---

name: scenario-maya-rigging description: 'Use when rigging a character, creature or prop in Maya: "rig this character", placing or orienting joints, Orient Joint doing nothing, IK/FK switch and matching, pole vector flips or elbow drift, space switching, foot roll, spline IK spine, offsetParentMatrix, blendMatrix or matrix rigging, controls and lockdown for animators, Quick Rig, HumanIK, AdvancedSkeleton or mGear, a game skeleton for Unreal or Unity, or a slow rig (parallel evaluation, profiling, GPU Override not kicking in).' license: MIT

Maya rigging

Expert rigging means every value is explained and every behavior measured: joints at real pivots with computed orientation, controls at zero with their placement in offsetParentMatrix, IK/FK and space switches that move nothing, and a rig that survives scale, distance, several characters per shot and an animator keying everything. The rig is the output of a build script and data. 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).

Module scripts/mx_rig.py (sys.path.insert(0, "<skills>/scenario-maya-rigging/scripts"); import mx_rig as R): not yet run in Maya; its pure math passed offline (test_rig_offline.py). With Maya installed, run tests/code/maya-rigging/run_all.sh and fix what test_rig_probe contradicts.

Stance (the expert delta)

• Clean values, one place for each. Rotations zero at rest, orientation in jointOrient (or nowhere: Fragapane), control placement in offsetParentMatrix (antCGi; Maya Learning Channel). OPM composes, world = local * OPM * dagParentWorld (2027 transform node): feed .matrix for a relative drive, or multiply a worldMatrix by the DAG parent's worldInverseMatrix, then zero translate, rotate and jointOrient or they apply twice (Maya Learning Channel; antCGi's leg rotating backwards).
• Compute, never eyeball. -orientJoint silently skips joints with rotations or no child (2027 cmds.joint); a secondary world axis along the bone is ill-defined (antCGi's toes). R.joint_chain computes jointOrient from positions; a pole on the chain plane adds zero drift (R.pole_vector_position) [added].
• The animator defines done. Zero pop switching IK/FK at neutral, only the active mode visible, everything else locked because animators "select hierarchy and key all", limits at breaking points, behavior mirroring on limbs, same orientation on eyes and face (antCGi).
• Matrices before constraint stacks, and know the blend. OPM, blendMatrix and pickMatrix keep channels free (Maya Learning Channel cut 20 joints from one hand). blendMatrix is an ordered stack, each target overriding the ones before (equal thirds: 1, 1/2, 1/3); parentMatrix (2025) normalizes and stores per-target offsets (2027 Help). Spaces blend offset targets, never the space itself, or the IK flips (antCGi).
• The built rig is disposable. Guides, build script, serialized weights and settings; fixes go back into the data (Miquel Campos, mGear: "we serialize everything and we build clean every single time"; Bungie). Metadata and joint labels, not names, identify chains (Bungie). A skinCluster reads only influence world matrices and bindPreMatrix: reset the bind pose, never rebuild joints for it (Fragapane).
• Structure decides speed. The Evaluation Graph leaves static curves out and rebuilds at the first differing key; without curves a rig "is going to be paralyzed terribly" (Fragapane; Using Parallel Maya 2027). Split components so cheap global work finishes early, even by duplicating cheap math (Fragapane); no Python nodes, scene-querying expressions or driven frozen; hosts out of the limbs (Miquel Campos); DG, Serial and Parallel agree before optimizing.
• Test at the extremes. Scale 1.24 and 1.6 (antCGi), a million units away and upside down (Fragapane), three referenced characters, a quick bind while building (antCGi).
• Deformation skeleton apart from the control rig. Engines see joints, skin and blendshapes only: one root at the origin, engine names, a joint budget (antCGi, Unreal and Unity docs, Bungie).

Establish first

Input	Changes	Default when silent ([added] unless attributed)
Target	film keyframe rig vs engine skeleton (names, axes, influences, one root)	feature keyframe rig, custom
Camera distance, style	auto-rigger or custom; twist and corrective joints; stylized: crude proportion rig and art-direction paintovers first, final-aim notes from day one (Miquel Campos)	hero: custom; crowd: Quick Rig
Skeleton convention	reusing engine animation needs its names and hierarchy	X down the bone, bend on local Z, positive Z flexes
Animator conventions	IK/FK default, spaces, colors, pickers	arms FK, legs IK; spaces world, cog, chest, head; ask
Shot load	characters per shot, dense meshes	three referenced characters
Face, mocap, budget	face to scenario-maya-deformation; HumanIK; joint and influence counts	jaw, eyes, aim control; ask the lead (antCGi: "it pays to be frugal")

Choose the path

Quick Rig + HumanIK	AdvancedSkeleton	mGear Shifter	Custom (mx_rig)
biped facing +Z; game or background character "not looked at closely" (Maya Learning Channel); HumanIK when animation is retargeted (2027 Help); 15 nodes, strict T-stance	body plus face from a fit skeleton, rebuilt by toggling fit; engine export via Convert to BlendShapes Only	many characters, a team, Unreal exporter; confirm a 2027 build for the OS first [verify]	hero rig, strict engine skeleton, rigs the agent must rebuild and verify: this workflow

Workflow

1. Gate the model. mx_validate.validate(profile="model") clean; cm; height to spec (about 180); feet at y = 0; centered on X; frozen; eyes dead ahead; loops at knuckles, elbows, knees; mouth bag for a face (antCGi). Send problems back as a list; never edit a client model unasked. GATE: no fail; wire sheet looked at.
2. Place joints at pivots on edge loops: spine joint under the rib cage (antCGi), wrist decoupled from the palm (Fragapane), elbows pre-bent back, knees forward; the first upper-arm and thigh twist joint stays at the top with the shoulder or hip, mid twist joints are placed by position so they keep the parent's orientation (antCGi). GATE: R.chain_plane does not raise on any limb; joints over wireframe looked at.
3. Build and orient. R.joint_chain(spec, parent, up="plane") for limbs, up=(0, 0, 1), up_axis="y" for spine and neck, world=[...] for root, cog, head, eyes and feet (antCGi: animators, engine foot placement), end_aim= for clavicles; existing skeletons: R.orient_with_cmds ('zup' for legs and spine). R.mirror_chain: behavior for limbs, orientation for eyes and face; rotate orders on controls (R.rotate_order(twist, main); Fragapane); R.set_preferred_angles(root). GATE: R.rig_check(joints=bind) clean; mirror within 1e-3 cm; all digits curl on one axis and sign, thumb fixed, so one attribute makes a fist (antCGi).
4. Hand the bind skeleton to scenario-maya-deformation (it binds while you build controls): joints only, one hierarchy, labels, rig_check JSON.
5. Control rig. R.ikfk_limb(bind[:3], name, rig_parent, ctrl_parent, own_solver=char + "_ikRPsolver"): handles of a type share one solver node, so several characters in a scene conflict (Fragapane: "can make a mess"; 2027 Help). Legs: extra=[ball, toe], orient_end=False, then R.reverse_foot with R.foot_pivots_from_mesh; spine P9; R.space_switch(ctrl, [None, cog, chest], names=[...]); R.ikfk_match(meta, to="fk"|"ik", key=True) uses stored offsets (antCGi instead adds a pole locator under the FK elbow and an ankle locator oriented like the IK foot). GATE: R.neutral_switch_test and pole drift below 1e-3; R.space_test as expected; FK and IK posed apart, R.blend_sweep_test finds no flip; roll sweep plants heel, ball, toe in turn; no shared-solver error; controls at zero.
6. Scale and stretch. Root scale only on groups that do not inherit it; stretch base lengths times root scaleY; IK spine joints under a world-pivot scale group, as segmentScaleCompensate blocks scale (antCGi). GATE: R.scale_test at 1.24 and 1.6, neutral and stretched, below 1e-3 cm.
7. Lock down. R.lockdown; FK hinges keep their bend axis; limits; controller tags; selection sets. GATE: R.lockdown_audit and R.opm_audit(rig) empty; the rig still works.
8. Performance. Key first (R.rom_keys, two different values per used attribute): a keyless rig profiles fast and lies. R.eval_census(rig), R.eval_ab(bind, meshes, frames); R.prepare_eval_graph(root=rig) tags every control (with the animator's Preferences > Settings > Animation > Include controllers in evaluation graph, first keys stop rebuilding the graph) and turns on the curve manager for the session. R.gpu_override_census(meshes), then the GUI Evaluation HUD must show a non-zero k count, else deformerEvaluator -meshes says why. Profile DG, Serial, Parallel and GPU Override with 1 and 3 referenced characters; keep dated profiles (Miquel Campos). GATE: no census error; modes agree within 1e-5; each dense mesh on the GPU or its reason written.
9. Review. R.rom_keys, then mx_review.playblast (GUI) or R.pose_review (headless); R.precision_test at shot distances; references/critique.md. GATE: frames looked at; elbows, knees, shoulders, wrists, feet read right.
10. Deliver. New version, validate(profile="rig") clean, test JSON, playblast, what was not verified.

Numbers

Value	Relative to	Source
Twist joints: upper arm 2 (top, mid), lower arm 1, thigh 2, calf 1, plus wrist and foot	bone	antCGi
Pole locator 30 units out by hand; computed pole distance = longer segment	elbow	antCGi; [added]
Switch, drift, scale error below 1e-3 cm; mid-blend path within max(1 deg, 5%) of the direct angle	bind joints	[added] gates
Foot break 25 deg: ball = roll to 25, then 50 - roll; toe = roll - 25; limits footRoll -40..80, footBank -60..60	footRoll	antCGi
Mid spine twist falloff 1, 0.5, 0.25	ribbon joints	antCGi
Precision: problems at 1e5 units, flicker at 1e6	rig top node	Fragapane
Hosts out of limbs: 20 to 17 ms per frame	one character	Miquel Campos
GPU Override: over 2000 vertices (NVIDIA), 500 (AMD); MAYA_OPENCL_DEFORMER_MIN_VERTS, 0 sends all	per mesh	Using Parallel Maya 2027
blendMatrix equal shares: 1, 1/2, 1/3 ...; counting the inputMatrix: 1/2, 1/3 ...	ordered stack	2027 Help; [added]
HumanIK 15 nodes; Unity 4 influences; Unreal FBX 2020.2	engine contract	docs

Quality gates

Code (mx_rig): rig_check (rest rotations, orientation, controls, poles, shared solvers, OPM doubling, evaluation census, lockdown), neutral_switch_test, blend_sweep_test, space_test, scale_test, precision_test, eval_ab, gpu_override_census, game_skeleton_check; scenario-maya-expert's validate(profile="rig").

Visual: placement over wireframe front and side; ROM playblast of every control axis, front, side and three-quarter; foot roll scrubbed; scaled rig in an extreme pose. Look for knee pop, elbow flips, candy-wrapper wrists, collapsing shoulders, sliding feet.

Common mistakes

Mistake	What it looks like	Fix
Orient Joint on a joint with rotations	nothing changes, no error	freeze rotations first, or R.joint_chain
Secondary world axis along the bone ('yup' on a leg, world Z on forward toes)	thigh and calf axes disagree	an axis across the bone; R.orient_with_cmds refuses it
Straight elbow or knee	IK bends the wrong way or backwards	pre-bend, rebuild, set preferred angle
Pole off the chain plane; pole too close	elbow drifts when constrained; arm flips as the hand moves	R.pole_vector_position; move it farther behind
worldMatrix into OPM under a moving parent, or values left under an OPM drive	double transform, reversed rotation	parent's worldInverseMatrix or .matrix; R.bake_to_opm; R.opm_audit
Equal blendMatrix weights for an average	the last target wins	1, 1/2, 1/3, or a parentMatrix node
Blending to a space's own matrix	control snaps onto the target, IK flips	offset targets (R.space_switch)
Handles on the scene's shared solver	referenced characters conflict; one edit changes every rig	own_solver per character
Settings host inside the limb	slow playback (node-level cycle)	host on the limb's parent space
Flip mid IK/FK blend	toes or wrist spin halfway	constraint blend: interpType Shortest (antCGi); matrix blend: measure with R.blend_sweep_test
Profiling a keyless rig	numbers collapse once animated; stutter on first keys	keys first; controller tags plus prepopulation
GPU Override HUD "Enabled (0 k)"	nothing deforms on the GPU	R.gpu_override_census, deformerEvaluator -meshes
Scale wired twice, or stretch not normalized	double scaling; limbs stretch when the rig scales	scale non-inheriting groups only; base length times root scaleY
Groups, handles, locators unlocked	"key all" breaks the rig	R.lockdown
Spline curve under its start joint; helpers under bind joints	dependency cycle; engine export errors	keep both outside the bind skeleton

Handoffs

• Receives from scenario-maya-modeling or scenario-maya-retopology-uv: a mesh passing validate(profile="model"), cm, Y-up, facing +Z, feet at y = 0, centered, frozen, symmetric, pre-bent limbs, eyes forward.
• Delivers to scenario-maya-deformation: the bind skeleton (joints only, one root, oriented, rotations zero, labels), rig_check JSON, and whether bind joints blend by constraint or through OPM (its pose readers depend on it); receives weights as data and correctives to reconnect.
• Delivers to scenario-maya-animation: the rig to reference, controls at zero and tagged, validate(profile="rig") clean, switch list, ROM playblast, known limits.
• Game rigs to scenario-maya-pipeline-scripting: game_skeleton_check clean, bind blends by constraint unless an OPM export was proven; they export the FBX.

Maya 2027 notes

• 2026 renames: addDL, multiplyDL, pointMatrixMultDL; plain math names create unitless nodes; condition, multiplyDivide, reverse, blendMatrix unchanged.
• parentMatrix (2025) has Snap and Initialize Target Offset; axisFromMatrix replaces the per-axis nodes.
• Orient Joint: negative primary axes fixed in 2027.1; Auto orient secondary axis since 2025.
• ikHandle lists only RP, SC and spline; load spring and 2-bone solvers first. FBIK forces Serial: use HumanIK.
• GPU Override: 2027.2 known issue, a crash when scrubbing with it on (disable it or the invisibility evaluator); the paper covers OpenCL on AMD and NVIDIA only, so read the HUD on Apple GPUs [verify]. Evaluator settings are per session, not saved in the scene.
• Skin Tools layers block classic weight tools until deleted (scenario-maya-deformation). PySide6 only; PyMEL is not bundled.

References

• references/procedures.md: code per stage and its test; before writing rig code.
• references/expert-notes.md: judgment by expert, disagreements; when a decision is not covered here.
• references/critique.md: the rubric; at every gate.
• references/gui-paths.md: menus, Profiler, HUD; when driving the GUI.
• references/sources.md: sources, URLs, timestamps.