fix(livekit-agents): don't drop conversation work started by a non-blocking tool

[fxhxdxd]

Fixes #6658

The trigger is the await ctx.update("") at the top of the reported tool, not update_agent() itself.

ctx.update() resolves the executor's first_update_fut, and _ToolExecutor.execute() ends with return await first_update_fut. So dispatch gets its answer and the turn finishes while exe_task, the rest of the tool body, is still running. The ctx.foreground() block and the session.update_agent() inside it happen after the run is already over.

RunResult stops recording as soon as _done_fut resolves, and that happens when every watched handle is done. _enqueue_reply registers self._reply_task with the run state, but nothing registered exe_task, so the run had no idea the tool was still going and Agent2 never got a chance to register its on_enter speech handle.

It works in production because the session outlives run(). Under the test framework run() returns, the session closes, and the handoff is dropped:

WARNING  livekit.agents:agent_session.py:1657 session is closing, skipping start activity of forwarded

The fix

Let an active RunResult wait on the tool body, mirroring what _enqueue_reply already does for the deferred reply.

The wait is bounded. ctx.update() is also how a tool answers early and keeps doing long background work, and that kind of tool should not hold run() open until it finishes. So the run watches a bounded waiter rather than exe_task itself, same shape as the realtime auto tool reply in agent_activity.py. Timeout is a module constant, NON_BLOCKING_TOOL_RUN_TIMEOUT = 5.0, matching the 5s already used there. Happy to change the value or make it configurable.

Nothing changes outside session.run(): the watch only happens when there is an active run state, same guard the auto reply uses.

Tests

tests/test_run_result_handoff_in_tool.py, four cases.

The reported shape, a tool that calls ctx.update("") then hands off inside ctx.foreground():

• one asserts the forwarded agent's ChatMessageEvent reaches result.events
• one walks the sequence the docs suggest: function call, function call output, agent handoff, assistant message

Both fail on main (only the two function-call events arrive) and pass with the change. Drop the ctx.update("") line and they pass either way, which isolates the trigger.

Then the two cases the bound exists for:

• a tool that reports progress and then sleeps for an hour must not gate the run. This one fails with a TimeoutError if the watch is unbounded.
• a tool that raises after going non-blocking is reported through the normal tool lifecycle, so the waiter must not turn it into a failed run

Full make unit-tests locally: 1260 passed. The 9 errors I see are all in tests/test_room.py and are identical on a clean checkout, so they are a local env thing, not this patch. ruff, ruff format and mypy -p livekit.agents are clean.