feat(bedrock): inject thinking for clear_thinking context_management on Messages API

Bedrock rejects clear_thinking_20251015 unless thinking is enabled or adaptive.
Inject minimal extended thinking and interleaved-thinking beta when Claude Code
sends context_management without thinking. Adds unit tests.

Made-with: Cursor

litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py

@@ -13,6 +13,8 @@ from typing import (
 import httpx
 
 from litellm.anthropic_beta_headers_manager import filter_and_transform_beta_headers
+from litellm.constants import BEDROCK_MIN_THINKING_BUDGET_TOKENS
+from litellm.litellm_core_utils.litellm_logging import verbose_logger
 from litellm.llms.anthropic.common_utils import AnthropicModelInfo
 from litellm.llms.anthropic.experimental_pass_through.messages.transformation import (
     AnthropicMessagesConfig,
@@ -213,6 +215,70 @@ class AmazonAnthropicClaudeMessagesConfig(
 
         return any(pattern in model_lower for pattern in supported_patterns)
 
+    def _ensure_thinking_for_clear_thinking_context_management(
+        self,
+        anthropic_messages_request: Dict,
+        model: str,
+    ) -> bool:
+        """
+        Bedrock rejects ``clear_thinking_20251015`` context-management edits unless
+        extended thinking is ``enabled`` or ``adaptive``. Claude Code often sends
+        context management without a top-level ``thinking`` field.
+
+        When we detect that edit type on a model that supports extended thinking on
+        Bedrock, inject a minimal ``thinking`` config so the request succeeds.
+
+        Returns:
+            True if ``thinking`` was added or upgraded for this fix (caller may
+            need to add the interleaved-thinking beta header).
+        """
+        cm = anthropic_messages_request.get("context_management")
+        if not isinstance(cm, dict):
+            return False
+        edits = cm.get("edits")
+        if not isinstance(edits, list):
+            return False
+        needs_thinking = any(
+            isinstance(e, dict) and e.get("type") == "clear_thinking_20251015"
+            for e in edits
+        )
+        if not needs_thinking:
+            return False
+        if not self._supports_extended_thinking_on_bedrock(model):
+            return False
+
+        thinking = anthropic_messages_request.get("thinking")
+        if isinstance(thinking, dict):
+            t = thinking.get("type")
+            if t in ("enabled", "adaptive"):
+                return False
+            # ``disabled`` or unknown — replace with enabled so clear_thinking is valid
+            verbose_logger.debug(
+                "Bedrock clear_thinking_20251015: replacing thinking=%s with minimal enabled thinking",
+                thinking,
+            )
+
+        max_tokens = anthropic_messages_request.get("max_tokens")
+        budget = BEDROCK_MIN_THINKING_BUDGET_TOKENS
+        if isinstance(max_tokens, int) and max_tokens <= budget:
+            verbose_logger.warning(
+                "Bedrock clear_thinking_20251015: max_tokens=%s is not greater than "
+                "minimum thinking budget (%s); cannot inject thinking safely",
+                max_tokens,
+                budget,
+            )
+            return False
+
+        anthropic_messages_request["thinking"] = {
+            "type": "enabled",
+            "budget_tokens": budget,
+        }
+        verbose_logger.debug(
+            "Bedrock clear_thinking_20251015: injected thinking with budget_tokens=%s",
+            budget,
+        )
+        return True
+
     def _is_claude_opus_4_5(self, model: str) -> bool:
         """
         Check if the model is Claude Opus 4.5.
@@ -414,6 +480,13 @@ class AmazonAnthropicClaudeMessagesConfig(
         if "model" in anthropic_messages_request:
             anthropic_messages_request.pop("model", None)
 
+        injected_thinking_for_clear_thinking = (
+            self._ensure_thinking_for_clear_thinking_context_management(
+                anthropic_messages_request=anthropic_messages_request,
+                model=model,
+            )
+        )
+
         # 4. Remove `ttl` field from cache_control in messages (Bedrock doesn't support it for older models)
         self._remove_ttl_from_cache_control(
             anthropic_messages_request=anthropic_messages_request, model=model
@@ -463,6 +536,9 @@ class AmazonAnthropicClaudeMessagesConfig(
         )
         beta_set.update(auto_betas)
 
+        if injected_thinking_for_clear_thinking:
+            beta_set.add("interleaved-thinking-2025-05-14")
+
         self._get_tool_search_beta_header_for_bedrock(
             model=model,
             tool_search_used=tool_search_used,

tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py

@@ -18,6 +18,7 @@ from litellm.llms.bedrock.common_utils import (
     normalize_tool_input_schema_types_for_bedrock_invoke,
     remove_custom_field_from_tools,
 )
+from litellm.constants import BEDROCK_MIN_THINKING_BUDGET_TOKENS
 from litellm.llms.bedrock.messages.invoke_transformations.anthropic_claude3_transformation import (
     AmazonAnthropicClaudeMessagesConfig,
     AmazonAnthropicClaudeMessagesStreamDecoder,
@@ -384,6 +385,58 @@ def test_bedrock_invoke_messages_transform_adds_name_when_tool_missing_name():
     assert result["tools"][0]["name"] == "litellm_unnamed_tool_0"
 
 
+def test_bedrock_invoke_messages_injects_thinking_for_clear_thinking_context_management():
+    """
+    Bedrock requires extended thinking when ``clear_thinking_20251015`` appears in
+    ``context_management`` (Claude Code sends CM without ``thinking``).
+    """
+    from litellm.types.router import GenericLiteLLMParams
+
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+    optional_params = {
+        "max_tokens": 32000,
+        "stream": False,
+        "context_management": {
+            "edits": [{"type": "clear_thinking_20251015", "keep": "all"}]
+        },
+    }
+    result = cfg.transform_anthropic_messages_request(
+        model="global.anthropic.claude-sonnet-4-6-v1:0",
+        messages=[{"role": "user", "content": "hi"}],
+        anthropic_messages_optional_request_params=copy.deepcopy(optional_params),
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+    assert result["thinking"]["type"] == "enabled"
+    assert result["thinking"]["budget_tokens"] == BEDROCK_MIN_THINKING_BUDGET_TOKENS
+    betas = result.get("anthropic_beta") or []
+    assert "interleaved-thinking-2025-05-14" in betas
+
+
+def test_bedrock_invoke_messages_skips_thinking_injection_when_already_enabled():
+    from litellm.types.router import GenericLiteLLMParams
+
+    cfg = AmazonAnthropicClaudeMessagesConfig()
+    optional_params = {
+        "max_tokens": 32000,
+        "stream": False,
+        "thinking": {"type": "enabled", "budget_tokens": 2048},
+        "context_management": {
+            "edits": [{"type": "clear_thinking_20251015", "keep": "all"}]
+        },
+    }
+    result = cfg.transform_anthropic_messages_request(
+        model="global.anthropic.claude-sonnet-4-6-v1:0",
+        messages=[{"role": "user", "content": "hi"}],
+        anthropic_messages_optional_request_params=copy.deepcopy(optional_params),
+        litellm_params=GenericLiteLLMParams(),
+        headers={},
+    )
+    assert result["thinking"]["budget_tokens"] == 2048
+    betas = result.get("anthropic_beta") or []
+    assert "interleaved-thinking-2025-05-14" not in betas
+
+
 def test_bedrock_invoke_messages_transform_converts_custom_tool_schema_type_to_object():
     """
     End-to-end: AmazonAnthropicClaudeMessagesConfig must emit Bedrock Invoke bodies