Use one workspace for prompt-only generation and reference-based edits. The page decides between async generations and async edits automatically.
Reference uploads reuse the shared asset pipeline and are rehydrated into Plato multipart binary payloads on the server.
Tasks are stored in the shared generation history, but this route filters them to GPT-Image-2.
Outputs currently render from provider-hosted URLs so the first model page can ship without a new media mirror layer.
This list is filtered to GPT-Image-2 model-page image tasks only.