Rework telemetry to support durable execution case

[anbuzin]

1. convert Span to pydantic so they can travel as json
2. introduce Sink that all spans go to. default sink is the adapter registry. users can explicitly swap it out with a different sink using a context manager. in the workflow body:

sink = ai.telemetry.Collector()
with use_sink(sink):
    # do the stuff
    
await send_spans_step(sink.finished)  # send all spans in a step

3. spans no longer have methods that immediately send them to adapters. instead, you can mutate span's attributes, then call span.push to send the updated version to the sink.

the async with span("foo") as sp api didn't change much, actually, it's just different under the hood. the Adapter api hasn't changed at all.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ai-python	Ready	Preview, Comment	Jul 17, 2026 10:12pm