Nanocoder gets stuck and won't continue

[vscarpenter]

I just installed Nanocoder and I'm using it on my Mac with ollama and qwen3-coder:30b It seems like I'm missing something, but every time I instruct Nanocoder to do something, it keeps 'stuck' and I have to keep telling it to continue after every single command. Here are a few screenshots - am I missing something in my instructions?

How do I get Nanocoder just to execute my instructions and not require me to get nudging it along?

[vscarpenter]

One more screenshot - can't get past this 'duplicate tool call'

[will-lamerton]

Hi @vscarpenter - appreciate you opening this issue.
I'll look into this this afternoon (UK time) and come back to you! It definitely looks like something is up there.

[vscarpenter]

Thank you very much @will-lamerton - please let me know what I can do to help? Happy to try/debug/etc or just provide more information

[will-lamerton]

Hey @vscarpenter - thank you very much for the offer of help!

Essentially, local, non-tool-calling models like Qwen3 Coder struggle through the Ollama provider. Users have better results through using Qwen3 Coder through other providers or using tool calling models with Ollama like the regular Qwen3 or GPT-OSS.

However, this is obviously not ideal. Ollama is a key cornerstone to local AI usage so we want to support as many models as possible through it.

There are several improvement areas such as system prompt deficiencies and missing continuation logic and guidance. Both of which I have started to implement improvements for on this feature branch: https://github.com/Mote-Software/nanocoder/tree/fix/local-non-tool-calling-model-improvements

If you get a chance there are some changes and improvements up there already so feel free to test by pulling and building locally from that feature branch.

If not, I'll be testing thoroughly tomorrow myself :)

[vscarpenter]

Hi @will-lamerton and thank you very much for your response. I got the latest codebase locally and then checked out the fix branch you suggested "origin/fix/local-non-tool-calling-model-improvements" and I have to say the behavior is much different and better connecting to ollama locally and working over multi-step reasoning.

I'll continue to test and report back. Please let me know if you'd like to test anything specifically.