Cognitive flexibility is characterized by the ability to adapt behavior in response to changing environments. When the environment changes, previously appropriate behavior can lead to negative feedback. Surprising negative feedback – failure to receive a reward when strongly expected – may signal a change in the environment and the need to adjust action plans. Thus, processes that compute confidence and detect feedback at the timescale of individual decisions should interact with slower processes that infer environment changes. The neural mechanisms that underlie these hierarchical decisions remain unknown. We investigated this question using large-scale neural recordings in the medial and lateral frontal and posterior parietal regions of monkeys, combined with recurrent neural network (RNN) modeling. Monkeys performed a hierarchical decision-making task, discriminating the direction of random dots motion in an unstable task environment where stimulus-response mappings changed unpredictably every few trials. Negative feedback could signal either a mistake about motion direction or an environment change. We measured how monkeys resolved this feedback attribution problem based on their choice of either an environment switch target or direction targets following error feedback. The behavior matched a normative model, with monkeys integrating motion information within a trial to infer direction and choice confidence, and integrating confidence and feedback across trials to infer environmental changes. RNNs trained to implement this normative model replicated the monkey’s behavior. Both the RNN and monkey frontoparietal neural populations represented within-trial and across-trial integrations in orthogonal subspaces. Notably, larger movements along the within-trial integrator for easier motion stimuli led to greater shifts along the across-trial integrator following negative feedback. RNNs elucidated the relaxation dynamics induced by orthogonal line attractors that underlie this interaction. Our results reveal how distinct line attractors, integrating information across different timescales, interact to enable adaptive decision-making in dynamic environments.