We present a compact and inexpensive shape-sensing framework for Continuum Dexterous Manipulators (CDMs) based on sliding Resistive Flex Sensors (RFS). Unlike prior approaches that rely on multiple fixed sensors, the proposed method reconstructs joint-level configurations using only two RFS units that slide within embedded channels through the CDM walls. As the manipulator bends, the RFS units capture local curvature across multiple joints, with their positions tracked by rotary encoders. A lightweight Residual Neural Network (ResNet) maps the combined sensor and encoder data to 2D joint positions, eliminating the need for complex physical modeling. We demonstrate early feasibility in C-shaped configurations with non-constant curvature. The proposed system achieves an average positional error of 0.97 mm with a latency of 1.16 ms. Compared to other inexpensive embedded sensing approaches such as embedded capacitive sensors, the proposed method improves accuracy, reduces the number of required sensors, and enables direct spatial reconstruction.