The channel market model is a tool for making communication systems dependable. It is a generalization of the network stack model: Where the network stack model uses network graphs as the fundamental abstraction and layering as the compositional structure, the channel market model starts smaller, using channels as the fundamental abstraction, and builds more freely, using a marketplace for composition. In a channel market, a communication system is a channel transformer, which uses some of the channels offered in the market to implement new channels and offer them in turn. The model developed out of work on connection recovery for the Transmission Control Protocol (TCP), as a tool for understanding the complex dynamics of the standard network stack. In this dissertation, I apply the lessons learned from the channel market model back to TCP, and in particular to application-driven connection recovery. Application-driven connection recovery is a technique by which a fault-tolerant application can recover and migrate connections, leveraging middleware to avoid modifications to its TCP implementation. The middleware depends on very little state, making application-driven recovery a lightweight and fast technique. To demonstrate what is possible, I present recovery middleware using both formal and empirical methods. Formally, I present specifications of TCP and recovery middleware. The specification of TCP also serves as an introduction to the details of the protocol; to serve that purpose, iv the it follows a novel decomposition I developed for my own understanding while working on application-driven connection recovery. Using both specifications, I prove that the simple middleware is sufficient for a failing and recovering TCP to refine non-failing TCP. Empirically, I present TCPR, an implementation of recovery middleware. I describe the systems problems that arise from masking connection failure and migrating without modifying TCP or sockets, particularly where the common interfaces violate the TCP specification or unnecessarily restrict what state is available to an application. I also present the results of a study of the Border Gateway Protocol (BGP), highlighting the severity of the routing disruptions that can be avoided only with connection recovery. The channel market model's role in the presentation displays its usefulness in both ways that a scientific model can be useful: For understanding existing complexity (as in the decomposition of TCP), and for simplifying the design of the new (as in TCPR). I wrap up by presenting two design principles that have emerged from using channel markets: The separation of justification and the haggling principle. The channel market model and its design principles are useful tools beyond making TCP more dependable, and they stand waiting for future work. v