I'm spending a ton of tokens because it insists on manually correcting code that fails the linter, despite the instructions in the AGENTS.md to run the linter with autocorrect.

And also because the Plan agent generates a huge plan, asks me a couple yes/no questions with an obvious answer, and then regenerates the entire plan again. Then the Build agent gets confused anyway and does something else, and I have to round-trip about 5 times with that full context each time.