Safety through feedback in Constrained RL