Scenario Maya Groom

Hair and fur grooming (XGen Interactive Groom, Maya 2027)

Expert grooming sets the flow on a few guides first, then a clump hierarchy with loose roots and varied profiles, then breakup only where the reference shows it (tips, strays), judged in close, side-lit renders with the real hair shader. Interactive Groom brushes have no stroke API, so the agent grooms with data (guides from a flow sheet wired into the stack, masks written as images, modifiers set by value) and judges with numbers from exported strands plus renders. Code: scripts/mx_groom.py (import mx_groom as G). If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-maya-expert (execution channel, review loop, 2027 version traps).

Status (2026-09-24): every Maya call here is not yet run in Maya (2027 not installed). The pure layer of mx_groom passes offline (tests/code/maya-groom/test_mx_groom_offline.py). Names not quoted from the 2027 help resolve at run time from G.NAME_CANDIDATES.

Stance (the expert delta)

• Flow first, on the guides, copied from arrows drawn over the reference (Fernandez). Flow is pressure: neighbors share the compression curve, and opposing streams meet through a straighter band. Short fur turns sharply, long hair slowly. Decide whether the flow or a clump owns each big shape.
• Roots loose, every clump different, two levels (Fernandez). No attraction over about the first 14 % of the length, or the base pinches into bald lines. 3 to 5 small clumps per big one; a third level only for long hair or manes. Vary size and profile (G.CLUMP_PROFILES: spike, onion).
• Mask and tightness are different knobs (Fernandez 05, 06). The mask decides triangle versus square: square clumps are tight small clumps inside a looser big clump, both masks lowered. Tightness (IG Clump strength [added]) 0 keeps the clump's shape without converging; mask 0 removes it. Straight guides make them look alike, so test on noisy guides (G.clump_follow).
• The level that gets the noise decides what breaks (Fernandez 07). Clump-level noise (the secondary Clump's Noise, tip half) breaks the big clumps and keeps the small ones and the roots; a Noise modifier breaks the small clumps too, so it goes on few hairs (Giovannini: 0 on the body, 3 on a 5 % share). Vary the frequency; fray only on few hairs. Stray clumps come from holes in the primary's mask, not a new system.
• The hairline reads real through thinner edge strands plus transition and peach-fuzz descriptions, not density alone (Hadi Karimi). One density map is shared by every description; each hair type gets its own description.
• A short fill layer hides the scalp (Giovannini): body guides cut to 2/3, so long hair need not be "crazy dense".
• Motion drives a few wires, never the hair (Flood). Export the wires from the Linear Wire's inGuide. Update the Reference State at the rest pose on frame 1. Curve to Spline gets Align to Normals off.
• Standard Hair realism comes from melanin, roughness and IOR (Arnold hair doc). Keep diffuse 0, tints white and indirect scales at 1, and use ribbon mode (Min Pixel Width only works there). Blond hair gets Extra Depth. Cost has four knobs (MPW, transparency depth, specular samples, AA); opacity below 1 costs significantly.

Establish first

Write the answers to brief.json before any node.

Workflow

Headless mayapy (mx_run --plugins mtoa,xgen,abc, in that order) covers guides, masks, Alembic, Unreal tags, lint, analysis and Arnold reviews; the GUI bridge covers the menu queries and whatever the probe shows does not evaluate headless (IG computes on the GPU). references/procedures.md P16 is the whole job as one script with every gate asserted (G.gate raises).

0. Probe and recipes (once per install). G.probe() headless and in the GUI. Menu actions with no documented command are read, not clicked: through the bridge, G.menu_recipe(name) queries the item's command string (cmds.menuItem(item, q=True, command=True)) and stores it; Attribute Editor buttons (Make Wires Dynamic, Reference State Update, Rebuild) go through G.button_recipe(name, node). Export Cache opens a dialog: write its @path@ template from G.recipe_source. GATE: probe JSON saved; recipe_status() shows the recipes the plan needs as recorded.
1. Scalp. G.scalp_check(scalp, for_unreal=...). GATE: no error: line, border edges present (the hairline), UVs unique.
2. Flow sheet and counts. From the reference write the flow sheet (base, whorl, part, direction, length regions) and count big clumps and small per big. GATE: G.check_flow_sheet(flow) == [].
3. Guides. g = G.make_guides(scalp, flow, spacing=..., cvs=8, collide=head), few first (FlippedNormals); symmetric=True only on symmetric topology (Hadi). GATE: G.gate(G.measure_curves(...)["verdict"]), penetration 0, discontinuities only at parts and whorls.
4. Descriptions and wiring. G.create_description(scalp, "hair_body") per hair type (recipe, renamed). G.wire_guides(desc, g["curves"]): a Guide modifier low in the stack interpolates the guides, fed by a Curve to Spline on its inGuide; it holds the region maps parts need. A Curve to Spline on the description itself turns each curve into one hair and disables everything below it: only for exact strands or caches. GATE: G.wiring_verdict: flow-sheet error of the hair drops after wiring, and the InGuide_base cache holds every guide.
5. Stack, density and masks. Bottom to top: base (CV count per Numbers), Scale, Guide, primary and secondary Clump (G.build_clumps), strand Noise on a share, length breakup, Collision. G.write_hairline_mask for the hairline (band=True: transition), one map shared across descriptions (G.connect_texture), G.patches holes in the primary's mask for stray clumps. GATE: G.gate(G.lint()["lines"]); the coverage pass (showthrough %) and spline count within budget.
6. Clumps and breakup, one layer at a time (Giovannini). Once per install, P5b's two tests settle the Clump Scale convention, which knob acts as tightness and what each noise level breaks. Export mid-stack caches, then G.hierarchy_from_states (3 to 5), G.breakup_report and G.measure_abc. Render Clump before and after Noise once each (2027 help). GATE: breakup_verdict and strand_verdict clean; a side-lit close-up passes references/critique.md, triangles and squares in the reference's proportion.
7. Shader and render. G.hair_shader(desc, preset=...) (replaces the default hairPhysicalShader), G.curve_render(desc, "ribbon", min_pixel_width=...), G.cost_sweep over the four knobs with flicker. Save before rendering XGen. GATE: G.review(...) sheet judged with the critique, no shader warn:.
8. Motion. Static follow, keyed wires (Flood) or dynamics (Linear Wire, Make Wires Dynamic). Caches are versioned. GATE: renders at the rest, mid and last frames with no penetration.
9. Deliver (Handoffs).