maki - the efficient coder (AI agent)

25 points by msangi

sloane

nitpicking, but i think one of the examples is wrong/nonsense?

cfg, meta = await asyncio.gather(
    read(path='Cargo.toml'),
    bash(command='cargo metadata')
)
deps = json.loads(meta)

used = set(re.findall(
    r'use (\w+)::', cfg
))
declared = {
    p['name']
    for p in deps['packages']
}

stale = declared - used
for s in sorted(stale):
    print(s)

Cargo.toml would also give you only declared dependencies, and that regex wouldn’t match anything because the cargo manifest doesn’t use that syntax. I assume they meant to… read all the Rust source files in the repo? use a glob? run grep?

I love it! I'm currently trying to find out if I can change the effort on OpenAI? I can change the model with /model but is there a way to select between XHIGH and HIGH ?

(sorry for removed comment I thought I posted 4 times but it was a glitch)

sy

Just a bit of feedback, the ~~right click to copy~~ select to copy functionality seems to break my usual select to copy into X clipboard and only allows you to use -selection c -- so I cannot actually paste what I copy in any terminal, so I need to paste it into notepad with Ctrl+V and then copy it there, to be able to paste its suggestions into a terminal that can execute them. I also can't Ctrl+Insert & Shift+Insert as the selected text deselects itself on copy.

So, I need to proxy between the clipboards with xclip -selection c -out | xclip -- this works.

That being said, this is fantastic for Windows -- the copy functionality does work and I can't really blame you for Linux's confusing clipboard handling -- I really value having two clipboards though.

radio

This looks pretty great, but I don't understand the sandboxed Python.

First: how did you manage to run Python inside Rust in a sandboxed environment? That looks hard.

Second: what is it good for in practice?

square_usual
Second: what is it good for in practice?

I've investigated this for work so I can take a shot at answering this. There's two basic advantages:
1. Instead of having to spend lots of tokens on your tool definitions (claude code's basic tool set takes 8k tokens of context!), and convincing the LLM to use the tools correctly (e.g. in parallel, or in the right order; go to https://weaxsey.org/en/articles/2025-10-12/ and Cmd+F "parallel" to see how much they have to tell the LLM to do this right), you just ask the LLM to write a script, and it's usually pretty good at writing Python.
2. If you let it use Python, the LLM can do things you didn't anticipate, and this can be both good and bad. The good is that a good model can very intelligently figure out how to do something that your inbuilt tools don't let it do to get shit done, the bad being it can rm -rf / while doing so. Sandboxes ideally eliminate some of the risk, but this is always in a continuum (how do you make a sandbox that's not annoying that also eliminates the risk?), but in general letting the LLM write and execute scripts is lot more powerful.
sloane

not the author but i peeked at the source, the how is pydantic/monty.

presumably the “why” is something like a minimally safe sandbox to interact with the shell without touching it directly, but idk how much isolation you get in practice.

Comment removed by author

cohix

Just yesterday I wrote a note on my Phone to experiment building a lightweight agent harness with sandboxing built in. This is awesome, will give it a try.