fix(bedrock/thinking/websearch): beta headers, thinking fixes, API key loading, guardrails null safety

litellm/constants.py

@@ -1083,7 +1083,6 @@
     "openai.gpt-oss-120b-1:0",
     "anthropic.claude-haiku-4-5-20251001-v1:0",
     "anthropic.claude-sonnet-4-5-20250929-v1:0",
-    "anthropic.claude-opus-4-6-v1:0",
     "anthropic.claude-opus-4-6-v1",
     "anthropic.claude-sonnet-4-6",
     "anthropic.claude-opus-4-1-20250805-v1:0",

litellm/integrations/websearch_interception/handler.py

@@ -697,8 +697,10 @@ async def _execute_search(self, query: str) -> str:
                 )
                 llm_router = None
 
-            # Determine search provider from router's search_tools
+            # Determine search provider and credentials from router's search_tools
             search_provider: Optional[str] = None
+            api_key: Optional[str] = None
+            api_base: Optional[str] = None
             if llm_router is not None and hasattr(llm_router, "search_tools"):
                 if self.search_tool_name:
                     # Find specific search tool by name
@@ -709,9 +711,10 @@ async def _execute_search(self, query: str) -> str:
                     ]
                     if matching_tools:
                         search_tool = matching_tools[0]
-                        search_provider = search_tool.get("litellm_params", {}).get(
-                            "search_provider"
-                        )
+                        litellm_params = search_tool.get("litellm_params", {})
+                        search_provider = litellm_params.get("search_provider")
+                        api_key = litellm_params.get("api_key")
+                        api_base = litellm_params.get("api_base")
                         verbose_logger.debug(
                             f"WebSearchInterception: Found search tool '{self.search_tool_name}' "
                             f"with provider '{search_provider}'"
@@ -725,9 +728,10 @@ async def _execute_search(self, query: str) -> str:
                 # If no specific tool or not found, use first available
                 if not search_provider and llm_router.search_tools:
                     first_tool = llm_router.search_tools[0]
-                    search_provider = first_tool.get("litellm_params", {}).get(
-                        "search_provider"
-                    )
+                    litellm_params = first_tool.get("litellm_params", {})
+                    search_provider = litellm_params.get("search_provider")
+                    api_key = litellm_params.get("api_key")
+                    api_base = litellm_params.get("api_base")
                     verbose_logger.debug(
                         f"WebSearchInterception: Using first available search tool with provider '{search_provider}'"
                     )
@@ -743,7 +747,15 @@ async def _execute_search(self, query: str) -> str:
             verbose_logger.debug(
                 f"WebSearchInterception: Executing search for '{query}' using provider '{search_provider}'"
             )
-            result = await litellm.asearch(query=query, search_provider=search_provider)
+            search_kwargs: Dict[str, Any] = {
+                "query": query,
+                "search_provider": search_provider,
+            }
+            if api_key:
+                search_kwargs["api_key"] = api_key
+            if api_base:
+                search_kwargs["api_base"] = api_base
+            result = await litellm.asearch(**search_kwargs)
 
             # Format using transformation function
             search_result_text = WebSearchTransformation.format_search_response(result)

litellm/llms/anthropic/chat/transformation.py

@@ -63,6 +63,7 @@
     any_assistant_message_has_thinking_blocks,
     get_max_tokens,
     has_tool_call_blocks,
+    last_assistant_message_has_no_thinking_blocks,
     last_assistant_with_tool_calls_has_no_thinking_blocks,
     supports_reasoning,
     token_counter,
@@ -1341,7 +1342,9 @@ def transform_request(
                 )
 
         # 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"
+        # This prevents Anthropic errors:
+        # - "Expected thinking or redacted_thinking, but found tool_use" (assistant with tool_calls)
+        # - "Expected thinking or redacted_thinking, but found text" (assistant with text content)
         #
         # IMPORTANT: Only drop thinking if NO assistant messages have thinking_blocks.
         # If any message has thinking_blocks, we must keep thinking enabled, otherwise
@@ -1350,7 +1353,10 @@ def transform_request(
         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 (
+                last_assistant_with_tool_calls_has_no_thinking_blocks(messages)
+                or last_assistant_message_has_no_thinking_blocks(messages)
+            )
             and not any_assistant_message_has_thinking_blocks(messages)
         ):
             if litellm.modify_params:

litellm/llms/base_llm/chat/transformation.py

@@ -110,8 +110,9 @@ def get_json_schema_from_pydantic_object(
         return type_to_response_format_param(response_format=response_format)
 
     def is_thinking_enabled(self, non_default_params: dict) -> bool:
+        thinking_type = non_default_params.get("thinking", {}).get("type")
         return (
-            non_default_params.get("thinking", {}).get("type") == "enabled"
+            thinking_type in ("enabled", "adaptive")
             or non_default_params.get("reasoning_effort") is not None
         )
 

litellm/llms/bedrock/README.md

@@ -0,0 +1,67 @@
+# AWS Bedrock Provider
+
+This directory contains the AWS Bedrock provider implementation for LiteLLM.
+
+## Beta Headers Management
+
+### Overview
+
+Bedrock anthropic-beta header handling uses a centralized whitelist-based filter (`beta_headers_config.py`) across all three Bedrock APIs to ensure:
+- Only supported headers reach AWS (prevents API errors)
+- Consistent behavior across Invoke Chat, Invoke Messages, and Converse APIs
+- Zero maintenance when new Claude models are released
+
+### Key Features
+
+1. **Version-Based Filtering**: Headers specify minimum version (e.g., "requires Claude 4.5+") instead of hardcoded model lists
+2. **Family Restrictions**: Can limit headers to specific families (opus/sonnet/haiku)
+3. **Automatic Translation**: `advanced-tool-use` → `tool-search-tool` + `tool-examples` for backward compatibility
+
+### Adding New Beta Headers
+
+When AWS Bedrock adds support for a new Anthropic beta header, update `beta_headers_config.py`:
+
+```python
+# 1. Add to whitelist
+BEDROCK_CORE_SUPPORTED_BETAS.add("new-feature-2027-01-15")
+
+# 2. (Optional) Add version requirement
+BETA_HEADER_MINIMUM_VERSION["new-feature-2027-01-15"] = 5.0
+
+# 3. (Optional) Add family restriction
+BETA_HEADER_FAMILY_RESTRICTIONS["new-feature-2027-01-15"] = ["opus"]
+```
+
+Then add tests in `tests/test_litellm/llms/bedrock/test_beta_headers_config.py`.
+
+### Adding New Claude Models
+
+When Anthropic releases new models (e.g., Claude Opus 5):
+- **Required code changes**: ZERO ✅
+- The version-based filter automatically handles new models
+- No hardcoded lists to update
+
+### Testing
+
+```bash
+# Test beta headers filtering
+poetry run pytest tests/test_litellm/llms/bedrock/test_beta_headers_config.py -v
+
+# Test API integrations
+poetry run pytest tests/test_litellm/llms/bedrock/test_anthropic_beta_support.py -v
+
+# Test everything
+poetry run pytest tests/test_litellm/llms/bedrock/ -v
+```
+
+### Debug Logging
+
+Enable debug logging to see filtering decisions:
+```bash
+LITELLM_LOG=DEBUG
+```
+
+### References
+
+- [AWS Bedrock Documentation](https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages-request-response.html)
+- [Anthropic Beta Headers](https://docs.anthropic.com/claude/reference/versioning)