[Bug]: Guardrails against infinite loops

[JGKle]

Is there an existing issue for the same bug? (If one exists, thumbs up or comment on the issue instead).

• I have checked the existing issues.

Describe the bug and reproduction steps

The biggest "operational" issue I've noticed with OpenHands is that sometimes it gets stuck in a logical loop, and goes on infinitely, burning tokens until you notice & stop it. It seems like there should be some kind of guard rails against this. For example, I've seen it doing things like: Let me try to build the application; OK, first let's install (package X); Let me try to build the application; OK, first let's install (package X); and so on and so on, because one or the other didn't work.

Similar but not quite identical, I've seen it do things like "Let me fix the remaining indentation" many times in a row. I realized that it's actually continuing to go a few lines at a time changing the indentation, but this is a ton of back-and-forth / tokens, so it seems like it should be caught by a similar guardrail - i.e. at a certain point say "It looks like we might be stuck - would you like me to continue?" That way it won't just keep burning tokens on non-productive work.

OpenHands Installation

Docker command in README

OpenHands Version

0.33.0

Operating System

WSL on Windows

Logs, Errors, Screenshots, and Additional Context

[enyst]

We have guardrails against this kind of loops. It sounds like they may have broken or be out of date. Could you please tell what LLM you were using, and more importantly, do you have some logs or trajectories with an example?

The screenshot seems convincing, but I wonder if it was editing different parts of the file? A log with the actual LLM messages would be very helpful!

[JGKle]

Could you please tell what LLM you were using

Claude, can't recall if it was 3.5 or 3.7. But definitely Claude.

do you have some logs or trajectories with an example?

...err...probably, but it's below the first 9 now. So...kinda relates back to #7928 and #7940 - very not easy to find/to pull up prior chats :P

I wonder if it was editing different parts of the file?

Yes it was certainly editing different parts of the file (that's what I meant by "it's actually continuing to go a few lines at a time changing the indentation"). It was basically going down the file a bit at a time changing the indentation as it went.

To me this particular indentation example is a lot less severe than the other case, where it was trying to install something, try to build, try to install, try to build, as that would never end. This indentation one would eventually stop once it's done the whole file. But it still felt like guardrails should notice and be like "hmm, am I actually being productive here? Do you want me to keep going back and forth doing this?"

As for the more severe install/build/install/build case, I gave all my projects a microagent instructing it to never try to build, just edit code. So I don't expect to see that again myself. Just wanted to flag that it happened.

[enyst]

Got it. I think that's not really an easy problem to solve. We have guardrails that hardcode logic like the last steps are the same, or every other step is the same. Those are simpler cases.

In fact, some users might not think it's a problem. People might want the agent to keep trying to get the job done, at long as it's not stuck. 🤔

[JGKle]

We have guardrails that hardcode logic like the last steps are the same, or every other step is the same. Those are simpler cases.

Pretty sure this is what I saw with the "try to install, try to build, try to install, try to build" - unfortunately that example is long gone. & now all my microagents explicitly say not to install or build anything, so I probably won't see that again. Anyway, just wanted to flag it - sounds like you do have something there :)

In fact, some users might not think it's a problem. People might want the agent to keep trying to get the job done, at long as it's not stuck.

Yeah I can definitely see that for the indentation thing. For me it was more like "wtf, you're just burning thru tokens & costing me money to do code formatting. I can just run prettier!". So not sure what the right answer is here. Somehow I just expected it to like...not keep burning so many tokens indenting a couple lines at a time 🤣

[BradKML]

This stuff is coming back again, so I would like to see something similar to Task Master or some other tooling/MCP that makes it keep track of what is already done. In the case of debugging, structured memory would be useful.

[BradKML]

Cross-reference from another app, feels like a meta-pattern at this point RooCodeInc/Roo-Code#3519 RooCodeInc/Roo-Code#978

[github-actions]

This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.

[BradKML]

Please keep this up since it is important

[github-actions]

This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.