fix(vertex_ai): send systemInstruction on the cachedContents create body

The cachedContents create call was the last place in the provider still
sending the snake_case alias. CachedContent, the read type for the same
resource, already declares systemInstruction, so the request and response
types disagreed on one field.

litellm/llms/vertex_ai/context_caching/transformation.py

@@ -158,7 +158,7 @@ def cached_messages_end_on_supported_turn(cached_messages: Sequence[AllMessageVa
     The cachedContents API rejects contents ending on a model turn, which is how it
     classifies both assistant messages and tool results, with HTTP 400
     "Requests ending with a model turn are not supported". System messages are
-    extracted into system_instruction before contents are built, so the terminal
+    extracted into systemInstruction before contents are built, so the terminal
     turn is the last non-system message.
     """
     non_system_messages: Final = tuple(message for message in cached_messages if message.get("role") != "system")
@@ -206,6 +206,6 @@ def transform_openai_messages_to_gemini_context_caching(
         data["ttl"] = ttl
 
     if transformed_system_messages is not None:
-        data["system_instruction"] = transformed_system_messages
+        data["systemInstruction"] = transformed_system_messages
 
     return data

litellm/types/llms/vertex_ai.py

@@ -350,7 +350,7 @@ class RequestBody(TypedDict, total=False):
 
 class CachedContentRequestBody(TypedDict, total=False):
     contents: Required[list[ContentType]]
-    system_instruction: SystemInstructions
+    systemInstruction: SystemInstructions
     tools: Tools
     toolConfig: ToolConfig
     model: Required[str]  # Format: models/{model}

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

@@ -334,7 +334,7 @@ class TestTransformationWithTTL:
 
         assert "ttl" in result
         assert result["ttl"] == "7200s"
-        assert "system_instruction" in result
+        assert "systemInstruction" in result
 
         if custom_llm_provider == "gemini":
             assert result["model"] == "models/gemini-2.5-pro"

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

@@ -1599,6 +1599,29 @@ def test_cached_messages_end_on_supported_turn():
     assert cached_messages_end_on_supported_turn([]) is False
 
 
+@pytest.mark.parametrize("custom_llm_provider", ["gemini", "vertex_ai", "vertex_ai_beta"])
+def test_cached_content_system_instruction_uses_canonical_rest_key(custom_llm_provider):
+    """cachedContents.create takes systemInstruction; this was the last snake_case sender in the provider."""
+    from litellm.llms.vertex_ai.context_caching.transformation import (
+        transform_openai_messages_to_gemini_context_caching,
+    )
+
+    result = transform_openai_messages_to_gemini_context_caching(
+        model="gemini-2.5-pro",
+        messages=[
+            {"role": "system", "content": "You are a concise assistant."},
+            {"role": "user", "content": "Reply with exactly: ok"},
+        ],
+        custom_llm_provider=custom_llm_provider,
+        cache_key="test-cache-key",
+        vertex_project="test_project",
+        vertex_location="test_location",
+    )
+
+    assert "system_instruction" not in result
+    assert result["systemInstruction"]["parts"][0]["text"] == "You are a concise assistant."
+
+
 class TestCheckCachePagination:
     """Test pagination logic in check_cache and async_check_cache methods."""