[iamwavecut]

Summary

The opt-in MOSS-Image layout ("prompt_layout": true in the MOSS model config, still off) now follows the official prompt enhancer (skills/moss-image-prompt-enhancer). The prose route and the optimizer prompt without a layout render unchanged.

• Pixel draft. moss_layout holds canvas_width and canvas_height (the 1 MP canvas draw-api renders for each aspect ratio, listed in the prompt; moss_canvas() mirrors area_dimensions(ratio, 1024, 16)), summary, and items with [x_min, y_min, x_max, y_max] pixel boxes. No word budgets. Items are two schema shapes, so only a text item carries text: with text allowed on any item the model folded captions into panel objects, and draw-api dropped them.
• Request kind instead of a flag. moss_kind (poster, comic, meme, infographic, arrangement, card, cover, menu, other_design, scene); code keeps a layout only for the first five (MOSS_LAYOUT_KINDS). With the boolean flag and a narrowed rule the model still marked menus, cards and covers as designed.
• Layout before the prompts. moss_kind and moss_layout sort before outputs in the alphabetical grammar order, so the layout is drafted first, as the enhancer does. After outputs, the MOSS prompt (outputs[2]) of the VIP call ran to max_tokens (runs of \n; with a no-blank-line pattern, endless prose). slot_moss_negative stays after the prompts.
• Quotes. ASCII or curly quotes that wrap a whole layout text are removed; MOSS-Image drew them. Guillemets stay.

Evidence

• Production Gemma (as a client), VIP Krea 2 + FLUX.3 + MOSS with layout: fields after outputs 5 of 25 calls ran away; fields before outputs 0 of 48 VIP + 12 MOSS-only calls, median 5.7 s, max 14.1 s, no junk slot. The deployed prompt without a layout: 0 of 12. moss_kind was the same in every repeat for all 24 requests (posters → poster, memes → meme/comic, comics → comic, infographics → infographic, menu → other_design, card → card, cover → cover, photos → scene).
• MOSS A/B (Vast RTX 4090, production settings, 14 designed requests × prose / layout / layout+prose × 2 seeds, 84 images): letter defects prose 10 (9 of 26 images), layout 6, layout+prose 6. Layout keeps multi-line meme captions right ("в проде", "прошёл" where prose drew "в преде", "прошел"); its failures were boxes crowded into a strip (draw-api now rejects those, iamwavecut/aifarm-draw#38) and quotes around titles (removed here). prompt_layout stays off until the owner decides.

Tests

• cargo test --locked --workspace, cargo clippy --locked --workspace --all-targets --all-features -- -D warnings, cargo fmt --all -- --check: pass.
• New or changed: moss_layouts_use_the_pixel_canvas_draw_api_draws (canvas table, prompt text, schema), item shapes, moss_kind decoding (only layout kinds keep a layout), key order moss_kind < moss_layout < outputs < slot_moss_negative, quote stripping, and examples_write_their_keys_in_the_order_the_grammar_does for the moved example fields.

Server side: iamwavecut/aifarm-draw#38 (merge and deploy it first).

[github-actions]

	Warnings
⚠️	Prompt files are cache-sensitive contracts; confirm prompt shape and role markers were preserved intentionally.

	Messages
📖	Danger checked 3 changed file(s).

Generated by 🚫 dangerJS against e609805