I have been running few coding agents in my spare time on my personal projects and Junie does not look great when paired with local model
First of all, local models are capable but their ability to answer prompts in sustainable speeds degrades very quickly. I was experimenting with Qwen3.6-35B-A3B and Qwen3.5-35B-A3B. Both seems to be fairly capable, which is great for a local model that can run on decent GPU that is few years old (I have Radeon 7900 XTX with 24GB of VRAM) – that is outstanding considering what we thought about computers i.e. ten years ago. If someone then would say to me that we will be able to converse with our applications in human language and instruct them to do useful stuff, I would say: ‘Impossible!’. But we can and it does work but it have limitation, mainly in context size.
It works fine till you pass threshold of about 20k tokens. After that it slows down and after 50k tokens it will be timing out constantly. After 100k I will be running single prompt for hours.
I started new project, that I called SharpPad. Intention is that it will be interactive CLI for running C# code. I am using sometimes Python for some quick operations like splitting more complex string into usable data and outputting it into Json. I thought that it would be nice to have something similar for C#. As far as I know it does not exists or I was unable to find it. Anyway I can write my own tool now with assistance of coding agents. I decided to try that.
For now, I tried Junie and Mistral Vibe (pretty terrible name IMHO). Junie advantages are that it have great integration with Rider IDE that I have been using for several years now (about 8 I think).
On the bad side it seems terribly unresponsive and slow.
For example I had a bug that
In currently opened file there is GetInput method. This method handles backspace key. If it is pressed it removes part of text from class state but do not clear it from console. GetInput should be aware that previously printed text is longer and it need to be cleared.
It was working for about an 20, maybe 30 mins. It time outed once during this simple task.
That is pretty simple thing to do, and it fails to do it with local model.
Good thing it does have ability to automatically retry by itself without me forcing it to retry.
But still for such a simple task it is so slow that it is basically useless.
I have been running this particular coding agent for few weeks now, experimenting with different tasks here and there but result is fairly consistent: It times out very often. What helps with that is auto compaction threshold.
Sadly Junie does not have that setting. Seems like I will have to experiment a bit more with Mistral Vibe, or PI or OpenCode, that does have such settings.