Context-aware systems must manage the dynamic selection, activation, and execution of feature variants according to changing contexts, detected from data gathered from their surrounding execution environment. Many context-oriented programming languages focus only on the implementation level by providing appropriate language abstractions for implementing behavioural variations that can adapt dynamically to changing contexts. They often ignore or presuppose the existence of mechanisms to deal with earlier aspects such as the gathering of sensory input and context discovery. In this paper we propose a layered software architecture that reconciles all these aspects in a single implementation framework, which can be customized by application programmers into actual context-aware applications. This framework is currently being implemented in Ruby on top of a reimplementation of the Phenomenal Gem context-oriented language.