From 2c6577959ae5a5910a2f07038014418f74c216fc Mon Sep 17 00:00:00 2001
From: Darsh Joshi <itsdarshjoshi@gmail.com>
Date: Fri, 2 Oct 2026 19:03:45 -0400
Subject: [PATCH] fix(deepseek): keep tool result images instead of collapsing
 them to text

The DeepSeek transform only kept image content lists on user messages, so a
role=tool message carrying an image_url block was collapsed to its text before
the shared OpenAI transform ran. That transform already moves tool message
images into a user message, but by then the image was gone

Keep tool message image lists so the image reaches the model

Fixes #44211
---
 litellm/llms/deepseek/chat/transformation.py       | 14 ++++++++------
 .../chat/test_deepseek_chat_transformation.py      | 10 ++++++++++
 2 files changed, 18 insertions(+), 6 deletions(-)

diff --git a/litellm/llms/deepseek/chat/transformation.py b/litellm/llms/deepseek/chat/transformation.py
index 4e428a23392..653b5d66bf3 100644
--- a/litellm/llms/deepseek/chat/transformation.py
+++ b/litellm/llms/deepseek/chat/transformation.py
@@ -121,10 +121,12 @@ 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
@@ -160,13 +162,13 @@ 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
diff --git a/tests/unit/llms/deepseek/chat/test_deepseek_chat_transformation.py b/tests/unit/llms/deepseek/chat/test_deepseek_chat_transformation.py
index 3f93264d0d0..9c584ffa605 100644
--- a/tests/unit/llms/deepseek/chat/test_deepseek_chat_transformation.py
+++ b/tests/unit/llms/deepseek/chat/test_deepseek_chat_transformation.py
@@ -169,6 +169,16 @@ 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)
+
+        assert result[0]["role"] == "tool"
+        assert result[0]["tool_call_id"] == "call_1"
+        assert any(isinstance(message["content"], list) and image_block in message["content"] for message in result)
+
     def test_audio_block_collapsed_even_on_vision_model(self):
         messages = [
             {