address CI: black format, fix empty-dict test, read prompt_tokens_details off Usage, drop redundant or-0

- black-format posthog.py + litellm_logging.py (lint gate)
- update test_get_usage_as_dict for the new cache keys in the _empty sentinel (core-utils)
- _inject_cache_tokens reads prompt_tokens_details off the Usage object, not the passed dict — correct regardless of caller; tolerates wrapper or dict
- add direct _inject_cache_tokens unit test (satisfies logging-coverage gate)
- drop redundant 'or 0' in posthog.py per review

litellm/integrations/posthog.py

@@ -229,8 +229,12 @@ class PostHogLogger(CustomBatchLogger):
             properties["$ai_output_tokens"] = self._safe_get(
                 standard_logging_object, "completion_tokens", 0
             )
-        cache_read = self._safe_get(standard_logging_object, "cache_read_input_tokens", 0) or 0
-        cache_creation = self._safe_get(standard_logging_object, "cache_creation_input_tokens", 0) or 0
+        cache_read = self._safe_get(
+            standard_logging_object, "cache_read_input_tokens", 0
+        )
+        cache_creation = self._safe_get(
+            standard_logging_object, "cache_creation_input_tokens", 0
+        )
         if cache_read:
             properties["$ai_cache_read_input_tokens"] = cache_read
         if cache_creation:

litellm/litellm_core_utils/litellm_logging.py

@@ -5197,8 +5197,10 @@ class StandardLoggingPayloadSetup:
             )
         if isinstance(_raw, dict):
             if ResponseAPILoggingUtils._is_response_api_usage(_raw):
-                usage = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(
-                    _raw
+                usage = (
+                    ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(
+                        _raw
+                    )
                 )
                 return StandardLoggingPayloadSetup._inject_cache_tokens(
                     usage.model_dump(), usage
@@ -5223,13 +5225,23 @@ class StandardLoggingPayloadSetup:
           2. prompt_tokens_details.cached_tokens — set by OpenAI, Gemini, DeepSeek
 
         cache_creation_input_tokens: Anthropic and Bedrock only via private attr.
+
+        Reads prompt_tokens_details off the Usage object (not ``d``) so the
+        result is correct regardless of what the caller passes as ``d``.
+        prompt_tokens_details may be a Pydantic wrapper (attribute access) or a
+        dict, so tolerate both.
         """
-        ptd = d.get("prompt_tokens_details") or {}
-        ptd_cached = (ptd.get("cached_tokens") or 0) if isinstance(ptd, dict) else 0
+        ptd = getattr(usage, "prompt_tokens_details", None)
+        if isinstance(ptd, dict):
+            ptd_cached = ptd.get("cached_tokens") or 0
+        else:
+            ptd_cached = getattr(ptd, "cached_tokens", 0) or 0
         d["cache_read_input_tokens"] = (
             getattr(usage, "_cache_read_input_tokens", 0) or ptd_cached
         )
-        d["cache_creation_input_tokens"] = getattr(usage, "_cache_creation_input_tokens", 0) or 0
+        d["cache_creation_input_tokens"] = (
+            getattr(usage, "_cache_creation_input_tokens", 0) or 0
+        )
         return d
 
     @staticmethod
@@ -5927,7 +5939,9 @@ def get_standard_logging_object_payload(
             prompt_tokens=usage_dict.get("prompt_tokens", 0),
             completion_tokens=usage_dict.get("completion_tokens", 0),
             cache_read_input_tokens=usage_dict.get("cache_read_input_tokens", 0),
-            cache_creation_input_tokens=usage_dict.get("cache_creation_input_tokens", 0),
+            cache_creation_input_tokens=usage_dict.get(
+                "cache_creation_input_tokens", 0
+            ),
             request_tags=request_tags,
             end_user=end_user_id or "",
             api_base=StandardLoggingPayloadSetup.strip_trailing_slash(

tests/logging_callback_tests/test_standard_logging_payload.py

@@ -1105,8 +1105,12 @@ def test_merge_litellm_metadata_bedrock_passthrough_scenario():
 
 # --- cache token tests ---
 
-def _make_usage(cache_read: int = 0, cache_creation: int = 0, ptd_cached: int = 0) -> Usage:
+
+def _make_usage(
+    cache_read: int = 0, cache_creation: int = 0, ptd_cached: int = 0
+) -> Usage:
     from litellm.types.utils import PromptTokensDetailsWrapper
+
     u = Usage(prompt_tokens=100, completion_tokens=50, total_tokens=150)
     u._cache_read_input_tokens = cache_read
     u._cache_creation_input_tokens = cache_creation
@@ -1154,3 +1158,29 @@ def test_get_usage_as_dict_from_response_obj():
     d = StandardLoggingPayloadSetup.get_usage_as_dict(response_obj={"usage": usage})
     assert d["cache_read_input_tokens"] == 300
     assert d["cache_creation_input_tokens"] == 100
+
+
+def test_inject_cache_tokens_prefers_private_attr_then_prompt_tokens_details():
+    """Direct coverage for the _inject_cache_tokens helper: the private
+    Anthropic/Bedrock attrs win, with a prompt_tokens_details.cached_tokens
+    fallback for OpenAI/Gemini/DeepSeek."""
+    from litellm.types.utils import PromptTokensDetailsWrapper
+
+    # Private attrs present → used directly.
+    anthropic = Usage(prompt_tokens=10, completion_tokens=5, total_tokens=15)
+    anthropic._cache_read_input_tokens = 200
+    anthropic._cache_creation_input_tokens = 500
+    d = StandardLoggingPayloadSetup._inject_cache_tokens({}, anthropic)
+    assert d["cache_read_input_tokens"] == 200
+    assert d["cache_creation_input_tokens"] == 500
+
+    # No private attr → fall back to prompt_tokens_details.cached_tokens.
+    openai = Usage(
+        prompt_tokens=10,
+        completion_tokens=5,
+        total_tokens=15,
+        prompt_tokens_details=PromptTokensDetailsWrapper(cached_tokens=80),
+    )
+    d = StandardLoggingPayloadSetup._inject_cache_tokens({}, openai)
+    assert d["cache_read_input_tokens"] == 80
+    assert d["cache_creation_input_tokens"] == 0

tests/test_litellm/litellm_core_utils/test_litellm_logging.py

@@ -1726,11 +1726,23 @@ def test_get_usage_as_dict():
 
     # Test case 1: None response_obj returns empty usage dict
     result = StandardLoggingPayloadSetup.get_usage_as_dict(response_obj=None)
-    assert result == {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
+    assert result == {
+        "prompt_tokens": 0,
+        "completion_tokens": 0,
+        "total_tokens": 0,
+        "cache_read_input_tokens": 0,
+        "cache_creation_input_tokens": 0,
+    }
 
     # Test case 2: Empty response_obj returns empty usage dict
     result = StandardLoggingPayloadSetup.get_usage_as_dict(response_obj={})
-    assert result == {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
+    assert result == {
+        "prompt_tokens": 0,
+        "completion_tokens": 0,
+        "total_tokens": 0,
+        "cache_read_input_tokens": 0,
+        "cache_creation_input_tokens": 0,
+    }
 
     # Test case 3: combined_usage_object takes priority
     combined = Usage(prompt_tokens=10, completion_tokens=5, total_tokens=15)
@@ -1752,7 +1764,13 @@ def test_get_usage_as_dict():
     result = StandardLoggingPayloadSetup.get_usage_as_dict(
         response_obj={"id": "resp-1", "choices": []}
     )
-    assert result == {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
+    assert result == {
+        "prompt_tokens": 0,
+        "completion_tokens": 0,
+        "total_tokens": 0,
+        "cache_read_input_tokens": 0,
+        "cache_creation_input_tokens": 0,
+    }
 
 
 def test_append_system_prompt_messages():