[noooooooookro]

Fixes #7151

What changed

Google Realtime can send a single response containing both terminal server content and a tool call. The receive loop now defers generation finalization for that server content, keeping the function channel open until the tool call is delivered. Existing responses without a tool call retain their current completion behavior.

Validation

• python -m pytest tests/test_plugin_google_realtime.py -q (19 passed)
• ruff check livekit-plugins/livekit-plugins-google/livekit/plugins/google/realtime/realtime_api.py tests/test_plugin_google_realtime.py
• ruff format --check livekit-plugins/livekit-plugins-google/livekit/plugins/google/realtime/realtime_api.py tests/test_plugin_google_realtime.py

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[longcw]

The behaviour is right. Maybe a cleaner shape: finalize here instead of in the two handlers. Both _handle_server_content and _handle_tool_calls are only called from this loop, so the loop can own the decision and the defer_completion flag goes away.

if response.server_content:
    self._handle_server_content(response.server_content)
if response.tool_call:
    self._handle_tool_calls(response.tool_call)
if ((sc := response.server_content) and sc.turn_complete) or response.tool_call:
    self._mark_current_generation_done()

Then drop the turn_complete finalize at the end of _handle_server_content and the one at the end of _handle_tool_calls. Same order as now: content, then the calls, then close, in one iteration.

[himanshumehta]

@noooooooookro, I opened #8376 for the same crash. It implements @longcw's review suggestion on this PR: the receive loop finalizes the generation once, after both handlers, so the defer_completion flag isn't needed. It also adds a test that a lone tool_call still ends the generation, so BLOCKING tools can't stall. This PR has had no activity since Sep 9 and its CLA is unsigned. If you'd rather finish it here, say so and I'll close #8376 and post the tests on this PR instead.