No problem, and thanks for the response here! I sent two feedback threads (added a second test where a file succeeded to edit and the rest failed). It was actually quite entertaining seeing the model try troubleshoot stuff anyway.

I used github copilot's sonnet 3.7. I now tried copilot's sonnet 3.5 and it seems to work, so it was prob a 3.7 issue? It did not let me try zed's sonnets, so I don't know if there is a problem with zed's 3.7 (I thought I could still do 50 prompts with a free account, but maybe that's not for the agent?).