Add the actions an agent ran, with each element's ref replaced by its most stable locator (asked of the page, which still knows the element). Steps that failed are left out.
Actions pages reported within range that could not be recorded (over the size limits, malformed).
Start recording in every page; pages already loaded start capturing right away.
Drop what pages report until the returned function is called (an agent's action is running).
Follow a page from its creation (like the log collectors): bindings are armed while recording.
Records what users do in the app's pages (protocol 1.1,
record.*): while on, a CDP binding in the agent's world receives each click, field change and command key, with a locator that outlives refs; it exists from every new document's start, so navigations lose nothing. Actions an agent runs are added from their results instead (noteAgentRun), and whatever the page reports while an action runs is dropped: that input was the agent's, not a user's.