Scenario Blender Geometry Nodes

Blender Geometry Nodes (procedural systems and simulations)

Expert Geometry Nodes work means thinking in fields evaluated on domains, keeping systems declarative (data and instructions come in from outside), and shipping tools that artists drive from modifier inputs without opening the tree. For an agent it also means proving every tree with numbers read from the evaluated geometry and a render, because the node editor's red links and warnings are invisible from Python unless you ask for them. 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-blender-expert (execution channel, review loop, 5.2 API changes). Hair node groups (Interpolate, Clump, Curl, Hair Dynamics grooms): scenario-blender-hair.

Stance (the expert delta)

• A field is a deferred question, not data. Thommes: it is evaluated by the node that consumes it, on that node's domain and geometry; values exist only once named or consumed. Consequence [added, verified]: two coupled updates (position and velocity) must be captured together in one Capture Attribute, or the second reads the first's new value.
• Domain changes interpolate; everything becomes a Boolean by a rule. Erindale: float > 0 is true (vectors: any non-zero), so a subdivided vertex-group selection bleeds; decide with Float to Integer or Compare with epsilon. Prefer trait selections (position, normal, proximity, raycast) over index ones, unless you generated the geometry and its order is the design (the Blender Studio matrix video drives a stack from Index == 0).
• Simulation only when history matters. Thommes: a simulation is rules applied to state; an animation is a function of time. If nothing depends on the previous frame, drive it from Scene Time and get scrubbing and sub-frames free.
• Loops are a last resort. Thommes: built-in nodes working on all elements use acceleration structures, loops cannot. For-each is for black-box generators run once per element: field outside, single value inside.
• Everything is object-local. Thommes: a (0, 0, -1) vector rotates with the object. World directions go through Self Object > Object Info > Invert Matrix > Transform Direction.
• Tools, not trees. Thommes: expose every scene dependency and tweakable value as a group input, never hard-reference an object inside a reusable tree, and drive gizmos only from Group Input sockets, through invertible math, into the first input of each math node.
• 5.2 physics is declarative. Thommes (BCON26): forces, colliders and constraints travel as bundles and closures into a black-box solver. Customize from outside first (inputs, bundles, Custom Effector closures); make the asset local only to edit a built-in constraint. Any per-step custom attribute must be listed in Extra Sim Attributes, and topology-changing sims must append new elements at the end.

Establish first

Workflow

1. Plan the data flow. Name the component and domain each value lives on, where each field gets evaluated, and which primitive fits: plain fields first, Accumulate Field for per-group sums, repeat zone for N dependent passes, for-each for a black-box generator per element, simulation zone for state, closure to inject behavior, bundle to pass many named values. GATE: you can say for every field which node evaluates it.
2. Build with bx_gn.Builder (headless or live, idempotent: a tree of the same name is rebuilt in place). b.node(kind, prop=..., ins={...}) sets properties before inputs, links sources, assigns constants; b.zone(...) pairs zones and adds items; b.join([...]) keeps list order. GATE: b.done() returns (zones paired, every link created).
3. Attach and set inputs with G.attach(obj, tree, inputs={...}) or G.set_inputs(md, {...}); "attr:NAME" switches a field input to an attribute or vertex group. GATE: G.check(md)["ok"], no hard_refs on an asset.
4. Measure. G.stats(obj) (counts, attributes per domain, world bbox with instances), G.attribute, G.instance_transforms, G.positions. Then the two robustness probes: rotate the object 90 degrees (world-space effects keep direction) and double the input resolution (trait selections scale, index ones jump). GATE: numbers match the analytic expectation.
5. Look. G.snapshot(objs, path, views=..., frames=...) renders the current scene state with material colors and shadows (floating instances show); bx_review.review for realized meshes. Open every sheet. GATE: references/critique.md visual items pass.
6. Simulations. Step every frame in order with G.step(range(...)), or G.bake(obj) for random access; assert bounds, monotonic states, counts, no NaN. GATE: a direct jump to a late frame equals the stepped result only after a bake.
7. Package (assets): inputs grouped in panels with defaults, min/max and subtypes; tree.is_modifier = True; tree.asset_mark(); drop it on a different object (Thommes' Suzanne test). GATE: works on the second object with no tree edit.

Recipes in scripts/bx_gn.py, tested and rendered: scatter_tree (density x mask, world-space slope limit, upright or aligned, pick from object and collection), curve_array_tree (tangent-aligned array, alternating twist makes a chain), growth_tree (repeat-zone extrusion), particle_drop_tree (simulation zone with emission, world gravity, ground bounce). Prefer the shipped Essentials assets (Scatter on Surface, Array, Instance on Elements, Randomize Transforms) when their inputs cover the brief; build your own when you need world-space rules, masks or custom attributes.

Numbers