The agent alternates between LLM calls and carrying out plans. At each call the model reads the board, the effects of its recent actions and its current rules, then revises the rules and writes a short plan. In the action history, a ◆ LLM row is the first action after a call, and the dim rows after it carry out the rest of its plan.
If an answer comes back empty or breaks off before its plan, the model is called again before the next action, so one ◆ LLM row can stand for several calls; all of them are listed under All LLM calls below. If a complete answer has no plan, the harness makes one random move instead.
A plan may include RESET (⟲), which puts the board back to the start of the level. The level and the action count carry on.