So the reason that Claude code is capable of outputting valid json is because if the prompt text suggests it should be JSON then it enters a special loop in the main query engine that just validates it against JSON schema (it looks like the schema just validates that something in fact and object and its keys are strings) and then feeds the data with the error message back into itself until it is valid JSON or a retry limit is reached.

This code is so eye wateringly spaghetti so I am still trying to see if this is true, but this seems to be how it not only returns json to the user, but how it handles all LLM-to-JSON, including internal output from its tools. There appears to be an unconditional hook where if the JSON output tool is present in the session config at all, then all tool calls must be followed by the "force into JSON" loop.

If that's true, that's just mind blowingly expensive

edit: please note that unless I say otherwise all evaluations here are just from my skimming through the code on my phone and have not been validated in any way that should cause you to be upset with me for impugning the good name of anthropic

edit2: this is both much worse and not as bad as i thought on first read - https://neuromatch.social/@jonny/116326861737478342

reminder that anthropic ran (and is still running) an ENTIRE AD CAMPAIGN around "Claude code is written with claude code" and after the source was leaked that has got to be the funniest self-own in the history of advertising because OH BOY IT SHOWS.

it's hard to get across in microblogging format just how big of a dumpster fire this thing is, because what it "looks like" is "everything is done a dozen times in a dozen different ways, and everything is just sort of jammed in anywhere. to the degree there is any kind of coherent structure like 'tools' and 'agents' and whatnot, it's entirely undercut by how the entire rest of the code might have written in some special condition that completely changes how any such thing might work." I have read a lot of unrefined, straight from the LLM code, and Claude code is a masterclass in exactly what you get when you do that - an incomprehensible mess.

from @sushee over here, (can't attach images in quotes) and apparently discussed on HN so i'm late, but...

They REALLY ARE using REGEX to detect if a prompt is negative emotion. dogs you are LITERALLY RIDING ON A LANGUAGE MODEL what are you even DOING

OK i can't focus on work and keep looking at this repo.

So after every "subagent" runs, claude code creates another "agent" to check on whether the first "agent" did the thing it was supposed to. I don't know about you but i smell a bit of a problem, if you can't trust whether one "agent" with a very big fancy model did something, how in the fuck are you supposed to trust another "agent" running on the smallest crappiest model?

That's not the funny part, that's obvious and fundamental to the entire show here. HOWEVER RECALL the above JSON Schema Verification thing that is unconditionally added onto the end of every round of LLM calls. the mechanism for adding that hook is... JUST FUCKING ASKING THE MODEL TO CALL THAT TOOL. second pic is registering a hook s.t. "after some stop state happens, if there isn't a message indicating that we have successfully called the JSON validation thing, prompt the model saying "you must call the json validation thing"

this shit sucks so bad they can't even CALL THEIR OWN CODE FROM INSIDE THEIR OWN CODE.

Look at the comment on pic 3 - "e.g. agent finished without calling structured output tool" - that's common enough that they have a whole goddamn error category for it, and the way it's handled is by just pretending the job was cancelled and nothing happened.

So ars (first pic) ran a piece similar to the one that the rest of the tech journals did "claude code source leaked, whoopsie! programmers are taking a look at it, some are finding problems, but others are saying it's really awesome."

like "inspiring and humbling" is not the word dog. I don't spend time on fucking twitter anymore so i don't hang around people who might find this fucking dogshit tornado inspiring and humbling. Even more than the tornado, i am afraid of the people who look at the tornado and say "that's super fucking awesome, i can only hope to get sucked up and shredded like lettuce in a vortex of construction debris one day"

the (almost certainly generated) blog post is the standard kind of vacuuous linkedin shillposting that one has come to expect from the gambling addicts, but i think it's illustrative: the only thing they are impressed with is the number of lines. 500k lines of code for a graph processing loop in a TUI is NOT GOOD. The only comments they make on the actual code itself is "heavily architected" (what in the fuck does that mean), "modular" (no the fuck it is not), and it runs on bun rather than node (so??? they own it!!!! of course it does!!!). and then the predictable close of "oh and also i'm also writing exactly the same thing and come check out mine"

the only* people this shit impresses are people who don't know what they're looking at and just appreciate the size of it all, or have a bridge to sell.

* I got in trouble last time i said "only" - nothing in nature is ever "only this or that," i am speaking emphatically and figuratively. there are other kinds of people who are impressed with LLMs too. Please also note that my anger is directed towards the grifters profiting off of it and people who are pouring gas on the fire and enabling this catastrophe by giving it intellectual, social, and other cover. I know there are folks who just chat with the bots because they need someone to talk to, etcetera and so on. people in need who are just making use of whatever they can grab to hang on are not who I am criticizing, and never are.

If i can slip in a quick PSA while my typically sleepy notifications are exploding, these are all very annoying things to say and you might want to reconsider whether they're worth ever saying in a reply directed at someone else - who are they for? what do they add?

• "why are you surprised"/"even worse than {thing} itself is people being surprised at {thing}": unless the person is saying "i am surprised by this" they are likely not surprised by the thing. just saying something doesn't mean you are surprised by it, and people talking about something usually have paid attention to it before the moment you are encountering them. this is pointless hostility to people who are saying something you supposedly agree with so much that you think everyone should already believe it
• "it's always been like this": slightly different than above. unless someone is saying "this is literally new and nothing like this has happened before" or you are adding actual historical context that you know for sure they don't already know, you're basically saying "hey did you know this thing you care enough about to be paying attention to and talking about frequently has happened before now as well." this is so easy to frame in a way that says "yes and" rather than "i assume you dont know about the things i know about due to being very smart." eg. "dang not again, they keep doing {thing}"
• "{thing} might be bad, but {alternative/unrelated, unmentioned, non-mutually exclusive thing} is even worse": multiple things can be bad at the same time and not mentioning something does not mean i don't think it's also bad
• "funny how people who think {thing} is bad also think {alternative/unrelated, unmentioned thing} is good": closely related to the above, just because you have binarized your thinking does not mean everyone else has.

anyway if the mental image you are conjuring for your interlocuters positions them as always knowing less than you by default, that might be something to look into in yourself!

i sort of love how LLM comments sometimes tell entire stories that nobody asked. claude code even has specific system prompt language for this, but they always end up making comments about what something used to do like "now we do x instead of y" like... ok? that is why i am reading current version of code!

so claude code is just not capable of rescuing itself from its own context - if an entry in its context window throws an error, it just keep throwing that error forever until you clear it. good stuff.

(and, of course we read the entire file before checking this, rather than just reading the first 5 bytes)

this is super minor, and i've seen this in human code plenty of times, but this is the norm of this app verging on being formal code style.

so you have a file reading tool, you need to declare what kinds of file extensions it supports. that's very normal. claude code takes the interesting strategy of defining what extensions it doesn't read. that's also defensible, there are a zillion text extensions. i've seen strategies that just read an initial range of bytes and see if some proportion of them are ascii or unicode.

where does this get declared? why of course in as many places as there are rules. hasBinaryExtension() comes from constants/files.ts, isPDFExtension() comes from utils/pdfUtils.ts (which checks if the file extension is a member of the set {'pdf'}), and IMAGE_EXTENSIONS is declared in the FileReadTool.ts file.

of course, elsewhere we also have IMAGE_EXTENSION_REGEX from utils/imagePaste (sometimes used directly, other times with its wrapper isImageFilePath), TEXT_FILE_EXTENSIONS in utils/claudemd.ts. and we also have many inlined mime type lists and sets. and all of these somehow manage to implement the check differently. so rather than having, for example, a getFileType() function, we have both exactly the same and kinda the same logic redone in place every time it is done, which is hundreds of times. but that's none of my business, that's just how code works now and i need to get with the times.

continuing thoughts in: https://neuromatch.social/@jonny/116328409651740378

one thing that is clear from reading a lot of LLM code - and this is obvious from the nature of the models and their application - is that it is big on the form of what it loves to call "architecture" even if in toto it makes no fucking sense.

So here you have some accessor function isPDFExtension that checks if some string is a member of the set DOCUMENT_EXTENSIONS (which is a constant with a single member "pdf"). That is an extremely reasonable pattern: you have a bunch of disjoint sets of different kinds of extensions - binary extensions, image extensions, etc. and then you can do set operations like unions and differences and intersections and whatnot to create a bunch of derived functions that can handle dynamic operations that you couldn't do well with a bunch of consts. then just make the functional form the standard calling pattern (and even make a top-level wrapper like getFileType) and you have the oft fabled "abstraction." that's a reasonable ass system that provides a stable calling surface and a stable declaration surface. hell it would probably even help the LLM code if it was already in place because it's a predictable rules-based system.

but what the LLMs do is in one narrow slice of time implement the "is member of set {pdf}" version robustly one time, and then they implement the regex pattern version flexibly another time, and then they implement the any str.endswith() version modularly another time, and so on. Of course usually in-place, and different file naming patterns are part of the architecture when it's feeling a little too spicy to stay in place.

This is an important feature of the gambling addiction formulation of these tools: only the margin matters, the last generation. it carefully regulates what it shows you to create a space of potential reward and closes the gap. It's episodic TV, gameshows for code: someone wins every week, but we get cycles in cycles of seeming progression that always leave one stone conspicuously unturned. The intermediate comments from the LLM where it discovers prior structure and boldly decides to forge ahead brand new are also part of the reward cycle: we are going up, forever. cleaning up after ourselves is down there.

Tech debt is when you have banked a lot of story hours and are finally due for a big cathartic shift and set the LLM loose for "the big cleanup." this is also very similar to the tools that scam mobile games use (for those who don't know me, i spent roughly six months with daily scheduled (carefully titrated lmao) time playing the worst scam mobile chum games i could find to try and experience what the grip of that addition is like without uh losing a bunch of money).

Unlike slot machines or table games, which have a story horizon limited by how long you can sit in the same place, mobile games can establish a space of play that's broader and more continuous. so they always combine several shepherd's tone reward ladders at once - you have hit the session-length intermittent reward cap in the arena modality which gets you coins, so you need to go "recharge" by playing the versus modality which gets you gems. (Typically these are also mixed - one modality gets you some proportion of resource x, y, z, another gets you a different proportion, and those are usually unstable).

Of course it doesn't fucking matter what the modality is. they are all the same. in the scam mobile games sometimes this is literally the case, where if you decompile them, they have different menu wrappings that all direct into the same scene. you're still playing the game, that's all that matters. The goal of the game design is to chain together several time cycles so that you can win->lose in one, win->lose in another... and then by the time you have made the rounds you come back to the first and you are refreshed and it's new. So you have momentary mana wheels, daily earnings caps, weekly competitions, seasonal storylines, and all-time leaderboards.

That's exactly the cycle that programming with LLMs tap into. You have momentary issues, and daily project boards, and weekly sprints, and all-time star counts, and so on. Accumulate tech debt by new features, release that with "cleanup," transition to "security audit." Each is actually the same, but the present themselves as the continuation of and solution to the others. That overlaps with the token limitations, and the claude code source is actually littered with lots of helpful panic nudges for letting you know that you're reaching another threshold. The difference is that in true gambling the limit is purely artificial - the coins are an integer in some database. with LLMs the limitation is physical - compute costs fucking money baby. but so is the reward. it's the same in the game, and the whales come around one way or another.

A series of flashing lights and pictures, set membership, regex, green checks, the feeling of going very fast but never making it anywhere. except in code you do make it somewhere, it's just that the horizon falls away behind you and the places you were before disappear. and sooner or later only anthropic can really afford to keep the agents running 24/7 tending to the slop heap - the house always wins.

If you are reading an image and near your estimated token limit, first try to compressImageBufferWithTokenLimit, then if that fails with any kind of error, try and use sharp directly and resize it to 400x400, cropping. finally, fuck it, just throw the buffer at the API.

of course compressImageBufferWithTokenLimit is also compression with sharp, and is also a series of fallback operations. We start by trying to detect the image encoding that we so painstakingly learned from... the file extension... but if we can't fuck it that shit is a jpeg now.

then, even if it's fine and we don't need to do anything, we still re-compress it (wait, no even though it's named createCompressedImageResult, it does nothing). Otherwise, we yolo our way through another layer of fallbacks, progressive resizing, palletized PNGs, back to JPEG again, and then on to "ultra compressed JPEG" which is... incredibly... exactly the same as the top-level in-place code in the parent function

while two of the legs return a createImageReponse, the first leg returns a compressedImageResponse but then unpacks that back into an object literal that's almost exactly the same except we call it type instead of mediaType.

here, if i fold all the return blocks and decrease my font size as small as it goes i can fit all the compression invocations in the first of three top-level compression fallback trees in a single screenshot, but since it is so small i just have to circle them in red like it's a football diagram.

this function is named "maybeResizeAndDownsampleImageBuffer" and boy that is a hell of a maybe!

and what if i told you that if it passes a page range to its pdf reader, it first extracts those pages to separate images and then calls this function in a loop on each of the pages. so you have the privilege of compressing n_pages images n_pages * 13 times.

this function is used 13 times: in the file reader, in the mcp result handler, in the bash tool, and in the clipboard handler - each of which has their entire own surrounding image handling routines that are each hundreds of lines of similar but still very different fallback code to do exactly the same thing.

so that's where all the five hundred thousand lines come from - fallback conditions and then more fallback conditions to compensate for the variable output of all the other fallback conditions. thirteen butts pooping, back and forth, forever.

there is a callback feature "file read listeners" which is only called if the file type is a text document, gated for anthropic employees only, such that whenever a text file is read (any part of any text file, which often happens in a rapid series with subranges when it does 'explore' mode, rather than just like grepping), another subagent running sonnet is spun off to update a "magic doc" markdown file that summarizes the file that's read - that's one "magic doc" per file, not one magic doc.

I have yet to get into the tool/agent graph situation in earnest, but keep in mind that this is an entirely single-use and completely different means of spawning a graph of subagents off a given tool call than is used anywhere else.

Spoiler alert for what i'm gonna check out next is that claude code has no fucking tool calling execution model it just calls whatever the fuck it wants wherever the fuck it wants. Tools are or less a convenient fiction. I have only read one completely (file read) and skimmed a dozen more but they essentially share nothing in common except for a humongous list of often-single-use params and the return type of "any object with a single key and whatever else"

i'm in hell. this is hell.

i have been writing a graph processing library for about a year now and if i was a fucking AI grifter here is where i would plug it as like "actually a graph processor library" and "could do all of what claude code does without fucking being the worst nightmare on ice money can buy."

I say that not as self promo, but as a way of saying how in the FUCK do you FUCK UP graph processing this badly. these people make like tens of times more money than i do but their work is just tamping down a volley of dessicated backpacking poops into muskets and then free firing it into the fucking economy

alrighty so that's one of 43 tools read, the tools directory being 38494 source lines out of 390592 source lines, 513221 total lines. I need to go to bed. This is the most fabulously, flamboyantly bad code i have ever encountered.

Worth noting I was reading the file reading tool because i thought it would be the simplest possible thing one could do because it basically shouldn't be doing anything except preparing and sending strings or bytes to the backend.

I expected to get some sense of "ok what is the format of the data as it's passed around within the program, surely text strings are a basic unit of currency. No dice. Fewer than no dice. Negative dice somehow.

I seriously need to work on my actual job today but i am giving myself 15 minutes to peek at the agent tool prompts as a treat.

"regulations are written in blood" seems like too dramatic of a way to phrase it, but these system prompts are very revealing about the intrinsically busted nature of using these tools for anything deterministic (read: anything you actually want to happen). Each guard in the prompt presumably refers to something that has happened before, but also, since the prompts actually don't work to prevent the thing they are describing, they are also documentation of bugs that are almost certain to happen again. Many of the prompt guards form pairs with attempted code mitigations (or, they would be pairs if the code was written with any amount of sense, it's really like... polycules...), so they are useful to guide what kind of fucked up shit you should be looking for.

so this is part of the prompt for the "agent tool" that launches forked agents (that receive the parent context, "subagents" don't). The purpose of the forked agent is to do some additional tool calls and get some summary for a small subproblem within the main context. Apparently it is difficult to make this actually happen though, as the parent LLM likes to launch the forked agent and just hallucinate a response as if the forked agent had already completed.