Is there any working coding agent for local #LLM? #opencode is insane, it shows models that just don't exist, #claude is poisoned, #cline is completely broken and just talks in circles, maybe try repl or goose?
@thijs with small models like Qwen 3.5 4B based ones, I find pi coding agent working great. But you have to size well your context.
@panmaster I will give it a try, thanks!
@panmaster it is nice but because the name has been in popular use by Raspberry Pi for so long it is really hard to find good info for it, a really poor decision on the name if you ask me
@thijs and he knows it, I read that somewhere but can’t find the article. The site to avoid any ambiguity: https://shittycodingagent.ai/
This morning I tried with Gemma 4 E2B: faster but I need to use it more to conclude anything.
pi.dev

A terminal-based coding agent

@panmaster I just tried Gemma4 but I got a ton of errors on tool handling, maybe a LM Studio issue then
@thijs I use llama.cpp with CUDA and have no such errors about tool handling