In my prompt, I had requested Claude 3.7 to do a website check to verify that the error was either resolved or to get more information. Instead of checking the link and doing the curl check as it was supposed to, it just did an echo print and simulated that it actually did the search.
I had something similar happen yesterday when claude was working through a login/auth error. It was going through a login issue and was supposed to run a script to test if the login was fixed. After about two “max 25 tool notices”, I see a message saying that the changelog was being edited, which it’s not supposed to do until it fixes the problem. Then, it did an echo print saying that it had successfully fixed the issue by simulating what the script would say if the problem was actually resolved.
I’ve been using Cursor since August of last year, and I’ve also used Claude in other IDEs. I’ve never seen it actively lie and act like a problem has been resolved, then create an echo print to simulate that it either did something or actually solved an issue. This is a new one for me.