Dagger Cloud

[silk iris]

CI was faster too: https://github.com/dagger/dagger/pull/11336

Not the 3x improvement I got locally though, most likely due to CI infra having high perf disks (?)

[toxic helm]

new high score

[silk iris]

yes that's "1 containerd" 😄

i had to patch their bolt db settings. fortunately, containerd v2 allows you to customize that. Unfortunately, upgrading to containerd v2 is a tedious PITA, working through that now

[toxic helm]

doug-able tedious, or too hard?

[silk iris]

I was wondering that... last time I tried agents for tasks that involve "insanely tedious updates to hundreds of files" it went very poorly, I just ran out of tokens cause it tries to read everything

But maybe I could go meta and have an agent write a script that does the update?

[toxic helm]

modules/dev/ hooks up a gopls MCP server, I've seen it pretty easily chase down compile errors

which might help? or if it's scriptable yeah, that'd be good too

[silk iris]

this is the diff when buildkit did it: https://github.com/moby/buildkit/commit/5be7edb69c0e27b5f11e9949615d20e0d0789e6c#diff-33ef32bf6c23acb95f5902d7097b7a1d5128ca061167ec0716715b0b9eeaa5f6

it's like 95% just go imports but then a handle of subtleties

[silk iris]

noice, that sounds promising, I'll try it

[toxic helm]

can always race with it in the background 😛

[toxic helm]

@silk iris ah one flaw in this plan may be how on-rails modules/dev is, it can't run arbitrary shell commands at the moment which might be kinda what you need for something like this

(i may be spying on your traces)

[silk iris]

yeah it seems to be getting scared lol

"That's a lot of files" 😂

[toxic helm]

also the word jumbling is an API issue, fixes on refresh, though the web UI UX kinda sucks for LLM sessions at the moment anyway, hasn't quite kept up

for adding new dependencies i've resorted to underscore-importing it with a comment somewhere and then go mod tidying so it doesn't get lost - if it ran into that issue

or we could add a shell function to modules/dev - it's not hard, i just don't like how often it resorts to it instead of the curated tools

it's also kinda sad because it'll see that it can't run commands, and then write a script instead, and then realize it can't run the script, and then move along like a cat does when you catch it doing something embarrassing

[silk iris]

yeah it's really trying it's heart out right now

I'm also asking codex if it can write a script

might also be a situation where it'd be nice to have "top-of-the-line reasoning model w/ web search and everything enabled write a prompt for the dumber model" sort of functionality

[toxic helm]

i hooked up Kagi MCP to the agent I use for Dang - but it's kind of a pain, their search API is "email us" + insert credit card

[toxic helm]

that would be sick though

for complex stuff I've done similar, but using the same model, just split into two phases: 1. research and write a plan to current-task.md with a checklist at the bottom, 2. start a new agent which loads it into its sytem prompt, and tell it to get to work

[silk iris]

Yeah the orchestration part is pretty interesting sometime. I also have codex and claude fix each others code sometimes ("this other idiot wrote awful code, please fix it" back and forth)