Merge pull request #18787 from aproorg/fix/bedrock-thinking-tool-call-2

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

litellm/llms/bedrock/chat/converse_transformation.py

@@ -53,7 +53,13 @@ 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,
+    any_assistant_message_has_thinking_blocks,
+    has_tool_call_blocks,
+    last_assistant_with_tool_calls_has_no_thinking_blocks,
+    supports_reasoning,
+)
 
 from ..common_utils import (
     BedrockError,
@@ -773,7 +779,7 @@ 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
@@ -1070,9 +1076,28 @@ 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"
+        #
+        # IMPORTANT: Only drop thinking if NO assistant messages have thinking_blocks.
+        # If any message has thinking_blocks, we must keep thinking enabled, otherwise
+        # 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)
+            and not any_assistant_message_has_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", [])
@@ -1459,11 +1484,11 @@ 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"]
@@ -838,7 +860,7 @@ 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] = []
@@ -936,13 +958,13 @@ 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

@@ -4186,7 +4186,14 @@ 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,