fix(caching): only skip injection on message or content marks

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/integrations/anthropic_cache_control_hook.py

@@ -518,22 +518,26 @@ class AnthropicCacheControlHook(CustomPromptManagement):
         return []
 
     @staticmethod
-    def _count_cache_control_blocks(message: object) -> int:
+    def _count_message_and_content_breakpoints(message: object) -> int:
         message_count: Final = 1 if _carries_cache_breakpoint(message) else 0
         content: Final = _as_object_list(_attribute_or_key(message, "content"))
         content_count: Final = sum(1 for block in content if _carries_cache_breakpoint(block)) if content else 0
+        return message_count + content_count
+
+    @staticmethod
+    def _count_cache_control_blocks(message: object) -> int:
         tool_calls: Final = _as_object_list(_attribute_or_key(message, "tool_calls"))
         tool_call_count: Final = (
             sum(1 for tool_call in tool_calls if _tool_call_cache_control_is_forwarded(tool_call, message))
             if tool_calls
             else 0
         )
-        return message_count + content_count + tool_call_count
+        return AnthropicCacheControlHook._count_message_and_content_breakpoints(message) + tool_call_count
 
     @staticmethod
     def _message_has_cache_control(message: AllMessageValues) -> bool:
-        """Return True if the message already carries any cache_control."""
+        """Return True if message-level or content-block cache_control is present."""
-        return AnthropicCacheControlHook._count_cache_control_blocks(message) > 0
+        return AnthropicCacheControlHook._count_message_and_content_breakpoints(message) > 0
 
     @staticmethod
     def _safe_insert_cache_control_in_message(

tests/unit/integrations/test_anthropic_cache_control_hook.py

@@ -1202,6 +1202,42 @@ def test_cache_control_hook_counts_pydantic_message_tool_call_marks():
     assert AnthropicCacheControlHook.count_request_cache_breakpoints(cast(list[AllMessageValues], [message])) == 1
 
 
+def test_injection_adds_message_mark_without_overwriting_tool_call_ttl():
+    hook: Final = AnthropicCacheControlHook()
+    tool_call_ttl: Final = {"type": "ephemeral", "ttl": "1h"}
+    messages: Final = [
+        {
+            "role": "assistant",
+            "content": "ok",
+            "tool_calls": [
+                {
+                    "id": "call_1",
+                    "type": "function",
+                    "function": {"name": "lookup", "arguments": "{}"},
+                    "cache_control": tool_call_ttl,
+                }
+            ],
+        }
+    ]
+
+    _, processed, _ = hook.get_chat_completion_prompt(
+        model="bedrock/us.anthropic.claude-opus-4-6-v1:0",
+        messages=messages,
+        non_default_params={"cache_control_injection_points": [{"location": "message", "index": -1}]},
+        prompt_id=None,
+        prompt_variables=None,
+        dynamic_callback_params={},
+    )
+
+    assistant_message: Final = processed[0]
+    assistant_tool_calls: Final = assistant_message.get("tool_calls")
+    assert assistant_message.get("cache_control") == {"type": "ephemeral"}
+    assert isinstance(assistant_tool_calls, list)
+    tool_call: Final = assistant_tool_calls[0]
+    assert tool_call.get("cache_control") == tool_call_ttl
+    assert AnthropicCacheControlHook.count_request_cache_breakpoints(processed) == 2
+
+
 def test_cache_control_hook_caps_at_four_blocks_with_client_cache_control():
     """Regression for LIT-3667 / Anthropic 'A maximum of 4 blocks ... Found 5'.