Fix content input conversion for OpenAI Responses API

[Copilot]

Summary

Fixed a bug where the OpenAI Responses API fails when processing Image and text inputs. The issue was that content items were not being converted from the Chat API format to the Responses API format.

Changes Made

• Added _convert_content_item_to_responses_format() helper function to convert both images and text from Chat API to Responses API format
• Updated _convert_chat_request_to_responses_request() to use the new helper function
• Added comprehensive tests for both base64 and URL images
• Verified async version works correctly
• All existing tests pass - no regressions
• Linting checks pass
• CodeQL security scan - no vulnerabilities found

Technical Details

The Chat API uses these formats:

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

The Responses API expects these formats per OpenAI documentation:

{
  "type": "input_text",
  "text": "..."
}
{
  "type": "input_image",
  "image_url": "data:image/png;base64,..."
}

The fix converts both text and image content items to the correct Responses API format:

• Text items: type: "text" → type: "input_text"
• Image items: type: "image_url" → type: "input_image" with image_url as a direct property

Testing

• Added test_responses_api_converts_images_correctly() to test conversion logic for both text and images
• Added test_responses_api_with_image_input() to test complete flow with LM class
• All existing Responses API tests pass
• Manually tested sync and async code paths
• Verified format matches OpenAI documentation

Security Summary

CodeQL security scan completed with no vulnerabilities found.

Fixes #8985

Original prompt

---

This section details on the original issue you should resolve

<issue_title>[Bug] OpenAI Responses model_type fails with Image inputs</issue_title>
<issue_description>### What happened?

When using the Responses API for OpenAI models, Image inputs aren't handled correctly.

Given the following code:

import python

class ImageDescriptionGeneratorSignature(dspy.Signature):
    image: dspy.Image = dspy.InputField(desc="")
    image_description: str = dspy.OutputField(desc="")

descriptor = dspy.Predict(ImageDescriptionGeneratorSignature)

lm = dspy.LM(
    model="openai/gpt-5-mini",
    api_key="MY-KEY",
    max_tokens=16000,
    temperature=1.0,
    model_type="responses"
)

dspy.configure(lm=lm)
img = dspy.Image("overload.jpg")

result = descriptor(image=img)
print(result.image_description)

This throws a LiteLLM error:

BadRequestError: litellm.BadRequestError: OpenAIException - {
  "error": {
    "message": "Invalid value: 'text'. Supported values are: 'input_text', 'input_image', 'output_text', 'refusal', 'input_file', 'computer_screenshot', and 'summary_text'.",
    "type": "invalid_request_error",
    "param": "input[0].content[1].type",
    "code": "invalid_value"
  }
}

To fix the issue, we need to make sure we're formatting images correctly for the Responses API.

Steps to reproduce

import python

class ImageDescriptionGeneratorSignature(dspy.Signature):
image: dspy.Image = dspy.InputField(desc="")
image_description: str = dspy.OutputField(desc="")

descriptor = dspy.Predict(ImageDescriptionGeneratorSignature)

lm = dspy.LM(
model="openai/gpt-5-mini",
api_key="MY-KEY",
max_tokens=16000,
temperature=1.0,
model_type="responses"
)

dspy.configure(lm=lm)
img = dspy.Image("overload.jpg")

result = descriptor(image=img)
print(result.image_description)

DSPy version

3.0.4b1</issue_description>

<agent_instructions>Can you support the request conversion for image inputs? Refer to https://platform.openai.com/docs/api-reference/responses</agent_instructions>

Comments on the Issue (you are @copilot in this section)

• Fixes [Bug] OpenAI Responses model_type fails with Image inputs #8985

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[TomeHirata]

@copilot can we remove docstring for the test cases?

[Copilot]

Done in 43c8fc9

[TomeHirata]

@copilot Does this match the documented interface?
https://platform.openai.com/docs/api-reference/responses/create?lang=curl

[Copilot]

You're absolutely right - I was using the wrong format. Fixed in 6da81b3 to match the OpenAI documentation. The Responses API expects {"type": "input_image", "image_url": "..."} with image_url as a direct property, not a nested source object.

[TomeHirata]

@copilot I this we also need to convert text item from type: text to type: input_text. See https://platform.openai.com/docs/api-reference/responses

[Copilot]

Good catch! Fixed in 0055def - now converting text items from type: "text" to type: "input_text" to match the OpenAI Responses API format.