fix: restore speaking state when releasing playout pause

[chenghao-mou]

Stacked on #7083.

A false-interruption pause can start just before a parent response invokes a transfer tool. #7083 releases the pause when the tool calls disallow_interruptions(), but audio resumed while the session still reported listening and recognition reported no agent speech.

State sequence:

H1 is the parent response that includes the tool call. H2 is the transfer message produced by session.say().

Event	Before #7083	#7083 only	#7083 + #7087
H1 starts playing	Session: speaking _agent_speaking=True	Same	Same
False-interruption pause starts	H1 pauses Session: listening _agent_speaking=False Saved state: speaking	Same	Same
disallow_interruptions()	H1 stays paused State stays listening / False	H1 resumes State incorrectly stays listening / False	State returns to speaking / True Then H1 resumes
Remaining H1 audio	Blocked in this path	Audible while state is listening / False	Audible while state is speaking / True
H2 starts	Blocked behind H1	Starts after H1 First frame corrects the state	Starts after H1 No state correction needed

Affected application pattern:

@function_tool
async def transfer_to_live_agent_tool(run_context: RunContext[SessionData]) -> None:
    run_context.disallow_interruptions()
    run_context.session.input.set_audio_enabled(False)
    await run_context.session.say(
        text=_TRANSFER_MSG,
        allow_interruptions=False,
    )
    await transfer_to_live_agent_and_shutdown(...)

Deterministic reproduction

The committed test delays the transfer tool until a parent response is playing, then starts a false-interruption pause before the tool runs:

actions = FakeActions()
actions.add_user_speech(0.5, 2.5, "Connect me to an agent.")
actions.add_tts(5.0, input="Happy to connect you.", ttfb=0.1, duration=0.2)
actions.add_user_speech(4.0, 8.0, "")
actions.add_tts(1.0, input=TRANSFER_MESSAGE)

session = create_session(actions, can_pause_audio=True)
await run_session(session, DelayedTransferAgent())

Run it from the repository root:

uv run pytest tests/test_disallow_interruptions_pause.py \
  -k restores_paused_parent_speaking_state --unit

Without the fix, the callback at audio_output.resume() observes ("listening", False). With the fix, it observes ("speaking", True).

Addresses AGT-3421

Initial prompt and agent context

Model: GPT-5.6

wait, did you make the change in 7083? Can you separate the change into another PR that targets 7083?

Low: #7087’s test checks state after disallow_interruptions() returns. It does not prove that state restoration occurs before audio_output.resume(). A resume callback assertion would preserve this ordering invariant. See test_disallow_interruptions_pause.py.

also you should update the pr description because it was not clear enough. Maybe some event sequence + example function code and script (folded) for reproduction.

False SOS pauses playout

vad speech might not be false.

VAD SOS itself does not pause, it is min interruption duration that pauses or min words.

you can just say false interruption pause starts

[devin-ai-integration]

✅ Devin Review: No Issues Found

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