Files
make-cad/docs/updates/2.3.0.md
T
Ni Jingzhe afb63fc375 release: 2.3.0 (includes the never-published 2.2.9)
DECISION: release the one-solid Part, the Windows build fix and the three issue fixes (#87, #88, #89) as 2.3.0. Not 2.2.10:
the owner chose 2.3.0. 2.2.9 is skipped; it was prepared on dev but never tagged or published (no v2.2.9 tag, nothing on PyPI,
the dry run of its workflow published nothing), so there is nothing to withdraw.

WHY A MINOR VERSION
- It removes interface of the withdrawn 2.2.8 (`bodies=`, `Part.multi_body`, profile 3), adds keywords (`clean=` on three
  booleans), and widens the exporters' input.
- It changes a default: `cut_rsolid`, `cut_rsolid_list` and `intersect_rsolid` unify their result, so the topology and the
  fingerprint of a body that used to keep a split face change.
- It breaks a recorded format on purpose: `clean` is a required parameter of the three recorded operations, a graph recorded
  before 2.3.0 is rejected on replay and has to be recorded again (the owner's rule: no branches for old formats).

WHAT CHANGES
- pyproject.toml, uv.lock (the simplecadapi package entry), docs/skill/SKILL.md and docs/skill/references/SDK_OVERVIEW.md:
  2.2.9 -> 2.3.0.
- CHANGELOG.md: the 2.2.9 entry becomes 2.3.0 (2026-11-12) and gains the four changes; "Fixed" gains the blend validity check.
- docs/updates/2.3.0.md and 2.3.0.zh-CN.md replace the 2.2.9 notes (deleted): everything the 2.2.9 notes said, then the booleans,
  the exporters and read-only runs, the fingerprints, and the Windows build. The release workflow turns them into the release
  page.

Checked: test_skill_lint and test_auto_docs_gen pass; the release workflow checks that the tag equals the pyproject version
(v2.3.0). test_skill_build::test_files_without_directives_copied_verbatim fails on the base commit as well (a Windows encoding
problem of this machine). The code of dev was green in CI at 8c482366 before this commit; this commit changes no code.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
2026-11-12 14:47:57 +08:00

7.3 KiB

2.3.0 Update Notes

2.3.0 follows 2.2.7. 2.2.8 was withdrawn (yanked on PyPI, its GitHub release deleted) and a 2.2.9 that replaced it was prepared but never published, so everything of both is in 2.3.0. It is a minor version because it removes interface of the withdrawn 2.2.8, adds keywords, and changes what cut_rsolid and intersect_rsolid return by default.

A Part holds exactly one solid

2.2.8 let a Part hold several bodies, and a Part like that could not take part in an assembly. In 2.3.0 a Part holds exactly one solid again, and the editor keeps what it needed from the change: several solids can sit side by side in a notebook while a model is built.

The editor has to let several solids exist in one notebook with no assembly constraint between them: to pattern a tool and then cut, there is a moment where more than one tool and the body exist together. In 2.2.8 that was solved inside the Part. But a Part is the unit an assembly places, mates and gives connectors to, and a Part of several bodies could do none of that:

  • connectors on it were refused, so no mate or public connector could reach it;
  • explode_rassembly raised TypeError on any assembly that held one;
  • the older scene compiler read the single body of a Part in four places.

The solids side by side are intermediate values of a notebook, not product items. So they are notebook values now, and the Part is one solid. The full account is in design-docs/part-holds-one-solid.md.

  • make_part_rpart(part_id, body, name=None) takes one Solid, a Body (a solid with a display name, colour and opacity), or a list of exactly one piece of a split cut. Several solids raise ValueError, which says what to do: one Part per solid in an Assembly (identity placements when they only sit together), or union / cut them into one solid first.
  • Removed (they only existed in the withdrawn 2.2.8): the bodies= argument, Part.multi_body, scad.part_shape, inspect_body_overlaps_rlist, BodyOverlap, BodyOverlapWarning, and the part definition profile simplecad-part-definition-3 with solid_cache.bodies. A model JSON recorded with several solids in one Part is refused on replay, with the same guidance.
  • A part notebook with one open solid still gets the implicit Part. With several open solids there is no product: running it raises SeveralOpenBodiesError, which names the solids; make the Part from the one that is the part (scad.make_part_rpart(part_id=scad.notebook_id(), body=...)), or make a Part of each and an Assembly of them.
  • The editor draws such open solids side by side and its header says "N open solids"; use_as_product makes the Part from the solid you choose.

Booleans: cut_rsolid and intersect_rsolid clean their result

union_rsolid always unified the faces and edges that lie on one surface; cut_rsolid, cut_rsolid_list and intersect_rsolid did not. A cut through the side of a box with a cylinder leaves the cylinder wall split along its seam, and OCCT fillets an edge along such a split face wrongly: it returned an invalid solid with a wrong volume and no error (issue #88), or refused with a message that blamed the radius.

  • cut_rsolid, cut_rsolid_list, intersect_rsolid take clean=True (the default) and unify what they leave split. Tags and lineage follow the clean-up. If the unified result would be invalid, the raw result is kept. clean=False returns the raw result, as before.
  • Fewer faces and edges in the result of a cut that used to leave a split face: the geometry (volume, area, bounding box) is the same, the topology and so the fingerprints of such a body change. Rebuild stored products that contain one.
  • fillet_rsolid and chamfer_rsolid check their result: a solid OCCT calls invalid (while the input was valid) is an error naming the operation, not a returned value.
  • A fillet that fails on split faces says so (and that the radius is not the limit) instead of advising a smaller radius.
  • Recorded graphs: clean is a recorded parameter of make_cut_rsolid, make_cut_rsolid_list and make_intersect_rsolid. A graph recorded before 2.3.0 has none and is rejected on replay like any graph with a missing parameter: record it again.

Exporting without a .scadpkg file, and running where the source is read-only

  • The exporters take the result of a build. export_product_package_to_step / stl / obj / mjcf and the FreeCAD translator accept the result of @scad.part / @scad.assemble (or its definition) as well as a package, bytes or a path, and package it in memory without the scene (the tessellated projection no exporter reads). A STEP file from a freshly built part: scad.exporter.export_product_package_to_step(build_part(), "part.step"); the 120 x 120 x 10 plate of issue #89 takes 0.26 s where capture followed by export took 1.85 s.
  • run_notebook(path, cache=False) leaves the source directory alone. It no longer writes __marimo__/.../deps.json or creates the state directory beside the notebook, so it runs where the source is read-only (a container with the source mounted :ro). A run with the cache still writes its state there, and fails if it cannot.
  • A single-script task is a supported entry. The skill keeps "a model is a marimo notebook" (R3) for everything a person opens or reuses, and now allows a plain script for a task that asks for exactly that, with a template in domains/export-and-translation.md.

Fingerprints do not split a value on a quantization boundary

A body's geometry fingerprint quantizes every measured value to 1e-7 mm. A design value such as -3.34690815 mm lies exactly on a boundary of that rounding, and one unit in the last place of computation noise (a BREP round trip) sent it to the neighbouring integer, so the fingerprint changed and restoring a topology snapshot raised geometry_mismatch (issue #87). The quantizer now snaps the value to a grid far coarser than the noise and far finer than a quantum before rounding. Fingerprints change once for a body with a value on a boundary; a snapshot recorded before that does not restore and has to be captured again.

Fixed

  • On Windows, sca cell no longer kills the editor it talks to: the editor's process id is probed without signalling it (os.kill(pid, 0) on Windows terminates the process).
  • Building from source: GCC 14 (operators_promote.cpp includes <algorithm>), MinGW (_USE_MATH_DEFINES, and build_occt.py picks Ninja when a MinGW g++ is the compiler), and the macOS wheels are built on macos-15 (the wheels still target macOS 14.0).

For contributors

  • A plain uv sync builds the package on Windows. Without Visual Studio the build used to fall to the NMake generator and fail, and every contributor set CMAKE_GENERATOR, CC, CXX and CMAKE_ARGS by hand. The build backend (tools/build_backend/scad_build.py) now picks Ninja and the gcc on the path when Visual Studio is absent, finds OpenCASCADE (SCAD_OCCT_PREFIX, C:/occt, or a per-user cache it builds once) and limits the jobs by memory. Anything you set yourself still wins. uv rebuilds only when src/cpp changes ([tool.uv] cache-keys), not on a version bump.
  • AGENT.md records how to rebuild the C++ module on Windows and how commits are organised: one decision per commit, with the reasons from the discussion that led to it.