fix: #2163 Preserve non-text tool outputs in LiteLLM and chatcmpl converters (#2214)

Co-authored-by: Kazuhiro Sera <[email protected]>

Author: ihower

src/agents/extensions/models/litellm_model.py

@@ -280,7 +280,9 @@ async def _fetch_response(
         )
 
         converted_messages = Converter.items_to_messages(
-            input, preserve_thinking_blocks=preserve_thinking_blocks
+            input,
+            preserve_thinking_blocks=preserve_thinking_blocks,
+            preserve_tool_output_all_content=True,
         )
 
         # Fix for interleaved thinking bug: reorder messages to ensure tool_use comes before tool_result  # noqa: E501

src/agents/models/chatcmpl_converter.py

@@ -340,6 +340,7 @@ def items_to_messages(
         cls,
         items: str | Iterable[TResponseInputItem],
         preserve_thinking_blocks: bool = False,
+        preserve_tool_output_all_content: bool = False,
     ) -> list[ChatCompletionMessageParam]:
         """
         Convert a sequence of 'Item' objects into a list of ChatCompletionMessageParam.
@@ -350,6 +351,12 @@ def items_to_messages(
                 for reasoning models like Claude 4 Sonnet/Opus which support interleaved
                 thinking. When True, thinking blocks are reconstructed and included in
                 assistant messages with tool calls.
+            preserve_tool_output_all_content: Whether to preserve non-text content (like images)
+                in tool outputs. When False (default), only text content is extracted.
+                OpenAI Chat Completions API doesn't support non-text content in tool results.
+                When True, all content types including images are preserved. This is useful
+                for model providers (e.g. Anthropic via LiteLLM) that support processing
+                non-text content in tool results.
 
         Rules:
         - EasyInputMessage or InputMessage (role=user) => ChatCompletionUserMessageParam
@@ -541,10 +548,14 @@ def ensure_assistant_message() -> ChatCompletionAssistantMessageParam:
                 output_content = cast(
                     Union[str, Iterable[ResponseInputContentWithAudioParam]], func_output["output"]
                 )
+                if preserve_tool_output_all_content:
+                    tool_result_content = cls.extract_all_content(output_content)
+                else:
+                    tool_result_content = cls.extract_text_content(output_content)  # type: ignore[assignment]
                 msg: ChatCompletionToolMessageParam = {
                     "role": "tool",
                     "tool_call_id": func_output["call_id"],
-                    "content": cls.extract_text_content(output_content),
+                    "content": tool_result_content,  # type: ignore[typeddict-item]
                 }
                 result.append(msg)