[Bug]: Chat-to-Responses API Bridge Doesn't Transform Tool Output Content Types

[alexanderkrauck]

What happened?

Bug: Chat-to-Responses API Bridge Doesn't Transform Tool Output Content Types

Summary

When using models that route through the Responses API (e.g., gpt-5.1-codex, o1-pro, o3-pro, o4-*), multimodal tool results fail with a 400 Bad Request because tool output content is passed without type transformation.

Error Message

{
  "message": {
    "content": "Error occurred: LiteLLM API error: BadRequestError: litellm.BadRequestError: {\"error\":{\"message\":\"litellm.BadRequestError: OpenAIException - {\\n  \\\"error\\\": {\\n    \\\"message\\\": \\\"Invalid value: 'text'. Supported values are: 'input_text', 'input_image', 'output_text', 'refusal', 'input_file', 'computer_screenshot', and 'summary_text'.\\\",\\n    \\\"type\\\": \\\"invalid_request_error\\\",\\n    \\\"param\\\": \\\"input[25].output[0].type\\\",\\n    \\\"code\\\": \\\"invalid_value\\\"\\n  }\\n}. Received Model Group=gpt-5.1-codex\\nAvailable Model Group Fallbacks=None\",\"type\":null,\"param\":null,\"code\":\"400\"}}",
    "role": "assistant"
  }
}

Parsed error:

• Param: input[25].output[0].type — 26th message (tool output), first content item
• Invalid value: 'text' (Chat Completions format)
• Expected: 'output_text', 'input_image', etc. (Responses API format)

Root Cause & Bug Location

File: litellm/completion_extras/litellm_responses_transformation/transformation.py
Method: convert_chat_completion_messages_to_responses_api
Lines: 162-170

elif role == "tool":
    # Convert tool message to function call output format
    input_items.append(
        {
            "type": "function_call_output",
            "call_id": tool_call_id,
            "output": content,  # ❌ BUG: content passed without transformation
        }
    )

The content is passed directly without calling the existing _convert_content_to_responses_format() method.

Contrast with user/assistant messages (lines 186-196) which correctly transform content:

elif content is not None:
    input_items.append(
        {
            "type": "message",
            "role": role,
            "content": self._convert_content_to_responses_format(  # ✅ Correct
                content, cast(str, role)
            ),
        }
    )

Content Type Mapping Required

Chat Completions API	Responses API
{"type": "text", "text": "..."}	{"type": "output_text", "text": "..."}
{"type": "image_url", "image_url": {"url": "..."}}	{"type": "input_image", "image_url": "..."}

Note: The image_url value must also be flattened from {"url": "..."} object to just the URL string.

Reproduction

import litellm
messages = [

{"role": "user", "content": "Take a screenshot"},

{

"role": "assistant",

"tool_calls": [{

"id": "call_abc123",

"type": "function",

"function": {"name": "screenshot", "arguments": "{}"}

}]

},

{

"role": "tool",

"tool_call_id": "call_abc123",

"content": [

{"type": "text", "text": "Screenshot captured"},

{"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo..."}}

]

}

]
This fails with 400 Bad Request
response = await litellm.acompletion(

model="openai/gpt-5.1-codex",  # Or any Responses API model

messages=messages

)

Suggested Fix

Transform tool output content when it's a list:

elif role == "tool":
    # Convert tool message to function call output format
    output = content
    if isinstance(content, list):
        output = self._convert_tool_output_to_responses_format(content)
    input_items.append(
        {
            "type": "function_call_output",
            "call_id": tool_call_id,
            "output": output,
        }
    )

With a new helper method:

def _convert_tool_output_to_responses_format(
    self, content: List[Dict[str, Any]]
) -> List[Dict[str, Any]]:
    """Convert tool output content types from Chat Completions to Responses API format."""
    result = []
    for item in content:
        if not isinstance(item, dict):
            result.append(item)
            continue
    item_type = item.get("type")
    if item_type == "text":
        result.append({"type": "output_text", "text": item.get("text", "")})
    elif item_type == "image_url":
        image_url = item.get("image_url")
        url = image_url.get("url") if isinstance(image_url, dict) else image_url
        result.append({"type": "input_image", "image_url": url})
    else:
        result.append(item)
return result

Alternatively, the existing _convert_content_to_responses_format method could be extended to handle tool outputs, though note that tool text outputs should use output_text (not input_text).

Affected Models

Models using OpenAI's Responses API internally:

• gpt-5.1-codex, gpt-5.1-codex-max, codex-mini-latest
• o1-pro, o3-pro
• o4-* series

Related Issues

• [Bug]: [bridge for /chat/completion -> /responses API] Images Sent via URL Not Working in /chat/completions for OpenAI o1-pro, o3-pro, codex-mini-latest #11820 - Similar issue for image URLs in user messages (fixed in Completion-To-Responses Bridge: Support passing image url's #11833)
• The fix in Completion-To-Responses Bridge: Support passing image url's #11833 updated user message handling but tool messages were not addressed

Environment

• LiteLLM Version: Latest (Dec 2025)
• Python: 3.11+
• API Endpoint: /chat/completions → /responses bridge

Relevant log output

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

1.80.0 proxy server

Twitter / LinkedIn details

No response