Streaming

AnthropicAdapter and OpenAIAdapter tee stream=True calls instead of forcing non-streaming (Groq and OpenRouter inherit this since they subclass OpenAIAdapter). Chunks flow through to your agent unmodified while the assembled text/tokens are recorded, with raw_response={"__stream__": True, "chunks": [...]}.

Replayed streams

Replay rebuilds the recorded chunks into real SDK chunk/event objects and yields them back incrementally — the agent consumes them exactly like a live stream, with zero API calls. Replaying a streaming recording against a non-streaming request (or vice versa) raises ReplayError with a "shape mismatch" message.

Unconsumed-stream finalization

A stream that is never iterated and never closed is finalized automatically at recorder/asserter exit, but consuming or closing it promptly is still recommended so events appear in call order.

Limitations

Not yet supported:

• The client.messages.stream() context-manager helper.
• Streamed OpenAI Responses-API calls.

Mistral still forces stream=False on every call.

See examples/demo_streaming.py for a full runnable walkthrough of recording and replaying a streaming agent, and examples/demo_async.py for the async-client version.