I just tried this with a smaller "thinking" model (deepseek distill, running locally) and boy are you right. It keeps flipping between which direction it should turn, second guessing its thought process and then getting sidetracked with a different approach.