Fix SupervisorAgent crash when agent response is a dict

[gavksingh]

Summary

• Fixed 'dict' object has no attribute 'content' error in SupervisorAgent.send_message() that occurs when a team agent's process_request() returns a dict instead of a ConversationMessage
• Added type guards in both send_message() and process_agent_streaming_response() to handle both ConversationMessage and dict response formats
• Added 6 new tests covering dict response handling, streaming with dict final_message, and regression tests

Details

The send_message() method directly accessed response.content[0].get('text', '') assuming response is always a ConversationMessage. However, in certain scenarios (e.g., tool use with some providers, or follow-up messages in multi-turn conversations), the response can be a plain dict with {'role': ..., 'content': [...]} structure.

The fix adds isinstance checks to handle both formats gracefully:

• ConversationMessage — access .content attribute as before
• dict — access via .get('content', [])
• Other types — fall back to str(response)

Tests added

Added to: python/src/tests/agents/test_supervisor_agent.py (6 new tests)

• test_send_message_dict_response — the bug fix test: agent returns dict, no AttributeError, extracts text correctly
• test_send_message_dict_string_content — dict with string content items handled
• test_send_message_conversation_message_regression — normal ConversationMessage still works (isinstance path)
• test_process_agent_streaming_response_dict_final_message — streaming with dict final_message handled
• test_process_agent_streaming_response_conversation_message — streaming with ConversationMessage (regression)
• test_send_messages_with_dict_response_agent — end-to-end through send_messages

Test plan

• All 18 tests pass (12 existing + 6 new) (pytest -v)
• Dict response handling works for both dict-with-dict-content and dict-with-string-content
• Streaming dict final_message handled correctly
• ConversationMessage responses still work (regression tests)

Fixes #268