refactor(vertex_ai): narrow transformation scope and preserve schema defaults

- Restrict recursive camelCase transformation to 'contents', 'systemInstruction', and 'tools' to avoid breaking existing users.
- Add logic to preserve user-supplied 'default' values in JSON Schema properties.
- Fix 'labels' preservation to only trigger at top-level.
- Ensure type-safe return for context caching with explicit cast.

litellm/llms/vertex_ai/context_caching/transformation.py

@@ -208,5 +208,11 @@ def transform_openai_messages_to_gemini_context_caching(
                 {"system_instruction": data["system_instruction"]}, parent_key=None
             )["systemInstruction"]
             del data["system_instruction"]
+        if "tools" in data:
+            data["tools"] = _transform_part_to_httpx_format(
+                {"tools": data["tools"]}, parent_key=None
+            )["tools"]
 
-    return data
+    from typing import cast
+
+    return cast(CachedContentRequestBody, data)

litellm/llms/vertex_ai/gemini/transformation.py

@@ -267,9 +267,12 @@ def _transform_part_to_httpx_format(part: dict, parent_key: Optional[str] = None
             elif k == "properties":
                 # 'properties' keys in a JSON Schema are user-defined
                 should_preserve_keys = True
-            elif k == "labels":
+            elif k == "labels" and parent_key is None:
                 # 'labels' is a top-level map for Vertex AI billing/tracking, keys are user-defined
                 should_preserve_keys = True
+            elif k == "default" and parent_key in ["properties", None]:
+                # 'default' values in a JSON Schema property are user-defined and should NOT be transformed
+                should_preserve_keys = True
 
         if should_preserve_keys:
             if k == "properties":
@@ -836,6 +839,10 @@ def _transform_request_body(  # noqa: PLR0915
                 {"system_instruction": data["system_instruction"]}, parent_key=None
             )["systemInstruction"]
             del data["system_instruction"]
+        if "tools" in data:
+            data["tools"] = _transform_part_to_httpx_format(
+                {"tools": data["tools"]}, parent_key=None
+            )["tools"]
 
     return data
 

tests/litellm/llms/vertex_ai/gemini/test_transformation.py

@@ -364,3 +364,64 @@ def test_vertex_transformation_field_casing():
     transformed = _transform_part_to_httpx_format(part)
     assert "labels" in transformed
     assert "response" in transformed["labels"]
+
+    # 6. Test default preservation in schema
+    part = {"properties": {"my_field": {"type": "object", "default": {"snake_case_key": 1}}}}
+    transformed = _transform_part_to_httpx_format(part, parent_key="properties")
+    assert "properties" in transformed
+    assert "my_field" in transformed["properties"]
+    assert "default" in transformed["properties"]["my_field"]
+    assert "snake_case_key" in transformed["properties"]["my_field"]["default"]
+
+    # 7. Test labels at non-top-level are camelCased
+    part = {"some_inner_object": {"labels": {"user_key": "value"}}}
+    transformed = _transform_part_to_httpx_format(part, parent_key="something")
+    assert "someInnerObject" in transformed
+    assert "labels" in transformed["someInnerObject"]
+    assert "userKey" in transformed["someInnerObject"]["labels"]
+
+
+@pytest.mark.asyncio
+async def test_vertex_request_body_tools_transformation():
+    from litellm.llms.vertex_ai.gemini.transformation import _transform_request_body
+    from litellm.types.utils import LlmProviders
+
+    messages = [{"role": "user", "content": "hi"}]
+    tools = [
+        {
+            "function_declarations": [
+                {
+                    "name": "my_func",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"my_param": {"type": "string", "format": "email"}},
+                    },
+                }
+            ]
+        }
+    ]
+
+    # Test for Vertex AI - should transform tools
+    transformed = _transform_request_body(
+        messages=messages,
+        model="gemini-1.5-flash",
+        optional_params={"tools": tools},
+        custom_llm_provider="vertex_ai",
+        litellm_params={},
+        cached_content=None,
+    )
+
+    assert "tools" in transformed
+    assert "functionDeclarations" in transformed["tools"][0]
+
+    # Test for Gemini (AI Studio) - should NOT transform tools
+    original = _transform_request_body(
+        messages=messages,
+        model="gemini-1.5-flash",
+        optional_params={"tools": tools},
+        custom_llm_provider=LlmProviders.GEMINI.value,
+        litellm_params={},
+        cached_content=None,
+    )
+    assert "tools" in original
+    assert "function_declarations" in original["tools"][0]

tests/test_litellm/llms/vertex_ai/context_caching/test_context_caching_ttl.py

@@ -342,7 +342,9 @@ class TestTransformationWithTTL:
         
         assert "ttl" in result
         assert result["ttl"] == "7200s"
-        assert "system_instruction" in result
+        
+        system_instruction_key = "systemInstruction" if custom_llm_provider in ["vertex_ai", "vertex_ai_beta"] else "system_instruction"
+        assert system_instruction_key in result
         
         if custom_llm_provider == "gemini":
             assert result["model"] == "models/gemini-2.5-pro"

tests/test_litellm/llms/vertex_ai/test_vertex.py

@@ -1555,5 +1555,7 @@ def test_system_prompt_only_adds_blank_user_message():
     #########################################################
     # system message was passed in
     #########################################################
-    assert len(data["system_instruction"]) == 1
-    assert data["system_instruction"]["parts"][0]["text"] == SYSTEM_INSTRUCTION
+    assert (
+        len(data["systemInstruction"]["parts"]) == 1
+    )  # was renamed to camelCase for Vertex AI
+    assert data["systemInstruction"]["parts"][0]["text"] == SYSTEM_INSTRUCTION