Run an Architect turn (Chat Completions)

OpenAI Chat Completions over one turn of the Architect on a project. The turn builds and edits the project’s scene. The reply is content; the readable reasoning (the model’s thinking and one line per step) is reasoning_content; a turn that changed the scene is saved as a named version, cited as a url_citation in annotations and named in metadata.pascal_version*. The last user message is the request, with its image_url and file parts; earlier messages the project’s conversation already holds are not repeated, and system and developer messages become instructions for this turn. With stream: true the response is a text/event-stream of chat.completion.chunk events (delta.reasoning_content, delta.content, the annotation on the last chunk, a usage chunk with stream_options.include_usage) ending with data: [DONE], with : keepalive comments during silences. A turn can run for minutes; closing the connection stops it. Without stream, the headers leave within 15 s and spaces keep the connection open until the JSON body, so a turn that fails after that answers 200 with the error as the body. A question the Architect asks is the reply, and the next request answers it: 1, 2… picks an option. metadata.commit_message names the version the turn saves; metadata.pascal_max_credits caps the credits the turn may spend, and a turn that reaches it ends with finish_reason: "length". Errors carry x-should-retry: false: a turn that started may already have edited the scene. A turn needs at least one credit in the project’s credit pool.