test(caching): add regression coverage for provider_specific_fields cache round trip

tests/test_litellm/caching/test_cache_provider_specific_fields.py

@@ -0,0 +1,71 @@
+"""
+Regression test for https://github.com/BerriAI/litellm/issues/13048
+
+When caching is enabled, a cached completion response must preserve
+`provider_specific_fields` (e.g. Anthropic web_search citations) after the
+store -> load -> reconstruct round trip.
+"""
+
+import pytest
+
+import litellm
+from litellm import Cache, ModelResponse
+from litellm.types.utils import Choices, Message
+
+
+def _make_response_with_citations() -> ModelResponse:
+    return ModelResponse(
+        id="chatcmpl-test",
+        model="claude-sonnet-4-20250514",
+        choices=[
+            Choices(
+                finish_reason="stop",
+                index=0,
+                message=Message(
+                    content="Pears are fruits.",
+                    role="assistant",
+                    provider_specific_fields={
+                        "citations": [
+                            {
+                                "type": "web_search_result_location",
+                                "cited_text": "Pears are fruits.",
+                                "url": "https://en.wikipedia.org/wiki/Pear",
+                            }
+                        ]
+                    },
+                ),
+            )
+        ],
+    )
+
+
+@pytest.mark.parametrize("cache_type", ["local", "disk"])
+def test_cache_round_trip_preserves_provider_specific_fields(cache_type, tmp_path):
+    litellm.cache = Cache(type=cache_type, disk_cache_dir=str(tmp_path / "litellm_cache"))
+    try:
+        kwargs = {
+            "model": "claude-sonnet-4-20250514",
+            "messages": [{"role": "user", "content": "Search the web and tell me about pears."}],
+        }
+        original = _make_response_with_citations()
+
+        litellm.cache.add_cache(result=original, **kwargs)
+
+        cached = litellm.cache.get_cache(**kwargs)
+        assert cached is not None, "expected a cache hit"
+
+        # Reconstruct the same way `_convert_cached_result_to_model_response` does
+        from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import (
+            convert_to_model_response_object,
+        )
+
+        reconstructed = convert_to_model_response_object(
+            response_object=cached,
+            model_response_object=ModelResponse(),
+        )
+
+        psf = reconstructed.choices[0].message.provider_specific_fields
+        assert psf is not None, "provider_specific_fields were dropped by the cache round trip"
+        assert "citations" in psf, f"citations missing from provider_specific_fields: {psf}"
+    finally:
+        litellm.cache = None