---
title: "Streaming and lifecycle events"
description: "Render live model output and observe tools, approvals, and delegated work."
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.cortavyn.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Streaming and lifecycle events

Use `stream(DeepRequest)` for a cold run and `resumeStream(threadId, decisions)` after an approval. Both return a `Flow.Publisher<DeepEvent>` and finish with `Completed`, `Interrupted`, or `Failed`. If the configured model implements `StreamingChatModel`, the same stream carries incremental `TextDelta` events; non-streaming models still publish the remaining lifecycle events.

```java
agent.stream(new DeepRequest("research-42", "Summarize the workspace findings."))
    .subscribe(new Flow.Subscriber<DeepEvent>() {
        public void onSubscribe(Flow.Subscription subscription) { subscription.request(Long.MAX_VALUE); }
        public void onNext(DeepEvent event) { render(event); }
        public void onError(Throwable error) { report(error); }
        public void onComplete() { }
    });
```

`DeepEvent` makes the lifecycle observable without depending on a provider-specific event format:

| Event | Meaning |
| --- | --- |
| `TextDelta` | Incremental assistant text from a streaming model turn. |
| `Message` | A complete assistant model message. |
| `ToolCallRequested` / `ToolResult` | A tool call and its resulting message. |
| `TodosUpdated` | The durable task plan changed. |
| `ContextOffloaded` | A large tool result was preserved in the workspace. |
| `SubagentStarted` / `SubagentCompleted` | A delegated task started or returned its report. |
| `ApprovalRequested` | One or more sensitive actions need a decision. |
| `Completed` / `Interrupted` / `Failed` | The terminal outcome. |

For separate nested-agent observability, subscribe to `agent.childStreams()`. Each `DeepAgentChildStream` provides its own event publisher plus focused message, tool-call, output, and nested-child publishers.

Normal `invoke` and `resume` also create internal graph checkpoints. Use `history(threadId)` to inspect those checkpoints; a pending approval is represented as a durable graph interrupt. See [context management](/deep-agent/context-management) for long-running conversation limits.

Source: https://docs.cortavyn.org/deep-agent/runtime-and-streaming/index.mdx