Known limits
This chapter lists what the kernel does not do or does not do reliably, ordered by how soon a user is likely to meet it. Each item says what happens and where the evidence is: a sweep result, an #[ignore]d test with its reason, or a run made for this book. Counts from sweeps are snapshots of a past build (see corpora and sweeps); re-run the sweep for the current state.
1. Things met in the first hour
Unknown fields are ignored. Part parsing is deliberately lenient. A misspelled field is skipped, and the feature runs with its default. Checked:
{"type":"extrude","id":"F2","op":"add","sketch":"F1","end":"blind","depht":8}
regenerates with no error and no warning, and the plate is 0.000001 mm thick (the default depth). A field with an unknown name (bogusField) is also accepted silently. The only structural errors are a part that is not an object, a wrong format, and features that is not an array; an unknown feature type is a per-feature error (unknownFeature, checked). Reporting ignored fields as errors or warnings is not implemented. Until then, compare the result with check or massProps instead of trusting an empty errors list.
A valid-looking reply can be wrong. The sweeps find wrongVolume, replies without an error whose volume differs from an independent computation. They occur in random overlapping curved primitives, outward shells and three-section lofts (details below). A volume sanity check exists for booleans (booleanVolumeMismatch) but does not catch everything. Call validate and compare massProps with what you expect.
Large blends are refused. A fillet radius that approaches half the shortest side of a box, or half the height of a cylinder, answers blendOverflow, although the geometry is valid and an independent geometry library builds it. In the fillet sweep this was 32 of 640 cases, starting at radii of about 6 mm on the sweep’s random sizes. Checked: a 9 mm fillet on the top long edge of the 8 mm-thick 40 x 20 x 8 plate answers blendOverflow (“spring leaves its face away from its edges”).
Blends at corners are limited. Variable-radius blends (var_box in the sweep: 9 of 40 ok, 28 wrongVolume, 3 blendFailed) are the weakest group. Unequal radii at a vertex where three edges meet end in a mitre, not a patch: valid, but not tangent-continuous, and flagged with a cornerNotTangent warning. A three-edge corner with radii 10, 5 and 5 (one case in the test-script corpus), face-face fillets, adjacent conic or curvature-continuous fillets (the ignored test conic_ring), and fillets of a ring at a blade root (the ignored test closed_blunt_root_ring_fillets_cleanly) answer blendFailed or blendUnsupported. Partial cylinder and cone sectors with fillets are blendUnsupported in the test-script corpus.
A fillet of a rounded-rectangle loop can stop responding. In the fillet sweep, 2 of 40 rrect_loop cases (3 with a second seed) returned no reply at all (kernel-null) for a top-loop fillet with propagate. A time budget may not stop such a request; run untrusted documents under a host-side timeout as well. See robustness.
2. Shells, offsets and thickness
| Case | Result (shell sweep, 40 cases per template) |
|---|---|
| Inward shell of a box, one or two faces removed | Works; thick walls refused with shellTooThick (expected) |
| Outward shell with the top removed | All 40 wrongVolume (an extra roof slab remains) |
| Shell of a drafted cylinder (cone) | 38 wrongVolume, 2 draftTooLarge |
| Closed shell of a two-box L | 19 wrongVolume, 21 refused |
| Tube with the top removed | 15 shellFailed |
| Thicken of planar, cylindrical and conical sheets | Works |
Walls made of general B-spline surfaces are offset by fitting a grid and report a sampled radius; extruded spline walls are exact. Shell and thicken of a hull with a point-ended loft work only when the cavity can be cut off short of the apex; a plain cone loft and walls thicker than the keel radius are refused with a hint.
3. Booleans on curved shapes
Coincident, tangent, contained and sliver relations between boxes, cylinders and spheres are handled (145 of 147 cases in the coincidence corpus without any tolerance; the two failures are near-coincident cylinders 1e-3 apart, and with a tolerance of 1e-3 or autoFuzzy all 147 pass). Random overlaps of curved primitives are the weak spot. In the boolean sweep, 29 of the 175 cases that were not an expected no-effect or empty result (320 minus 80 okNoEffect and 65 okEmpty) failed (22 wrongVolume, 5 invalid, 2 booleanFailed):
- anything with a torus (10 cases, relative error 1.7e-3 to 0.95);
- cone against cone and cone against cylinder;
- sphere against torus or cone;
- chains of three bodies with a cone, torus or sphere.
In the test-script corpus, the clusters are a sphere against a box corner, perpendicular cylinders intersected, cone and cylinder near a seam, and staged blends followed by a union (86 failures among 1664 cases). Tool bodies with 100 instances (a long pattern) make a boolean take 5 to 30 s.
4. Lofts, sweeps and free-form surfaces
- Three-section circle lofts differ from a smooth thru-sections surface: relative volume error 3e-3 to 0.15 in the shape sweep (16 of 30); in the mass-property sweep all 20 three-section lofts differ from the thru-sections reference, by up to 6.7 percent in volume, 3.7 percent in area, 1.6 percent in centre of mass and 10 percent in inertia. The mass-property reply is exact for the shape the kernel builds; the shape itself is a different surface from the reference. Two-section lofts, rotated rectangles and point-ended lofts agree.
- No closed loft, no sweep normal control; multi-section sweeps are partial; mitre joints that involve arcs, twist and multi-region profiles on multi-piece paths are not built.
- Surface-surface intersection on arbitrary NURBS is the weak spot of the free-form side. There are no network or constrained plate surfaces, and no mesh-to-B-rep conversion or sewing from meshes.
- Booleans on free-form walls: spline pockets, bosses and rings work on spline walls; other tilted plane or extrusion sections still march and can be slow.
- Thread features: pipe-thread tapers, threads on conical faces and rounded roots are not modelled.
5. Import
See the import chapters for details; the main points:
| Format | Main limits |
|---|---|
| STEP | No assemblies, placements, colours or names. Shell-based surface models and open shells are not imported. In four samples of 200 published files, 185 to 188 per sample imported valid and the rest were invalid (mostly TESSELLATION_OPEN and EDGE_OFF_SURFACE). Torus-fillet shell orientation and sphere pole boundaries are open |
| SolidWorks | Geometry of the active configuration only; the feature tree is read as a list, and sldprtHistory rebuilds it as editable features where it can (a part it cannot rebuild completely keeps the file’s exact solid for the rest); the (deltas) transmit is not applied; sheet and wire bodies refused (unsupported.body); self-intersecting tori and some B-surface domains unsupported; blends and intersection curves are fitted (reported, refused above 0.05 mm) and some then fail the strict validate edge check |
| SketchUp | A surface body of planar faces; no curves, dimensions, scenes or text; refused above 50 000 faces unless maxFaces is raised |
| ACIS SAT | Exact splines only; no periodic splines, elliptical cones or reflecting transforms; no export. The bundled sphere and torus files (one face, no edges) import with volume 0 and VOLUME_NONPOSITIVE |
| Other formats | No STL, DXF, DWG, IGES or glTF import or export (listed as planned) |
Feature recognition rebuilds a feature tree for simple prismatic and revolved parts: 27 of 66 prismatic sample parts were exact or close. Curved-wall holes, lofts, shells and variable blends are not recognised.
6. Output
- Drawings (drawings): geometry only, as polylines; no dimensions, annotations or sheet layout; hidden-line handling of curved occluders is approximate, and thread symbols are not drawn.
- STEP export writes one product with geometry only: no colours, names, layers or assemblies.
- Display mesh: the default tolerance is
max(1e-3, 4e-4 x bounding-box diagonal)mm. In the tessellation sweep 84 of 180 cases had a mesh volume more than 0.1 percent off the exact one, at worst 0.8 percent for a torus and 0.4 percent for a cylinder. Exact quantities come frommassProps. A rare crack (TESSELLATION_OPEN) remains for one text case in about 1500 and for some imported parts; one imported part (eyebolt) loses its torus eye from the mesh (an ignored test). - Text (text and fonts) is planar and single-line; no wrapping onto curved faces, no complex scripts.
7. Features not built
Not implemented: sheet metal flanges on a curved edge, louvres, forming tools, a jog and a mitre flange (the rest of sheet metal is built: see sheet metal); mate kinds profileCenter, path, linearCoupler, cam and universalJoint; hinge angle limits; upToFace for revolve and sweep, offsetFromFace on curved faces and upToBody; thin on sweep and loft and variable wall thickness on a thin extrude; moving a hole sideways with moveFace; a true n-ary union (booleans are sequential).
Six feature types are accepted as names (they are not unknownFeature) but are not built. dome, splitLine, fillpattern, surfaceOffset and cutWithSurface always fail with TODO(rebuild): <type> needs ck-ops (checked: each as an empty feature after the plate of the STEP chapter), and spattern does the same once it has a sketch (without one it answers missingSketch). Some options of built features are refused with the same prefix: a drafted mid or throughBoth extrude, a lip on curved or several rim faces, an extend with a shape other than same, and a surface sweep of an open profile. See regeneration and patterns and transforms.
8. Speed and concurrency
- The kernel is single-threaded. A trial with native threads gave 1.38 times; multithreaded booleans are open.
- Free-form parts (springs, propellers, shell-like lofts) and long patterns take seconds to tens of seconds in WebAssembly. The default per-feature budget is 120 s.
- Native runs are about two to three times faster than the module.
- Feature recognition of gear bodies takes 2 to 10 s.
9. Ignored tests
cargo test -- --ignored lists the documented open items (the others are developer probes that need an environment variable):
| Test | Reason in the attribute |
|---|---|
closed_blunt_root_ring_fillets_cleanly (ck-model) | The ring path runs but the ball-centre solve diverges at the blade root |
marine_fuse_open_edge_spacings_build (ck-model) | Three near-tangent rail spacings still leave an open edge |
the_concave_corner_patch_tessellates_watertight (ck-model) | The sphere patch seam is sampled finer than the cylinder’s end arc: 30 of 462 mesh edges are T-junctions (chord sag below 0.04 mm at r = 2) |
conic_ring (ck-ops) | Adjacent conic or curvature fillets at a corner of two spline sections answer blendFailed |
eyebolt_torus_fragments_tessellate (ck-io) | The torus eye of this imported part is dropped from the display mesh |
10. Common refusal codes
These codes mean the request was understood and refused; the kernel state is unchanged.
| Code | Typical cause |
|---|---|
unknownFeature | A feature type the kernel does not have |
blendOverflow, blendTooLarge | Blend radius or distance too large for the faces |
blendUnsupported, blendFailed | A blend configuration not built, or whose construction failed |
booleanFailed, booleanVolumeMismatch | The boolean could not assemble a result, or the result’s volume failed the sanity check |
shellTooThick, shellFailed, shellUnsupported | Wall thickness, construction or surface type |
draftUnsupported, draftTooLarge | Draft of a face type or angle that is not supported |
thickenFailed, thickenReach | Thicken could not close the offset sheet |
loft*, sweep* | Section, guide or path problems (profile open, guide misses, path invalid) |
cutNoEffect | A cut changed nothing |
featureBudgetExceeded, regenBudgetExceeded, booleanBudgetExceeded | A time budget ran out |
The full list with parameters is in errors and feature errors and diagnostics.