From 59f29ca6dfb51f5e45fdd4b8a764e7f048caea2d Mon Sep 17 00:00:00 2001
From: Simon Sorg <simonsorg13@gmail.com>
Date: Mon, 21 Sep 2026 21:51:39 +0200
Subject: [PATCH] fix(responses): preserve multimodal cache breakpoints

---
 .../transformation.py                         | 18 +++++---
 ...responses_transformation_transformation.py | 46 +++++++++++--------
 2 files changed, 39 insertions(+), 25 deletions(-)

diff --git a/litellm/completion_extras/litellm_responses_transformation/transformation.py b/litellm/completion_extras/litellm_responses_transformation/transformation.py
index e2fe45102d4..8020cf41e68 100644
--- a/litellm/completion_extras/litellm_responses_transformation/transformation.py
+++ b/litellm/completion_extras/litellm_responses_transformation/transformation.py
@@ -1069,11 +1069,14 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
                         verbose_logger.debug("Chat provider:   text -> %s", converted)
                     elif original_type == "image_url":
                         # Map to responses API image format
-                        converted = cast(
-                            dict,
-                            self._convert_content_to_responses_format_image(
-                                cast(ChatCompletionImageObject, item), role
+                        converted = with_prompt_cache_breakpoint(
+                            cast(
+                                dict,
+                                self._convert_content_to_responses_format_image(
+                                    cast(ChatCompletionImageObject, item), role
+                                ),
                             ),
+                            item.get("prompt_cache_breakpoint"),
                         )
                         result.append(converted)
                         verbose_logger.debug("Chat provider:   image_url -> %s", converted)
@@ -1085,8 +1088,11 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
                             result.append(converted)
                             verbose_logger.debug("Chat provider:   image -> %s", converted)
                         elif item_type == "file":
-                            converted = _input_file_from_file_value(
-                                cast("ChatCompletionFileObject", item).get("file"),  # cast-ok: type tag checked
+                            converted = with_prompt_cache_breakpoint(
+                                _input_file_from_file_value(
+                                    cast("ChatCompletionFileObject", item).get("file"),  # cast-ok: type tag checked
+                                ),
+                                item.get("prompt_cache_breakpoint"),
                             )
                             result.append(converted)
                             verbose_logger.debug("Chat provider:   file -> %s", converted)
diff --git a/tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py b/tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
index 78aac112b2a..90144e47e7f 100644
--- a/tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
+++ b/tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py
@@ -4185,24 +4185,38 @@ def _system_input_item(text: str) -> dict[str, object]:
     return {"type": "message", "role": "system", "content": [{"type": "input_text", "text": text}]}
 
 
-def test_prompt_cache_breakpoint_survives_chat_to_responses_conversion() -> None:
+@pytest.mark.parametrize(
+    ("content_block", "expected_content"),
+    [
+        (
+            {"type": "text", "text": "Stable prefix"},
+            {"type": "input_text", "text": "Stable prefix"},
+        ),
+        (
+            {"type": "image_url", "image_url": "https://example.com/image.png"},
+            {"type": "input_image", "image_url": "https://example.com/image.png", "detail": "auto"},
+        ),
+        (
+            {"type": "file", "file": {"file_id": "file-123"}},
+            {"type": "input_file", "file_id": "file-123"},
+        ),
+    ],
+    ids=("text", "image_url", "file"),
+)
+def test_prompt_cache_breakpoint_survives_chat_to_responses_conversion(
+    content_block: dict[str, object], expected_content: dict[str, object]
+) -> None:
     handler: Final = LiteLLMResponsesTransformationHandler()
     cache_breakpoint: Final = {"mode": "explicit"}
+    marked_content: Final = {**content_block, "prompt_cache_breakpoint": cache_breakpoint}
 
     request: Final = handler.transform_request(
         model="gpt-5.6-sol",
         messages=[
             {
-                "role": "system",
-                "content": [
-                    {
-                        "type": "text",
-                        "text": "Stable prefix",
-                        "prompt_cache_breakpoint": cache_breakpoint,
-                    }
-                ],
-            },
-            {"role": "user", "content": "Use a tool"},
+                "role": "user",
+                "content": [marked_content],
+            }
         ],
         optional_params={"prompt_cache_options": cache_breakpoint},
         litellm_params={},
@@ -4212,14 +4226,8 @@ def test_prompt_cache_breakpoint_survives_chat_to_responses_conversion() -> None
 
     assert request["input"][0] == {
         "type": "message",
-        "role": "system",
-        "content": [
-            {
-                "type": "input_text",
-                "text": "Stable prefix",
-                "prompt_cache_breakpoint": cache_breakpoint,
-            }
-        ],
+        "role": "user",
+        "content": [{**expected_content, "prompt_cache_breakpoint": cache_breakpoint}],
     }
     assert request["prompt_cache_options"] == cache_breakpoint