awaiting nested AgentTask in a tool call within parent Agent's on_enter() stops working since v1.6.1

[cchen-dialpad]

Bug Description

I'm trying to migrate to LiveKit agents 1.6.1+ versions, and found some incompatibility between the newer versions and our previous evaluation/testing frameworks based on: https://docs.livekit.io/agents/start/testing/ . Specifically, the behavior of how to await nested AgentTask from a tool call within the parent Agent's on_enter() is changing and we would like your guidance.

I have a minimal reproducible script based on an earlier issue: #5371 (comment) (a related previous PR: #5377) See below for the script and error info.

The following error is obtained with livekit-agents v1.6.4, but the same error appeared since v1.6.1.

ERROR    livekit.agents:log.py:22 Error in _llm_inference_task
Traceback (most recent call last):
  File "/Users/cj/Library/Caches/pypoetry/virtualenvs/agent-eval-2eDZ1Xtk-py3.11/lib/python3.11/site-packages/livekit/agents/utils/log.py", line 17, in async_fn_logs
    return await fn(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/cj/Library/Caches/pypoetry/virtualenvs/agent-eval-2eDZ1Xtk-py3.11/lib/python3.11/site-packages/opentelemetry/util/_decorator.py", line 71, in async_wrapper
    return await func(*args, **kwargs)  # type: ignore
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/cj/Library/Caches/pypoetry/virtualenvs/agent-eval-2eDZ1Xtk-py3.11/lib/python3.11/site-packages/livekit/agents/voice/generation.py", line 145, in _llm_inference_task
    async for chunk in llm_node:
  File "/Users/cj/Library/Caches/pypoetry/virtualenvs/agent-eval-2eDZ1Xtk-py3.11/lib/python3.11/site-packages/livekit/agents/voice/agent.py", line 476, in llm_node
    activity = agent._get_activity_or_raise()
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/cj/Library/Caches/pypoetry/virtualenvs/agent-eval-2eDZ1Xtk-py3.11/lib/python3.11/site-packages/livekit/agents/voice/agent.py", line 415, in _get_activity_or_raise
    raise RuntimeError("no activity context found, the agent is not running")
RuntimeError: no activity context found, the agent is not running
WARNING  livekit.agents:agent_session.py:1463 session is closing, skipping start activity of agent_task1
WARNING  livekit.agents:generation.py:729 ToolError while executing tool: activity doesn't start for agent_task1, likely due to session closing
WARNING  livekit.agents:agent_activity.py:1474 attempting to schedule a new SpeechHandle, but the scheduling_task is not running, the speech will be cancelled
FAILED

Based on the pattern in livekit-agents’ own test for nested AgentTask on Github: https://github.com/livekit/agents/blob/main/tests/test_nested_agent_task.py , I modified the script accordingly to make it work with v1.6.1+. The change is to remove the on_enter() method from TestAgent.

Expected Behavior

The original script should continue to work, or if that's an intended behavioral change (breaking existing tests), a clarification paragraph could be added to the official documentation at https://docs.livekit.io/agents/start/testing/ and https://docs.livekit.io/agents/logic/tasks/

Reproduction Steps

The minimal reproducible script: #5371 (comment) (A related previous PR: #5377)

"""
LIVEKIT_EVALS_VERBOSE=1 pytest -s -o log_cli=true test_nested_task.py
Works for livekit-agents versions: 1.5.19rc1, 1.6.0, but not with 1.6.1

1.6.1 fails maybe due to this PR: https://github.com/livekit/agents/pull/6071

"""

import asyncio
# from typing import override

from livekit.agents import Agent, AgentSession, AgentTask, function_tool
from livekit.agents.llm.tool_context import NamedToolChoice
from livekit.plugins import openai
import pytest


class AgentTask1(AgentTask):
    def __init__(self) -> None:
        super().__init__(instructions="say hello")
        print('after init AgentTask1')

    # @override
    async def on_enter(self) -> None:
        print("AgentTask1 on_enter")
        tool = NamedToolChoice(type="function", function={"name": "run_agent_task_2"})
        await self.session.generate_reply(tool_choice=tool)

    @function_tool
    async def run_agent_task_2(self) -> None:
        print("run_agent_task_2")
        await AgentTask2()
        self.complete(None)


class AgentTask2(AgentTask):
    def __init__(self) -> None:
        super().__init__(instructions="say hello")

    # @override
    async def on_enter(self) -> None:
        print("AgentTask2 on_enter")
        tool = NamedToolChoice(type="function", function={"name": "run_agent_task_3"})
        await self.session.generate_reply(tool_choice=tool)

    @function_tool
    async def run_agent_task_3(self) -> None:
        print("run_agent_task_3")
        await AgentTask3()
        self.complete(None)


class AgentTask3(AgentTask):
    def __init__(self) -> None:
        super().__init__(instructions="say hello")

    # @override
    async def on_enter(self) -> None:
        print("AgentTask3 on_enter")
        await self.session.generate_reply(instructions="say hello")

    @function_tool
    async def goodbye(self) -> None:
        print("goodbye")
        self.complete(None)


class TestAgent(Agent):
    def __init__(self) -> None:
        super().__init__(instructions="say hello")

    # @override
    async def on_enter(self) -> None:
        print("\nTestAgent on_enter")
        tool = NamedToolChoice(type="function", function={"name": "run_agent_task_1"})
        await self.session.generate_reply(tool_choice=tool)

    @function_tool
    async def run_agent_task_1(self) -> None:
        print("run_agent_task_1")
        await AgentTask1()


@pytest.mark.asyncio
async def test_assistant_greeting() -> None:
    llm = openai.LLM(model="gpt-5.4-mini")
    async with AgentSession(llm=llm) as session:
        await session.start(TestAgent())
        print("user turn 1")
        await session.run(user_input="hi")
        print("user turn 2")
        await session.run(user_input="hi there")  # hangs here
        print("user turn 3")
        await session.run(user_input="sayonara")

Based on the pattern in livekit-agents’ own test for nested AgentTask on Github: https://github.com/livekit/agents/blob/main/tests/test_nested_agent_task.py , I modified the script accordingly to make it work with v1.6.1+. The change is to remove the on_enter() method from TestAgent.

Operating System

macOS Sequoia

Models Used

llm = openai.LLM(model="gpt-5.4-mini")

Package Versions

livekit-agents == v1.6.1 or newer

Session/Room/Call IDs

No response

Proposed Solution

Additional Context

No response

Screenshots and Recordings

No response