For all of the speak about frontier AI, it’s humorous how usually the actually massive reveals nonetheless come from the software program equal of leaving your keys within the door.
On March 31, 2026, builders observed {that a} printed npm package deal for Anthropic’s Claude Code appeared to incorporate a supply map pointing to a downloadable archive of the device’s unobfuscated TypeScript supply. The discovery was first amplified by Chaofan Shou on X, then rapidly mirrored throughout GitHub, Reddit, and the broader AI-dev web.
Within hours, what had been a closed business coding agent was being picked aside in public like a brand new iPhone on teardown day.
The apparent story is that Anthropic had a foul morning. The extra fascinating story is what the leak seems to disclose: Claude Code is not only “Claude, but in a terminal.” It seems a lot nearer to an working system for software program work, with permissions, reminiscence layers, background duties, IDE bridges, MCP plumbing, and multi-agent orchestration all stacked across the mannequin.
That issues as a result of the AI coding race is now not nearly who has the neatest mannequin. It’s about who has one of the best harness.
Not a chatbot wrapper
According to the general public mirror at nirholas/claude-code, the leaked Claude Code CLI spans roughly 1,900 recordsdata and greater than 512,000 strains of code in strict TypeScript, constructed with Bun, React, and Ink for the terminal UI.
The repo’s structure docs describe a sprawling system: an enormous QueryEngine, a centralized device registry, dozens of slash instructions, persistent reminiscence, IDE bridge help, MCP integrations, distant periods, plugins, expertise, and a activity layer for background and parallel work. In different phrases: not a chatbot wrapper. A product stack.
That’s the true leak.
We’ve spent the final yr watching coding brokers get pitched as if the magic have been all within the mannequin. Make the mannequin higher, give it a shell, sprinkle in a code editor, and voilà: software program engineer. But the Claude Code mirror suggests the exhausting half is every thing across the mannequin. Tool permissions. Context loading. State administration. Session restoration. Memory hygiene. Background jobs. Team coordination. The stuff that sounds boring proper up till your agent nukes a repo, loses the thread, or hallucinates its method right into a refactor from hell.
The reminiscence story is the true story
And if one theme retains surfacing from builders analyzing the leaked code, it’s reminiscence.
One of the sharpest breakdowns got here from Himanshu on X, who argued that Claude Code’s reminiscence system seems designed much less like a large pocket book and extra like a constrained retrieval structure. The key thought is deceptively easy: reminiscence is handled as an index, not storage.
A light-weight MEMORY.md or CLAUDE.md-style layer stays resident, but it surely largely factors to information moderately than making an attempt to include it. Topic recordsdata are fetched on demand. Full transcripts aren’t hauled again into context wholesale. If one thing is derivable from the codebase, it usually shouldn’t be saved in any respect.
That’s a refined however necessary design alternative. Most individuals think about “agent memory” as an even bigger backpack. Claude Code’s reminiscence, a minimum of from the general public evaluation, seems extra like a submitting system with a strict librarian.
The remainder of the sample is much more revealing. The reminiscence system described by analysts seems to be bandwidth-aware, skeptical, and aggressively self-editing. It doesn’t simply append extra notes ceaselessly. It rewrites. It deduplicates. It prunes contradictions. It treats stale reminiscence as a legal responsibility, not an asset.
And crucially, it appears to separate reminiscence consolidation from the primary agent context, which is precisely the type of element you solely construct after studying the exhausting method that autonomous…
