@hendrik considering I am a complete #dart newbie I think it did pretty well. Even when it messed up it was able to auto iterate until make passed without me having to keep re-prompting.

I'm still using #ellama for other interactions like for example reviewing a patch series before posting. This #eca workflow is really tuned for the edit/compile/test cycle of writing new code.

The next time I play with it I want to try local inference and see how that performs with local models in control.

The experience is very different from the #ellama integration I currently use for general queries. The principle interface is still a chat window but rather than copy and pasting code you can watch the #LLM's internal monologue and then approve requests to edit files and run tools.

The first time I hit a compile error I just told it the build failed and for it to fix the problem. It's quite something watching it invoke make, read the error and then iterate until the problem is fixed.

2/n

@fidel I've mainly been focused on #codellama as the intent seems to be targeted at coding. Both 7b and 34b. 70b is beyond my capacity.

I find it difficult to get an idea of what's happening behind the scenes. It would be quite nice if #ellama placed the request in a buffer so you could see the conversation.