Layerhand - AI retouching that returns a layered PSD, not a flat JPEG

Most AI photo tools hand you pixels you cannot change, so I stopped generating images and taught Astra to use a real image editor instead. What inspired the build was a constraint that looked like a fence. GPT-6 Astra is text-output only: it cannot generate an image. Route around that, and the limitation becomes the product. An image model returns flattened pixels. An agent operating a real editor returns an edit history: named layers, masks, and adjustments a retoucher can still change. No diffusion model can emit that artifact, because a layer stack is not an image. The problem is not quality. Commercial retouching is done by hand, inside a dense professional GUI, one photograph at a time. Almost every AI tool aimed at that work returns a flattened JPEG. That is fine until the client asks for one more change, and they always do. A flat file cannot absorb the request, so the retoucher starts again. The AI saved nothing, because it did not produce the document the workflow actually runs on. The approach evolved through ten rounds of ideation. Desktop computer use died on scope. Hosted-browser computer use survived, then mid-turn steering survived as the only other primitive a user can feel. Crossing them pointed at Photopea: web professional software with no public API. Driving its toolbar through forty-plus steps is the long-horizon GUI task the OSWorld jump measures. Swapping Astra for another model does not degrade Layerhand. It removes the product, because the agent cannot finish the run, and there is no layered file. So the work narrowed to three properties: the output is editable, the work is visible, and the work is steerable. You watch Photopea being operated. You type a correction mid-run and later steps bend instead of restarting. You download a PSD that still opens in Photoshop, Affinity Photo, GIMP, or back in Photopea. Three free runs, no account. A run takes a few minutes, and the step count is capped.