forgecad-build-model — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited forgecad-build-model (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Create new ForgeCAD models in the user's active ForgeCAD project.
Unless the user asks otherwise, the output is a manufacture-realistic prototype: a model someone could fabricate, buy parts for, assemble, inspect, and iterate in a real shop — not a concept sketch, not a universal 3D-printing exercise, not a claim of certified production readiness.
blockout → rough massing; production-realistic → DFM and production-intent materials; printable → make the selected printed parts honest; visual-CAD → clearly visual, not pretending build-ready.New .forge.js files go under date-based directories (today's date) in the user's current ForgeCAD project or a clearly named local folder:
YYYY/MM/DD/file.forge.js — single-file model
YYYY/MM/DD/folder/main.forge.js — multi-file entry point (always main.forge.js)
YYYY/MM/DD/folder/parts/*.forge.js — standalone/importable parts
YYYY/MM/DD/folder/lib/*.js — pure helpers/constants, no geometryparametric-lego.forge.js). Each part file must run standalone and import via require('./parts/name.forge.js', params)..js only for constants, math, tables, formatting — never geometry.matchTo().mkdir -p YYYY/MM/DD/[folder].param()/Param.bool() for tunable dimensions; pick the manufacturing process before styling; build real internals; follow the contracts below.forgecad run <file> (main.forge.js for multi-file).main.forge.js.Manufacturing process is a choice, not an assumption. Never treat every model as printable. Pick process cues from the load path and operating story: machined, bent sheet, tube-and-plate, wood/composite, molded-look, printed, or hybrid purchased-hardware construction (rideables: metal/composite structure + purchased wheels/bearings; furniture: real joinery). Print-specific features (slicer clearances, heat-set inserts, layer-oriented ribs) only when the process includes printed parts.
Visual style: expensive and credible, not generically colorful. Restrained material-driven palette (ivory, charcoal, satin black, brushed aluminum, brass, muted burgundy/green/navy, smoked polymer, natural wood); match color to material/process so metal, polymer, rubber, PCB, and wood read differently. Bright color only as small accents (controls, seals, indicators). Keep seams, fasteners, gaskets, and purchased parts legible.
Form is part of credibility. Real products carry deliberate edge treatment — chamfered or rounded profiles where hands, seals, or tooling meet the part, draft on molded faces, consistent proportions and design language across parts. A knife-sharp box reads as a blockout, not a product. Get the rounding from profile-level geometry (rounded sketch corners, chamfered profiles) per the fillet caveat below.
Variants are parameter-selected. Sizes/styles/revisions go behind one choice parameter (Variant, Preset); return only the selected variant. Comparison lineups only behind an explicit debug parameter so they can never pollute collision findings.
Deliver the real closed artifact — covers installed, parts in assembled positions. The forgecad skill carries the no-labels/no-cutaway rule (no explanatory labels, arrows, or legends in production geometry; never a cutaway, sectioned, or exploded default); it binds here. Markings only when the real artifact has them (serial plates, gauge ticks, molded icons) — sparse, process-appropriate. Explain roles via named return objects and verify.*; review annotations go in Viewport.label() or a debug mode, never exported geometry.
Internal geometry is part of the model. If the real artifact has internals, model them as real geometry even when hidden: cavities, ribs, bosses, screw holes, bearing seats, electronics/battery volumes, wire channels, mechanism clearances and stops. Verify hidden structure with exploration tools — alternate views, inspect sections, --hide, transparent shells, named ghosts — never by mutilating the returned model.
For any mechanism (linkage, hinge, slider, suspension, gripper, drawer), the rig is the source of truth: build and prove the motion structure first, then attach geometry — never retrofit motion onto finished shells.
verify.* checks all pass: forgecad run model.forge.js --joint "theta=45"
forgecad render 3d model.forge.js /tmp/theta-45.png --joint "theta=45" --camera isoUse real joint names; repeat --joint per control. A pose that fails to solve, clamps unexpectedly, breaks a connector check, or adds a collision means the rig is not ready for final geometry.
rotate() that makes one pose look assembled.Assembly so Motion controls re-run the real solve. Never bake a posed SolvedAssembly to make one pose look right.verify.*: convergence, connector origins coinciding, link lengths holding, end effectors reaching targets, running clearances positive.A mechanical script is not done when it merely looks assembled. Every visible piece needs a physical reason to be where it is: fused material, contact faces, a screw stack, a pin in a bore, a tab in a slot, a gasket on a land, a bearing in a seat, a cable in a channel, or a named intentional ghost.
For multi-part assemblies, the component model is mandatory:
shape plus connectors and metadata, e.g. return { shape, boltPattern, pinionZ }. Declare mating faces with .withConnectors({}); axes point outward, with prismatic slide axes as the explicit exception.connect(), match(), or matchTo(). Final translate() calls are not assembly contracts.require('./part.forge.js', params) overrides and up through returned metadata. Siblings never import each other; the parent routes shared decisions and measured outputs.shared-dims.js just to coordinate siblings.Reject these shortcuts on sight: sibling require(), assembly-space coordinates inside a part, translate() used to position a structural assembly member, console.log + if validation instead of verify.*, and bare connector.neutral() outside a reusable component library with compatibility checks.
matchTo(), verify each mate with verify.connectorDistance.addFixed/addRevolute/addPrismatic with a hand-built frame:) are scaffolding, not contracts. Before delivery, convert mating interfaces to connectors with connect()/match(), or prove the manual joint with forgecad debug assembly --fail-on warning and documented geometry.verify.*, not comments: verify.clearanceBetween for seated fits and clearance bands, verify.minClearance/verify.notColliding for keep-out and running gaps, verify.connectorDistance for connector mates. Part counts and generic dimensions never prove an interface.lib.fastenerSet(), lib.boltPattern(), real bores and pockets, connectors + matchTo()) — not finished backplates, brackets, or hinge leaves.Treat fillet()/chamfer() as experimental (Manifold can be incorrect, OCCT slow); prefer profile-level rounding and inspect before relying on the result.
When the user supplies mesh or CAD files to design around (a motor, an off-the-shelf housing, a scanned part), the import IS a component of the assembly — keep it, don't rebuild it parametrically (rebuilding is forgecad-reconstruct-cad-file, a different request).
Import.mesh() for STL/OBJ/3MF, Import.step() for STEP (OCCT backend), normalize scale and orientation, recenter to a sane local origin, then treat it exactly like a purchased part — connectors at its real mating features, positioned by the assembly via matchTo().forgecad run --details, or inspect section --ray across a bore or face pair. For 3MF, account for every build item printed by forgecad run before flattening.verify.clearanceBetween/verify.connectorDistance against the import like any other body. Collision Policy applies to imports too.Each returned part is real matter. Expected final collision count: zero.
verify.intentionalOverlap on the exact visible object pair with the physical reason. The mechanical-integrity gate honors a declaration only when that pair has a confirmed exact collision; unused or non-visible declarations still fail.difference(), primitives before union(), exploratory layouts) must be consumed, hidden, named as ghosts, or isolated with --focus/--hide so final findings stay meaningful.Prove technical validity and visual plausibility before declaring done. Apply to any model with multiple bodies, surface details, cables, rails, handles, product skins, or hidden mating geometry.
forgecad inspect physical components and require the count to match. Unexpected islands, accidental fusion, or bbox-only "touching" are model bugs.assembly(): forgecad debug assembly model.forge.js --fail-on warning. Fix warnings (multiple roots, manual joint contracts, disconnected bodies, unused connectors, collisions); a truly intentional one gets a visible reason in code.forgecad inspect mechanical-integrity . --collisions is the shareability gate — it fails on missing verify.* interface checks, fragmented named groups, uncontracted manual assemblies, positive-volume collisions, timeouts, runtime failures. Do not share while red unless the user asked for a blockout.--joint and/or in-script solve(state) checks, with convergence, attachment, and clearances holding at every pose.forgecad inspect fit interference; read the manifest collision count AND the evidence PNGs. Zero unexpected findings per Collision Policy; visually confirm where any findings appear.scene() rig (Scene Presentation below) — default flat lighting in a final render is a finding. One whole-model context view plus views chosen from this object's failure modes — opposing, underside, interior-facing, or grazing angles that catch internals showing through openings, covers that don't close, bad boolean cuts. Per meaningful interface: one contextual and one focused/isolated view. Risk prompts:inspect sections for hidden geometryA manufacture-realistic model must yield a package a shop can consume, not just a clean viewport.
bom() (exact spec, quantity, purpose) so the BOM lives in the model — forgecad export report must reproduce it without prose supplements.dim() annotations on the dimensions a builder must hit: overall envelope, critical interfaces, mating bores and bolt patterns.export stl/export 3mf for printed parts; export step for machined parts and CAD interchange; sheet-metal parts must unfold to a valid flat pattern (export cutting-layout for sheet goods). step, report, and cutting-layout need a Production license — if unavailable, run the free exports and name the gated commands that complete the package instead of failing.You are building blind unless you render. forgecad run passing only means the code didn't crash — it cannot tell you a hole is misplaced, a rib pokes through a cover, or a part doesn't fit. Render from angles chosen for the model's actual geometry and read every PNG. For command syntax, evidence selection, and manifest reading, use the forgecad-inspect-model skill and the CLI docs — this skill fixes only the cadence and the gates.
Render after every feature addition, boolean cut, symmetric copy placement, and the last feature. Inspect after adding hidden geometry a surface render cannot prove, after adding or moving mating parts, ghosts, connectors, thin walls, or screw holes, and before delivery with thresholds set for the material/process.
Keep inspection scenes small. Return one selected configuration; include only the parts needed to prove the current risk (if a check concerns three objects, inspect those three, not the whole shop floor); prefer --focus/--hide and parameter-selected diagnostic modes over permanent extra objects; collapse proven subassemblies into fewer named objects where identity doesn't matter for collisions, masks, or contracts. If you cannot hold the scene in your head, you cannot debug it honestly.
Ghost parts for fit checks. When a part holds or contains another object, render both with the contained object as a compact transparent named ghost — e.g. a box() at the seat position with .color('#ff4444').material({ opacity: 0.4 }), returned as { name: 'Servo Ghost', shape: ghost }.
Use verify.* for dimensions and clearances that decide acceptance; console.log() only for explanatory traces (shown under "Script output:" in forgecad run).
You cannot target a complex model in one pass. Decompose, solve the smallest piece, verify, compose upward:
forgecad run, then render and read the PNG. Fix while the scope is tiny.For any model with more than ~3 distinct geometric features, plan the decomposition explicitly before writing geometry.
Always set up scene() — default lighting looks flat. Worked recipes (studio and matte-industrial setups, named views, plinth) live in guides/scene-presentation.md via the forgecad skill; the schema is in the viewport docs. Hard-won cliffs:
lights replaces ALL defaults — always include an ambient light or the scene goes black.castShadow: true) + weaker cool fill/rim. Keep environment.intensity low — environment fill kills shadows.toneMappingExposure by ~0.05 before redoing the rig; avoid big ambient jumps.fov 35–50, target at the visual center of mass. Ground plane with shadows for grounded objects.studio for metallic/jewelry, warehouse/apartment for organic/matte, warehouse + strong directionals for mechanical, night + bloom/vignette for dramatic.~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.