Bug

_sanitize_openai_conversation_item() in agents/run_internal/session_persistence.py:568 unconditionally strips id from all conversation items before calling session.add_items():

def _sanitize_openai_conversation_item(item):
    if isinstance(item, dict):
        clean_item = cast(dict[str, Any], strip_internal_input_item_metadata(item))
        clean_item.pop("id", None)       # strips id from ALL items
        clean_item.pop("provider_data", None)
        return cast(TResponseInputItem, clean_item)
    return item

The OpenAI Conversations API requires id on built-in tool-call item types. The SDK's own response param type definitions confirm this:

• ResponseFileSearchToolCallParam — id: Required[str] (openai/types/responses/response_file_search_tool_call_param.py:43)
• ResponseCodeInterpreterToolCallParam — id: Required[str] (openai/types/responses/response_code_interpreter_tool_call_param.py:37)
• ResponseFunctionWebSearchParam — id: Required[str] (openai/types/responses/response_function_web_search_param.py:79)

When the sanitizer strips id from these items, the Conversations API returns 400 Missing required parameter: 'items[0].id'.

Reproduction

1. Create an agent with FileSearchTool and a real vector store
2. Use OpenAIConversationsSession for conversation persistence
3. On turn 1, trigger the agent to use file_search (produces file_search_call items)
4. On turn 2, send any follow-up message
5. The SDK collects turn 1's output items, sanitizes them (stripping id), and calls add_items()
6. The Conversations API rejects the batch because file_search_call is missing id

Failing batch shape

[
  {"type": "file_search_call", ...},       // id: Required — stripped by sanitizer → 400 error
  {"type": "function_call", ...},          // id: Optional — works without id
  {"type": "function_call_output", ...}    // id: Optional — works without id
]

Note: function_call has id as optional in the SDK type definitions (response_function_tool_call_param.py:29). The confirmed required-id types are file_search_call, code_interpreter_call, and web_search_call.

Workaround

Override the sanitizer to only strip id from item types where it's optional (role == "user", type == "function_call_output"), preserving it for everything else.

Environment

• openai-agents==0.13.5
• openai==2.30.0
• Python 3.13.6
• Agent uses: FileSearchTool, CodeInterpreterTool, custom @function_tool tools
• Session: OpenAIConversationsSession (subclassed for batching)