fix(anthropic): preserve server_tool_use and web_search_tool_result in multi-turn conversations

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -1623,7 +1623,8 @@ def convert_function_to_anthropic_tool_invoke(
 
 def convert_to_anthropic_tool_invoke(
     tool_calls: List[ChatCompletionAssistantToolCall],
-) -> List[AnthropicMessagesToolUseParam]:
+    web_search_results: Optional[List[Any]] = None,
+) -> List[Union[AnthropicMessagesToolUseParam, Dict[str, Any]]]:
     """
     OpenAI tool invokes:
     {
@@ -1659,38 +1660,68 @@ def convert_to_anthropic_tool_invoke(
         }
       ]
     }
+
+    For server-side tools (web_search), we need to reconstruct:
+    - server_tool_use blocks (id starts with "srvtoolu_")
+    - web_search_tool_result blocks (from provider_specific_fields)
+
+    Fixes: https://github.com/BerriAI/litellm/issues/17737
     """
-    anthropic_tool_invoke = []
+    anthropic_tool_invoke: List[Union[AnthropicMessagesToolUseParam, Dict[str, Any]]] = []
 
     for tool in tool_calls:
         if not get_attribute_or_key(tool, "type") == "function":
             continue
 
-        _anthropic_tool_use_param = AnthropicMessagesToolUseParam(
-            type="tool_use",
-            id=cast(str, get_attribute_or_key(tool, "id")),
-            name=cast(
-                str,
-                get_attribute_or_key(get_attribute_or_key(tool, "function"), "name"),
-            ),
-            input=json.loads(
-                get_attribute_or_key(
-                    get_attribute_or_key(tool, "function"), "arguments"
-                )
-            ),
+        tool_id = cast(str, get_attribute_or_key(tool, "id"))
+        tool_name = cast(
+            str,
+            get_attribute_or_key(get_attribute_or_key(tool, "function"), "name"),
         )
-
-        _content_element = add_cache_control_to_content(
-            anthropic_content_element=_anthropic_tool_use_param,
-            original_content_element=dict(tool),
+        tool_input = json.loads(
+            get_attribute_or_key(
+                get_attribute_or_key(tool, "function"), "arguments"
+            )
         )
 
-        if "cache_control" in _content_element:
-            _anthropic_tool_use_param["cache_control"] = _content_element[
-                "cache_control"
-            ]
+        # Check if this is a server-side tool (web_search, tool_search, etc.)
+        # Server tool IDs start with "srvtoolu_"
+        if tool_id.startswith("srvtoolu_"):
+            # Create server_tool_use block instead of tool_use
+            _anthropic_server_tool_use: Dict[str, Any] = {
+                "type": "server_tool_use",
+                "id": tool_id,
+                "name": tool_name,
+                "input": tool_input,
+            }
+            anthropic_tool_invoke.append(_anthropic_server_tool_use)
+
+            # Add corresponding web_search_tool_result if available
+            if web_search_results:
+                for result in web_search_results:
+                    if result.get("tool_use_id") == tool_id:
+                        anthropic_tool_invoke.append(result)
+                        break
+        else:
+            # Regular tool_use
+            _anthropic_tool_use_param = AnthropicMessagesToolUseParam(
+                type="tool_use",
+                id=tool_id,
+                name=tool_name,
+                input=tool_input,
+            )
+
+            _content_element = add_cache_control_to_content(
+                anthropic_content_element=_anthropic_tool_use_param,
+                original_content_element=dict(tool),
+            )
 
-        anthropic_tool_invoke.append(_anthropic_tool_use_param)
+            if "cache_control" in _content_element:
+                _anthropic_tool_use_param["cache_control"] = _content_element[
+                    "cache_control"
+                ]
+
+            anthropic_tool_invoke.append(_anthropic_tool_use_param)
 
     return anthropic_tool_invoke
 
@@ -2052,8 +2083,15 @@ def anthropic_messages_pt(  # noqa: PLR0915
             if (
                 assistant_tool_calls is not None
             ):  # support assistant tool invoke conversion
+                # Get web_search_results from provider_specific_fields for server_tool_use reconstruction
+                # Fixes: https://github.com/BerriAI/litellm/issues/17737
+                _provider_specific_fields = assistant_content_block.get("provider_specific_fields") or {}
+                _web_search_results = _provider_specific_fields.get("web_search_results")
                 assistant_content.extend(
-                    convert_to_anthropic_tool_invoke(assistant_tool_calls)
+                    convert_to_anthropic_tool_invoke(
+                        assistant_tool_calls,
+                        web_search_results=_web_search_results,
+                    )
                 )
 
             assistant_function_call = assistant_content_block.get("function_call")

litellm/llms/anthropic/chat/transformation.py

@@ -1082,6 +1082,7 @@ def extract_response_content(self, completion_response: dict) -> Tuple[
         ],
         Optional[str],
         List[ChatCompletionToolCallChunk],
+        Optional[List[Any]],
     ]:
         text_content = ""
         citations: Optional[List[Any]] = None
@@ -1092,6 +1093,7 @@ def extract_response_content(self, completion_response: dict) -> Tuple[
         ] = None
         reasoning_content: Optional[str] = None
         tool_calls: List[ChatCompletionToolCallChunk] = []
+        web_search_results: Optional[List[Any]] = None
         for idx, content in enumerate(completion_response["content"]):
             if content["type"] == "text":
                 text_content += content["text"]
@@ -1117,6 +1119,11 @@ def extract_response_content(self, completion_response: dict) -> Tuple[
                 # This block contains tool_references that were discovered
                 # We don't need to include this in the response as it's internal metadata
                 pass
+            ## WEB SEARCH TOOL RESULT - preserve web search results for multi-turn conversations
+            elif content["type"] == "web_search_tool_result":
+                if web_search_results is None:
+                    web_search_results = []
+                web_search_results.append(content)
             elif content.get("thinking", None) is not None:
                 if thinking_blocks is None:
                     thinking_blocks = []
@@ -1148,7 +1155,7 @@ def extract_response_content(self, completion_response: dict) -> Tuple[
                 if thinking_content is not None:
                     reasoning_content += thinking_content
 
-        return text_content, citations, thinking_blocks, reasoning_content, tool_calls
+        return text_content, citations, thinking_blocks, reasoning_content, tool_calls, web_search_results
 
     def calculate_usage(
         self,
@@ -1288,6 +1295,7 @@ def transform_parsed_response(
                 thinking_blocks,
                 reasoning_content,
                 tool_calls,
+                web_search_results,
             ) = self.extract_response_content(completion_response=completion_response)
 
             if (
@@ -1307,6 +1315,8 @@ def transform_parsed_response(
             }
             if context_management is not None:
                 provider_specific_fields["context_management"] = context_management
+            if web_search_results is not None:
+                provider_specific_fields["web_search_results"] = web_search_results
 
             _message = litellm.Message(
                 tool_calls=tool_calls,

tests/llm_translation/test_prompt_factory.py

@@ -18,6 +18,7 @@
     anthropic_pt,
     claude_2_1_pt,
     convert_to_anthropic_image_obj,
+    convert_to_anthropic_tool_invoke,
     convert_url_to_base64,
     create_anthropic_image_param,
     llama_2_chat_pt,
@@ -947,3 +948,217 @@ def test_ollama_pt():
     ]
     prompt = ollama_pt(model="ollama/llama3.1", messages=messages)
     print(prompt)
+
+
+# ============ Server Tool Use Reconstruction Tests ============
+# Fixes: https://github.com/BerriAI/litellm/issues/17737
+
+
+def test_convert_to_anthropic_tool_invoke_regular_tool():
+    """Test that regular tool_use is converted correctly."""
+    tool_calls = [
+        {
+            "id": "toolu_01ABC123",
+            "type": "function",
+            "function": {
+                "name": "get_weather",
+                "arguments": '{"location": "San Francisco"}'
+            }
+        }
+    ]
+
+    result = convert_to_anthropic_tool_invoke(tool_calls)
+
+    assert len(result) == 1
+    assert result[0]["type"] == "tool_use"
+    assert result[0]["id"] == "toolu_01ABC123"
+    assert result[0]["name"] == "get_weather"
+    assert result[0]["input"] == {"location": "San Francisco"}
+
+
+def test_convert_to_anthropic_tool_invoke_server_tool():
+    """
+    Test that server_tool_use (srvtoolu_) is reconstructed as server_tool_use.
+
+    Fixes: https://github.com/BerriAI/litellm/issues/17737
+    """
+    tool_calls = [
+        {
+            "id": "srvtoolu_01ABC123",
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "arguments": '{"query": "elephant weight"}'
+            }
+        }
+    ]
+
+    result = convert_to_anthropic_tool_invoke(tool_calls)
+
+    assert len(result) == 1
+    assert result[0]["type"] == "server_tool_use"  # NOT tool_use
+    assert result[0]["id"] == "srvtoolu_01ABC123"
+    assert result[0]["name"] == "web_search"
+    assert result[0]["input"] == {"query": "elephant weight"}
+
+
+def test_convert_to_anthropic_tool_invoke_with_web_search_results():
+    """
+    Test that web_search_tool_result is included after server_tool_use.
+
+    Fixes: https://github.com/BerriAI/litellm/issues/17737
+    """
+    tool_calls = [
+        {
+            "id": "srvtoolu_01ABC123",
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "arguments": '{"query": "elephant weight"}'
+            }
+        }
+    ]
+
+    web_search_results = [
+        {
+            "type": "web_search_tool_result",
+            "tool_use_id": "srvtoolu_01ABC123",
+            "content": [
+                {
+                    "type": "web_search_result",
+                    "url": "https://example.com",
+                    "title": "Elephant Facts",
+                    "snippet": "Elephants weigh 5000 kg"
+                }
+            ]
+        }
+    ]
+
+    result = convert_to_anthropic_tool_invoke(tool_calls, web_search_results=web_search_results)
+
+    assert len(result) == 2
+    # First: server_tool_use
+    assert result[0]["type"] == "server_tool_use"
+    assert result[0]["id"] == "srvtoolu_01ABC123"
+    # Second: web_search_tool_result
+    assert result[1]["type"] == "web_search_tool_result"
+    assert result[1]["tool_use_id"] == "srvtoolu_01ABC123"
+
+
+def test_convert_to_anthropic_tool_invoke_mixed_tools():
+    """
+    Test that mixed server and regular tools are reconstructed correctly.
+
+    Fixes: https://github.com/BerriAI/litellm/issues/17737
+    """
+    tool_calls = [
+        {
+            "id": "srvtoolu_01ABC123",
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "arguments": '{"query": "elephant weight"}'
+            }
+        },
+        {
+            "id": "toolu_01XYZ789",
+            "type": "function",
+            "function": {
+                "name": "add_numbers",
+                "arguments": '{"a": 5000, "b": 100}'
+            }
+        }
+    ]
+
+    web_search_results = [
+        {
+            "type": "web_search_tool_result",
+            "tool_use_id": "srvtoolu_01ABC123",
+            "content": [{"url": "https://example.com", "title": "Test"}]
+        }
+    ]
+
+    result = convert_to_anthropic_tool_invoke(tool_calls, web_search_results=web_search_results)
+
+    assert len(result) == 3
+    # First: server_tool_use
+    assert result[0]["type"] == "server_tool_use"
+    assert result[0]["id"] == "srvtoolu_01ABC123"
+    # Second: web_search_tool_result
+    assert result[1]["type"] == "web_search_tool_result"
+    # Third: regular tool_use
+    assert result[2]["type"] == "tool_use"
+    assert result[2]["id"] == "toolu_01XYZ789"
+
+
+def test_anthropic_messages_pt_with_server_tool_use():
+    """
+    Test that anthropic_messages_pt correctly reconstructs server_tool_use from provider_specific_fields.
+
+    Fixes: https://github.com/BerriAI/litellm/issues/17737
+    """
+    messages = [
+        {"role": "user", "content": "Search for elephant weight and add 100"},
+        {
+            "role": "assistant",
+            "content": "Let me search for that.",
+            "tool_calls": [
+                {
+                    "id": "srvtoolu_01ABC123",
+                    "type": "function",
+                    "function": {
+                        "name": "web_search",
+                        "arguments": '{"query": "elephant weight"}'
+                    }
+                },
+                {
+                    "id": "toolu_01XYZ789",
+                    "type": "function",
+                    "function": {
+                        "name": "add_numbers",
+                        "arguments": '{"a": 5000, "b": 100}'
+                    }
+                }
+            ],
+            "provider_specific_fields": {
+                "web_search_results": [
+                    {
+                        "type": "web_search_tool_result",
+                        "tool_use_id": "srvtoolu_01ABC123",
+                        "content": [{"url": "https://example.com", "title": "Test", "snippet": "5000 kg"}]
+                    }
+                ]
+            }
+        },
+        {
+            "role": "tool",
+            "tool_call_id": "toolu_01XYZ789",
+            "content": "5100"
+        }
+    ]
+
+    result = anthropic_messages_pt(messages, model="claude-sonnet-4-5", llm_provider="anthropic")
+
+    # Find the assistant message
+    assistant_msg = next(m for m in result if m["role"] == "assistant")
+    content = assistant_msg["content"]
+
+    # Should have: text, server_tool_use, web_search_tool_result, tool_use
+    types = [c.get("type") for c in content]
+    assert "text" in types
+    assert "server_tool_use" in types
+    assert "web_search_tool_result" in types
+    assert "tool_use" in types
+
+    # Verify server_tool_use
+    server_tool = next(c for c in content if c.get("type") == "server_tool_use")
+    assert server_tool["id"] == "srvtoolu_01ABC123"
+
+    # Verify web_search_tool_result comes after server_tool_use
+    server_idx = types.index("server_tool_use")
+    web_result_idx = types.index("web_search_tool_result")
+    assert web_result_idx == server_idx + 1
+
+    # Verify regular tool_use
+    tool_use = next(c for c in content if c.get("type") == "tool_use")
+    assert tool_use["id"] == "toolu_01XYZ789"