In the case of an impasse, the system begins selective backtracking, using an internal episodic memory (separate from the general episodic memory, but similar in structure). All planning events are recorded here, allowing the control to retreat to a prior decision point to resolve the problem.
If a query comes in through the text interpreter, the planner will start up in inference mode. Here, plans can only contain inference steps, ignoring the actions and events which are also available in the general case. This simplifies the style of control in that no separate reasoning module is required.
The planner also provides mechanisms for replanning. The dynamic nature of the environment, the planning result may become invalidated before execution by the interpreter. For instance, a new piece of information, received through perception or conversation, could potentially indicate plan failure. The old information superceded by the new data is said to be "terminated". Plans contain protected pieces of data which are vital to plan success. If the old information was protected, then indeed the plan is invalidated, resulting in pruning (using the internal episodic memory) and regeneration of the plan.
However, new information can also advance the plan, by indicating the external achievement of on of the agent's goals. The basic agent matches new data against existing goals, and on a successful match, destroys the corresponding goal's plan. It then reinvokes the planner, which may not always result in an improvement in the overall plan, but it will not result in a deterioration (except for the time lost to planning).
The basic agent will offer natural language output through the text generator if either of the above events takes place. It will also notify a supervisor of a goal it cannot accomplish, offering an excuse for its failure. Such a blocking goal is maintained, in the hope that future events or information will permit its satisfaction.