Is RTK extension worth adding?

[gterras]

Hi,

First of all thanks for this wonderful software, I can't state how usable it is for real-world work even with local AI on a low tier GPU.

https://github.com/rtk-ai/rtk
https://github.com/sherif-fanous/pi-rtk

This lib might be a good fit little-coder (from what I know about it, which is not a lot), do you have thoughts on this?

In any case upvoting the ability to add pi extensions described in #28.

Thanks

[itayinbarr]

@gterras — answering the concrete half first: you can add pi-rtk yourself now, without waiting on me. That's the #28 ask you upvoted, and it shipped in v1.12.0 — either little-coder --with-pi-extensions (pi discovers its own from ~/.pi/agent/extensions), or drop it in ~/.config/little-coder/extensions/, which loads after the bundled set and survives upgrades. /extensions shows what actually loaded. See docs/extensions.md.

On whether it should be bundled — my instinct is no, and the reason is specific to this project rather than a judgement about RTK. little-coder's whole argument is that a small local model fails from harness mismatch rather than from lack of capability, and the lever is a small, fixed tool surface: cold start sits near 7k tokens, and every tool added is schema the model has to hold in a context window it does not have to spare. Adding a toolkit to the default set spends that budget for everyone, including people who never use it.

That is exactly what the extension directory is for, and I would rather learn from people running it than guess. If you try it, I would genuinely like to know: does it help or hurt on a small local model, and how much does it add to cold-start context? /extensions plus the %/ctx figure in the status line will tell you both. That is the sort of measured report that moves something into the default set — same bar as community models in #63.

And thank you for the kind words about real-world use on a low-tier GPU. That is the case the project is for.