ADR-0003 — Permissive licenses only, enforced by CI§

• Status: Accepted
• Date: 2026-08-06

Context§

Surveying the browser CAD/BIM ecosystem turned up an uncomfortable and non-obvious pattern: in four separate categories, the best-in-class option is copyleft or not open source at all.

The tldraw case is worth calling out because it is the easiest to get wrong. It looks like an open-source library, has an npm package and public docs, and is genuinely excellent. Its actual license is development-only by default; production requires a paid key, the hobby tier keeps a "made with tldraw" watermark, and the startup license is roughly $6,000. Adopting it as a foundational canvas dependency and discovering the terms later would be a strategic problem, not a licensing detail.

MassingViewer publishes @massing/* packages to npm. That makes the exposure worse than for an application: a copyleft dependency does not merely constrain us, it constrains everyone who installs us.

There is also an in-organisation hazard, which is the one that actually happens. Two sibling repos in MassingCloud are not permissively licensed:

• MassingCloud/massing-cloud — GPL-2.0
• MassingCloud/massing-families — license "Other" (unresolved)

Contamination from those arrives by copy-paste between sibling checkouts, not via npm install, so a dependency scanner alone would not see it.

Decision§

Permissive licenses only, enforced by scripts/check-licenses.mjs as a required CI check on every PR, walking the full transitive tree.

Allowed: MIT, MIT-0, ISC, BSD-2-Clause, BSD-3-Clause, Apache-2.0, BSL-1.0 (Boost), MPL-2.0, 0BSD, Unlicense, CC0-1.0, Zlib, Python-2.0, BlueOak-1.0.0, CC-BY-4.0.

Refused: GPL-*, AGPL-*, LGPL-*, SSPL, BUSL, CC-BY-NC-*, Elastic, Commons-Clause — and, equally, "UNKNOWN", "Other", "SEE LICENSE IN …" and "UNLICENSED". An unresolvable license is not a neutral state; it is an unreviewed one, and treating it as acceptable is exactly how the entries above get in.

MPL-2.0 is allowed deliberately. It is file-level weak copyleft, compatible with a permissively-licensed project, and it is what web-ifc and @ifc-lite/* use — the latter being a serious candidate for the 2D drawing pipeline (see docs/adr/0004-2d-drawing-engine.md). The obligation it creates is that modifications to MPL-licensed files stay MPL and are published. The practical rule that follows: consume MPL dependencies as dependencies and adapt at the boundary; never patch them in-tree.

Banned by name, with the reason attached§

Some packages pass a license check on their manifest but should still be refused, and some are simply the attractive choice in a category where the trap is invisible from npm install. scripts/check-licenses.mjs bans these by name and carries the reason inline, so nobody has to re-derive it:

tldraw, @tldraw/tldraw, xeokit-sdk, mupdf, mupdf-js, libredwg, @mlightcad/libredwg-web, @mlightcad/libredwg-converter, ifcopenshell, web-ifc-three.

ifcopenshell deserves a note: it is LGPL-3.0, it is excellent, and it is the reference implementation for authoring IFC correctly. It stays server-side, behind RemoteKernel, and never crosses into a bundle. LocalKernel uses web-ifc (MPL-2.0) + manifold-3d (Apache-2.0) + clipper2-wasm (BSL-1.0) instead.

Three enforcement layers§

1. scripts/check-licenses.mjs — the transitive tree, plus the name bans, plus a source-header scan for the two in-org copyleft repos.
2. GitHub's dependency-review-action on PRs, configured with the same deny list — a second, independent line of defence that catches what a manifest-based scan misses.
3. THIRD-PARTY-NOTICES.md is generated and staleness-gated. An attribution file that has drifted from the dependency tree satisfies nobody's legal review.

Exceptions live in ALLOWED_EXCEPTIONS in scripts/check-licenses.mjs and require both a written reason and an issue link — the same discipline massing applies to its .gitleaksignore entries. An exception without both is itself a failure.

Consequences§

• No DWG. DXF only, via MIT libraries. DXF covers most real-world 2D exchange; DWG is available later either through an arms-length GPL converter running as a separate process, or by licensing the Open Design Alliance SDK. Neither is a 0.x problem.
• No exact BREP hidden-line removal in v1. OpenCascade is LGPL-2.1 and a very large WASM payload. v1 ships z-sort HLR, declared honestly in DrawingProvenance.approximations so a drawing says what it approximated rather than implying precision it does not have.
• No parametric constraint solver in v1. planegcs is LGPL-2.1. Grid, snap and alignment inference cover architectural massing intent; revisit when a true sketcher ships.
• Reading GPL/AGPL source for patterns is fine and encouraged — Chili3D's builder chain and Open CAD Studio's ribbon are the best available references. Copying code from them is not. The distinction is real and the rule is: read, then write your own.
• The gate must be able to fail, not merely pass. Its refusal paths are verified against synthetic GPL-3.0, (MIT AND GPL-3.0), and (MIT OR Apache-2.0) manifests — the OR case must be accepted and the AND case refused, and getting either backwards would be silently wrong in opposite directions.