step: 1
state: observed
action: read_task_cards
observation: {"ok": true, "matches": []}
decision: needs_input
