PR for Claude Delegation Tool is ready: https://github.com/OpenHands/software-agent-sdk/pull/2100 Let me know what you think!

Is the review correct that

stop task

doesn’t stop it? I think maybe it matters

I already had tests for that here Moreover, I added the integration tests to confirm that

LocalConversation

is saving everything on disk and the implementation is not buggy 🙂 We are having a lot of differents review type. Probably because now we are choosing randomly the reviewer...

Ah, interesting. FWIW mine was OpenHands V1 with GPT-5.2 "roasted" prompt.

I re-run the openhands-bot after adding integration tests to cover more cases and now seems happy about that: https://github.com/OpenHands/software-agent-sdk/pull/2100#pullrequestreview-3820256721

I asked to try to prove the stopping issue or analyze your new tests to see if they cover it

Don't worry 🙂 The class is quite complicated, and just by reading through the code now, I think I might have forgotten something... I'm already on it 😉 Thanks for checking...

Thank you! I think it gets a bit confusing/crowded lol, with the successive reviews, this one was the GPT-5.2 "roasted" that might be worth checking the rest too https://github.com/OpenHands/software-agent-sdk/pull/2100#issuecomment-3920406890

Hey @calm-artist-7206 I did another round of checks based on the roasted review, and here's a summary: 1. "Resume is structurally broken (it can't work as written). This statement is wrong, but I can understand why the model said that. In fact, it could not know that

LocalConversation

was saving the conversation under the hood, and I had no integration tests for that at the time of the review. Now there are integration tests in place — see

TestRealConversationPersistence

. It is a little difficult to create good tests, but in this one the conversation is not mocked (just the LLM calls). 2. "task_stop doesn't stop anything (it just lies about it)". This was true, and your script proved it. Now it is fixed, and I also converted the script into an integration test to validate the behavior. 3. "Tool descriptions are actively misleading, which matters for LLM tools". Corrected. More details on that in the PR. 😉 Thanks for you support!