COBRA is a snake‑shaped robot that can move, push, and steer a box without separate arms. The robot can use any part of its body for propulsion or contact. A contact‑implicit dynamic model that links robot–ground, robot–object, and object–ground interactions is used inside a time‑stepping framework. The model feeds a full‑order trajectory optimizer guided by a geometric planner and a simple proportional controller. The base gait is a sinusoidal sidewinding pattern generated by a central pattern generator. Steerable pitch and yaw amplitudes and a steering bias control heading, giving a warm start for the optimizer. In 20 trials, COBRA moved a box from a random start position to a goal within 0.2 m of the target, achieving a 100 % success rate. The planner’s reduced‑order representation follows the observed box trajectory accurately, even during sharp heading reversals. The system relies on external motion capture and object detection for pose estimation. Onboard visual‑inertial sensing is not yet used. The contact‑implicit optimization used for the object subsystem omits full robot dynamics, solving directly for contact forces. The gait controller was only tested on flat, rigid ground; uneven terrain performance is unknown. Lastly, the full joint optimization that handles both gait and object push has not yet been tested in closed‑loop on hardware. These findings lay the groundwork for future autonomous snake robot manipulation.