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

# Streaming and human approval

> Consume state, events, model messages, and custom progress in real time, then resume safely

## Stream modes

| Mode       | Payload                                     | Typical use                 |
| ---------- | ------------------------------------------- | --------------------------- |
| `values`   | Complete state after each commit            | UI state mirrors, debugging |
| `updates`  | Node/superstep state deltas                 | Lightweight progress        |
| `events`   | Lifecycle events with sequence and metadata | Tracing, audit, Studio      |
| `custom`   | Raw values from `stream_writer(value)`      | Tokens, business progress   |
| `messages` | `(AIMessageChunk/AIMessage, metadata)`      | Chat token streams          |

```python theme={null}
async for mode, chunk in graph.astream(
    input_state,
    context={"tenant": "acme"},
    stream_mode=["updates", "custom", "messages"],
):
    print(mode, chunk)
```

Custom output wakes the consumer immediately; it does not wait for the node to finish or the superstep to commit:

```python theme={null}
def retrieve(state, runtime):
    runtime.stream_writer({"stage": "searching"})
    docs = search(state["query"])
    runtime.stream_writer({"stage": "found", "count": len(docs)})
    return {"documents": docs}
```

Streaming is an observation channel and does not participate in reducers. Closing the iterator cancels unfinished tasks and propagates the cancellation token.

## Dynamic interrupts

```python theme={null}
from lingxigraph import interrupt


def approve(state):
    decision = interrupt({
        "kind": "approval",
        "action": state["proposed_action"],
    })
    return {"approved": bool(decision["approved"])}
```

The graph needs a checkpointer and stable `thread_id`. At `interrupt()`, it saves a checkpoint and pauses. On resume, the node replays from its beginning, so side effects before the interrupt must also be idempotent.

```python theme={null}
from lingxigraph import Command

config = {"configurable": {"thread_id": "ticket-42"}}
graph.invoke(initial, config=config)
graph.invoke(Command(resume={"approved": True}), config=config)
```

In Agent Server, the original run becomes `paused`. Resume creates a new run pinned to the same graph/version/config/context:

```bash theme={null}
curl -X POST http://localhost:8124/v1/runs/$RUN_ID/resume \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"resume":{"approved":true}}'
```

<Warning>
  Show reviewers only the fields they need. Never put secrets, tokens, or complete sensitive state in interrupts, events, or traces.
</Warning>