Streaming

“Streaming” in Shiro means seeing the run while it is still moving. Provider token deltas are one signal. Tool starts, approval waits, handoffs, and completion events are equally important — and they are what Studio timelines are built from.

Why this is broader than provider streaming#

If you only stream model text, operators miss:

• which tool is hanging
• that an approval is blocking progress
• that a handoff changed the active agent

Subscribe to the engine’s event bus / TraceManager, or open Studio while developing.

Observing events#

import { Engine, ShiroEventType, TraceManager } from "@shiro-sdk/core";

const traces = new TraceManager();
const engine = new Engine({ events: traces });

// Execute as usual — exporters and Studio consume the same event stream.
await engine.execute(agent, input);

Provider adapters may also implement stream() for token-level deltas. Use that for UI typing effects; use runtime events for operational truth.

When to care#

Best practices#

• Render event types differently from text deltas
• Keep runId on every UI update
• Treat approval waits as first-class states
• Do not assume the first event is model text

Common mistakes#

• Streaming provider text only and calling the run “observable”
• Hiding tool failures until the final message
• Dropping run ids from client events

Related#

• Tracing
• Studio
• Providers