A repair agent forms an idea about how the code works, turns it into an implementation plan, then writes a patch. Hand it a specific wrong idea — the wrong invariant, the wrong place to hold state — and where that idea sits decides almost everything. Over 1,844 requests on two deliberately constructed Python tasks, the same misconception failed 21.9% of attempts from the hypothesis and 54.7% from the plan on one task, and 53.4% against 83.0% on the other. Adding the matching wrong hypothesis on top of a plan that already carried it bought almost nothing.