[Bug]: `TerminateTarget` ignored by `AutoPattern`

[UMainLove]

Describe the bug

DESCRIPTION

When using AutoPattern with set_after_work(TerminateTarget()) on an agent, the termination signal is silently ignored. The workflow runs to max_rounds instead of stopping after the agent speaks.

SEVERITY - HIGH 🔴

TerminateTarget is a core API primitive. It silently does nothing instead of failing visibly, making the bug hard to diagnose.

MITIGATIONS

The workaround found consists in a custom is_termination_msg callback on the GroupChatManager via group_manager_args, which bypasses the after_work mechanism entirely.

Steps to reproduce

AG2 version: 0.11.4

python

from autogen import ConversableAgent
from autogen.agentchat.group.patterns import AutoPattern
from autogen.agentchat.group.multi_agent_chat import initiate_group_chat
from autogen.agentchat.group.targets.transition_target import TerminateTarget

llm_config = {"model": "factory-model_name-version-release", "api_type": "factory"}

agent_a = ConversableAgent(
    name="agent_a",
    system_message="Say hello in one sentence.",
    llm_config=llm_config,
    human_input_mode="NEVER",
)

agent_b = ConversableAgent(
    name="agent_b",
    system_message="Say goodbye in one sentence.",
    llm_config=llm_config,
    human_input_mode="NEVER",
)

# agent_b should terminate the workflow after speaking

agent_b.handoffs.set_after_work(TerminateTarget())
user = ConversableAgent(name="user", llm_config=None, human_input_mode="TERMINATE")

pattern = AutoPattern(
    initial_agent=agent_a,
    agents=[agent_a, agent_b],
    user_agent=user,
    group_manager_args={"llm_config": llm_config},
  )

result, ctx, last = initiate_group_chat(
    pattern=pattern,
    messages="Start the conversation.",
    max_rounds=10,
)

# Check how many rounds actually ran                                                                                                                                           
print(f"Messages: {len(result.chat_history)}")

# Expected: ~3-4 (user prompt, agent_a, agent_b, done)                                                                                                                         
# Actual: runs to max_rounds=10

Model Used

claude-sonnet-4-5-20250929 - claude-haiku-4-5-20251001

Expected Behavior

After agent_b speaks, TerminateTarget should stop the workflow. The chat history should contain ~3-4 messages.

Screenshots and logs

No response

Additional Information

SUGGESTED FIX

Use a sentinel to distinguish "no after_work configured" from "terminate":

python

_NO_AFTER_WORK = object()
                                                                                                                                                                                 
def _evaluate_after_works_conditions(...):
  if not hasattr(agent, "handoffs") or not agent.handoffs.after_works:
      return _NO_AFTER_WORK  # ← sentinel instead of None

# ... rest unchanged ...
                                                                                                                                                                                 
# In determine_next_agent:                                                                                                                                                   
after_works_result = _evaluate_after_works_conditions(...)                                                                                                                     
  if after_works_result is not _NO_AFTER_WORK:
      return after_works_result  # None (terminate) passes through correctly

This lets None from TerminateTarget propagate to group_transition() → _prepare_and_select_agents() → NoEligibleSpeakerError → clean termination.

NOTE: OnContextCondition with AgentNameTarget works correctly because it resolves to a non-None agent reference. Only TerminateTarget is affected because its termination signal maps to None.