ogeom

A boundary-representation CAD kernel, written from scratch in Rust.

Status: geometry, topology, construction, tessellation, intersection, booleans, blends and chamfers, offsets, shells, sweeps and lofts, healing, product structure, 2D drawings with hidden line removal and sections, and exchange — STEP and IGES in both directions, STEP carrying assemblies, colours, semantic PMI and saved views, plus the native format, STL, DXF, glTF, OBJ, PLY, VRML and 3MF — are built and tested.

How complete that is against a real kernel is measured, and gated. The target is docs/SCOPE.md: parity with the reference kernel's modelling modules, and nothing else. docs/PARITY.md audits every public header of those modules against 97 named capabilities — every verdict citing symbols and tests the build verifies — and tools/check.sh fails if the audit and the code drift apart. No capability is absent; the partial rows say exactly what restriction each one carries.

Documentation: the guide and the API reference. Every code block in the guide is included from a test this repository runs, so an example that stops being true fails the build.

What this is

An independent solid modeling kernel: parametric curves and surfaces, shared B-rep topology, booleans, blending, offsetting, tessellation, assemblies, drawing generation and data exchange. A kernel and nothing else — what that excludes, and why, is docs/SCOPE.md.

It is not a binding, a wrapper, or a port. It depends on no existing CAD kernel, vendors none, and links against none.

The data model

Every crate implements one shared B-rep model. Geometry (curves, surfaces) and topology (vertices, edges, faces, shells, solids) live in shared arenas; a shape is a cheap handle into them, and the same node can appear many times — placed, mirrored, instanced — without being copied. The model rests on a small set of invariants, each cheap to honour from the start and effectively impossible to retrofit across a kernel's worth of algorithms later:

a shape is a (topology node, location, orientation) triple, cheap to copy;
location is a chain of transforms, not a flat matrix, so instancing works;
orientation composes multiplicatively on descent;
identity is a trichotomy — same / equal / partner — each with a matching hasher;
tolerances are per entity and only ever widen;
an edge carries a list of representations, including one pcurve per adjacent face;
every operation emits history.

Two places where the conventional design is wrong and this one diverges:

Stable entity identity. Kernels that identify topology by pointer produce new entities on every operation, so downstream references die — that is the topological naming problem, and applications have spent years working around it. Here, provenance is recorded at creation.
Predicates behind a trait, so the robustness strategy can change without rewriting algorithms.

Layout

Every crate listed has working code in it.

• crates/ogeom-core        arenas, identity, errors, tolerances, predicates
• crates/ogeom-math        primitives, transforms, B-spline basis, solvers
• crates/ogeom-geom        curves and surfaces behind adaptor traits
• crates/ogeom-topo        the B-rep data model
• crates/ogeom-algo        construction, primitives, measurement, classification
• crates/ogeom-mesh        tessellation
• crates/ogeom-io          native format, STEP, IGES, glTF, STL, DXF, OBJ, PLY, …
• crates/ogeom-intersect   curve/curve, curve/surface, surface/surface
• crates/ogeom-bool        general fuse and the boolean filters
• crates/ogeom-heal        healing, simplification, upgrading
• crates/ogeom-fillet      blends and chamfers
• crates/ogeom-offset      offset, shell, sweep, loft, form features
• crates/ogeom-hlr         hidden line removal, sections, 2D drawings
• crates/ogeom-doc         product structure, appearance, semantic PMI
• crates/ogeom             the public umbrella API
• tools/ogeom-cli          command-line front end
• tools/ogeom-view         software renderer, so a wrong result is visible
• tools/ogeom-bench        benchmarks, watched by the gate but not gating
• tools/apisurf            reference-index and API-usage tooling (analysis only)
• tools/parity.py          generates and checks the parity audit

Working code that is not kernel lives in outside/, which is a separate workspace the kernel excludes. docs/SCOPE.md says what the boundary is and outside/README.md says why each crate is on the far side of it.

• outside/crates/ogeom-recognize   feature recognition and process planning
• outside/crates/ogeom-reverse     mesh → B-rep, canonical surface recognition
• outside/crates/ogeom-select      BVH picking, selection, draft and thickness
• outside/crates/ogeom-sketch      2D geometric constraint solver

Building

cargo build --workspace
cargo test  --workspace

The full gate — format, lints, repeated tests, docs, the book and the parity audit — is one script, and it is what CI runs:

bash tools/check.sh   # needs mdbook: cargo install mdbook --locked

outside/ is excluded from that workspace and builds on its own:

cd outside && cargo test --workspace

License