responses.create input types are not Responses API content parts #13

ResponseInputItem.content is typed as ResponseInputTextContent (type: "text") or ResponseInputImageContent (type: "image" with an Anthropic-style source). Neither is a Responses API content part. The spec uses:

• {"type": "input_text", "text": ...}
• {"type": "input_image", "image_url": ..., "detail": ...} or {"type": "input_image", "file_id": ...}
• {"type": "input_file", "file_id" | "file_data" | "file_url", "filename": ...}

A type-checked caller who follows the Responses API docs gets a type error. A caller who follows the SDK types sends parts that OpenAI rejects.

Fix: add ResponseInputText, ResponseInputImage and ResponseInputFile TypedDicts, and accept them in ResponseInputItem.content. Keep the old two in the union, so existing code still type-checks. Parts are sent unchanged either way.

Note: the gateway currently passes these parts through to OpenAI channels, but drops them when it converts a request for other providers. That is tracked separately, on the gateway side.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Uy5vXsRpMhug7v3XwTMZHv