fix(autofix): Handle empty messages and move cleaning calls inside LLMClient

src/seer/automation/agent/client.py

@@ -474,6 +474,10 @@ def generate_text(
         default_temperature = defaults.temperature if defaults else None
         # More defaults to come
 
+        messages = LlmClient.clean_message_content(messages if messages else [])
+        if not tools:
+            messages = LlmClient.clean_tool_call_assistant_messages(messages)
+
         if model.provider_name == LlmProviderType.OPENAI:
             model = cast(OpenAiProvider, model)
 
@@ -516,6 +520,9 @@ def generate_structured(
         if run_name:
             langfuse_context.update_current_observation(name=run_name + " - Generate Structured")
 
+        messages = LlmClient.clean_message_content(messages if messages else [])
+        messages = LlmClient.clean_tool_call_assistant_messages(messages)
+
         if model.provider_name == LlmProviderType.OPENAI:
             model = cast(OpenAiProvider, model)
             return model.generate_structured(
@@ -550,6 +557,15 @@ def clean_tool_call_assistant_messages(messages: list[Message]) -> list[Message]
                 new_messages.append(message)
         return new_messages
 
+    @staticmethod
+    def clean_message_content(messages: list[Message]) -> list[Message]:
+        new_messages = []
+        for message in messages:
+            if not message.content:
+                message.content = "."
+            new_messages.append(message)
+        return new_messages
+
 
 @module.provider
 def provide_llm_client() -> LlmClient:

src/seer/automation/autofix/components/coding/component.py

@@ -99,12 +99,9 @@ def _handle_simple_fix(
     ):
         state = self.context.state.get()
 
-        # Clean memory of tool messages since we're running without tools
-        cleaned_memory = LlmClient.clean_tool_call_assistant_messages(memory)
-
         agent = AutofixAgent(
             config=AgentConfig(interactive=True),
-            memory=cleaned_memory,
+            memory=memory,
             context=self.context,
             name="Plan+Code Simple fixer",
         )
@@ -219,11 +216,8 @@ def _is_feedback_obvious(self, memory: list[Message], llm_client: LlmClient = in
             class NeedToSearchCodebaseOutput(BaseModel):
                 need_to_search_codebase: bool
 
-            # Clean message roles to ensure compatibility with OpenAI's API
-            cleaned_memory = LlmClient.clean_tool_call_assistant_messages(memory)
-
             output = llm_client.generate_structured(
-                messages=cleaned_memory,
+                messages=memory,
                 prompt="Given the above instruction, do you need to search the codebase for more context or have an immediate answer?",
                 model=OpenAiProvider.model("gpt-4o-mini"),
                 response_format=NeedToSearchCodebaseOutput,

src/seer/automation/autofix/components/insight_sharing/component.py

@@ -126,11 +126,7 @@ def invoke(
                 insight=insight,
                 latest_thought=request.latest_thought,
             )
-            memory = []
-            for message in llm_client.clean_tool_call_assistant_messages(request.memory):
-                if message.role != "system":
-                    memory.append(message)
-
+            memory = [msg for msg in request.memory if msg.role != "system"]
             completion = llm_client.generate_structured(
                 messages=memory,
                 prompt=prompt_two,

src/seer/automation/autofix/components/root_cause/component.py

@@ -111,7 +111,7 @@ def invoke(
                 self.context.event_manager.add_log("Cleaning up the findings...")
 
                 formatted_response = llm_client.generate_structured(
-                    messages=LlmClient.clean_tool_call_assistant_messages(agent.memory),
+                    messages=agent.memory,
                     prompt=RootCauseAnalysisPrompts.root_cause_formatter_msg(),
                     model=OpenAiProvider.model("gpt-4o-2024-08-06"),
                     response_format=MultipleRootCauseAnalysisOutputPrompt,

tests/automation/agent/test_client.py

@@ -297,6 +297,18 @@ def test_clean_tool_call_assistant_messages():
     assert cleaned_messages[4].role == "assistant"
 
 
+def test_clean_message_content():
+    messages = [
+        Message(role="user", content=""),
+    ]
+
+    cleaned_messages = LlmClient.clean_message_content(messages)
+
+    assert len(cleaned_messages) == 1
+    assert cleaned_messages[0].role == "user"
+    assert cleaned_messages[0].content == "."
+
+
 def test_openai_generate_structured_refusal(mock_openai_client):
     llm_client = LlmClient()
     model = OpenAiProvider.model("gpt-3.5-turbo")