It's the harness giving the LLM contradictory instructions.

What you don't see is Claude Code sending to the LLM "Your are done with plan mode, get started with build now" vs the user's "no".