In 3D interactive fiction systems, a virtual camera must film the behaviors of multiple autonomous characters as they unpredictably interact with one another, are modified by the viewer, and manipulate artifacts in 3D worlds with complex scene geometries. It must continuously plan camera movements to clearly shoot the salient visual features of each relevant character. To address these issues, we have developed a 3D interactive fiction system with a narrative planner that, together with a bank of autonomous character directors, creates cinematic goals for a constraint-based realtime 3D virtual cinematography planner. As interactive narratives unfold, a cinematic goal selector creates view constraints to film the most salient activities performed by the characters. These constraints are then passed to a camera planner, which employs a partial constraint-based approach to compute the position and orientation of the virtual camera. This framework has been implemented in a prototype 3D interactive fiction system, Cops&Robbers, a testbed with multiple characters interacting in an intricate cityscape.