fix(bedrock): handle thinking with tool calls for Claude 4 models

litellm/llms/bedrock/chat/converse_transformation.py

@@ -53,7 +53,12 @@ from litellm.types.utils import (
     PromptTokensDetailsWrapper,
     Usage,
 )
-from litellm.utils import add_dummy_tool, has_tool_call_blocks, supports_reasoning
+from litellm.utils import (
+    add_dummy_tool,
+    has_tool_call_blocks,
+    last_assistant_with_tool_calls_has_no_thinking_blocks,
+    supports_reasoning,
+)
 
 from ..common_utils import (
     BedrockError,
@@ -729,7 +734,7 @@ class AmazonConverseConfig(BaseConfig):
             return optional_params
 
         """
-        Follow similar approach to anthropic - translate to a single tool call. 
+        Follow similar approach to anthropic - translate to a single tool call.
 
         When using tools in this way: - https://docs.anthropic.com/en/docs/build-with-claude/tool-use#json-mode
         - You usually want to provide a single tool
@@ -912,16 +917,16 @@ class AmazonConverseConfig(BaseConfig):
         inference_params = {
             k: v for k, v in inference_params.items() if k in total_supported_params
         }
-        
+
         # Only set the topK value in for models that support it
         additional_request_params.update(
             self._handle_top_k_value(model, inference_params)
         )
-        
+
         # Filter out internal/MCP-related parameters that shouldn't be sent to the API
         # These are LiteLLM internal parameters, not API parameters
         additional_request_params = filter_internal_params(additional_request_params)
-        
+
         # Filter out non-serializable objects (exceptions, callables, logging objects, etc.)
         # from additional_request_params to prevent JSON serialization errors
         # This filters: Exception objects, callable objects (functions), Logging objects, etc.
@@ -1021,9 +1026,24 @@ class AmazonConverseConfig(BaseConfig):
                     llm_provider="bedrock",
                 )
 
+        # Drop thinking param if thinking is enabled but thinking_blocks are missing
+        # This prevents the error: "Expected thinking or redacted_thinking, but found tool_use"
+        # Related issues: https://github.com/BerriAI/litellm/issues/14194
+        if (
+            optional_params.get("thinking") is not None
+            and messages is not None
+            and last_assistant_with_tool_calls_has_no_thinking_blocks(messages)
+        ):
+            if litellm.modify_params:
+                optional_params.pop("thinking", None)
+                litellm.verbose_logger.warning(
+                    "Dropping 'thinking' param because the last assistant message with tool_calls "
+                    "has no thinking_blocks. The model won't use extended thinking for this turn."
+                )
+
         # Prepare and separate parameters
-        inference_params, additional_request_params, request_metadata = (
-            self._prepare_request_params(optional_params, model)
+        inference_params, additional_request_params, request_metadata = self._prepare_request_params(
+            optional_params, model
         )
 
         original_tools = inference_params.pop("tools", [])
@@ -1410,11 +1430,11 @@ class AmazonConverseConfig(BaseConfig):
             )
 
         """
-        Bedrock Response Object has optional message block 
+        Bedrock Response Object has optional message block
 
         completion_response["output"].get("message", None)
 
-        A message block looks like this (Example 1): 
+        A message block looks like this (Example 1):
         "output": {
             "message": {
                 "role": "assistant",

litellm/llms/bedrock/chat/invoke_handler.py

@@ -374,6 +374,29 @@ class BedrockLLM(BaseAWSLLM):
     def __init__(self) -> None:
         super().__init__()
 
+    @staticmethod
+    def is_claude_messages_api_model(model: str) -> bool:
+        """
+        Check if the model uses the Claude Messages API (Claude 3+).
+
+        Handles:
+        - Regional prefixes: eu.anthropic.claude-*, us.anthropic.claude-*
+        - Claude 3 models: claude-3-haiku, claude-3-sonnet, claude-3-opus, claude-3-5-*, claude-3-7-*
+        - Claude 4 models: claude-opus-4, claude-sonnet-4, claude-haiku-4
+        """
+        # Normalize model string to lowercase for matching
+        model_lower = model.lower()
+
+        # Claude 3+ indicators (all use Messages API)
+        messages_api_indicators = [
+            "claude-3",      # Claude 3.x models
+            "claude-opus-4", # Claude Opus 4
+            "claude-sonnet-4", # Claude Sonnet 4
+            "claude-haiku-4",  # Claude Haiku 4
+        ]
+
+        return any(indicator in model_lower for indicator in messages_api_indicators)
+
     def convert_messages_to_prompt(
         self, model, messages, provider, custom_prompt_dict
     ) -> Tuple[str, Optional[list]]:
@@ -465,7 +488,7 @@ class BedrockLLM(BaseAWSLLM):
                         completion_response["generations"][0]["finish_reason"]
                     )
             elif provider == "anthropic":
-                if model.startswith("anthropic.claude-3"):
+                if self.is_claude_messages_api_model(model):
                     json_schemas: dict = {}
                     _is_function_call = False
                     ## Handle Tool Calling
@@ -595,13 +618,12 @@ class BedrockLLM(BaseAWSLLM):
                         outputText = choice["message"].get("content")
                     elif "text" in choice:  # fallback for completion format
                         outputText = choice["text"]
-                    
                     # Set finish reason
                     if "finish_reason" in choice:
                         model_response.choices[0].finish_reason = map_finish_reason(
                             choice["finish_reason"]
                         )
-                    
+
                     # Set usage if available
                     if "usage" in completion_response:
                         usage = completion_response["usage"]
@@ -842,7 +864,7 @@ class BedrockLLM(BaseAWSLLM):
                     ] = True  # cohere requires stream = True in inference params
                 data = json.dumps({"prompt": prompt, **inference_params})
         elif provider == "anthropic":
-            if model.startswith("anthropic.claude-3"):
+            if self.is_claude_messages_api_model(model):
                 # Separate system prompt from rest of message
                 system_prompt_idx: list[int] = []
                 system_messages: list[str] = []
@@ -940,13 +962,13 @@ class BedrockLLM(BaseAWSLLM):
             # Use AmazonBedrockOpenAIConfig for proper OpenAI transformation
             openai_config = AmazonBedrockOpenAIConfig()
             supported_params = openai_config.get_supported_openai_params(model=model)
-            
+
             # Filter to only supported OpenAI params
             filtered_params = {
-                k: v for k, v in inference_params.items() 
+                k: v for k, v in inference_params.items()
                 if k in supported_params
             }
-            
+
             # OpenAI uses messages format, not prompt
             data = json.dumps({"messages": messages, **filtered_params})
         else:

litellm/utils.py

@@ -621,7 +621,7 @@ def load_credentials_from_list(kwargs: dict):
     """
     # Access CredentialAccessor via module to trigger lazy loading if needed
     CredentialAccessor = getattr(sys.modules[__name__], 'CredentialAccessor')
-    
+
     credential_name = kwargs.get("litellm_credential_name")
     if credential_name and litellm.credential_list:
         credential_accessor = CredentialAccessor.get_credential_values(credential_name)
@@ -648,7 +648,7 @@ def _is_gemini_model(model: Optional[str], custom_llm_provider: Optional[str]) -
         if custom_llm_provider in ["vertex_ai", "vertex_ai_beta"]:
             return model is not None and "gemini" in model.lower()
         return True
-    
+
     # Check if model name contains gemini
     return model is not None and "gemini" in model.lower()
 
@@ -670,7 +670,7 @@ def _process_assistant_message_tool_calls(
     """
     role = msg_copy.get("role")
     tool_calls = msg_copy.get("tool_calls")
-    
+
     if role == "assistant" and isinstance(tool_calls, list):
         new_tool_calls = []
         for tc in tool_calls:
@@ -683,17 +683,17 @@ def _process_assistant_message_tool_calls(
             else:
                 new_tool_calls.append(tc)
                 continue
-            
+
             # Remove thought signature from ID if present
             if isinstance(tc_dict.get("id"), str):
                 if thought_signature_separator in tc_dict["id"]:
                     tc_dict["id"] = _remove_thought_signature_from_id(
                         tc_dict["id"], thought_signature_separator
                     )
-            
+
             new_tool_calls.append(tc_dict)
         msg_copy["tool_calls"] = new_tool_calls
-    
+
     return msg_copy
 
 
@@ -708,7 +708,7 @@ def _process_tool_message_id(msg_copy: dict, thought_signature_separator: str) -
             msg_copy["tool_call_id"] = _remove_thought_signature_from_id(
                 msg_copy["tool_call_id"], thought_signature_separator
             )
-    
+
     return msg_copy
 
 
@@ -719,7 +719,7 @@ def _remove_thought_signatures_from_messages(
     Remove thought signatures from tool call IDs in all messages.
     """
     processed_messages = []
-    
+
     for msg in messages:
         # Handle Pydantic models (convert to dict)
         if hasattr(msg, "model_dump"):
@@ -730,17 +730,17 @@ def _remove_thought_signatures_from_messages(
             # Unknown type, keep as is
             processed_messages.append(msg)
             continue
-        
+
         # Process assistant messages with tool_calls
         msg_dict = _process_assistant_message_tool_calls(
             msg_dict, thought_signature_separator
         )
-        
+
         # Process tool messages with tool_call_id
         msg_dict = _process_tool_message_id(msg_dict, thought_signature_separator)
-        
+
         processed_messages.append(msg_dict)
-    
+
     return processed_messages
 
 
@@ -960,7 +960,7 @@ def function_setup(  # noqa: PLR0915
                     input=buffer.getvalue(),
                     model=model,
                 )
-            
+
             ### REMOVE THOUGHT SIGNATURES FROM TOOL CALL IDS FOR NON-GEMINI MODELS ###
             # Gemini models embed thought signatures in tool call IDs. When sending
             # messages with tool calls to non-Gemini providers, we need to remove these
@@ -976,7 +976,7 @@ def function_setup(  # noqa: PLR0915
 
                     # Get custom_llm_provider to determine target provider
                     custom_llm_provider = kwargs.get("custom_llm_provider")
-                    
+
                     # If custom_llm_provider not in kwargs, try to determine it from the model
                     if not custom_llm_provider and model:
                         try:
@@ -987,18 +987,18 @@ def function_setup(  # noqa: PLR0915
                         except Exception:
                             # If we can't determine the provider, skip this processing
                             pass
-                    
+
                     # Only process if target is NOT a Gemini model
                     if not _is_gemini_model(model, custom_llm_provider):
                         verbose_logger.debug(
                             "Removing thought signatures from tool call IDs for non-Gemini model"
                         )
-                        
+
                         # Process messages to remove thought signatures
                         processed_messages = _remove_thought_signatures_from_messages(
                             messages, THOUGHT_SIGNATURE_SEPARATOR
                         )
-                        
+
                         # Update messages in kwargs or args
                         if "messages" in kwargs:
                             kwargs["messages"] = processed_messages
@@ -2977,7 +2977,7 @@ def get_optional_params_embeddings(  # noqa: PLR0915
 ):
     # Lazy load get_supported_openai_params
     get_supported_openai_params = getattr(sys.modules[__name__], 'get_supported_openai_params')
-    
+
     # retrieve all parameters passed to the function
     passed_params = locals()
     custom_llm_provider = passed_params.pop("custom_llm_provider", None)
@@ -4063,7 +4063,14 @@ def get_optional_params(  # noqa: PLR0915
                 ),
             )
         elif "anthropic" in bedrock_base_model and bedrock_route == "invoke":
-            if bedrock_base_model.startswith("anthropic.claude-3"):
+            # Check for Claude 3+ models (Messages API) including regional prefixes and Claude 4
+            # Models like eu.anthropic.claude-opus-4-5, us.anthropic.claude-3-5-sonnet, etc.
+            bedrock_base_model_lower = bedrock_base_model.lower()
+            is_messages_api_model = any(
+                indicator in bedrock_base_model_lower
+                for indicator in ["claude-3", "claude-opus-4", "claude-sonnet-4", "claude-haiku-4"]
+            )
+            if is_messages_api_model:
                 optional_params = (
                     litellm.AmazonAnthropicClaudeConfig().map_openai_params(
                         non_default_params=non_default_params,
@@ -6911,7 +6918,7 @@ def get_valid_models(
         # init litellm_params
         #################################
         from litellm.types.router import LiteLLM_Params
-        
+
         if litellm_params is None:
             litellm_params = LiteLLM_Params(model="")
         if api_key is not None:
@@ -7513,7 +7520,7 @@ class ProviderConfigManager:
             return litellm.IBMWatsonXAIConfig()
         elif litellm.LlmProviders.EMPOWER == provider:
             return litellm.EmpowerChatConfig()
-        elif litellm.LlmProviders.MINIMAX == provider:            
+        elif litellm.LlmProviders.MINIMAX == provider:
             return litellm.MinimaxChatConfig()
         elif litellm.LlmProviders.GITHUB == provider:
             return litellm.GithubChatConfig()
@@ -8314,8 +8321,7 @@ class ProviderConfigManager:
             from litellm.llms.vertex_ai.ocr.common_utils import get_vertex_ai_ocr_config
 
             return get_vertex_ai_ocr_config(model=model)
-        
-        MistralOCRConfig = getattr(sys.modules[__name__], 'MistralOCRConfig')
+MistralOCRConfig = getattr(sys.modules[__name__], 'MistralOCRConfig')
         PROVIDER_TO_CONFIG_MAP = {
             litellm.LlmProviders.MISTRAL: MistralOCRConfig,
         }
@@ -8752,12 +8758,12 @@ def __getattr__(name: str) -> Any:
     """Lazy import handler for utils module with cached registry for improved performance."""
     # Use cached registry from _lazy_imports instead of importing tuples every time
     from litellm._lazy_imports import _get_lazy_import_registry
-    
+
     registry = _get_lazy_import_registry()
-    
+
     # Check if name is in registry and call the cached handler function
     if name in registry:
         handler_func = registry[name]
         return handler_func(name)
-    
+
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")