A single parsed CUA action. Coordinates are already scaled to the caller's
viewport. text is the action's primary string argument, keyed by type:
TYPE->text, VISIT_URL->url, WEB_SEARCH->query, TERMINATE->answer,
ASK_USER/READ_PAGE_ANSWER->question, PAUSE_MEMORIZE->fact, KEY->space-joined
keys. reasoning holds any chain-of-thought preceding the tool_call.