Let’s say the agent makes 10 changes in 10 different code blocks. You realize 7 8 9 and 10 are unwarranted. You already pressed accept all. You have no choice but to restore checkpoint at the prompt that started the 10/10 changes, and start over entirely. You have entirely removed the checkpoint system leaving it in an entirely broken and dysfunctional state. This is magnified to the 11th degree with the fact that YOU KNOW sonnet 3.7 is over zealous. I am forced to reprompt over and over again, hoping i catch the correct ‘stop’ point staring at my screen for minutes at a time, completely devaluing the entire point of agentic ai. I’m having to redo 1000 line code edits that take 4 minutes to complete. REPEATEDLY. Staring at the screen every moment, not missing it for a single second, to ensure I hit the stop button at the right time. Is this really the desired state of agentic ai? Otherwise I can go file by file at the end of it accepting and denying. Might as well disable the accept all button, as it’s a grenade waiting to happen.
I can ctrl z the file until hopefully nailing the desired state, as there were many changes this is subject to human error, or hope my last git commit lets me discard changes. These are only solutions if you hit the grenade that is accept all changes under the present paradigm. Even better! is that ctrl z BREAKS! and files will get stuck INBETWEEN checkpoints if you press the reject button.