21e0e99c86
Model renders (screenshot with source model, ui.render) show the section the window shows: the CPU renderer drew the cut face on the far walls at their own depth in a plain grey hatch, so the inside of the body, its edges and the grid showed through it and the cut barely read as a section, it had no outline, perspective views cut along a curve, and an open Section Analysis dialog's preview (or a sketch's Slice) was ignored. The cut face now writes the section plane's depth where it is the visible surface (the CPU counterpart of the viewport's stencil cap), takes the viewport's cap tint and hatch, the cut outline is drawn (section_segments moved from the UI to engine::view and shared), world positions and lines are clipped perspective-correctly, and ui.render in the app uses the viewport's section plane
CypherNaugh_0x
2026-10-11 10:16:49 +02:00
cc22672041
Sketch commands return the ids of the points, curves, constraints and dimensions they create as created: {points, curves, constraints, dimensions}, worked out in Session::execute from the sketch before and after, so a script can dimension and constrain new geometry right away; ids stay stable and are not reused after a delete (undo rewinds the counters with the design). The MCP execute tool describes the field
CypherNaugh_0x
2026-10-11 05:25:40 +02:00
2406957039
A loft gives the same body whatever the order of its sections: the ruled side faces were parameterised from the first section to the last, and a loft cut that touches the body along an edge (a fall lofted from a shelled basin's whole outline down to a drain pad) worked for one order and failed for the other, with an Option::unwrap() panic inside the kernel's intersection or "shell not oriented and closed" from the planar fallback. The loft now builds its sections in one canonical order (first-to-last step with a positive largest component), so either order gives the same, working body. Tests: the basin fall cut, joined and as a new body in both orders and with the pad in every corner (engine), and the kernel loft's volume and cut in both orders, including large-below and planes facing down
CypherNaugh_0x
2026-10-11 05:09:53 +02:00
5073f5dd72
A failed boolean no longer blames "coincident faces on curved bodies" whatever the cause: the exact polygon fallback runs after any failed B-rep boolean (faces that touch, or cross at a shallow angle) and takes planar faces only, but its refusal named coincident faces and curved bodies regardless. It now says which operand is in the way and why: "the tool has 2 curved faces (cylinder)" (kinds: cylinder, cone, sphere, torus, freeform), a face that could not be meshed, a planar face whose mesh is not flat, or a mesh too detailed. Test: planar_fallback_names_the_curved_faces
CypherNaugh_0x
2026-10-11 05:25:30 +02:00
848e7aec0e
feat: pick sketch profiles by their bounding curve ids so features keep the same region when parameters move the geometry, matching curves cut into pieces, refusing ambiguous or missing picks instead of taking another region, and storing point picks by curves
CypherNaugh_0x
2026-10-11 10:17:25 +02:00
3afd6ae447
feat: add multi-language (i18n) support with Japanese UI
Mithril-3d
2026-10-11 12:41:41 +09:00
7c667eda53
test: verify constrained line geometry after preselection and redo
lillith
2026-10-11 01:24:42 +02:00
adb1e5fec5
timeline.inspect returns a feature's definition in the shape timeline.edit's set takes, so agents no longer guess field names: timeline.edit had no documented format ({"extent": {"distance": …}} worked, but only by trying). The new read-only command (feature: id or name) returns type, definition (the editable fields and their current values; bookkeeping and large payloads such as sketch geometry and imported STEP text are left out and listed under omitted), inputs (label, parameter name, expression, unit) and the feature's error and warning. timeline.edit's params doc points to it, and the book's Solid features chapter documents the edit format with the fields of the common features. Tests: inspect, edit by the field names it shows, inspect again (extrude, fillet, sketch); no undo step; hostile feature arguments are refused
CypherNaugh_0x
2026-10-11 05:01:44 +02:00
3573b96c7a
fix: shell bodies where four or more planar faces meet at a corner (a floor falling to a drain from the trough edges) by splitting such corners into several joined by short edges instead of refusing because the moved planes miss a common point
CypherNaugh_0x
2026-10-11 10:01:46 +02:00
c67ad22129
A sketch fillet keeps the dimensions of the corner it rounds: filleting the corner of a spout (a vertical riser and a horizontal top, the top's length dimensioned to the corner and the corner's height from the origin) dropped the height dimension with the corner point and left the length on the shortened straight part, so the top came out a fillet radius too long. Like Fusion, a corner that dimensions or constraints measure to now stays as a virtual sharp corner, a free point held on both curves' extensions (two point-on-curve constraints), with those dimensions kept and a line length to the corner turned into a distance from the line's far end to it; chamfers do the same. sketch.fillet and the chamfers report virtual_corners and moved_to_virtual_corner, and the viewport draws a virtual corner as an open square with dashed extension lines from the curves. Tests: the spout keeps its 80 x 100 size and both dimensions still drive it, the virtual corner survives save/load and undo/redo, a fillet with nothing on its corner leaves no point behind, and nearly straight or folded corners with dimensions never panic
CypherNaugh_0x
2026-10-11 05:01:40 +02:00
75e9472a47
modified: CONTRIBUTING.md new file: flake.lock new file: flake.nix
GU Junhua
2026-10-11 16:07:46 +08:00
8c1b1478ce
Joins and cuts of round pads, bosses and holes across a floor that falls a degree or two no longer fail: the B-rep results were right, but the plausibility check meshed them without refinement, and where the shallow cut trims a round side to a thin band the plain mesh spans it with triangles cutting through the cylinder, so the check saw missing material (0.8 to 1.6 % of samples misclassified), rejected the result, and the planar fallback then refused the round tool. The check now meshes the operands and the result with refinement on curved faces (as measure does). Test: round_pads_crossing_a_shallow_fall (a 1° fall and a 2° four-face fall toward a drain)
CypherNaugh_0x
2026-10-11 05:33:50 +02:00
6b46915507
docs: align testing guide with the CI profile
lillith
2026-10-11 00:32:48 +02:00
a6bb8fc924
feat: stabilize face attachments and closed spline workflows
bherbruck
2026-10-11 05:32:46 -04:00
c1d819f235
feat: preserve body history and placed modelling workflows
bherbruck
2026-10-11 03:59:12 -04:00
ff1453f868
View cube face labels are larger: one font size for all six, the widest (BOTTOM) spanning 90% of its face, instead of a fixed 9.5 px that left the names small on the cube. Tests: BOTTOM spans 90% of the face seen straight on, and every label stays within its face from any direction
N3MI-DG
2026-10-11 10:06:05 +11:00
e3ee1b359e
Fix cubic input-name repair when loading large designs
bherbruck
2026-10-10 19:04:19 -04:00
f33efa2b6e
The timeline scrolls sideways when its features don't fit the window: with more features than the window is wide, the features past the right edge were clipped and could not be reached
CypherNaugh_0x
2026-10-11 00:13:53 +02:00
c61b77613a
solid.cylinder, box, sphere, torus and coil place by expressions: base, corner and center take numbers or length expressions per coordinate ("basin_width/2", "10 mm") and follow parameter changes, renames and component frames; before, a point that was not three plain numbers (or an unreadable axis) was silently replaced by the origin (or Z). The features now store the point as expressions (plain numbers still saved as numbers, so existing designs read and write the same) and evaluate it on every recompute; an unreadable or out-of-range point or axis is refused with an error. Tests: primitive points take expressions, follow parameters, renames and a moved component, and bad points are refused
CypherNaugh_0x
2026-10-11 00:01:16 +02:00
9c6f9410d2
Fix equal-distance chamfers on sloped end faces
CypherNaugh_0x
2026-10-11 00:30:27 +02:00
ade30e9c4e
Merge origin/main (with #86): sketch.dimension reports the solved value in the dimension's unit; the spout test expects the solved 30 and 60 degrees
CypherNaugh_0x
2026-10-11 00:13:51 +02:00
9c09e82cff
sketch.dimension reports measured on the solved sketch: it reported the value the geometry had before the new dimension was solved, so dimensioning a 99.84 mm trough line to 100 answered measured: 99.84 and looked like the dimension had not taken. The command now measures the new dimension again once the sketch is solved (driven dimensions too), through a new solvecraft_sketch::measure_dimension that shares the solver's own measuring. Tests: lengths, a driven distance and a diameter report their solved values
CypherNaugh_0x
2026-10-11 00:01:19 +02:00
1227fab860
Extrude's start_offset says which way it goes: along the sketch normal, also with direction: negative (where a positive offset starts on the far side of the sketch from the extrusion)
CypherNaugh_0x
2026-10-10 23:33:27 +02:00
9a2a3cef8b
A sweep along a path with sharp corners follows the corners: sweeping a rim profile along three lines (an open rectangle) gave a body of volume -0.0006 mm3 with its centre of mass at x = -3.5e9, and no error
CypherNaugh_0x
2026-10-10 23:33:27 +02:00
3086c32a9a
A round hole elsewhere on a body no longer breaks a loft cut beside it: a loft cut from a pocket floor (556.68 x 99 down 4 mm to 80 x 80), with a 35 mm hole 66 mm away that it never touches, failed with a kernel panic in TroughFall, while the same cut without the hole, or with the hole made afterwards, worked
CypherNaugh_0x
2026-10-10 23:32:48 +02:00
5249ae84ef
Moving a body that a boolean made no longer hangs: a boolean between a cylinder and a tilted plane (an elliptical edge) froze the app at 100 % CPU, because its retries move both bodies and the result back
CypherNaugh_0x
2026-10-10 23:31:46 +02:00
f3a689f284
A thin part trimmed by a long one is no longer refused as an implausible result: a 437 mm bar with a 3 mm concave arc in its profile, cut by a 920 mm prism with a top tilted 21 degrees, failed with "implausible result volume 3870" (it was 3694 before the cut)
CypherNaugh_0x
2026-10-10 23:17:02 +02:00
f17ff913d6
A panic while healing a boolean result no longer ends the boolean: a shallow, wide loft cut from a top face (556.68 x 99 down 4 mm to 80 x 80 in a 900 x 480 x 160 slab) failed with "called Option::unwrap() on a None value"
CypherNaugh_0x
2026-10-10 23:16:01 +02:00
ab23fc9906
A failed MCP batch goes back to where it started however long the undo history is: rollback undid as many steps as the undo history had grown, but the history is capped at 200 steps, so once it was full a failed batch reported undone: 0 and left its sketches behind (an empty Sketch1, two sketches named TapHole). New commands edit.checkpoint and edit.restore_checkpoint (forget: true drops one) remember and restore the design, the active sketch and the undo history; batch takes a checkpoint before its first step, restores it on a failure and forgets it on success. Against an app without checkpoints batch falls back to counting undo steps
CypherNaugh_0x
2026-10-10 20:21:47 +02:00
54a86dc6f5
Deleting a sketch deletes every feature built from it: a pipe whose path sketch was deleted stayed in the timeline as a broken feature ("unknown sketch"), and so did holes drilled at sketch points and patches filling sketch regions. Which sketches a feature uses now comes from one place, FeatureKind::input_sketches, a match over every kind with no wildcard arm, so a new feature kind does not compile until it says which sketches it uses; a test reads the FeatureKind definition and fails when a kind with a sketch field is listed as using none (it catches a sketch field added to an existing kind). timeline.dependents reports the same features
CypherNaugh_0x
2026-10-10 20:21:47 +02:00
acf4669f73
fix: select complete circular rims and curved surfaces in 3D
bherbruck
2026-10-10 17:13:36 -04:00
e042a7b555
ci: cache native Linux test compilation with sccache (#93)
bherbruck
2026-10-10 17:07:18 -04:00
96ee326424
ci: drop pages: write permission from Flatpak repository build job
Dylan M. Taylor
2026-10-10 16:18:23 -04:00
d6cdec40da
Fix modelling regressions and reconcile WIP UI work
bherbruck
2026-10-10 15:05:27 -04:00
49bd89de13
Parameter and feature edits report the features they restored from the evaluation cache: setting a parameter back (or a timeline.edit back) reused the earlier results for the same inputs, so the geometry changed back but recomputed said 0, which read as "nothing rebuilt". The evaluator now counts cache hits as last_restored, and parameters.change, timeline.edit, timeline.redefine and solid.press_pull on a fillet or chamfer face report it as restored next to recomputed (params docs, the MCP set_parameter description and docs/mcp.md say what each counts). Tests: setting a parameter back and a feature edit back report restored, in the evaluator and through the commands
CypherNaugh_0x
2026-10-10 21:00:32 +02:00
90ce5261b3
An angle dimension between two lines takes the reading nearest its value and reports degrees: without text_at the angle was measured between the lines' directions, so a riser pointing up and an outlet pointing down-left (meeting at 30°, directions 150° apart) asked for 30° swung the outlet round by 120°, which the fillets between them could not follow, and the dimension was refused as over-constraining with a degree of freedom left. Two lines make two angles, θ and 180° − θ; the dimension now takes the one nearer the value (the acute one without a value), and text_at still picks a sector. sketch.dimension reports measured in the dimension's unit (deg for angles, mm for lengths) with a unit field, instead of radians for angles
CypherNaugh_0x
2026-10-10 20:40:07 +02:00
81be4e74b8
browser.visibility hides and shows sketches, bodies, construction geometry and origin items (items: ids or names, or a whole folder; visible: bool), so agents can hide helper sketches over MCP. What is hidden moved from the egui UI state into the engine session (view state: not saved, not an undo step); the desktop eye toggles and menus use it, headless renders and fit leave hidden items out, and document.inspect, sketch.inspect and ui.inspect report visibility. ui.set still accepts hiddenBodies, hiddenOrigin, hiddenSketches and shownSketches. Tests: helper sketches hide and show in the engine (hostile input changes nothing) and through MCP only
CypherNaugh_0x
2026-10-10 20:02:06 +02:00
0321778d38
Perspective views no longer show hidden edges through solids: lines, decals and ghosts were pulled toward the eye by a fixed step in depth, and perspective depth is not linear in distance (near plane at about 0.1 % of the distance), so at the model that step spanned tens of millimetres and hole bottoms, the far side of bores and the back edges showed through the part (and through section caps). In perspective the point now moves toward the eye by a fraction of its distance (about 0.2 mm at 140 mm, like the orthographic step); orthographic is unchanged. Not covered by an automated test (no GPU in the test suite); checked in the app with before/after screenshots
CypherNaugh_0x
2026-10-10 19:23:49 +02:00
b5fd55cce6
Section Analysis cut faces hide the inside of the body: the cap was drawn on the far walls' back faces at their own depth, so every edge between the cut and the far wall (hole and boss edges, corner seams) showed through it like a pane of glass. While a section is shown the solids mark in the stencil where the cap is the visible surface, then draw the cap again there at the depth where the view ray meets the section plane (orthographic and perspective), so holes stay open and the inside is hidden. The native app asks for an 8-bit stencil; without one (eframe's web painter has none) the cap draws as before. Not covered by an automated test (no GPU in the test suite); checked in the app with before/after screenshots
CypherNaugh_0x
2026-10-10 19:19:46 +02:00
b1ec9fc1ce
Release: drop the riscv64 Linux build
bherbruck
2026-10-10 17:09:56 +00:00
29796762de
cargo xtask test runs every test binary at once (cargo test runs them one after another): 323 s to 75 s on 12 cores. cargo xtask ci runs the quick gates (fmt, assets, licences, layers) first, then clippy, then the tests with the wasm check and doc tests alongside; the wasm check uses the ci profile to reuse clippy's host build scripts and proc macros. GitHub CI runs its tests through cargo xtask test
CypherNaugh_0x
2026-10-10 18:51:18 +02:00
aa268f0f0e
Faster tests: hostile_params_never_panic shares its commands across threads (154 s to 20 s), the design-file fuzz test tries its seeded cases on all cores (21 s to 7.5 s), the kernel's tangent-primitive boolean test runs its three seeds at once, and the headless scenario harness skips the software viewport render it threw away every frame (scenarios 63 s to 26 s). kill_mid_save waits for the saver to reach a varying round instead of sleeping a fixed time, so it holds on a busy machine
CypherNaugh_0x
2026-10-10 18:51:18 +02:00
4a41de72ec
Autosave ids are unique within the process: an app restarted in the same second (the scenario harness crashes and restarts in one process) took the id {seconds}-{pid} of the crashed app, locked its entry again and so offered nothing to recover. The id now carries a per-process sequence number. Test: a restart in the same second still offers the crashed app's work
CypherNaugh_0x
2026-10-10 18:51:18 +02:00
524ef6b8c2
The up axis preference (Preferences > Design, ui.prefs up_axis) turns the view: the camera, home view, ground grid and orbit follow it, at startup too
bherbruck
2026-10-10 17:08:51 +00:00
74f70614cc
Section Analysis: cut faces in their body's colour (tinted toward the cap colour) and hatched; snapshots show the section too (the cut part removed, hatched cut faces)
bherbruck
2026-10-10 16:49:54 +00:00
c90070c497
Agent cursor speed: a slider from 1× to 5× with Instant at the far right. Instant no longer breaks: the command runs at once (as with the cursor off) and the pointer jumps to where it acted, with a click ripple. ui.agent_cursor speed takes 1–5 or "instant"; saved "slow" and "normal" read as 1×. Scenario covers 1×, 3× and Instant against the cursor-off result
bherbruck
2026-10-10 16:54:54 +00:00
5ec6d3562e
Version 0.1.3
bherbruck
2026-10-10 16:42:41 +00:00
9508ee749c
Shell refuses a wall thicker than a rim instead of returning a corrupt body: the topology-preserving face offset now checks every planar face's moved loops, and a loop that turns inside out or two loops that cross (a rim or ledge narrower than the move, e.g. a bowl opening sliding past the slab's outer edge) is an error naming the spot. Before, the self-crossing cavity went into the boolean and the shelled body came back with roughly twice its volume and its centre of mass outside the body, so every later cut failed its volume check
CypherNaugh_0x
2026-10-10 17:55:31 +02:00
a0f456d4f4
Clippy passes on macOS and Windows again: the options rebinding for the XWayland default is compiled only where the Linux block that mutates it is, instead of leaving an unused mut that -D warnings rejects (#75)
CypherNaugh_0x
2026-10-10 18:09:07 +02:00
9a22d9df92
Section Analysis takes a world coordinate: inspect.section {plane, at} places the cut at that coordinate on the named plane's axis (XY: Z, XZ: Y, YZ: X), so XZ's -Y normal no longer flips the sign; the params doc gives each named plane's normal; the result reports axis and at for named planes. at with offset, or with an {origin, normal} plane, is an error
CypherNaugh_0x
2026-10-10 17:10:04 +02:00
b8bd43f797
chore: bump GitHub Actions to fix Node 20 deprecation warnings
Dylan M. Taylor
2026-10-10 15:55:50 +00:00
5d5ff44db2
Drawing connects to existing points: a new point landing exactly on an existing point (the origin too) gets a coincident constraint in the engine, for every drawing command and every caller (MCP, scripts, clicks), in the same undo step; connect: false opts out. Re-tying an existing coincident returns it ("existing": true) instead of failing; the UI's after-the-fact tie_snaps is gone
CypherNaugh_0x
2026-10-10 16:25:41 +02:00
448ea769eb
CI builds tests and lints with a ci profile instead of --release
bherbruck
2026-10-10 15:06:58 +00:00
49d6682fe7
Rectangles name their sides and corners: sketch.rectangle.two_point and .center return sides {bottom, right, top, left} (line ids) and corners {bottom_left, bottom_right, top_right, top_left} (point ids), so they can be dimensioned without a sketch.inspect round trip
CypherNaugh_0x
2026-10-10 14:50:32 +02:00
d57e90a32a
View cube labels are printed on their faces: each face's name is laid out flat and mapped through the cube's projection, so it turns and foreshortens with the face and fades out as the face turns edge-on (instead of floating upright at the face's centre and vanishing past 70 degrees). TOP and BOTTOM read upright in the plan views clicking them turns to. Tests: every label reads upright in its own view and stays within its face from every direction
David Gray
2026-10-10 21:06:34 +11:00
52d2936f94
Linux: default to XWayland for file drag-and-drop on Wayland desktops
Dylan M. Taylor
2026-10-10 09:02:59 -04:00
03790f5509
View cube, move triad plane squares and the Offset Plane preview no longer flash lines while orbiting: faces turned nearly edge-on are drawn with projected_polygon (an unfeathered fill, outlined edge by edge) instead of egui's convex_polygon, whose miter joins shot spikes hundreds of pixels long at the slivers' sharp ends. Tests check everything drawn stays within each shape from 4050 view directions
David Gray
2026-10-10 20:28:08 +11:00
be095132b9
Re-land of the shell branches on the renamed ids (#6): the occurrence Move window goes through frame.rs like every window
bherbruck
2026-10-09 23:00:50 +00:00
8b60d77dad
Scenario s40 (toolbar customisation survives a restart) runs now
bherbruck
2026-10-08 19:05:57 +00:00