One of the most critical challenges in human-robot collaborative work settings is ensuring the health and safety of the involved human workers. We propose to integrate task-level planning with semantically represented workplace safety rules that are published by regulatory bodies, meaning that our system can adapt to produce different variants of a product while respecting workplace safety regulation. Our prototype system interacts with human workers and machine agents via Activity Streams and a speech synthesis interface, and we have shown that its SPIN reasoning engine can scale to scenarios that incorporate complex products and many agents. The current system state and action logs of the agents and products are easily observable using a dashboard interface. The semantic models were evaluated by five experts in workplace safety and process engineering who expressed confidence about using, maintaining, and even extending the models themselves after only negligible training, a crucial factor for the real-world adoption of such systems.