Endless try Loop detection

[datenzar]

copying request from shabc via Discord:

i am trying small and simple winfoem really small,
and the model (qwen 2.5 coder 32b 16k-ctx-win) is looping forever because of 2 warning of possible null,
i gave it an hour and it loops without realising that it tried already the same solutions,

i will add more context window hopping this can solve it, but i think some code to realise we are in a loop should be placed and admit defit and allow the user to take control, or at least allow the user to enter a prompt in the middle of the task (while running) so it can add new light to the AI retrys (for instance: you tried already this solution... maybe try more ideas ...)

[datenzar]

I was just about to open the same feature request before I stumbled over your message. I'm running roo code against a self-hosted ollama model and gave it a task. I then left the machine since I didn't expect a timely response.
When I returned, I found Roo Code in a Loop asking the same question 30+ times (I'm glad its winter here so my computer was heating up the room 😊).

I stumbled over such issues already in the past and would absolutely appreciate, if RC could detect loops and abort exection (mayby even trigger a notification to my mobile about the abort, but notifications might be a different topic...)

[harrisdan]

I'm dealing with this too. My Sonnet costs are out of control last month so I've been trying and having some good progress with FuseAI/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview-GGUF. However, twice now it's gotten stuck in a loop. I wonder if it would be possible to detect the same exact response twice in a row from the model and perhaps inject something into the prompt to get it back on track.

[BradKML]

This issue is also known in OpenHands as well, so there has tom be a common cause with this when it comes to trying alternate solutions

[BradKML]

@dosu please figure out how in multiple software applications (e.g. Cline, RooCode, and OpenHands), LLMs often get stuck in a loop. instead of trying alternative solutions All-Hands-AI/OpenHands#7960 (comment) cline#2307 cline#1094

[BradKML]

@dosu what are the common causes of LLM agent tools constantly repeating the same mistake, even when given a history/log to reflect on the mistake (for RooCode/Cline AND OpenHands)?

[dosubot[bot]]

A common cause of LLM agent tools repeating the same mistake—even when provided with history or logs—is that the model fails to effectively use or interpret its own context window. This can happen for several reasons: the model may not recognize that it is repeating itself, it may not have a robust mechanism for tracking attempted solutions, or it may lack strategies for escaping dead-ends when stuck. This issue can affect both RooCode/Cline and OpenHands, especially when the agent relies on the LLM to self-reflect or innovate without explicit loop detection logic.

For Roo Code, a mechanism has been implemented to detect and prevent tool execution loops. When the system identifies a repetitive cycle, it intervenes by prompting for user input, which helps prevent infinite loops and keeps the application responsive. This improves stability and user experience by reducing the risk of the tool getting stuck in unintentional loops [source].

There is no specific documentation available about how OpenHands handles loop detection or repetitive behavior, so it is unclear if similar safeguards are in place.

In general, these issues can be mitigated by implementing explicit loop detection (such as checking for repeated actions or responses), prompting for user intervention when a loop is detected, and designing the agent to try more diverse strategies when it encounters repeated failures. Roo Code’s recent updates reflect these best practices, but if you are using other frameworks or tools, you may need to add similar logic yourself.

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}