovery, and coordination layer. The patterns for how it manages context, specifically, are things I've been trying to figure out for my own agent pipelines.

For anyone building agentic systems, this is now the reference architecture whether Anthropic intended it or not. How they solved context management, how tool safety works in practice, how multi-agent coordination actually functions at production scale. Study the memory system especially. The self-healing pattern, where the agent detects when context is degrading and proactively manages it, is something I haven't seen documented this well anywhere else.

Anthropic hasn't commented beyond pulling the affected version. The code is out there. The community is already learning from it.