feat(beta/tools): add speak_to_ivr tool auto-injected with ivr_detection

[ibiscp]

Problem

ivr_detection=True automatically injects send_dtmf_events for key presses, but provides no equivalent for verbal responses. IVR systems use both modalities: pressing digits and saying phrases ("say your name", "say yes or no", "say your account number"). Without a built-in tool, every developer building an IVR agent must implement the same boilerplate:

@function_tool()
async def speak_to_ivr(self, context: RunContext, text: str) -> None:
    handle = self.session.say(text, allow_interruptions=False)
    await context.wait_for_playout()
    await handle
    raise StopResponse()

And then instruct the agent via system prompt to always call it after every verbal response. This is the same class of problem as the missing StopResponse in send_dtmf_events — the tool exists conceptually but the correct behavior isn't encoded by default.

Fix

Add speak_to_ivr as a built-in function_tool in beta/tools/ and auto-inject it alongside send_dtmf_events via IVRActivity.tools. The tool calls session.say() with allow_interruptions=False (appropriate for IVR contexts where the agent must finish speaking before the IVR processes the response), waits for playout, and raises StopResponse() to yield the turn back to the IVR.

send_dtmf_events and speak_to_ivr are symmetric IVR interaction primitives — key presses and verbal responses. Both should be available out of the box when ivr_detection=True.

Impact

• No breaking changes. speak_to_ivr is additive.
• Agents using ivr_detection=True will now have speak_to_ivr available automatically. Agents that already define their own equivalent tool should remove it to avoid duplication.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

View in Devin Review to see 4 additional findings.