This message was deleted.

Cline is second for me with the Plan and Act model. I still use OH from time to time but only for casual work.

Hey @jolly-magazine-94144 thanks for the feedback. Could you give a bit more detail about 2 and 3? We may be able to improve here easily

Sure. Claude Code’s subagents just take things much much further and don’t cop out quickly. For instance I just had an issue where I submit on the client side a request to a backend API and get hit by a signature mismatch. The signature is generated by an external SDK. I put both OH and CC on the same debugging task with the same prompts. Specifically, I instructed both to use either web search or the context7 mcp to get the latest docs for that SDK. Both CC and OH did web search but OH totally didn’t pick up the mcp tool call request. CC also took things further and went into the sdk’s declaration files to find out exactly what the types. Just one example of many. 99% of the time I’d go for CC over OH. Just my cents. It took me one week to convert.

OK, this sounds like something we should be benchmarking, thanks! My impression was not that claude code was that much better than OpenHands when I used it, but if it indeed is that’d be good to know. CC @millions-crayon-32880 to think about if this is possible

@jolly-magazine-94144 thanks for expanding on the use case! Are you able to share the specific commit and the prompt you used? We are improving the system prompt and would love to add this as a test case

Also it might be interesting to see if we could actually evaluate Claude code in some systematic way

Yes, we should be able to run swe bench splits. Will look into it

Are you able to share the specific commit and the prompt you used? We are improving the system prompt and would love to add this as a test case

Sounds exciting, and after 4.0, worth a look. 🙏 👀 Usually, prompting PRs are on github, with evals either between discussions, or after hashing them out in the open. Sometimes Xingyao and others asked for feedback on prompt updates here on slack. Sometimes they're in PR-review. A couple small ones I shared in the slack threads that triggered them. Is there a draft PR you could share?