NNRP/1 Quick Start

From a host-integration perspective, the minimal working path is:

1. Establish a reliable byte-stream connection and complete CLIENT_HELLO / SERVER_HELLO_ACK.
2. Create a session with SESSION_OPEN / SESSION_OPEN_ACK, including the default profile, schema, and budget window.
3. Send FRAME_SUBMIT or another submit-class message to create an operation inside that session.
4. Run a result pump in parallel and keep receiving RESULT_PUSH / RESULT_DROP / RESULT_HINT / FLOW_UPDATE.
5. Adjust your sending behavior according to result state and FLOW_UPDATE.

The three most common mistakes are:

1. Treating NNRP as synchronous: the correct integration runs three paths in parallel—a loop that sends work, a loop that reads results, and a path that processes FLOW_UPDATE and other control messages.
2. Confusing session and operation: a session is a persistent context container (like a conversation), an operation is a single task inside it. Sessions are opened once; operations are submitted many times.
3. Skipping profile/schema binding: payload field layouts are defined by the profile and schema, not by any language runtime's private conventions. Guessing at byte offsets is unreliable.

Minimum host requirements:

1. Keep a send loop and a result-reading loop that run independently without blocking each other.
2. Record each session's profile_id / schema_id / schema_version so schema mismatches are easy to diagnose.
3. Handle the four visible result types: partial (intermediate, safe to display), terminal (final frame), drop (discarded), degraded (lower-quality output).
4. Apply FLOW_UPDATE at the right scope—connection-level updates affect all sessions, session-level ones affect only that session.