Just a TL;DR and RFC: I think long-running tasks are here, not on the horizon, not in the future. Currently, some CLI like

codex-cli

appears capable of running 60k events without noticeable slowdown and without losing track. I think OpenHands cannot, it slows down very noticeably around 1k in CLI, and/or crashes sandboxes on cloud. • original proposal to avoid O(n)+ code: https://github.com/OpenHands/software-agent-sdk/issues/1824 • on SDK, I fixed most occurrences of unbounded events; I believe we have a PR on those remaining in condenser; IIRC maybe Robert found another in agent-server (will check) • on CLI and Web, we have issues posted • on CLI, a PR has taken it into account. Today, some issues related to this problem have been closed from Linear without justification. That may have been an unintended mistake, with the best intentions, as @elegant-branch-54423 explained elsewhere. I just want to say: if we’re not fixing this, I think we may have just said “OpenHands is unable to compete on long-running tasks and we’re OK with that”. I don’t love the sound of that. Am I mistaken though? IMHO, it’s very important. Reasonable minds can disagree, of course. I do need to attract your attention on this, I have reopened the issues and would argue that we should not just give up, at least not without reason/investigation.

Hi Engel, I think closing the issues was a logistics mistake, instead of going into a category corresponding to “backlog” on linear they went into “canceled”, which resulted in the github issues being auto-closed. We still want to fix them and I think this is somewhat important, although it’s probably a better idea to split large chunks of work into smaller chunks of work i guess?

Whew, thank you. You wouldn’t believe what a relief this is.

Hi Engel, re: long running conversations, I actually received a support email today from an OSS user who said they noticed OH wasn't able to support long running tasks vs. Codex. I think this is important to look into, thank you for raising this.

I feel fairly confident that it's even more difficult on the cloud. How were they using OH?

This was a Local GUI user. Again, anecdotal feedback at this point (no hard metrics or stats shared), but thought it was coincidental given your feedback