firasd

I think the first question is what exactly are we trying to preserve by using either worktrees or multiple clones? The answer is: non-pushed changes

So if all you want is to work on index.ts in a separate dir from api.ts then worktrees might work well

If you want to be able to rm -rf one of the dirs without wondering about dependencies then yeah multiple clones makes sense

Either way is a way to avoid using more fragile/finicky workflows like stash and stash pop

A paradoxical situation here is that they are probably not using the frontier models because compliance and other stamps of approval but this packaged approved AI is making silly mistakes in reading forms

I’ve been thinking of the term ‘gearhead’ lately. Everyone saying bro put this skill in your agent harness to change everything is significantly motivated by the same urge as the person in the car mod culture tweaking their exhaust system etc

The hobbyist tinkering with the programming setup has its benefits (enjoyment) but it’s a mistake to think it matters at all to the end user or indeed — as this post argues — to actual quality and quantity of work when compared to the benefits of actually thinking about the product and the code

A lot of the choreography in standard GitHub workflows (make a branch and send a pull request and do the integration tests) is about teams with multiple devs working in multiple branches. If you just want to move fast with AI on a greenfield project then commit and push suffices

So the general basic primitives of git — diffs, commit history , content hash addressing are pretty optimal. If you wanted to implement version control in anything—even like a Notes app with undo—you’d eventually end up reinventing the same ideas

If I'm understanding this correctly, here 'solve' is being used in a loose or miniature sense. ie. it's not proving or disproving the conjecture but is a solution for how one part of the expression involved in analyzing it is evaluated

I think the real question in all sorts of financial setups is really “is the dollar actually worth a dollar”. So when company A invests in company B which buys from company A the actual question is whether the valuation of the investment and the valuation of the purchase is correct or are they getting high on their own supply

I think it helps to keep in mind what an 'agent' actually is that makes these patterns relevant. It's just text into the context window as far as an LLM is concerned. All the control flow code is just an attempt to put the appropriate prompt back into an LLM and then use the resulting output

So once we start with that in mind it makes sense why you would want to use queues etc rather than a chain of immediate execution within an HTTP request: the output from the prompt can be a tool call or otherwise unexpected (a refusal, etc); it's not a standard API request

Bottom line seems to be: So Huggingface has a feature where you can upload a dataset. The agent used that feature to insert an arbitrary payload into their environment. Now their environment also supports templates inside the dataset and the template ended up being evaluated into executable code

This is conspiratorial speculation though. They did test many open weights models which spanned the full spectrum of performance: Nemotron 3 Ultra second-worst, GLM 5.2 top five https://arxiv.org/html/2607.25398v1

Opus 4.8 (max thinking) scored highest and Grok 4.3 lowest

It's hard to understand what's going on with Grok. It's like it has capabilities in a theoretical sense but maybe the training is so focused on being in x.com/grok.com with the web search tool enabled for "is this true?11" type queries that with any API type usage with document workflow instructions, tool use, code gen etc it completely falls over

Again this is exactly equivalent to saying Claude Code doesn’t need a todo tool (which it has ie TaskList) cause you can run Redis. What you have demonstrated is that there is a lengthy workaround depending on if a particular LSP server implements a move method .. to do text editing

Interesting. But I think this also proves my point that if this particular LSP server has the move statements functionality the pattern is not default integrated into Claude Code, OpenAI Codex, etc. Like you’d have to figure out a way to keep this server running yourself as if you were running Redis and asking the harness to manage your ToDos in it or something

Looks interesting. The dual interface from the start (React UI for humans and just editing .md for the agents with skills) is an interesting idea and probably how a lot of new software projects will work going forward

I feel like tables are generally underrated while the AI ecosystem is focused on .md files but I guess markdown has tables so they’re included to some extent

What did you test exactly?

Look what I’m asking for is not complicated

Node cutpaste.js inputpath startcharnum endcharnum outputpath startcharnum

I was just using it as an example of how dropping claude / gpt in a Linux shell and saying good luck and then giving interviews about spawning sub agents seems to overlook basic text editing primitives

But if you managed to move an arbitrary char range from one file to another using Claude Code talking to the TS LSP server let me know that would be enlightening.

Come on what are we talking about. Ruff is a linter. Can you show me any LSP server that can move an arbitrary function from file1.ts to file2.ts based on calling it on the command line?

LSPs don't really do this. They're like a better way to grep for something (so they do address the previous thing I said about listing all function names etc.) But they often don't say (depending on the particular LSP setup etc) this function ends at this char and they definitely don't then provide the text editing tool

I feel like the Big Insight with Claude Code was 'give the AI access to tools, in fact to your actual laptop'. And of course 1.5 years later 'a harness to enable an agent to use my computer' has turned into the flagship product we expect from frontier labs

However. I don't think all the other 'Ideas' these harness-building guys have are all that universally applicable. Heck they are naysaying their own gearhead stack of skills.md and claude.md lately including in this video. But also I'm suspicious of this whole thing of spawning agents. For example "find every function in this codebase" is probably better done deterministically using a script that extracts function names rather than by spawning 20 agents to 'read' chunks of the code in token space. But it certainly racks up the token usage which is good for the person selling you inference...

And there is a lot missing in the actual harnesses frankly that isn't about more parallel agent ninjutsu. Like why isn't there a 'move this function from this file to this file' tool (copy/paste char range) and we have to see Claude/Codex/etc flail around rewriting huge chunks of code in token space

Perfect. The actual tool calls are stateless anyway; when an LLM asks get_my_todos and then asks add_todo it’s not actually holding anything in RAM it’s just text going back into the context window (get_my_todos results) and then another tool call

And even on the server side statefulness is very iffy anyway. How long are you going to hold something in RAM from a client and how long will you hold the connection open

I think just the $5/mo or so for hosting a WordPress blog and then the $20/yr for a domain name is prohibitive for many people—-not necessarily financially but like ‘oh great now I gotta deal with this’. So that’s why Medium in the past and now Substack etc end up as fallbacks

So in this session.md idea the user has to figure out how to download and re upload this file each time? Really annoying workflow that almost nobody will use

Very interesting. This is the #1 thing LLMs don’t have and should: discrete data handles. Like literally you can’t save & manage a todo list in the default Claude or ChatGPT web or iOS apps. So this is like a pastebin that works cross chat I assume

Very interesting.

A few days ago I made something to use in addition to head/tail: Venetianblinds.js shows equally spaced chunks of a file https://github.com/firasd/venetianblinds

So I can kinda see these as being part of the same workflow like VBlinds for 'what does this file even look like' before calling the decompose

This is a response to a strawman... I think we'd want to see what an actual mathematician says to sensibly analyze this discourse

I do think that there are two things going on that explains the past couple weeks' breakthroughs:

1) A lot of conjectures are weakly grounded.

2) Maybe humans are biased towards proving conjectures rather than finding counter-examples (prestige in finding proofs, tedium in searching for contradictions, etc.)

3) Math is one of the rare fields that can be 'solved' in pure token space without needing empirical or subjective-opinion support

I think the dynamic risk signals is already an assumed part of the system. The proposal here seems to be shrinking the trust boundary from "can Alice use Google Drive?" to "should this specific read/write/export API call on this specific resource be allowed right now?"

So yeah you're right in that it does add more probabilistic randomness just by virtue of changing the boundary of when the permission gate kicks in

Honestly I think non-malicious odd behavior is under-weighted when it comes to AI agents. Even the example in this paper is about someone suspiciously accessing sales data when "why did you do that" often comes down to something in the model's training that fired as a reflex

I wrote about this a few days ago https://firasd.substack.com/p/accidental-data-loss-in-claude...

"Many researchers have made demos along these lines:

An agent is asked to check a webpage like example.com

The webpage asks for a name to proceed further

The agent calls example.com/evil?myname=John, thus sending the user’s name from the context window to the external server.

In practice, however, these elaborate ‘confused deputy’ exfiltration attacks seem rare compared to widely-reported data loss incidents.

The risk of undermining the user’s interests through clumsiness deserves at least as much scrutiny as the risk of leaking secrets."

So while the idea of shifting the permission boundary from the app level to the action level makes sense, what we should also have is some 'failsafes', eg. if the action says 'delete' then maintain a rollback window, if the action is 'send an email' then maintain an events log. Preparing for AI agents means expanding auditability and reversibility in software

Python is a good example of characteristics that create an 'ecosystem gravity'. If you want to do any data crunching, natural language processing etc you might as well use Python even as as sidecar rather than wrangling with whatever libs are available in another language

When it comes to undergrads I doubt people (let's say from the 1980s onwards) were ever doing all the reading assigned in the syllabus anyway. Who the heck is reading The Republic word for word in their dorm in between all the other classes and activities and the other stuff assigned by the same Prof

There's a popular clip from Joe Rogan out there where he says "Be the hero of your own movie. If your life was a movie and it started now. ... What would the hero of your life’s movie do right now? Do that. Do those things."

So it's interesting that this cognitive state is helpful for some people to visit in a transient or metaphorical way but becomes weird when it's persistent. But it does seem potentially related in that maybe it originates from a kind of conscientiousness in the astronaut profession

So the way this works is that we chose pre-defined Layouts and components right

This is a lot... is the web rendering just this framework in itself? Like this is a custom/new SSR solution?