Merge d620e29f54 into 083961731c

litellm/llms/deepseek/chat/transformation.py

@@ -148,10 +148,12 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
         DeepSeek vision models accept image_url content blocks in user
         messages (https://api-docs.deepseek.com/guides/vision), so those
         content lists are forwarded as-is, with any search_results text
-        appended as a trailing text block. Every other message keeps the
-        historical string collapse (which also folds search_results text
-        into string content); a list with no extractable text stays
-        unchanged, matching what DeepSeek historically received.
+        appended as a trailing text block. Tool message image lists are kept
+        too, so the parent transform can move their images into a user
+        message. Every other message keeps the historical string collapse
+        (which also folds search_results text into string content); a list
+        with no extractable text stays unchanged, matching what DeepSeek
+        historically received.
         """
         forward_images: Final = any(
             isinstance(message.get("content"), list) for message in messages
@@ -187,13 +189,13 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
 
     def _is_vision_forwardable_content(self, message: AllMessageValues, content: Sequence[object]) -> bool:
         """
-        True only for a user message whose content list holds well-formed
+        True only for a user or tool message whose content list holds well-formed
         text and image_url blocks with at least one image; a block missing
         its payload falls back to the string collapse instead of crashing
         or reaching the wire malformed. The model capability gate lives in
         the caller.
         """
-        if message.get("role") != "user":
+        if message.get("role") not in ("user", "tool"):
             return False
         if not all(self._is_forwardable_block(block) for block in content):
             return False

tests/unit/llms/deepseek/chat/test_deepseek_chat_transformation.py

@@ -295,6 +295,33 @@ class TestDeepSeekVisionMultimodalContent:
 
             assert result[0]["content"] == "what is in this image?"
 
+    def test_tool_result_image_reaches_vision_model(self):
+        tool_message = {**self._image_message(role="tool"), "tool_call_id": "call_1"}
+        image_block = tool_message["content"][1]
+
+        result = self.config._transform_messages([tool_message], model=self.VISION_MODEL)
+
+        tool_result, image_carrier = result
+        assert tool_result["role"] == "tool"
+        assert tool_result["tool_call_id"] == "call_1"
+        assert image_block not in tool_result["content"]
+        assert image_carrier["role"] == "user"
+        assert image_block in image_carrier["content"]
+
+    def test_tool_result_image_carrier_placed_directly_after_tool_result(self):
+        for next_message in (
+            {"role": "user", "content": "so what do you see?"},
+            {"role": "assistant", "content": "I see an image."},
+        ):
+            tool_message = {**self._image_message(role="tool"), "tool_call_id": "call_1"}
+            image_block = tool_message["content"][1]
+
+            result = self.config._transform_messages([tool_message, next_message], model=self.VISION_MODEL)
+
+            assert [m["role"] for m in result] == ["tool", "user", next_message["role"]]
+            assert image_block in result[1]["content"]
+            assert result[2] == next_message
+
     def test_audio_block_collapsed_even_on_vision_model(self):
         messages = [
             {