summarization manager - add summary prompt to messages

Author: pgrayy

src/strands/agent/conversation_manager/summarizing_conversation_manager.py

@@ -81,6 +81,7 @@ def __init__(
         self.summarization_agent = summarization_agent
         self.summarization_system_prompt = summarization_system_prompt
         self._summary_message: Optional[Message] = None
+        self._summary_prompt: Message = {"content": [{"text": "Please summarize this conversation"}], "role": "user"}
 
     @override
     def restore_from_session(self, state: dict[str, Any]) -> Optional[list[Message]]:
@@ -94,7 +95,7 @@ def restore_from_session(self, state: dict[str, Any]) -> Optional[list[Message]]
         """
         super().restore_from_session(state)
         self._summary_message = state.get("summary_message")
-        return [self._summary_message] if self._summary_message else None
+        return [self._summary_prompt, self._summary_message] if self._summary_message else None
 
     def get_state(self) -> dict[str, Any]:
         """Returns a dictionary representation of the state for the Summarizing Conversation Manager."""
@@ -152,15 +153,15 @@ def reduce_context(self, agent: "Agent", e: Optional[Exception] = None, **kwargs
 
             # Keep track of the number of messages that have been summarized thus far.
             self.removed_message_count += len(messages_to_summarize)
-            # If there is a summary message, don't count it in the removed_message_count.
+            # If there is a summary message, don't count it or the summary prompt in the removed_message_count.
             if self._summary_message:
-                self.removed_message_count -= 1
+                self.removed_message_count -= 2
 
             # Generate summary
             self._summary_message = self._generate_summary(messages_to_summarize, agent)
 
             # Replace the summarized messages with the summary
-            agent.messages[:] = [self._summary_message] + remaining_messages
+            agent.messages[:] = [self._summary_prompt, self._summary_message] + remaining_messages
 
         except Exception as summarization_error:
             logger.error("Summarization failed: %s", summarization_error)
@@ -200,8 +201,7 @@ def _generate_summary(self, messages: List[Message], agent: "Agent") -> Message:
             summarization_agent.messages = messages
 
             # Use the agent to generate summary with rich content (can use tools if needed)
-            result = summarization_agent("Please summarize this conversation.")
-
+            result = summarization_agent(self._summary_prompt["content"])
             return result.message
 
         finally:

tests/strands/agent/test_summarizing_conversation_manager.py

@@ -95,13 +95,16 @@ def test_reduce_context_with_summarization(summarizing_manager, mock_agent):
 
     summarizing_manager.reduce_context(mock_agent)
 
-    # Should have: 1 summary message + 2 preserved recent messages + remaining from summarization
-    assert len(mock_agent.messages) == 4
+    # Should have: 1 summary prompt + 1 summary message + 2 preserved recent messages + remaining from summarization
+    assert len(mock_agent.messages) == 5
 
-    # First message should be the summary
-    assert mock_agent.messages[0]["role"] == "assistant"
-    first_content = mock_agent.messages[0]["content"][0]
-    assert "text" in first_content and "This is a summary of the conversation." in first_content["text"]
+    # First message should be the summary prompt
+    assert mock_agent.messages[0] == {"content": [{"text": "Please summarize this conversation"}], "role": "user"}
+    # Second message should be the summary
+    assert mock_agent.messages[1] == {
+        "content": [{"text": "This is a summary of the conversation."}],
+        "role": "assistant",
+    }
 
     # Recent messages should be preserved
     assert "Message 3" in str(mock_agent.messages[-2]["content"])
@@ -434,17 +437,18 @@ def test_reduce_context_tool_pair_adjustment_works_with_forward_search():
     # messages_to_summarize_count = (3 - 1) * 0.5 = 1
     # But split point adjustment will move forward from the toolUse, potentially increasing count
     manager.reduce_context(mock_agent)
-    # Should have summary + remaining messages
-    assert len(mock_agent.messages) == 2
-
-    # First message should be the summary
-    assert mock_agent.messages[0]["role"] == "assistant"
-    summary_content = mock_agent.messages[0]["content"][0]
-    assert "text" in summary_content and "This is a summary of the conversation." in summary_content["text"]
-
+    # Should have summary prompt + summary + remaining messages
+    assert len(mock_agent.messages) == 3
+
+    # First message should be the summary prompt
+    assert mock_agent.messages[0] == {"content": [{"text": "Please summarize this conversation"}], "role": "user"}
+    # Second message should be the summary
+    assert mock_agent.messages[1] == {
+        "content": [{"text": "This is a summary of the conversation."}],
+        "role": "assistant",
+    }
     # Last message should be the preserved recent message
-    assert mock_agent.messages[1]["role"] == "user"
-    assert mock_agent.messages[1]["content"][0]["text"] == "Latest message"
+    assert mock_agent.messages[2] == {"content": [{"text": "Latest message"}], "role": "user"}
 
 
 def test_adjust_split_point_exceeds_message_length(summarizing_manager):
@@ -590,21 +594,19 @@ def test_summarizing_conversation_manager_properly_records_removed_message_count
     assert manager.removed_message_count == 0
 
     manager.reduce_context(agent)
-    # Assert the oldest message is the sumamry message
     assert manager._summary_message["content"][0]["text"] == "Summary"
     # There are 8 messages in the agent messages array, since half will be summarized,
-    # 4 will remain plus 1 summary message = 5
-    assert (len(agent.messages)) == 5
+    # 4 will remain plus 1 summary prompt and 1 summary message = 6
+    assert len(agent.messages) == 6
     # Half of the messages were summarized and removed: 8/2 = 4
     assert manager.removed_message_count == 4
 
     manager.reduce_context(agent)
     assert manager._summary_message["content"][0]["text"] == "Summary"
-    # After the first summary, 5 messages remain. Summarizing again will lead to:
-    # 5 - (int(5/2)) (messages to be sumamrized) + 1 (new summary message) = 5 - 2 + 1 = 4
-    assert (len(agent.messages)) == 4
-    # Half of the messages were summarized and removed: int(5/2) = 2
-    # However, one of the messages that was summarized was the previous summary message,
-    # so we dont count this toward the total:
-    # 4 (Previously removed messages) + 2 (removed messages) - 1 (Previous summary message) = 5
+    # After the first summary, 6 messages remain. Summarizing again will lead to:
+    # 6 - (6/2) (messages to be sumamrized) + 1 (summary prompt) + 1 (new summary message) = 6 - 3 + 1 + 1 = 5
+    assert len(agent.messages) == 5
+    # Half of the messages were summarized and removed: (6/2) = 3
+    # However, the summary prompt and previous summary were also summarized but we don't count this in the total:
+    # 4 (Previously removed messages) + 3 (removed messages) - 1 (summary prompt) - 1 (Previous summary message) = 5
     assert manager.removed_message_count == 5

tests/strands/session/test_repository_session_manager.py

@@ -151,10 +151,9 @@ def test_initialize_restores_existing_agent_with_summarizing_conversation_manage
 
     # Verify agent state restored
     assert agent.state.get("key") == "value"
-    # The session message plus the summary message
-    assert len(agent.messages) == 2
-    assert agent.messages[1]["role"] == "user"
-    assert agent.messages[1]["content"][0]["text"] == "Hello"
+    # The session message plus the summary prompt plus the summary message
+    assert len(agent.messages) == 3
+    assert agent.messages[2] == {"content": [{"text": "Hello"}], "role": "user"}
     assert agent.conversation_manager.removed_message_count == 1
 
 

tests_integ/test_summarizing_conversation_manager_integration.py

@@ -151,15 +151,18 @@ def test_summarization_with_context_overflow(model):
 
     # Verify summarization occurred
     assert len(agent.messages) < initial_message_count
-    # Should have: 1 summary + remaining messages
+    # Should have: 1 summary prompt + 1 summary + remaining messages
     # With 6 messages, summary_ratio=0.5, preserve_recent_messages=2:
     # messages_to_summarize = min(6 * 0.5, 6 - 2) = min(3, 4) = 3
-    # So we summarize 3 messages, leaving 3 remaining + 1 summary = 4 total
-    expected_total_messages = 4
+    # So we summarize 3 messages, leaving 3 remaining + 1 summary prompt + 1 summary = 5 total
+    expected_total_messages = 5
     assert len(agent.messages) == expected_total_messages
 
-    # First message should be the summary (assistant message)
-    summary_message = agent.messages[0]
+    # First message should be the summary prompt
+    summary_prompt = agent.messages[0]
+    assert summary_prompt == {"content": [{"text": "Please summarize this conversation"}], "role": "user"}
+    # Second message should be the summary
+    summary_message = agent.messages[1]
     assert summary_message["role"] == "assistant"
     assert len(summary_message["content"]) > 0
 
@@ -361,7 +364,7 @@ def test_dedicated_summarization_agent(model, summarization_model):
     assert len(agent.messages) < original_length
 
     # Get the summary message
-    summary_message = agent.messages[0]
+    summary_message = agent.messages[1]
     assert summary_message["role"] == "assistant"
 
     # Extract summary text