bug: Response output repeats infinitely due to missing stream termination

[JKGeovision]

Hello,

I discovered a bug where the ai response output repeats the same text block dozens of times in a single turn. This happens when the model finishes generating a response but the streaming connection is not properly terminated.

Steps to reproduce:

1. Start a conversation with cptr
2. Ask a question that triggers a multi-step response (e.g. running commands, creating files)
3. After the model finishes generating, the last response block repeats infinitely
4. The chat becomes unusable until manually stopped

I suspect the issue is that the harness does not correctly detect or signal stream termination.

Thank you again for your work and I wish you all a great day,
Joshua Krimmer

[owui-terminator]

🔍 Related Issues Found

I found some existing issues that might be related. Please check if any of these are duplicates or contain helpful solutions:

1. 🟣 #96 bug: OpenAI-compatible streaming usage can end responses early
   This issue is about streaming lifecycle handling in the same chat_task path. Although it reports early termination on usage events rather than infinite repetition, both involve incorrect stream end detection and response finalization in streaming responses.
   by sandrogauci

---

💡 If your issue is a duplicate, please close it and add any additional details to the existing issue instead.

This comment was generated automatically. React with 👍 if helpful, 👎 if not.

[tjbck]

Should be addressed!

[JKGeovision]

Hello,
I have tested the changes but noticed that the model still executes the same tool multiple times even though it confirms it is done.

Thanks again for your work and have a great day,
Joshua Krimmer

[tjbck]

@JKGeovision which model/ai agent are you using here? any other details you think are relevant that you can provide us with?

[JKGeovision]

Hello,
thanks for the fast response. I am using a local llama.cpp connection and use Qwen3.6 35B with Q6 as the model. I am also using the same model on other harnesses or chat apps like Open WebUI and never observed the model repeating itself like this before. Is there a log or something like that which could contain useful information?
Thanks again for your work and have a great day,
Joshua Krimmer

[tjbck]

@JKGeovision are you still facing this issue?

[cclecle]

Same problem here, confirmed with 6 different models. These models are running on llamacpp, working perfectly through OWUI. Seeing many repeatition loops using 'computer'.

[tjbck]

@cclecle could you confirm with the latest release?

[cclecle]

I just tried, and it was still happening, but it was a continuation of a previously broken chat. I am trying with new chat(s) now.. !
I also configured the connexion as 'llamacpp' now, it was left to default.

[cclecle]

Still happening :-/
It has something to do with tool calls I think.
Analysing the issue with an agent atm...

Edit: ChatGPT Diagnosis, I let the maintainers use it or not.

Repeated Assistant Responses After Tool Calls.pdf

Edit 2: I was running in non-thinking mode, now with thinking enabled (no code change), I am not seeing repetition anymore.. will try more and update here.

[tjbck]

@cclecle could you confirm with the latest?

[cclecle]

Not able to reproduce anymore, both in thinking and non-thinking mode :)
Will test more during the weekend, but it looks very promising.
thank you !