Dagger is inscrutable to LLMs

[steady jackal]

I can't productively use AI to develop and debug my dagger scripts because the trace is hidden behind a UI that needs to be visually navigated. Do you have any tips?

[ancient spruce]

You mean an LLM calling your Dagger script to do something and debug its output? Does using --progress dots make it any better? That should give the LLM more context about the pipeline without exploding the context.

[steady jackal]

I mean having the LLM write the dagger script and fix mistakes based on its output

[pure vale]

I don’t understand what you mean, the output in the CLI is the same as what you’d see in the TUI or dagger cloud UI.

can you give some specific example of how the LLM is calling the dagger function and what output it is or is not receiving.

If possible, a code repro would be best!

[grave walrus]

What "dagger script" is the LLM writing? Are you trying to have it write a ci pipeline for your app?

[steady jackal]

Yes, CI. A fairly complex pipeline.

[fallow tulip]

you should definitely talk to @crisp wyvern and @fringe pendant about the daggerizer project 😁

The way they do it, is to wrap the dagger-generating agent itself in a dagger module. All the primitives are available in the dagger API.

[steady jackal]

@crisp wyvern Any repos for me to study?

[crisp wyvern]

I'll have something to share soon 🙂 I'm curious about what Lev said though, is the issue that the AI agent you're using isn't able to see the dagger output locally?