Problem

The GraphFlowManager was allowing _StopAgent to speak and add termination messages to the conversation context, which polluted the message history that other agents could see. This violated the principle of clean agent interactions.

Before this fix:

# _StopAgent would produce messages like this in the conversation:
result.messages = [
    TextMessage(source="user", content="Task"),
    TextMessage(source="A", content="Response from A"),
    TextMessage(source="B", content="Response from B"), 
    StopMessage(source="DiGraphStopAgent", content="Digraph execution is complete")  # ❌ Pollutes context
]

After this fix:

# Clean conversation with no stop agent messages:
result.messages = [
    TextMessage(source="user", content="Task"),
    TextMessage(source="A", content="Response from A"),
    TextMessage(source="B", content="Response from B")  # ✅ Clean context
]
# Termination still occurs but without polluting conversation

Solution

1. Removed _StopAgent entirely - No longer added to participants list
2. Updated GraphFlowManager.select_speaker() - Directly signals termination when self._ready is empty instead of selecting stop agent
3. Simplified GraphFlow constructor - Removed StopMessageTermination and stop agent logic
4. Preserved all functionality - Same termination behavior and message content, just without the polluting agent message

Key Changes

• GraphFlowManager.select_speaker(): Now handles termination directly when no ready agents remain
• GraphFlow.__init__(): No longer adds _StopAgent to participants or uses StopMessageTermination
• Removed unused code: Deleted _StopAgent class and cleaned up imports
• Updated tests: All tests now expect clean conversation history without stop agent messages

Benefits

✅ Clean conversation context - No stop agent messages pollute agent contexts
✅ Same termination behavior - Graph execution still stops correctly when complete
✅ Preserved functionality - All existing features work exactly the same
✅ Better performance - One fewer participant to manage
✅ Cleaner codebase - Removed unnecessary _StopAgent complexity

Verification

Created comprehensive tests that verify:

• No StopMessage instances appear in conversation history
• Agents don't see stop agent messages in their context
• Termination still occurs at the correct time
• All existing functionality preserved (resume, state management, etc.)

Fixes #6751.

💬 Share your feedback on Copilot coding agent for the chance to win a $200 gift card! Click here to start the survey.