Can I get some feedback on the API compliance monitor: https://github.com/OpenHands/software-agent-sdk/pull/2204 Should limit the issue Robert calls out above until we can get the root cause identified

Please correct me if I make the wrong assumption, but could we do this on the

view

, rather than on

state._events

?

That's true, but the

view

properties assume the event list is well-formed by the time the view is made, that way they can just focus on maintaining those properties instead of enforcing them. That way the views don't need all the events every time they're made/updated.

You may want to see this too • https://github.com/OpenHands/software-agent-sdk/pull/2256 Also, IMHO we could maybe consider to merge your PR on this • https://github.com/OpenHands/software-agent-sdk/pull/2141 As you can see, Xingyao is trying something interesting and problematic: to somehow fix the duplicate tool result. But the reason for that appears to be that it made it into the event log already.

I see how that PR started, too: trying to capture the duplicate tool result idea as a

ViewProperty

. Unfortunately those properties can only be used to define contiguous groups (you have to limit their expressivity or the condenser has an impossible job), and the failure case might have the

AgentErrorEvent

pretty far from the actual tool call observation. So I'm not surprised that the transform in the

events_to_messages

function is where Xingyao landed.

I'm a bit unsure on the idea that it will never happen in the first place. I mean, it's an event log in the SDK in a single place, so it feels like we could fix it completely... on the other hand, we're playing with weird timing issues, and in V0 I recall we had a similar problem: something like, execution doesn't return for too long, then it does, but in the meantime we had to give the LLM something so it's an AgentError, so now it exists as an Event but the Obs returns in a split second I mean, OK, if I put it this way it feels fixable, I just have some feeling that we might be better off to accept weird things happen

I think we're on the same page. We'll never eliminate the weirdness, but we'll be well-served by making the agent code as predicable and reliable as possible.