Add support for GPT-Realtime-2.0

[ShayneP]

Summary

Update realtime agent output handling for Realtime 2.0 responses and fix transcript synchronization races around overlapping segment lifecycle events.

Realtime 2.0 can produce multiple message items for a single response. The Agents output stack exposes playback through a shared AudioOutput segment contract, so this PR forwards realtime message outputs sequentially through the sink. That keeps playback-start/playback-finished events attributable to the correct message and avoids adding or truncating the wrong assistant item during interruption.

This PR also hardens TranscriptSynchronizer segment handling when audio/text flush timing is not perfectly aligned.

Changes

• Support multiple realtime message items in one generation.
• Forward each realtime message’s audio/text output to completion before starting the next message’s sink forwarding.
• Wait for per-message audio playout before registering the next message’s playback listener.
• Preserve existing single-message realtime behavior for current providers.
• Fix transcript synchronizer stale-duration handling by resetting per-segment pushed audio duration during flush.
• Keep pending/finalizing transcript segment impls alive until their text/audio inputs are complete.
• Emit interrupted or already-ready playback-finished events synchronously to preserve existing state transition ordering.
• Delay non-interrupted synced playback-finished events only when needed to include complete synchronized transcript text.
• Apply pause/resume to active, pending, and finalizing transcript segments.
• Add focused regression tests for transcript segment advancement, delayed text completion, and pause/resume behavior.

Compatibility

This is intended to be public-API compatible. It does not change the AudioOutput or TextOutput interfaces or event payload shapes.

Existing realtime models emitted a single message item per response, so the sequential forwarding path preserves prior behavior while correctly handling the new Realtime 2.0 multi-message response shape.

The transcript synchronizer changes are internal lifecycle fixes. The main observable timing change is that a non-interrupted synced playback_finished event may wait for text/audio inputs to complete when playback finishes before transcript input drains. Interrupted and already-ready playback events remain synchronous.

Testing

• make check
• make unit-tests
• uv run pytest tests/test_agent_session.py tests/test_transcript_synchronizer.py -q
• git diff --check

make unit-tests result: 635 passed, 2 skipped

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[tinalenguyen]

tested it and it works well, though i think the start/stop speaking_at times recorded in the chat messages aren't accurate 🤔 all messages in the same generation would have the same started_speaking_at timestamps for the chatmessages

[longcw]

should we call perform_audio_forwarding once for multiple segment instead of calling it multiple times in one response, like merge the streams from generation_ev.message_stream? in that way we don't need to change the output and synchronizer?

[longcw]

btw, I didn't see multiple messages in a single response during my testing, is there a way to trigger that or it's just random?

update: I saw multiple segment with asking it to do so, like tell me a story with two parts