On the cloud, I set condenser max size to 700, but it removes history after ~140. Please understand that, from user perspective, this feels bugged. It just… feels bugged. I set options intently, and the cloud is doing… whatever else. 🤷‍♂️ I understand from Graham’s recent talk that there have been (are?) running experiments with condenser size live, overwriting existing user settings without notice? I’m not sure I was aware of that… if I was, I’m sorry, I might have forgotten. That might help explain why GPT-5 feels a bit gimped on the cloud. It does need more than Sonnet, IMHO. (700 is probably too much, but some ~300 I expected to work fairly smoothly, depending on task, idk) If I understand correctly, these cloud experiments are coming to an end? I admit, I hope they do, because I’m afraid this… is not good UX. 😓 I think that the way it can come across, to the user, is that the agent behaves erratically at times. Without them knowing what or why. That can be frustrating. (With the caveat that, to be honest, I’m not unbiased on this issue 😅, some details might matter more for me. But I’m not sure it’s a one-off either.) The more important point is that, it seems to me, this is the same kind of issue that people are starting to dislike deeply about Cursor and the like, when they just strangle models under the hood without user’s knowledge.

@wide-doctor-37830 any ideas what's happening here? are there logs we could look for?

To clarify, I was referring to Graham's talk where he mentioned experiments in varying condenser size (the figure on the right side)

I don't think this is directly a problem with the condenser. Best guess is the cloud experiments are overwriting Engel's settings. Not sure how those are logged to double-check -- @quick-window-91705 do you know a way to verify this hypothesis?

I can make this a priority I think transitioning to BYOK has caused experiments to overstep; most likely yes the experiments are taking priority over user settings - will fix this asap

ahh thanks team!

https://github.com/OpenHands/OpenHands/pull/11713 - this is a quick fix i'll also work on a more long term solution for running experiments without violating user expectations