Browser actions the controller wants to execute
OptionalCheckpointThe exact name of a tour checkpoint the controller believes it has just reached. When it names a checkpoint with pending visual criteria, the engine forces a judge call this step scoped to THAT checkpoint's criteria — verifying the section against the frame it's actually on. Undefined when not on a tour or no section was reached.
OptionalEvaluationSelf-tracked agent state, all optional & parser-tolerant:
Evaluation — did the PREVIOUS step's action achieve its intent
(self-verification; pairs with Rule 11).Memory — durable notes the model carries forward (≤~200 chars).Plan — a short checklist with the current item marked.
Memory and Plan are echoed back into the next prompt so history is
self-describing and the model stops re-deriving intent every step.OptionalMemoryOptionalPlanRaw text response from the LLM (for debugging/logging)
The LLM's reasoning for its decisions this step
OptionalRequestWhether the controller is requesting immediate judge evaluation. When true, the judge will be consulted immediately regardless of JudgeFrequency settings. Use this when the controller believes the goal has been accomplished or needs confirmation of progress.
Tool calls the controller wants to make (executed before browser actions)
Parsed output from the controller LLM.
The engine parses the LLM's text/structured output into this format. Contains the reasoning, browser actions to execute, and any tool calls.