Streaming

Agent.stream() returns a reactive Flux<Event> that lets callers observe every reasoning step, tool-call result, and final reply in real time, rather than waiting for call() to return a single Msg.

Basic Usage

// Works with both ReActAgent and HarnessAgent
Flux<Event> events = agent.stream(
        List.of(Msg.builder().role(MsgRole.USER).textContent("Analyze this log").build()),
        StreamOptions.defaults());

// Subscribe and print each event
events.subscribe(event ->
    System.out.printf("[%s|last=%s] %s%n",
            event.getType(), event.isLast(),
            event.getMessage().getTextContent()));

// Blocking collect (for testing / batch pipelines)
List<Event> all = events.collectList().block();

With RuntimeContext (HarnessAgent):

RuntimeContext ctx = RuntimeContext.builder()
        .sessionId("my-session").userId("alice").build();

Flux<Event> events = agent.stream(msgs, StreamOptions.defaults(), ctx);

StreamOptions.defaults() enables all event types in incremental mode.

Event Types (EventType)

Type	When	Typical content
`REASONING`	Each reasoning turn (may have multiple chunks)	Model text / thinking chain / tool_use calls
`TOOL_RESULT`	After each tool execution	`ToolResultBlock` (tool name, id, output)
`HINT`	After RAG / memory retrieval	Context text injected into the model
`SUMMARY`	When `maxIters` is reached	Iteration summary text
`AGENT_RESULT`	Final reply ready	Same as `call()` return value; not included in stream by default
`ALL`	Placeholder for all the above (except `AGENT_RESULT`)	—

StreamOptions options = StreamOptions.builder()
        .eventTypes(EventType.REASONING, EventType.TOOL_RESULT)
        .build();

Distinguish chunks from final results

For the same REASONING message the model first streams several incremental chunks, then emits a final event with the complete text. Use Event.isLast():

events.filter(e -> e.getType() == EventType.REASONING)
      .subscribe(e -> {
          if (e.isLast()) {
              System.out.println("Reasoning done: " + e.getMessage().getTextContent());
          } else {
              System.out.print(e.getMessage().getTextContent());  // live delta
          }
      });

Incremental vs. Full-text Mode

`incremental`	Behaviour
`true` (default)	Each chunk contains only the new text delta; caller concatenates.
`false`	Each chunk carries the full accumulated text up to that point — convenient for direct rendering.

StreamOptions options = StreamOptions.builder()
        .incremental(false)   // recommended for UI rendering
        .build();

Reasoning / Summary Content Filtering

Some models emit both intermediate delta chunks and a final consolidated result inside the same REASONING event group. Control them independently:

StreamOptions options = StreamOptions.builder()
        .eventTypes(EventType.REASONING)
        .includeReasoningChunk(false)   // skip intermediate deltas
        .includeReasoningResult(true)   // keep only the final reasoning result
        .build();

Similarly for summaries:

StreamOptions options = StreamOptions.builder()
        .eventTypes(EventType.SUMMARY)
        .includeSummaryChunk(false)
        .includeSummaryResult(true)
        .build();

Error Handling

stream() follows standard Reactor error semantics; use onErrorResume to handle:

events.onErrorResume(e -> {
    log.error("Stream error", e);
    return Flux.empty();
}).subscribe(...);

When a tool execution fails, the framework converts the exception to a TOOL_RESULT error string. The onError signal is not propagated to the parent stream, so the overall stream is not terminated prematurely.

Integration with Spring WebFlux / SSE

Flux<Event> bridges directly into Server-Sent Events:

@GetMapping(value = "/chat", produces = MediaType.TEXT_EVENT_STREAM_VALUE)
public Flux<ServerSentEvent<String>> chat(@RequestParam String message) {
    return agent.stream(
                    List.of(Msg.builder().role(MsgRole.USER).textContent(message).build()),
                    StreamOptions.defaults())
            .map(event -> ServerSentEvent.<String>builder()
                    .event(event.getType().name().toLowerCase())
                    .data(event.getMessage().getTextContent())
                    .build());
}

When using HarnessAgent and you need to forward child-agent source metadata to the frontend, serialize event.getSource() as well — see Harness Subagent Streaming.

StreamOptions Reference

StreamOptions options = StreamOptions.builder()
        // Event types to receive (default: ALL, excluding AGENT_RESULT)
        .eventTypes(EventType.REASONING, EventType.TOOL_RESULT, EventType.AGENT_RESULT)

        // Incremental mode: true = push deltas (default), false = push full accumulated text
        .incremental(true)

        // Reasoning content filters (both default true)
        .includeReasoningChunk(true)
        .includeReasoningResult(true)

        // Summary content filters (both default true)
        .includeSummaryChunk(true)
        .includeSummaryResult(true)

        .build();

Hook — Insert custom logic at reasoning / tool-call lifecycle points
Model — Underlying streaming model configuration
Harness Subagent Streaming — Child-agent event forwarding and EventSource fields when using HarnessAgent

Quick Start

Features

Multi Agent

Harness

Basic Usage

Event Types (EventType)

Distinguish chunks from final results

Incremental vs. Full-text Mode

Reasoning / Summary Content Filtering

Error Handling

Integration with Spring WebFlux / SSE

StreamOptions Reference

​Basic Usage

​Event Types (EventType)

​Subscribe to specific types only

​Distinguish chunks from final results

​Incremental vs. Full-text Mode

​Reasoning / Summary Content Filtering

​Error Handling

​Integration with Spring WebFlux / SSE

​StreamOptions Reference

​Related Documents

Basic Usage

Event Types (EventType)

Subscribe to specific types only

Distinguish chunks from final results

Incremental vs. Full-text Mode

Reasoning / Summary Content Filtering

Error Handling

Integration with Spring WebFlux / SSE

StreamOptions Reference

Related Documents