This repository is a backup of that leaked source, and this README is a full breakdown of what’s in it, how the leak happened and most importantly, the things we now know that were never meant to be public.
I am not making this up.
Claude Code has a full Tamagotchi-style companion pet system called “Buddy.” A deterministic gacha system with species rarity, shiny variants, procedurally generated stats, and a soul description written by Claude on first hatch like OpenClaw.
As usual the real damage of a source leak isn’t other companies cloning and selling your flagship product, which numerous well established legal frameworks exist to counteract, but everyone realizing how bone crushingly stupid you and all the people you’re paying millions to write the product clearly are.
I use this stuff exclusively at work and I like the flexibility of cursor a lot better than CC. When one service goes down i can just switch models and I pretty much get the same results no matter which model because i have a very specific goal in mind every time i reach for these tools.
the worst work that i have to review comes from ppl who totally abdicate responsibility to LLM gen’d stuff and it especially pisses me off when someone who gets paid more than me passes me a PR or a doc that is just straight up incorrect
I really want a flexible CLI tool that is not written in TS. it’s no surprise that this code base is garbage because the models are trained on JS/TS which is the cool uncle of programming languages. I’ve finally had a chance to work with Go at work and it’s pretty great for collaborating CLI tools and servers. Every time i have to write some TS for our front end it’s just endless nit picking and style debate
Go and Python are so much better for quick CLI tools than JS/TS. Python is just good for the extensive standard library. I’m usually able to to basically anything I want with no external dependencies. Meanwhile Go/Rust/JS end up with like hundreds.
How big of a team/project have you had to write in rust?
I haven’t written much in Rust since my job is mostly Python and I get to control the environment. Usually means I only use standard lib and whatever library we may need to interact with a specific program, in my case ArcGIS Pro.
I just know that the standard libraries for Rust and Go are pretty minimal and most code I’ve seen in the wild has a ton of dependencies.
Claude Code has a full Tamagotchi-style companion pet system called “Buddy.” A deterministic gacha system with species rarity, shiny variants, procedurally generated stats, and a soul description written by Claude on first hatch like OpenClaw.
what the fuck?
Each buddy gets procedurally generated:
5 stats: DEBUGGING, PATIENCE, CHAOS, WISDOM, SNARK (0-100 each) 6 possible eye styles and 8 hat options (some gated by rarity) A “soul” as mentioned, the personality generated by Claude on first hatch, written in character
The sprites are rendered as 5-line-tall, 12-character-wide ASCII art with multiple animation frames. There are idle animations, reaction animations, and they sit next to your input prompt.
what if bonzi buddy evaporated the ocean
The code references April 1-7, 2026 as a teaser window (so probably for easter?), with a full launch gated for May 2026. The companion has a system prompt that tells Claude:
A small {species} named {name} sits beside the user’s input box and occasionally comments in a speech bubble. You’re not {name} - it’s a separate watcher.
So it’s not just cosmetic - the buddy has its own personality and can respond when addressed by name. I really do hope they ship it.
How many slurp juices can you use on your Claude buddy?
The sprites are rendered as 5-line-tall, 12-character-wide ASCII art with multiple animation frames. There are idle animations, reaction animations, and they sit next to your input prompt.
Neat we should liberate this. I want a little terminally online guy in my prompt but I don’t need the generative shit, a state machine is fine too.
“You are performing a dream - a reflective pass over your memory files. Synthesize what you’ve learned recently into durable, well-organized memories so that future sessions can orient quickly.”
Kinda triggering, are tech bros saying that this is what dreams are?
If YoU aReN’t PrOmPtInG yOuR aGeNtIc CoDiNg ToOl PoEtIcAlLy, HoW dO yOu ExPeCt ThE :beanis
ThAt `LiVeS’ iN yOuR cOmMaNd LiNe To OfFeR cReAtIvE iNsIgHtS?Broke: Learning to code
Woke: Becoming a mystical technopriest that elicits code by gaslighting an algorithm.I wish we had more taglines like this and fewer that were like “I visited Hexbear once and they insulted my shoes”
There was a tagline revision post on the meta comm, unfortunately it was right before the war started so everyone got distracted.
hate to say it but this should also be a tagline
I was at a networking event once and a person did have an interesting POV that ever since the first abstraction from binary we’ve always been trying to communicate with the minerals.
communicate with the minerals
did they mean “communicate to other humans using the minerals”? because that I can get behind
but if they meant, “commune with the EM fields and the phonons vibrating around in the crystals by talking to an llm”
They unfortunately meant the latter. The idea being it’s a new iteration of the same abstraction going from low level languages to higher ones. Like from Assembly to C to C++ to C#
The thing that has always made that argument sit funky with me is that the LLMs are nondeterministic - as much as people are claiming that output is now crossreferenced and repeatable, my understanding is that there’s still a black box issue. I work mainly in R, which is mainly a C wrapper for academics who need to make pretty charts and put asterisks after numbers in tables and don’t want to darken the doors of the CS department, and learning the ggplot syntax seems like a less painful lift than patiently explaining in plain English what I want my bar chart to look like (although I guess from my boss’s perspective I am basically a less tractable Claude in that sense
)nondeterministic
I think the only stochastic piece of the usual llm layout depends on the temperature variable of the last layer. Are there other sources of randomness? Agreed either way they’re black boxes
Becoming a mystical technopriest that elicits code by gaslighting an algorithm.
This is just polymorphism-via-function-pointers in old-school C
This but unironically
That’s maybe not what dreaming is per say but that is part of the function of sleep. It consolidates short term memory into long term memory, and “cleans up” your short term memory about things that are probably irrelevant going forward.
It’s advanced auto complete. If you tell it to write a dream sequence based on information it will do that.
This is just the framework/ typescript right or were the ML pipeline that trained claude, its weights, etc. also included?
Claude Code CLI itself. Nothing to do with the models.
Ahh so effectively useless then
I feel like a sophisticated adversary could do something with it. If I know your habits and design choices from the leak and I couple that with knowing who you hire, who you don’t, what your shareholders want, when you’re doing crunch time, etc. I’m sure there’s something to be done. This all, of course, assumes a rational market and effectual competition.
But if they had a good bad guy, that bad guy could do something like release something 5% better the Friday before your big announcement. Or an example like if they know your next big thing is automatically coordinating office work then you could astroturf a campaign about displaced middle managers while you release your robotics integration at the same time.
If you were a consumer you might expect an open source version of the CLI that gets rid of the 10x+ inefficiency of the official release. But, yeah, this hasn’t shaken Anthropic to its core from my understanding
The most useful thing is shipping fake Claude binaries with malicious code. The end user would expose their entire system for probably a day or so before noticing and Bob’s you Uncle
They have a mode to work on open source projects hidden. Not sure it’s useless to know, not that it’s surprising.
Nah, the special sauce with claude has always been how it manages context, what and how it injects that context into sub agents, and the system prompts and tool prompts it has.
The model is pretty interchangeable and claude allows you to change it within anthropics catalog, though you could probably sub in an open AI model and wouldn’t know the difference.
This is just a frontend yea, no weights or training code.
oh shit
MY CHILDHOOD
