fix(types): safe attribute access on PromptTokensDetailsWrapper deleted fields

__init__ deletes None-valued cache fields from the instance dict for payload
hygiene, but pydantic v2 then raises AttributeError on direct reads of those
deleted fields (issue #43756: first-turn responses with no cache activity
crash downstream consumers that read wrapper.cache_creation_tokens directly).

Add a __getattr__ fallback on the wrapper returning the declared class
default for deleted fields, keeping payload hygiene (model_dump still
excludes them) while eliminating the crash. The old hasattr-based
expectations in test_prompt_tokens_details_no_cache_write_tokens_when_absent
are updated to the new direct-read contract, plus regression tests for both
deleted-field reads and dump exclusion.

Fixes #43756

litellm/types/utils.py

@@ -1827,6 +1827,23 @@ class PromptTokensDetailsWrapper(
         if self.cached_tokens_details is None:
             del self.cached_tokens_details
 
+    def __getattr__(self, name: str):
+        """Return declared defaults for payload-hygiene deletions (issue #43756).
+
+        __init__ deletes None-valued fields from the instance dict so they stay
+        out of serialized payloads; pydantic v2 then raises on direct attribute
+        access to those deleted fields. This fallback returns the declared
+        class default instead, so downstream direct reads
+        (e.g. wrapper.cache_creation_tokens on a first-turn response with no
+        cache activity) behave like a zero-valued field rather than crashing.
+        """
+        field_info = type(self).model_fields.get(name)
+        if field_info is not None:
+            return field_info.get_default()
+        raise AttributeError(
+            f"{type(self).__name__!s} object has no attribute {name!r}"
+        )
+
 
 class ServerToolUse(BaseModel):
     web_search_requests: int | None = None

tests/unit/test_utils.py

@@ -154,11 +154,16 @@ def test_usage_anthropic_cache_creation_maps_to_cache_write_tokens():
 
 
 def test_prompt_tokens_details_no_cache_write_tokens_when_absent():
-    """A read-only cache hit (no cache write) must not surface cache-write fields."""
+    """A read-only cache hit (no cache write) must not surface cache-write fields
+    in the serialized payload; direct reads of deleted fields return the declared
+    defaults instead of raising AttributeError (issue #43756)."""
     details = PromptTokensDetailsWrapper(cached_tokens=800)
     assert details.cached_tokens == 800
-    assert not hasattr(details, "cache_write_tokens")
-    assert not hasattr(details, "cache_creation_tokens")
+    assert details.cache_write_tokens is None
+    assert details.cache_creation_tokens is None
+    dumped = details.model_dump()
+    assert "cache_write_tokens" not in dumped
+    assert "cache_creation_tokens" not in dumped
 
 
 def test_prompt_tokens_details_cache_write_creation_stay_in_sync_on_assignment():
@@ -174,6 +179,32 @@ def test_prompt_tokens_details_cache_write_creation_stay_in_sync_on_assignment()
     assert details.cache_write_tokens == details.cache_creation_tokens == 375
 
 
+def test_prompt_tokens_details_none_cache_fields_read_as_none():
+    """Direct reads of deleted None-valued fields must return the declared
+    defaults instead of raising AttributeError (issue #43756)."""
+    wrapper = PromptTokensDetailsWrapper(
+        prompt_tokens=100,
+        cache_creation_tokens=None,
+        cache_read_input_tokens=None,
+        cache_write_tokens=None,
+    )
+    assert wrapper.cache_creation_tokens is None
+    assert wrapper.cache_write_tokens is None
+
+
+def test_prompt_tokens_details_dump_excludes_deleted_none_fields():
+    """Payload hygiene survives: deleted None fields stay out of model_dump."""
+    wrapper = PromptTokensDetailsWrapper(
+        prompt_tokens=100,
+        cache_creation_tokens=None,
+        cache_read_input_tokens=None,
+        cache_write_tokens=None,
+    )
+    dumped = wrapper.model_dump()
+    assert "cache_write_tokens" not in dumped
+    assert "cache_creation_tokens" not in dumped
+
+
 def test_potential_model_names_keeps_provider_prefixed_candidate():
     """A provider whose own model ids repeat the litellm provider name (Perplexity's
     Agent API serves `perplexity/glm-5.2`, mapped as `perplexity/perplexity/glm-5.2`)