Observability

Runlet emits structured runtime events through an event sink.

Basic observer

Use InMemoryObserver to capture emitted events:

from runlet import InMemoryObserver, Runtime


observer = InMemoryObserver()
runtime = Runtime(event_sink=observer)

Common event types

• run.started
• context.budget_checked
• model.requested
• model.completed
• model.stream.started
• model.stream.delta
• model.stream.reasoning_delta
• model.stream.completed
• tool.started
• tool.completed
• human.requested
• human.responded
• human.response_rejected
• run.interrupted
• run.resumed
• run.completed
• policy.stopped

Human-in-the-loop events

When a tool requires approval or calls ask_human(), Runlet saves a checkpoint and then emits human.requested followed by run.interrupted. When the application resumes the checkpoint, it emits human.responded and run.resumed. Invalid or stale responses emit human.response_rejected.

Human event payloads identify the request, checkpoint, request kind, and where applicable the response action. They intentionally omit prompts, choices, and submitted values so applications can observe the flow without exposing human input content in routine event sinks.

Example

result = await runtime.run(agent, "用一句中文介绍 Runlet。")

for event in observer.events:
    print(event.type, event.payload)

Why events matter

Events give applications a stable visibility layer without coupling tracing, logging, or metrics to control flow.