This paper presents a Therapist-in-the-Loop (TIL) framework for robotics-assisted mirror rehabilitation therapy integrated with adaptive Assist-as-Needed (ANN) training, to be adjusted based on the impairment and disability level of the patient's affected limb. Closed-loop system stability has been investigated using a combination of the Circle Criterion and the Small-Gain Theorem to account both for time-delay and the time-varying adaptive ANN training. Experiments to investigate the performance of the proposed framework are reported.