NNRP/1

This is the only current public version line. The sidebar keeps the “Preview” label only because the stable release marker is not frozen yet; for readers and integrators, this page is simply the entry for the current NNRP/1 contract.

If you only remember a few things, start with these four:

1. tensor and token are peer standard profiles. The public layer no longer bakes in fields for any single use case.
2. Connection, session, and operation are three distinct layers. A real-time interaction is not "one request, one response"—it is a continuous submit loop, a result loop, and a control-message path running in parallel.
3. The transport layer is not hard-wired. The protocol supports path probing before the handshake and session migration at runtime, without dropping the connection.
4. The common header is fixed at 40 bytes with self-describing length fields, so the same parsing logic works over QUIC, TCP+TLS, or any other byte-stream transport.

Current Public Contract

1. A payload-agnostic public layer

The public layer handles submission, results, flow control, and interpretation context—things that are common to all real-time AI tasks. tensor- and token-specific fields live inside their respective profiles and schemas, not in the common header. That means adding a new profile does not require touching the public layer.

2. A three-layer execution model with explicit flow control

Connection, session, and operation each have a clear scope: the connection is the transport container, the session holds context and default parameters, and the operation is the lifecycle of one submitted task.

Flow control is part of the protocol: FLOW_UPDATE can target the connection level, a session, or a single operation, and it explicitly tells the host to slow down, pause, or resume—rather than each implementation guessing from timeouts.

3. Transport strategy and session continuity

The protocol does not lock to one transport. Hosts can probe multiple path candidates (QUIC, TCP+TLS, etc.) before committing to a connection, and can migrate a live session to a different transport path later without tearing down the session.

More details on the transport strategy and probing page.

Quick Start Follow the recommended minimal integration path for the current public version. Session and Operation Model See how connection, session, and operation are separated and why the model is not synchronous request-response. Transport Strategy and Probing Understand why the protocol must support transport probing, dynamic path selection, and session migration. Cache Capabilities and Leases See why cache objects, object references, and lease contracts need to stay in the public protocol surface. Schema / Profile Registry See why type-system extension should go through a registry rather than keep inflating public enums. Flow Control and Priority See why backpressure, credit, and scheduling preference cannot stay hidden inside private local logic.