fix(responses): extract cache_control_injection_points from extra_body

When cache_control_injection_points is passed via extra_body in the
Responses API, the prompt management hook system cannot find it because
it looks for top-level kwargs, not params nested inside extra_body.

Extract recognised DynamicPromptManagementParam keys from extra_body
into kwargs before the hook dispatch runs, in both the sync and async
Responses API paths.

Fixes #25335

litellm/responses/main.py

@@ -80,6 +80,25 @@ def _has_file_search_tool(tools: Optional[Any]) -> bool:
     return any(isinstance(t, dict) and t.get("type") == "file_search" for t in tools)
 
 
+def _extract_dynamic_params_from_extra_body(
+    extra_body: Dict[str, Any], kwargs: Dict[str, Any]
+) -> None:
+    """Extract dynamic prompt management params from extra_body into kwargs.
+
+    The Responses API receives params like ``cache_control_injection_points``
+    inside ``extra_body``, but the prompt management hook system expects them
+    as top-level kwargs.  Move recognised keys so the hooks can find them.
+    """
+    if not isinstance(extra_body, dict):
+        return
+
+    from litellm.types.utils import DynamicPromptManagementParamLiteral
+
+    for param in DynamicPromptManagementParamLiteral.list_all_params():
+        if param in extra_body and param not in kwargs:
+            kwargs[param] = extra_body.pop(param)
+
+
 def mock_responses_api_response(
     mock_response: str = "In a peaceful grove beneath a silver moon, a unicorn named Lumina discovered a hidden pool that reflected the stars. As she dipped her horn into the water, the pool began to shimmer, revealing a pathway to a magical realm of endless night skies. Filled with wonder, Lumina whispered a wish for all who dream to find their own hidden magic, and as she glanced back, her hoofprints sparkled like stardust.",
 ):
@@ -484,6 +503,10 @@ async def aresponses(
         prompt_variables = cast(Optional[dict], kwargs.get("prompt_variables", None))
         original_model = model
 
+        # Extract dynamic prompt management params from extra_body
+        if extra_body:
+            _extract_dynamic_params_from_extra_body(extra_body, kwargs)
+
         if isinstance(
             litellm_logging_obj, LiteLLMLoggingObj
         ) and litellm_logging_obj.should_run_prompt_management_hooks(
@@ -783,6 +806,11 @@ def responses(
             local_vars=local_vars,
         )
 
+        # Extract dynamic prompt management params (e.g. cache_control_injection_points)
+        # from extra_body into kwargs so the hook system can find them.
+        if extra_body:
+            _extract_dynamic_params_from_extra_body(extra_body, kwargs)
+
         #########################################################
         # PROMPT MANAGEMENT
         # If aresponses() already ran the async hook, it pops prompt_id and

tests/test_litellm/integrations/test_anthropic_cache_control_hook.py

@@ -1073,3 +1073,45 @@ async def test_anthropic_cache_control_hook_string_negative_index():
                 f"Expected cachePoint in last message content, got: {last_message_content}. "
                 "String index '-1' was not parsed correctly (str.isdigit() returns False for negative strings)."
             )
+
+
+def test_extract_dynamic_params_from_extra_body():
+    """
+    Verify that cache_control_injection_points inside extra_body gets
+    extracted into kwargs so the prompt-management hook system can find it.
+
+    Regression test for https://github.com/BerriAI/litellm/issues/25335
+    """
+    from litellm.responses.main import _extract_dynamic_params_from_extra_body
+
+    injection_points = [
+        {"location": "message", "role": "system", "index": 0},
+        {"location": "message", "role": "user", "index": -1},
+    ]
+
+    extra_body: dict = {"cache_control_injection_points": injection_points, "other_key": "keep"}
+    kwargs: dict = {}
+
+    _extract_dynamic_params_from_extra_body(extra_body, kwargs)
+
+    # cache_control_injection_points should be moved to kwargs
+    assert kwargs["cache_control_injection_points"] == injection_points
+    # should be removed from extra_body
+    assert "cache_control_injection_points" not in extra_body
+    # other keys should remain
+    assert extra_body["other_key"] == "keep"
+
+
+def test_extract_dynamic_params_does_not_overwrite_existing_kwargs():
+    """If kwargs already has the key, extra_body should not overwrite it."""
+    from litellm.responses.main import _extract_dynamic_params_from_extra_body
+
+    extra_body: dict = {"cache_control_injection_points": [{"from": "extra_body"}]}
+    kwargs: dict = {"cache_control_injection_points": [{"from": "kwargs"}]}
+
+    _extract_dynamic_params_from_extra_body(extra_body, kwargs)
+
+    # kwargs value should remain unchanged
+    assert kwargs["cache_control_injection_points"] == [{"from": "kwargs"}]
+    # extra_body should still have it since it wasn't consumed
+    assert "cache_control_injection_points" in extra_body