@wide-doctor-37830 sadly, another condensation corner cases 😞 Do you think if we can implement a mechanism where we do compact similar to claude-code (instead of reduce events by half, we can just condense everything into one new message) 👀 https://github.com/OpenHands/software-agent-sdk/issues/1518#issuecomment-3695021678

We can tweak the condensation triggers to get something like what Claude Code does if we need to, yeah. There are other fixes I'd like to try first if possible though: • Parameter tweaking: the

keep_first

setting we inherited from v0 is not quite the behavior we want -- in the case of the event dump it's trying to keep the first few messages in the tool loop, but it can't start a condensation in the middle of one so it fails. We could lower

keep_first

to just keep up to the first user message, and/or raise

max_size

so we're likely to actually get a full tool loop in before condensation. • Delay condensation: in the case that we hit one of our condensation triggers (tokens/events) but can't actually forget anything, put the condensation on hold until the first opportunity. The suggested change in the bug report gets halfway there but I think there's also an unrelated bug where we'd start to condense a tool loop, continue the conversation, then get Anthropic yelling at us again because we condensed the thinking block and it still wants to keep the tool loop going. My hypothesis is that what Claude Code does is more likely to get the agent off-track than our rolling condensations. We would probably be well-served by making the condensation parameters a little less aggressive though. I'm traveling the next two days so I'll only be sporadically available, but back at on Wednesday if you want to assign me the issue to look at then. In the meantime maybe we can just set

keep_first=2

(or it might be

1

, can't remember the indexing atm) and

max_size=180

or

240

or something else large.

strong +1 to at least considering CC’s approach and seeing where we can improve in the most recent updates of CC with opus 4.5 I was extremely impressed by the detail of its condensation messages, to the point where I’ve started saving a few condensation messages to study from and try and suggest improvements to our approach at some point in the future

@adamant-controller-19704 would love to hear any suggestions you might have! Just want to clarify any apparent resistance I have to CC's approach isn't so much the "what/how" as the "when", but if we need to modify that too then all good. We just need to be careful about pushing condensation back too far as that has an impact on the cost of long-running conversations and we don't have Anthropic's usage/subscription model to hide token costs.

(I would say that when I’ve used CC I’ve had a less good experience after compaction than with OpenHands condensation, but I’m happy to try new things)

@wide-doctor-37830 that makes total sense!

Yeah, thanks a lot Alona, any insights are super-helpful!

of course! thanks yall!

On this note, I saw practically only good experience with codex-cli's approach. One thing I noticed is that the LLM is told to make sure that the state of the repo is what it thinks it is - so the first thing it does is look at branch/diff and maybe more, to get tabs on the current situation again. I guess we may need less of this if it has ~half, though occasionally it may prevent the strange cases when it feels like OH is restarting (they seem rare these days though)

👆 yeah, and on the plus side, the approach of condense + starting everything from scratch could help avoid some of the edge cases we were trying to fix lately (e.g., mixture of thinking blocks, wrong order of action-observation event pairs). But it is definitely a relatively bigger effort

I agree, but I think starting everything from scratch has a fairly high likelihood of confusing the agent. It’d be nice if we still keep the option to not do that

yeah i think fallback is definitely reasonable

@nutritious-balloon-65098 Create an issue on the software-agent-sdk repo based on this discussion that when we encounter a malformed message error from the language model, we should fall back to condensing the entire trajectory into a single message, similarly to how Claude Code does it. Assign this to

csmith49

I'm on it! Graham Neubig can track my progress here.