In face-to-face communication, the communicative function of the spoken text is clarified through supporting verbal and nonverbal discourse devices. In computer-mediated communication, the mediating channel may not be able to carry all those devices. To ensure the original intent gets communicated effectively, discourse tags can be embedded in a message to encode the communicative function of text given the context in which it was produced. The receiving client can then generate its own supporting discourse devices from the received tags, taking into account the receiver's context. Spark is a synchronous CMC architecture based on this concept of a message transformation, where an outgoing text message gets automatically annotated with discourse function markup that is then rendered as nonverbal discourse cues by a graphical avatar agent on the receiver side. A user study performed on a derived application for collaborative route planning demonstrates the strength of the approach.