Streaming IO

Pain Signals — You Need This Pattern When:

• TTFT (time to first token) matters; users wait silently for 30+ seconds
• The harness exposes a UI where typing-style output is expected
• Long generations need to be cancellable mid-stream (user changes mind, model goes off-rails)
• Tool calls should dispatch as soon as the model commits, not after the full turn completes
• Structured output is large; partial parse + early validation catches errors faster

Streaming changes the harness from request/response to event pipeline. Every assumption about "I have the full response now" must be revisited.

Core Principle

The model emits a stream of events; the harness threads those events through transformation, dispatch, and rendering. Each event has a type and may carry partial content.