neuron-streaming

Neuron AI Streaming

This skill covers how output leaves a running agent or workflow and reaches a consumer: the native chunk stream, portable progress events, protocol adapters for UI libraries, and channels for push delivery.

The Mental Model

A Workflow segment produces a sequence of live objects while it runs. Four independent concerns decide what happens to them:

Concern	Question it answers	Component
Source	What is emitted?	Provider chunks and events yielded by nodes
Shape	What does the wire format look like?	StreamAdapterInterface via setStreamAdapter()
Destination	Where does it go?	Pull iteration over the generator, or eager delivery through StreamingChannelInterface via setChannel()
Encoding	How does it become bytes?	SSEEncoder on the HTTP edge; a channel encodes for its own transport

Adapter and channel compose. The adapter decides the shape, the channel the destination. The Workflow converts each item once into a ProtocolEvent. run() consumes the pipeline eagerly; events() and Agent::stream() always return lazy generators. Consumption also delivers adapted events to a configured channel. It never frames bytes: the edge that owns the transport encodes it.

Live output is ephemeral. Nothing yielded during a segment is stored in persistence or replayed when a completed step is restored. Chat history is the record the UI reconciles from. Never make correctness depend on a client receiving a streamed item.

Resources constructed per segment

Use the streamAdapter() and channel() hooks, or pass a factory to setStreamAdapter() / setChannel(); the setters take factories only:

$agent->setStreamAdapter(fn () => new AGUIAdapter($input['threadId'], $input['runId'] ?? null));
$state = $agent->run($request);

These factories run only for an owned execution and return a new adapter or channel for each segment: both hold the state of one segment's stream. There is no Workflow-mutating preparation callback; saved outcomes and idle polls skip resource construction. The AG-UI run ID is the client's per-request ID from its input, not Neuron's durable run ID. Lazy calls capture input at creation and execute during iteration. Setters configure the definition; an active execution keeps its already resolved resources and graph. This rule applies to Workflow, Agent and RAG configuration.