Sampling-based motion planning algorithms with task-space guidance have been proved to work efficiently in solving task-constrained manipulator planning problem as the planned end-effector path can satisfy the task constraints directly. However, planning in task-space rather than in configuration space (C-space) may lead to discontinuous motion in joint space, especially the connection of the bidirectional tree. In this paper, a Progressive Constraint Extension Bi-direction Rapidly Exploring Random Trees (PCE-BiRRT) algorithm is presented, which adopts a progressive constraint expansion strategy to make sure that the sampling configurations can satisfy the task constraint directly. For the smoothness of the bidirectional-tree connection, the proposed algorithm deals with the "connection failure problem" of the bidirectional exploring tree by limiting the connecting range and choosing a property redundancy resolution. The effectiveness of the PCE-BiRRT algorithm is tested on the robot operating system (ROS) platform by an orientation-constrained case.