For systems to respond to scenarios that were unforeseen at design time, they must be capable of safely adapting, at runtime, the assumptions they make about the environment, the goals they are expected to achieve, and the strategy that guarantees the goals are fulfilled if the assumptions hold. Such adaptation often involves the system degrading its functionality, by weakening its environment assumptions and/or the goals it aims to meet, ideally in a graceful manner. However, finding weaker assumptions that account for the unanticipated behaviour and of goals that are achievable in the new environment in a systematic and safe way remains an open challenge. In this paper, we propose a novel framework that supports assumption and, if necessary, goal degradation to allow systems to cope with runtime assumption violations. The framework, which integrates into the MORPH reference architecture, combines symbolic learning and reactive synthesis to compute implementable controllers that may be deployed safely. We describe and implement an algorithm that illustrates the working of this framework. We further demonstrate in our evaluation its effectiveness and applicability to a series of benchmarks from the literature. The results show that the algorithm successfully learns realizable specifications that accommodate previously violating environment behaviour in almost all cases. Exceptions are discussed in the evaluation.