dagger llm

[gloomy parcel]

pr: https://github.com/dagger/dagger/pull/9773

[proven leaf]

Ok that is pretty fire

I wonder how we could control things like "export the result"? maybe give it a host object, maybe tweaking the API to restrict blast radius?

[proven leaf]

gonna play with it tonight

[gloomy parcel]

yeah i kind of had a moment where it clicked for me, the ability to change the toolset/scope throughout the prompt is pretty nifty

[proven leaf]

https://x.com/solomonstre/status/1897806156426625280

Solomon Hykes (@solomonstre) on X

"here's a container with github credentials. help me do issue triage. you'll need to install the github cli. figure it out"

X · Mar 7, 2025, 12:27 AM

@gloomy parcel someone just pointed out to me that goose actually has a CLI, so we're also competing with them 🙂

[gloomy parcel]

oh interesting

[proven leaf]

worth playing with it to see what ideas we could borrow

[proven leaf]

@gloomy parcel @orchid owl @stray junco let's talk about current state of the command and how to present it in docs & readme in its current state

[orchid owl]

is there any help text today for the slash commands?

[gloomy parcel]

no /help yet, but you can type / and tab through them, they each have a description

added a TODO

[orchid owl]

thanks! that works

[proven leaf]

@gloomy parcel how are you feeling about the idea to merge with shell?

[gloomy parcel]

my preference would be that they're both the same implementation, just dagger llm starts you in prompt mode and dagger shell starts you in shell mode

based purely on the idea that it's nice to have an quick LLM CLI handy and switching into prompt mode every time would be annoying

[proven leaf]

but would it be technically feasible to switch to shell mode from llm? I'm thinking about multi-object & the potential of just using shell syntax to set variables, then in prompt mode the llm automatically has access to the vars

[gloomy parcel]

yeah I think so - been wanting to play with that very idea, just been stuck in the quagmire of figuring out non-shitty scrollback UX (think I have something, finally)

there's already a /shell and you can reference vars in /with

just, no LLM -> shell vars yet

[orchid owl]

ah I didn't think about it like that. So could we potentially just switch modes on the fly like we do today between input / follow mode in shell?

[gloomy parcel]

there's a /shell for switching to shell mode and /prompt for switching to prompt mode

but they don't share context/history at the moment

[orchid owl]

got it. that's cool

[orchid owl]

gemini seems happier in this session with some more directed prompting

[orchid owl]

llama3.3, which is generally really good at this sort of thing, can't figure out chaining. It calls Directory.file, gets the ID, and then just makes up what i thinks a main.go could look like

llama3.3 = llama-3-70b

[stray junco]

shelL LM mode

[proven leaf]

@gloomy parcel @orchid owl want to talk live about it sometime today? I stressed myself out by tweeting about it 😅

[orchid owl]

haha yeah sure! For me it would help so we can get on the same page about where/how it fits in the llm docs we have at the moment and then we can ship it

[proven leaf]

yes exactly

[gloomy parcel]

sure, I'll be around

[proven leaf]

free in 30mn or so?

[gloomy parcel]

still planning to chat? just checking in, getting near EOD

[proven leaf]

yes sorry! ready now

[gloomy parcel]

cool, will join in a sec

[proven leaf]

(just joined team audio)

@orchid owl if you're still around