Render the scene server-side (real WebGL, headless) and return screenshot image(s) — your EYES on the scene. Use it after apply_ops batches and after probe_framing, and judge each image as a photographer and a web designer would: one clear subject, on a third (for a page's own camera, the whole card at a readable angle); the horizon off the middle; the eye at a standing height; text legible and on even margins; a way out in view; the colours serving the brief (one accent, on the primary action); nothing floating, clipped or overlapping. Beside each image comes `composition`, read from its own pixels: tonalRange (p5 to p95 of lightness; under 0.35 is flat), clipping (share blown to white, crushed to black), key (low, mid, high), a 3x3 thirds grid of lightness and detail with the busiest and brightest cells, the hues on screen by share, and findings in sentences. Fix with apply_ops. Discipline: at most 2 screenshot→fix rounds per user request, then report what you built. views (max 4, default = the visitor's entry view): pass "page:<pageIdOrSlug>" to frame one page/neighborhood, or {position:[x,y,z], target:[x,y,z], fov?} for an exact camera. Rendering is real but slow — ~10s for simple scenes, and the server gives up after 45 s per view by default; a 400-node scene with many textures needs more, so pass timeoutSeconds (up to 180) and a smaller maxDim (a 512-wide shot is a fine check of placement and readability and renders faster). A timeout comes back as an error naming it, never as an empty image; repeat shots of an unchanged scene at the same size are cached and instant. Rate-limited: prefer ONE well-chosen view; only multi-view when areas are far apart (four views at 180 s can hold the renderer for twelve minutes). If it returns a busy/unavailable error, keep building and retry once later — do not loop on it.
ConnectorNo auth
