BREAKING NEWS + ANALYSIS
Anthropic’s Accidental Code Leak Just Revealed Claude’s Secret Roadmap: Tamagotchi Pets, Dream Mode, and Multi-Agent Teams
512,000 lines of source code. 44 hidden feature flags. 20+ unshipped features. Here’s what Anthropic didn’t want you to see yet — and why it matters for anyone using Claude.
|
15 min read
|
Updated with Anthropic’s official response
What Happened in 30 Seconds
On March 31, 2026, Anthropic accidentally shipped a debugging file (source map) inside a public npm package for Claude Code. That file contained the entire 512,000-line TypeScript source code — including 44 hidden feature flags covering features that are fully built but not yet released. Within hours, the code was mirrored across GitHub and analyzed by thousands of developers worldwide. This wasn’t a hack. It was a packaging mistake. But what it revealed is a fascinating look at where Claude is heading next.
What’s in This Article
1. How the Leak Happened (It’s Embarrassingly Simple)
2. What Was Actually Inside: The Architecture
3. Hidden Feature #1: BUDDY — A Tamagotchi AI Pet System
4. Hidden Feature #2: KAIROS — Always-On Autonomous Mode
5. Hidden Feature #3: ULTRAPLAN & BRIDGE MODE — Multi-Agent Teams
6. The “Self-Healing Memory” System
7. Other Discoveries Worth Knowing
8. What This Means for Claude Users
1. How the Leak Happened (It’s Embarrassingly Simple)
1. How the Leak Happened (It’s Embarrassingly Simple)This wasn’t a sophisticated cyberattack. It wasn’t a disgruntled employee. It was someone forgetting to exclude a debugging file when publishing a software update.
When developers build JavaScript/TypeScript applications, the build process generates source map files (.map files). These files map the compressed, unreadable production code back to the original, readable source code — they’re essential for debugging but should never be included in public packages.
When Anthropic published version 2.1.88 of the @anthropic-ai/claude-code package to the npm registry, a 59.8 MB source map file was accidentally bundled in. That file pointed directly to a publicly accessible ZIP archive sitting on Anthropic’s own Cloudflare R2 storage — containing the complete, unobfuscated TypeScript source code.
At 4:23 AM ET, security researcher Chaofan Shou spotted the exposure and posted it on X. Within hours, the entire 512,000-line codebase was mirrored across GitHub, racking up tens of thousands of forks before Anthropic’s DMCA takedowns began. Hacker News, Reddit, and developer communities lit up with analysis.
Anthropic’s official response
“This was a release packaging issue caused by human error, not a security breach. No customer data or credentials were involved or exposed. We’re rolling out measures to prevent this from happening again.”
The irony that several analysts have pointed out: buried in the leaked code is an entire subsystem called “Undercover Mode” — specifically designed to prevent Claude from accidentally revealing internal codenames in git commits. Anthropic built a whole system to prevent AI leaks… then shipped the entire source code in a .map file.
This is also the second time this has happened. A nearly identical source map leak occurred with an earlier version of Claude Code back in February 2025, making this a repeat incident roughly 13 months later.
2. What Was Actually Inside: The Architecture
Before diving into the hidden features, it’s worth appreciating the sheer scale of what was exposed. Claude Code is not a “wrapper around an API.” It’s a full-blown software engineering platform.
By the Numbers
~1,900 files — covering application logic, infrastructure, tooling, documentation, and internal utilities.
512,000+ lines of TypeScript source code.
40+ built-in tools — each a discrete, permission-gated capability (file reading, bash execution, web fetch, LSP integration, etc.). The base tool definition alone spans 29,000 lines.
QueryEngine.ts — 46,000 lines — the largest single module. It handles all LLM API calls, streaming, caching, and orchestration. This is the brain of the operation.
44 feature flags — covering features that are fully built but not yet shipped to users. They compile to false in the external build.
Tech stack: Bun runtime (not Node.js), React + Ink for terminal UI rendering, Zod v4 for schema validation.
Now let’s get to the part everyone cares about: what’s coming next.
3. Hidden Feature #1: BUDDY — A Tamagotchi AI Pet System
Expected release: May 2026 (teasers in April)
Status: Fully built. Hidden behind compile-time flags in the external build.
Yes, you’re reading this correctly. Buried in the code is a complete virtual pet system — Tamagotchi-style — that lives in your terminal next to the Claude Code input prompt.
BUDDY is a speech-bubble companion that appears alongside your terminal prompt. It reacts to your coding activity, gains experience, and evolves. The system is surprisingly deep:
18 creatures — including Duck, Dragon, Axolotl, Capybara, Ghost, and more. Each has unique personality traits and dialogue lines.
Rarity tiers — Common, Uncommon, Rare, Epic, and Legendary (1% drop rate). It’s a gacha-style collection system built right into a developer tool.
5 stat attributes — DEBUGGING, PATIENCE, CHAOS, WISDOM, and SNARK. Stats grow based on your coding behavior — long debugging sessions boost your pet’s PATIENCE, for example.
Reactive behavior — Your Buddy reacts to what you’re doing. Long build times might trigger a bored animation. Successful deployments get celebratory responses.
This might sound frivolous, but it’s actually a clever retention mechanism. Developers who spend hours in the terminal get a persistent companion that makes the experience feel less isolating. GitHub did something similar with their Achievement badges — gamification that drives engagement without changing core functionality. Anthropic is applying the same psychology to their CLI.
4. Hidden Feature #2: KAIROS — Always-On Autonomous Mode
Status: Hidden behind PROACTIVE / KAIROS compile-time flags
Referenced over 150 times in the source code. No public release date announced.
This is the big one. KAIROS (Ancient Greek for “at the right time”) is a fundamental shift in how Claude Code would work — from a reactive tool you ask questions to, into an always-on background agent that thinks, plans, and acts on its own.
According to VentureBeat’s analysis of the leaked code, KAIROS represents what they call an “autonomous daemon mode.” Here’s what the code reveals:
Cross-Session Memory
Currently, each Claude Code session starts fresh — the AI doesn’t remember what happened yesterday. KAIROS would maintain persistent memory across sessions, building a continuously growing understanding of your project, your patterns, and your preferences.
“Dreaming” Mode (autoDream)
When you’re not actively using Claude, KAIROS would perform “memory consolidation” in the background — merging observations, resolving contradictions, and converting vague insights into concrete facts. Think of it as Claude organizing its own notes while you sleep. The code includes logic for the agent to clean up its own knowledge base, remove outdated information, and prepare for the next work session.
Proactive Execution
Instead of waiting for you to ask, KAIROS would monitor your behavior patterns and take action preemptively. Detect a failing test? Fix it before you notice. See a pattern of similar bugs? Refactor the underlying cause. Spot a dependency update? Prepare a migration plan. The system is designed to act like a proactive team member, not a reactive assistant.
If this ships in anything close to what the code describes, it would be the most significant upgrade to Claude Code since its launch. It’s essentially the difference between having a tool and having a colleague.
Why this matters beyond developers
KAIROS-style “always-on” AI agents represent where the entire industry is heading. If Anthropic ships this for Claude Code, expect the same pattern to appear in Cowork (for non-developers), Claude in Excel, and other products. The concept of AI that works while you’re away — organizing, planning, preparing — is the next major paradigm shift after chat-based AI.
5. Hidden Feature #3: ULTRAPLAN & BRIDGE MODE — Multi-Agent Teams
Two more hidden features reveal Anthropic’s ambition to move beyond single-agent AI into coordinated multi-agent systems.
HIDDEN FEATURE
ULTRAPLAN — Remote Solo Planning
A cloud-based mode where Claude runs an extended planning session entirely on its own — up to 30 minutes of autonomous thinking — without any user input. You’d describe the goal (“plan the migration to a new database architecture”), walk away, and come back to a complete, structured plan. This is significant because current AI interactions are limited to real-time, back-and-forth conversations. ULTRAPLAN breaks that constraint entirely.
HIDDEN FEATURE
BRIDGE MODE — Multi-Agent Coordinator
Multiple Claude instances working together as a team, each handling different aspects of a project simultaneously. One instance writes code, another reviews it, a third writes tests, and a coordinator manages the workflow. This is the multi-agent orchestration that the AI industry has been promising but rarely delivers in practice. The code suggests Anthropic has built the infrastructure for real, production-grade agent teamwork — not just demo-quality prototypes.
The combination of KAIROS + ULTRAPLAN + BRIDGE MODE paints a picture of where Claude Code is headed: an autonomous AI team that works around the clock, plans independently, and coordinates multiple specialized agents. If that sounds like science fiction, remember — the code is already written. It’s just hidden behind feature flags.
6. The “Self-Healing Memory” System
Beyond upcoming features, the leak revealed how Claude Code’s existing memory architecture works — and it’s more sophisticated than anyone expected.
The biggest challenge for AI coding agents is “context entropy” — as conversations get longer and projects get more complex, the AI gradually loses track of what’s happening and starts making mistakes. Anthropic’s solution is a three-layer memory architecture that VentureBeat describes as “self-healing”:
Layer 1: MEMORY.md (Always loaded) — A lightweight index of short pointers (~150 characters per line). This file doesn’t store actual data — it stores locations. Think of it as a table of contents that’s always in Claude’s context window, pointing to where detailed knowledge lives.
Layer 2: Topic files (On-demand) — Actual project knowledge is distributed across separate topic-specific files. Claude only loads the relevant file when it needs that information — not everything all at once. This prevents the context window from getting bloated with irrelevant information.
Layer 3: Raw transcripts (Searchable, never fully loaded) — Past conversation transcripts are stored but never read back into the context in full. Instead, Claude searches them (grep) for specific identifiers when needed. This means the AI can reference past conversations without wasting context on complete chat histories.
The critical design rule: Claude is instructed to treat its own memory as a “hint” — it must verify facts against the actual codebase before acting on what it “remembers.” This skeptical approach to its own knowledge prevents the model from confidently repeating outdated or incorrect information.
Practical takeaway for Claude Code users
This is why your CLAUDE.md file matters so much. It’s the core of Layer 1 — the index that Claude always has in its context. Keep it concise (pointers, not paragraphs), update it regularly, and let it serve as the entry point for Claude to find what it needs. This is the engineering principle behind why the harness engineering community emphasizes small, focused instruction files over massive documents.
7. Other Discoveries Worth Knowing
Beyond the headline features, developers combing through the code have found several other notable details:
The system prompt lives in the CLI, not on the server. This surprised many analysts. Claude Code’s system instructions — the rules governing how it behaves — are bundled into the distributed package rather than being fetched from Anthropic’s servers at runtime. This means anyone with access to the package could always read the instructions, even before this leak.
The Bash tool is the real crown jewel. According to multiple analysis posts, the most sophisticated engineering in the entire codebase isn’t the chat interface or the memory system — it’s how Claude Code handles terminal command execution. The permission system, sandboxing, error recovery, and safety guardrails around bash execution represent thousands of engineering hours.
187 loading spinner verbs. Developers found that Claude Code ships with 187 unique spinner messages that display while it’s processing — things like “Pondering solutions,” “Unraveling complexity,” “Consulting the oracle.” Someone at Anthropic had a lot of fun writing those.
Animal codenames everywhere. Internal projects use names like Tengu, Fennec, and Capybara. Features are called “Penguin Mode” and “Dream System.” The engineering culture that produced these names also produced a Tamagotchi pet system for a CLI tool — which tells you something about how Anthropic approaches product design.
Anthropic-only internal tools. The code includes tools that only load for Anthropic employees, suggesting there’s an entirely separate set of capabilities that internal teams use for testing, debugging, and development that external users never see.
8. What This Means for Claude Users
Whether you’re a developer using Claude Code or a non-technical user on Claude.ai, here’s what to take away from this:
For Claude Code users
Your CLAUDE.md file is more important than you thought. The memory architecture confirmed that this file serves as the always-loaded index for Claude’s understanding of your project. Invest time in keeping it concise, current, and well-organized.
Multi-agent workflows are coming. BRIDGE MODE suggests that Claude Code will eventually support coordinated teams of AI agents. If you’re building complex projects, start thinking about how to structure work that could be parallelized across multiple agents.
Background processing is the future. KAIROS and ULTRAPLAN both point toward Claude doing work while you’re away. Expect to see more “set it and forget it” features in Claude Code — similar to what Cowork already offers with scheduled tasks.
For all Claude users
Autonomous AI is no longer theoretical. The code reveals that Anthropic has already built — not just planned, but built — infrastructure for AI that acts independently, maintains long-term memory, and coordinates with other AI instances. These features will likely trickle down from Claude Code into Cowork and other products.
“Dream Mode” is a paradigm shift. The idea that AI could consolidate its own knowledge during downtime — cleaning up, organizing, and improving — represents a fundamentally different relationship between humans and AI. We’re moving from “tools I use” to “colleagues who prepare for our next interaction.”
For the industry
The engineering bar is extraordinarily high. 512,000 lines of production code, 40+ tools, a three-layer memory system, multi-agent orchestration — this is the infrastructure behind what looks like a “simple” CLI chat tool. It shows that the real competition in AI isn’t about model benchmarks. It’s about the engineering around the model — the harness.
The AI pet surprised everyone. In an industry obsessed with “productivity” and “enterprise,” Anthropic built a Tamagotchi with gacha mechanics into a developer tool. It signals that the company is thinking about engagement and emotional connection, not just utility. That’s an underrated competitive advantage.
⚠️ Security Note for Claude Code Users
A separate npm supply-chain attack on the axios package occurred hours before this leak. If you installed or updated Claude Code via npm on March 31, 2026 between 00:21 and 03:29 UTC, you may have pulled a malicious version containing a Remote Access Trojan (RAT).
Check immediately: Search your package-lock.json, yarn.lock, or bun.lockb for axios versions 1.14.1 or 0.30.4, or the dependency plain-crypto-js.
Going forward: Anthropic recommends using their native installer over npm to avoid supply-chain risks.
The Bottom Line
Anthropic’s accidental code leak is simultaneously embarrassing and revelatory. The embarrassment is obvious — a $380 billion company shipping debug files in production is a basic packaging mistake. But what the leak revealed about Claude’s roadmap is genuinely exciting.
BUDDY adds personality to a tool that developers spend hours with daily. KAIROS transforms AI from a reactive tool into a proactive colleague. ULTRAPLAN and BRIDGE MODE push toward autonomous, multi-agent systems that work independently. And the self-healing memory architecture shows engineering sophistication that competitors will now study closely.
The code is out there, and it’s not going back. What Anthropic does next — how quickly they ship these features and how they respond to the security questions — will determine whether this leak is remembered as a stumble or as the moment the world got an early look at the future of AI development tools.
Timeline of Events
Feb 2025 — First Claude Code source map leak via npm (nearly identical incident)
Mar 26, 2026 — Anthropic CMS misconfiguration exposes ~3,000 internal files about unreleased “Claude Mythos” model
Mar 31, 2026, ~4:23 AM ET — Researcher Chaofan Shou discovers source map in npm package v2.1.88, posts on X
Within hours — Code mirrored across GitHub; 41,500+ forks before takedowns begin
Same day — Anthropic removes source map, republishes clean version, begins DMCA takedowns
Ongoing — Developer analysis continues across Hacker News, Reddit, X, and dev blogs
⭐ Claude Just Killed the $600 AI Mac Mini Trend (For $20/mo)
