Claude Code Source Leak: a Tamagotchi-Style Pet and an Always-On Background Agent

The Claude Code source leak has exposed the inner workings of one of the most talked-about AI development tools. More than 512,000 lines of TypeScript, Anthropic's system prompts and traces of two unannounced features, a Tamagotchi-style virtual pet and an always-on background agent, became publicly accessible after a accidental mistake in update 2.1.88.

It all started with a routine release. After Claude Code version 2.1.88 came out, a user on X noticed that a source map had been included in the package by mistake. This is an auxiliary debugging file that contains the application's original TypeScript code. The user published the file's contents, and the material quickly spread through the developer community. Ars Technica and VentureBeat were among the first to report on the leak.

The data covers more than 512,000 lines of code. That is the tool's complete codebase, enough to study its entire architecture in detail.

Researchers found several layers of information in the code. First, the system instructions: prompts that Anthropic built into Claude Code. They describe how the model should behave when working with code, what style to follow in its answers and what restrictions to observe. Until now, the content of these instructions had remained closed. Second, the memory architecture became clear. It is now known how Claude Code stores context between sessions and draws on it during work, a question that has long interested developers who use the tool daily.

The main finds were two unannounced additions. The first is a virtual pet whose logic resembles the mechanics of classic Tamagotchi: an interactive element tied to the user's activity inside the tool. Anthropic apparently is experimenting with gamification mechanics, which build attachment to a product and encourage regular use.

The second is an always-on agent mode, able to work in the background without explicit commands from the user. This is a fundamentally different model of interacting with AI: not a reactive tool that answers requests, but a proactive assistant that tracks context on its own and acts on its own initiative.

Anthropic has not commented publicly on the incident. In subsequent updates the source map disappeared from the package, so the company quietly closed the leak channel. The fate of the discovered features remains unknown: they may ship in final form, change, or stay internal experiments.

Such incidents happen in the industry, but a leak of this scale, half a million lines from one of the most secretive AI companies, is truly rare. For competitors, it is a chance to study Anthropic's real technical decisions. For security researchers, it is valuable material. For ordinary users, it is a rare chance to look behind the scenes of a tool many use every day without knowing what happens inside.

Claude Code competes with Cursor, GitHub Copilot and Windsurf in the fast-growing market for AI coding assistants. If the always-on background agent ships in final form, it could become a significant differentiator, since most competitors still stick to the reactive model. The Tamagotchi pet, in turn, points to an attempt by Anthropic to add an element of personal attachment to the tool, turning a utility into something people come back to for more than just results.