As robotics expands into manufacturing, logistics, and services, long-horizon tasks (i.e., characterized by extended duration and complex subtask dependencies) pose significant perception and coordination challenges, especially in multi-robot systems. Current approaches, often reliant on single-source data or predefined task structures, struggle with task complexity and deployment heterogeneity, which hinders a transition from demonstrations to multi-robot collaboration. In this paper, we propose a multi-robot collaborative framework termed DeCo, which builds on single-robot long-horizon task demonstrations, systematically addressing task segmentation, subtask modeling, and deployment-aware execution to enable multi-robot collaboration. Our method achieves robust subtask boundary detection by integrating visual features, global optical flow, and local motion cues. Subtasks are represented via structured preconditions and effects, enabling automated construction of a task dependency graph for sequential or parallel execution. A trajectory adaptation strategy aligns robot orientations through coordinate transformations for diverse deployments. We further present DeCoBench, a benchmark comprising 48 long-horizon manipulation tasks to evaluate the transition from single-robot demonstrations to collaborative multi-robot execution. Experiments demonstrate superior segmentation accuracy, dependency modeling, and execution efficiency, validating the effectiveness of the framework in long-horizon tasks.