Sanitize MCP tool output items in openai-agents-sdk template

[dhruv0811]

Summary

• MCP tools (e.g. Genie) can return output items where the output field is a list instead of a string, causing Pydantic validation errors in ResponsesAgentResponse
• Adds sanitize_output_items / _sanitize_item to utils.py and wires it into both invoke and stream handlers

Temporary workaround until mlflow/mlflow#20777 is released.

[Copilot]

Pull request overview

This PR adds sanitization of MCP tool output items to fix Pydantic validation errors that occur when the output field is a list instead of a string. The changes introduce helper functions to convert non-string output values to JSON strings before passing them to ResponsesAgentResponse.

Changes:

• Added _sanitize_item and sanitize_output_items helper functions to convert list-type output fields to JSON strings
• Applied sanitization in both the invoke handler (for non-streaming responses) and the stream event handler (for streaming responses)
• Reorganized imports in agent.py to be alphabetically sorted

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
agent-openai-agents-sdk/agent_server/utils.py	Added sanitization functions _sanitize_item and sanitize_output_items to handle MCP tool outputs with list-type output fields; integrated _sanitize_item into the stream event processing
agent-openai-agents-sdk/agent_server/agent.py	Updated imports to include sanitize_output_items and applied it to the invoke response output; reorganized imports alphabetically

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The function mutates the input dictionary directly, which can cause unintended side effects if the same dictionary is used elsewhere. Consider creating a copy of the dictionary before modifying it to avoid mutating the caller's data. For example, create a shallow copy at the start of the function with item = input_item.copy() and then modify and return the copy.

Suggested change

	if isinstance(input_item.get("output"), list):
	input_item["output"] = json.dumps(input_item["output"])
	return input_item
	item = input_item.copy()
	if isinstance(item.get("output"), list):
	item["output"] = json.dumps(item["output"])
	return item

[Copilot]

The sanitization only handles the case where output is a list. However, other non-string types (e.g., dict, int, bool, None) could also cause Pydantic validation errors. Consider checking if output is not a string rather than checking if it's specifically a list. For example: if "output" in input_item and not isinstance(input_item["output"], str): to handle all non-string types consistently.

Suggested change

	a *list* of content objects instead of a plain string. MLflow's Pydantic
	models expect ``output`` to be a string, so this serialises any non-string
	values to JSON.
	TODO: Remove once https://github.com/mlflow/mlflow/pull/20777 is released.
	"""
	if isinstance(input_item.get("output"), list):
	input_item["output"] = json.dumps(input_item["output"])
	a collection or other non-string value instead of a plain string. MLflow's
	Pydantic models expect ``output`` to be a string, so this serialises any
	non-string values to JSON.
	TODO: Remove once https://github.com/mlflow/mlflow/pull/20777 is released.
	"""
	output_value = input_item.get("output")
	if "output" in input_item and not isinstance(output_value, str):
	input_item["output"] = json.dumps(output_value)