🤖 Implement Tool Message Validation in Client Logic

src/seer/automation/agent/client.py

@@ -321,10 +321,14 @@ def generate_text(
             system_prompt=system_prompt,
             tools=tools,
         )
+
+        # Validate message sequence before processing
+        self._validate_messages(messages)
 
         anthropic_client = self.get_client()
 
         completion = anthropic_client.messages.create(
+
             system=system_prompt or NOT_GIVEN,
             model=self.model_name,
             tools=cast(Iterable[ToolParam], tool_dicts) if tool_dicts else NOT_GIVEN,
@@ -375,15 +379,22 @@ def _format_claude_response_to_message(completion: anthropic.types.Message) -> M
     @staticmethod
     def to_message_param(message: Message) -> MessageParam:
         if message.role == "tool":
+            if message.tool_call_id:
+                # Only create tool_result if we have a valid tool_call_id
+                return MessageParam(
+                    role="user",
+                    content=[
+                        ToolResultBlockParam(
+                            type="tool_result",
+                            content=message.content or "",
+                            tool_use_id=message.tool_call_id,
+                        )
+                    ],
+                )
+            # Fallback to regular user message if no tool_call_id
             return MessageParam(
-                role="user",
-                content=[
-                    ToolResultBlockParam(
-                        type="tool_result",
-                        content=message.content or "",
-                        tool_use_id=message.tool_call_id or "",
-                    )
-                ],
+                role="user", 
+                content=[TextBlockParam(type="text", text=message.content or "")]
             )
         elif message.role == "tool_use":
             if not message.tool_calls:
@@ -550,6 +561,36 @@ def clean_tool_call_assistant_messages(messages: list[Message]) -> list[Message]
                 new_messages.append(message)
         return new_messages
 
+    @staticmethod 
+    def validate_tool_message_sequence(messages: list[Message] | None) -> list[Message]:
+        """
+        Validates that tool result messages have corresponding tool use messages before them.
+        Returns a new message list with invalid tool messages converted to regular messages.
+        """
+        if not messages:
+            return []
+
+        # Track valid tool use IDs we've seen
+        tool_use_ids = set()
+        validated_messages = []
+
+        for message in messages:
+            # Track valid tool use IDs
+            if message.role == "tool_use" and message.tool_calls:
+                for tool_call in message.tool_calls:
+                    if tool_call.id:
+                        tool_use_ids.add(tool_call.id)
+                validated_messages.append(message)
+            elif message.role == "tool" and message.tool_call_id:
+                if message.tool_call_id not in tool_use_ids:
+                    # Convert to regular user message if no matching tool use found
+                    validated_messages.append(Message(role="user", content=message.content))
+                else:
+                    validated_messages.append(message)
+            else:
+                validated_messages.append(message)
+        return validated_messages
+
 
 @module.provider
 def provide_llm_client() -> LlmClient:

tests/automation/agent/test_client.py

@@ -406,3 +406,68 @@ def test_anthropic_prep_message_and_tools():
         assert "input_schema" in tool_dicts[0]
 
     assert returned_system_prompt == system_prompt
+
+def test_validate_tool_message_sequence():
+    messages = [
+        Message(role="user", content="Hello"),
+        Message(
+            role="tool_use",
+            content="Using search",
+            tool_calls=[ToolCall(id="search_1", function="search", args='{"query": "test"}')],
+        ),
+        Message(role="tool", content="Search results", tool_call_id="search_1"),
+        Message(
+            role="tool_use",
+            content="Using another tool",
+            tool_calls=[ToolCall(id="tool_2", function="other", args='{"param": "value"}')],
+        ),
+        # Invalid tool result without matching tool_use
+        Message(role="tool", content="Invalid tool result", tool_call_id="invalid_id"),
+    ]
+
+    validated_messages = LlmClient.validate_tool_message_sequence(messages)
+
+    assert len(validated_messages) == 5
+    # First message unchanged
+    assert validated_messages[0].role == "user"
+    assert validated_messages[0].content == "Hello"
+
+    # Valid tool_use remains
+    assert validated_messages[1].role == "tool_use"
+    assert validated_messages[1].tool_calls[0].id == "search_1"
+
+    # Valid tool result remains
+    assert validated_messages[2].role == "tool"
+    assert validated_messages[2].tool_call_id == "search_1"
+
+    # Second valid tool_use remains
+    assert validated_messages[3].role == "tool_use"
+    assert validated_messages[3].tool_calls[0].id == "tool_2"
+
+    # Invalid tool result converted to user message
+    assert validated_messages[4].role == "user"
+    assert validated_messages[4].content == "Invalid tool result"
+    assert validated_messages[4].tool_call_id is None
+
+def test_anthropic_provider_handles_invalid_tool_sequence(mock_anthropic_client):
+    llm_client = LlmClient()
+    model = AnthropicProvider.model("claude-3-sonnet-20240229")
+
+    messages = [
+        Message(role="user", content="Hello"),
+        # This would previously cause an error due to tool_result without tool_use
+        Message(role="tool", content="Invalid tool result", tool_call_id="invalid_id"),
+    ]
+
+    mock_anthropic_client.messages.create.return_value = MockAnthropicResponse(
+        content="Response", role="assistant"
+    )
+
+    # This should now work without raising an error
+    response = llm_client.generate_text(
+        messages=messages,
+        model=model,
+    )
+
+    assert response.message.content == "Response"
+    assert response.message.role == "assistant"