fix(caching): skip injection on messages whose tool calls carry marks

Reverts 9f08d8aef8. A default 5m mark injected on assistant text lands
before the client's 1h tool_use mark, which Anthropic rejects with a 400
because a 1h breakpoint must not follow a 5m one. Keeping the full census
in the skip check leaves the client's tool_call breakpoint as the only
one on that message.

litellm/integrations/anthropic_cache_control_hook.py

@@ -518,26 +518,22 @@ class AnthropicCacheControlHook(CustomPromptManagement):
         return []
 
     @staticmethod
-    def _count_message_and_content_breakpoints(message: object) -> int:
+    def _count_cache_control_blocks(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 AnthropicCacheControlHook._count_message_and_content_breakpoints(message) + tool_call_count
+        return message_count + content_count + tool_call_count
 
     @staticmethod
     def _message_has_cache_control(message: AllMessageValues) -> bool:
-        """Return True if message-level or content-block cache_control is present."""
-        return AnthropicCacheControlHook._count_message_and_content_breakpoints(message) > 0
+        """Return True if the message already carries any cache_control."""
+        return AnthropicCacheControlHook._count_cache_control_blocks(message) > 0
 
     @staticmethod
     def _safe_insert_cache_control_in_message(

tests/unit/integrations/test_anthropic_cache_control_hook.py

@@ -1202,40 +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():
+def test_injection_skips_assistant_whose_tool_call_carries_a_longer_ttl_mark():
     hook: Final = AnthropicCacheControlHook()
     tool_call_ttl: Final = {"type": "ephemeral", "ttl": "1h"}
-    messages: Final = [
+    messages: Final[list[AllMessageValues]] = [
+        {"role": "user", "content": "What is the weather in Paris?"},
         {
             "role": "assistant",
-            "content": "ok",
+            "content": "Let me look that up.",
             "tool_calls": [
                 {
-                    "id": "call_1",
+                    "id": "toolu_01A",
                     "type": "function",
-                    "function": {"name": "lookup", "arguments": "{}"},
+                    "function": {"name": "get_weather", "arguments": '{"city": "Paris"}'},
                     "cache_control": tool_call_ttl,
                 }
             ],
-        }
+        },
+        {"role": "tool", "tool_call_id": "toolu_01A", "content": "18C and sunny"},
     ]
 
     _, processed, _ = hook.get_chat_completion_prompt(
-        model="bedrock/us.anthropic.claude-opus-4-6-v1:0",
+        model="anthropic/claude-haiku-4-5",
         messages=messages,
-        non_default_params={"cache_control_injection_points": [{"location": "message", "index": -1}]},
+        non_default_params={"cache_control_injection_points": [{"location": "message", "role": "assistant"}]},
         prompt_id=None,
         prompt_variables=None,
         dynamic_callback_params={},
     )
 
-    assistant_message: Final = processed[0]
+    assistant_message: Final = processed[1]
     assistant_tool_calls: Final = assistant_message.get("tool_calls")
-    assert assistant_message.get("cache_control") == {"type": "ephemeral"}
+    assert assistant_message.get("cache_control") is None
+    assert assistant_message.get("content") == "Let me look that up."
     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
+    assert assistant_tool_calls[0].get("cache_control") == tool_call_ttl
+    assert AnthropicCacheControlHook.count_request_cache_breakpoints(processed) == 1
 
 
 def test_cache_control_hook_caps_at_four_blocks_with_client_cache_control():