Developers must often diagnose anomalies in programs they only have a partial knowledge of. As a result, they must simultaneously reverse engineer parts of the system they are unfamiliar with while interpreting dynamic observation data (performance profiling traces, error-propagation channels, memory leaks), a task particularly difficult. To support developers in this kind of comprehension task, filtering and aggregation have long been suggested as key enabling strategies. Unfortunately, traditional approaches typically only provide a uniform level of aggregation, thus limiting the ability of developers to construct context-dependent representations of a program's execution. In this paper, we propose a localised approach to navigate and analyse the CPU usage of little-known programs and libraries. Our method exploits the structural information present in profiling call trees to selectively raise or lower the local abstraction level of the performance data. We explain the formalism underpinning our approach, describe a prototype, and present a preliminary user study that shows our tool has the potential to complement more traditional navigation approaches.