Discussion
models get regularly confused with the pi harnesses' built-in edit tool; multiple edits rarely succeed. it was only QFN wondering about its implementation in reasoning traces that tipped me off that it wasn't the model's fault.
on bsky sune marcher recommended this extension to me, and it works beautifully.
https://pi.dev/packages/pi-hashline-edit-pro
most of the edit woes are gone. i have seen QFN batch 19+ replaces successfully in one operation this way.
@SonnyBonds Qwen 3.8 Flash Next via Niko1221/Strata
yes, it falls back to bash and heredocs when the tools are not well documented or produce errors during usage.
@lritter Ah okay. I couldn't get Strata to run because lack of VRAM. (On the computer I'm at right now I do have a bit more VRAM, 8GB, but I still don't think that's enough.)
Regular Qwen 3.8 runs on Ollama here but it's fairly slow. It's the most capable I've tried though, but spends ages getting there. Not only does it evaluate slow, but it's really anxious in its "thinking", mulling over everything.
I've never actually used any of this to actually do anything, just messing around with it.
Install Fedizen
Get the full app experience
@lritter In my experiments I feel like the models often want to fall back to terminal commands to read and write stuff. A lot of sed etc going on. As I typically mess with them in Zed, that requires approval to run terminal stuff, I often have to go "don't use terminal, just use tools".
What model are you running?