Add multi-pattern orchestrator with plan-then-execute and supervisor (#739)

Introduce an agent orchestrator service that supports three
execution patterns (ReAct, plan-then-execute, supervisor) with
LLM-based meta-routing to select the appropriate pattern and task
type per request. Update the agent schema to support
orchestration fields (correlation, sub-agents, plan steps) and
remove legacy response fields (answer, thought, observation).

tests/contract/conftest.py

@@ -87,10 +87,11 @@ def sample_message_data():
             "history": []
         },
         "AgentResponse": {
-            "answer": "Machine learning is a subset of AI.",
+            "chunk_type": "answer",
+            "content": "Machine learning is a subset of AI.",
+            "end_of_message": True,
+            "end_of_dialog": True,
             "error": None,
-            "thought": "I need to provide information about machine learning.",
-            "observation": None
         },
         "Metadata": {
             "id": "test-doc-123",

tests/contract/test_message_contracts.py

@@ -212,10 +212,11 @@ class TestAgentMessageContracts:
         
         # Test required fields
         response = AgentResponse(**response_data)
-        assert hasattr(response, 'answer')
+        assert hasattr(response, 'chunk_type')
+        assert hasattr(response, 'content')
+        assert hasattr(response, 'end_of_message')
+        assert hasattr(response, 'end_of_dialog')
         assert hasattr(response, 'error')
-        assert hasattr(response, 'thought')
-        assert hasattr(response, 'observation')
 
     def test_agent_step_schema_contract(self):
         """Test AgentStep schema contract"""

tests/contract/test_translator_completion_flags.py

@@ -188,12 +188,10 @@ class TestAgentTranslatorCompletionFlags:
         # Arrange
         translator = TranslatorRegistry.get_response_translator("agent")
         response = AgentResponse(
-            answer="4",
-            error=None,
-            thought=None,
-            observation=None,
+            chunk_type="answer",
+            content="4",
             end_of_message=True,
-            end_of_dialog=True
+            end_of_dialog=True,
         )
 
         # Act
@@ -201,7 +199,7 @@ class TestAgentTranslatorCompletionFlags:
 
         # Assert
         assert is_final is True, "is_final must be True when end_of_dialog=True"
-        assert response_dict["answer"] == "4"
+        assert response_dict["content"] == "4"
         assert response_dict["end_of_dialog"] is True
 
     def test_agent_translator_is_final_with_end_of_dialog_false(self):
@@ -212,12 +210,10 @@ class TestAgentTranslatorCompletionFlags:
         # Arrange
         translator = TranslatorRegistry.get_response_translator("agent")
         response = AgentResponse(
-            answer=None,
-            error=None,
-            thought="I need to solve this.",
-            observation=None,
+            chunk_type="thought",
+            content="I need to solve this.",
             end_of_message=True,
-            end_of_dialog=False
+            end_of_dialog=False,
         )
 
         # Act
@@ -225,31 +221,9 @@ class TestAgentTranslatorCompletionFlags:
 
         # Assert
         assert is_final is False, "is_final must be False when end_of_dialog=False"
-        assert response_dict["thought"] == "I need to solve this."
+        assert response_dict["content"] == "I need to solve this."
         assert response_dict["end_of_dialog"] is False
 
-    def test_agent_translator_is_final_fallback_with_answer(self):
-        """
-        Test that AgentResponseTranslator returns is_final=True
-        when answer is present (fallback for legacy responses).
-        """
-        # Arrange
-        translator = TranslatorRegistry.get_response_translator("agent")
-        # Legacy response without end_of_dialog flag
-        response = AgentResponse(
-            answer="4",
-            error=None,
-            thought=None,
-            observation=None
-        )
-
-        # Act
-        response_dict, is_final = translator.from_response_with_completion(response)
-
-        # Assert
-        assert is_final is True, "is_final must be True when answer is present (legacy fallback)"
-        assert response_dict["answer"] == "4"
-
     def test_agent_translator_intermediate_message_is_not_final(self):
         """
         Test that intermediate messages (thought/observation) return is_final=False.
@@ -259,12 +233,10 @@ class TestAgentTranslatorCompletionFlags:
 
         # Test thought message
         thought_response = AgentResponse(
-            answer=None,
-            error=None,
-            thought="Processing...",
-            observation=None,
+            chunk_type="thought",
+            content="Processing...",
             end_of_message=True,
-            end_of_dialog=False
+            end_of_dialog=False,
         )
 
         # Act
@@ -275,12 +247,10 @@ class TestAgentTranslatorCompletionFlags:
 
         # Test observation message
         observation_response = AgentResponse(
-            answer=None,
-            error=None,
-            thought=None,
-            observation="Result found",
+            chunk_type="observation",
+            content="Result found",
             end_of_message=True,
-            end_of_dialog=False
+            end_of_dialog=False,
         )
 
         # Act
@@ -302,10 +272,6 @@ class TestAgentTranslatorCompletionFlags:
             content="",
             end_of_message=True,
             end_of_dialog=True,
-            answer=None,
-            error=None,
-            thought=None,
-            observation=None
         )
 
         # Act