fix: summarization trims tool call without trimming tool response (#2010)

Co-authored-by: cthomas <caren@letta.com>
Co-authored-by: Sarah Wooders <sarahwooders@gmail.com>

letta/services/agent_manager.py

@@ -866,8 +866,9 @@ class AgentManager:
     @enforce_types
     def trim_older_in_context_messages(self, num: int, agent_id: str, actor: PydanticUser) -> PydanticAgentState:
         message_ids = self.get_agent_by_id(agent_id=agent_id, actor=actor).message_ids
-        new_messages = [message_ids[0]] + message_ids[num:]  # 0 is system message
-        return self.set_in_context_messages(agent_id=agent_id, message_ids=new_messages, actor=actor)
+        newer_messages = self._trim_tool_response(agent_id=agent_id, actor=actor, message_ids=message_ids[num:])
+        trimmed_messages = [message_ids[0]] + newer_messages  # 0 is system message
+        return self.set_in_context_messages(agent_id=agent_id, message_ids=trimmed_messages, actor=actor)
 
     @enforce_types
     def trim_all_in_context_messages_except_system(self, agent_id: str, actor: PydanticUser) -> PydanticAgentState:
@@ -876,6 +877,16 @@ class AgentManager:
         new_messages = [message_ids[0]]  # 0 is system message
         return self.set_in_context_messages(agent_id=agent_id, message_ids=new_messages, actor=actor)
 
+    def _trim_tool_response(self, agent_id: str, actor: PydanticUser, message_ids: list[str]) -> PydanticAgentState:
+        """
+        Trims the tool response from the in-context messages if there is no tool call present in trimmed messages.
+        """
+        if message_ids:
+            messages = self.message_manager.get_messages_by_ids(message_ids=[message_ids[0]], actor=actor)
+            if messages and messages[0].role == "tool":
+                return message_ids[1:]
+        return message_ids
+
     @enforce_types
     def prepend_to_in_context_messages(self, messages: List[PydanticMessage], agent_id: str, actor: PydanticUser) -> PydanticAgentState:
         message_ids = self.get_agent_by_id(agent_id=agent_id, actor=actor).message_ids

tests/integration_test_summarizer.py

@@ -113,6 +113,50 @@ def test_cutoff_calculation(mocker):
     assert messages[cutoff - 1].role == MessageRole.user
 
 
+def test_cutoff_calculation_with_tool_call(mocker, client: LocalClient, agent_state):
+    """Test that trim_older_in_context_messages properly handles tool responses with _trim_tool_response."""
+    agent_state = client.get_agent(agent_id=agent_state.id)
+
+    # Setup
+    messages = [
+        generate_message("system"),
+        generate_message("user", text="First user message"),
+        generate_message(
+            "assistant", tool_calls=[{"id": "tool_call_1", "type": "function", "function": {"name": "test_function", "arguments": "{}"}}]
+        ),
+        generate_message("tool", text="First tool response"),
+        generate_message("assistant", text="First assistant response after tool"),
+        generate_message("user", text="Second user message"),
+        generate_message("assistant", text="Second assistant response"),
+    ]
+
+    def mock_get_messages_by_ids(message_ids, actor):
+        return [msg for msg in messages if msg.id in message_ids]
+
+    mocker.patch.object(client.server.agent_manager.message_manager, "get_messages_by_ids", side_effect=mock_get_messages_by_ids)
+
+    # Mock get_agent_by_id to return an agent with our message IDs
+    mock_agent = mocker.Mock()
+    mock_agent.message_ids = [msg.id for msg in messages]
+    mocker.patch.object(client.server.agent_manager, "get_agent_by_id", return_value=mock_agent)
+
+    # Mock set_in_context_messages to capture what messages are being set
+    mock_set_messages = mocker.patch.object(client.server.agent_manager, "set_in_context_messages", return_value=agent_state)
+
+    # Test Case: Trim to remove orphaned tool response
+    client.server.agent_manager.trim_older_in_context_messages(agent_id=agent_state.id, num=3, actor=client.user)
+
+    test1 = mock_set_messages.call_args_list[0][1]
+    assert len(test1["message_ids"]) == 4
+
+    mock_set_messages.reset_mock()
+
+    # Test Case: Does not result in trimming the orphaned tool response
+    client.server.agent_manager.trim_older_in_context_messages(agent_id=agent_state.id, num=2, actor=client.user)
+    test2 = mock_set_messages.call_args_list[0][1]
+    assert len(test2["message_ids"]) == 6
+
+
 def test_summarize_many_messages_basic(client, disable_e2b_api_key):
     small_context_llm_config = LLMConfig.default_config("gpt-4o-mini")
     small_context_llm_config.context_window = 3000