Stop vs disconnect - why canceling AI streaming is harder than it looks

The discussion highlights a critical issue in AI chat applications using the Vercel AI SDK, where invoking the `chat.stop()` function or closing a browser tab only terminates the HTTP connection without halting the server-side process, leading to continued billing for token consumption. This behavior is not erroneous but inherent to resumable stream architectures, which are designed to withstand disconnections without stopping the process. The solution involves creating a dedicated stop endpoint with features such as idempotency checking, persistence of partial assistant snapshots, and backend-specific cancellation logic, none of which the SDK inherently provides. In contrast, the Ably AI Transport offers a more efficient approach by leveraging a bidirectional session model, allowing explicit cancellation signals to be sent, thus handling race conditions and disconnect distinctions more effectively. This model eliminates the need for a separate stop endpoint by integrating cancel, stop, and redirect as first-class session signals, ensuring that intent, rather than mere disconnection events, guides the process termination.