* claim: #2987
* Flange on a plain solid says how to start sheet metal (#2987)
Flange, Flat Pattern, and Export DXF on a design with no sheet metal body
now say how to make one: Flange > Base, or Convert to Sheet Metal on a
plate's large flat face. An edge flange checks for a sheet metal body
before it adds a feature. The message is in the catalogs, and engine
messages that match a catalog entry are shown in the interface language.
* Dialogs show a refusal without the engine's parameter prefix (#2987)
The Flange dialog showed "invalid parameters for `sheet.flange`: " in front
of the new guidance. The dialog already names its command, so the red
refusal text leaves that prefix out and follows the interface language
when the message is in the catalogs. Create Issue still gets the full text.
* claim: #4116
* Web: a trackpad swipe pans instead of zooming (#4116)
Browsers deliver trackpad swipes as wheel events without scroll phases, so
the shared navigation treated every vertical swipe as a mouse-wheel zoom.
The web raw-input hook now brackets point-mode swipes (small or horizontal
deltas) with gesture start/end phases, ending after 100 ms idle or before
line/page wheel or pinch input. Wheel notches and pinch still zoom, and
camera input still goes through navigation::handle.
* claim: #221
* Units of new designs apply to the blank design already open (#221)
Changing Preferences > Units of new designs only took effect at the next
File > New, so the empty design open at launch stayed in millimetres. When
the preference changes (the dialog, Cancel, Restore Defaults, or
preferences.set) and the open design is blank and unsaved, it takes the new
units at once, without an undo step or unsaved changes. A design with
content or a file keeps its units.
* claim: #4794
* Hem: the direction arrow points the way the hem builds (#4794)
The Hem dialog shared the edge flange's arrow, which points out of the
sheet's larger neighbouring face. A hem folds back the other way, so the
arrow pointed away from the material it adds. Flip the hem's arrow so it
matches the geometry; edge flanges keep theirs.
A regression test compares the arrow with the side the finished hem adds
material on, for both Flip states, top and bottom edge picks, and both
dialog layouts.
* claim: #3220
* Model cache: keep fewer states of a design too big for its file instead of none (#3220)
A model cache over the container's budget keeps the final state, then the
slowest features' states while the bodies they add fit (a body several
states hold counts once), and is left out only when not even the final
state fits. Opening evaluates just the features whose state the file left
out, each from the state before it, so the model is whole afterwards.
* Model cache: keep each encoded body apart, copy only what a cache stores
Encoding a model too big for its file no longer copies every distinct body
into one buffer before thinning: the payload is copied from the encodings
once its size fits, so saving holds no more than before.
* Model cache tests: canonical runs in the timeline fixture; rustfmt
* Model cache: count the records a thinned cache keeps; always try the final state
Thinning counted only the bodies, so a model of many copies of one part
(records, not bodies, filling the room) shrank its budget to nothing and
saved no cache although its final state fitted. The choice now counts the
records of what it keeps, always keeps the final state, and says what it
counted, so each retry keeps less and the final state alone is always tried.
* docs: a thinned model cache counts its records and always tries the final state
* Model cache test: the slowest feature's state holds bodies 16 to 23
* Model cache: thinned opens with fingerprints that read the state before a feature
Since #3397 a feature's fingerprint reads where the bodies before it lie.
After a state the file left out, adopt takes the saved fingerprint and the
evaluation that works the state out checks it. A result evaluated ahead
is never taken as one the file left out. The thinned round trips include
the placed components' design, whose cut reads those placements.
---------
Co-authored-by: MateuszSychXpl <151755914+MateuszSychXpl@users.noreply.github.com>
* claim: #4880
* Rename a body: keep the names of siblings a later feature removed (#4880)
Renaming a body sets the stored names of all bodies its feature made, so
the siblings are not renumbered. Those names were read from the final
model, where a body that a later feature removed no longer exists: its
stored name was lost, it came back as `Ext (3)`, and the later Remove
failed with "no body `Body3`". Read the siblings from the feature's own
result instead, as the later features already are.
* Test: renaming one half of a split leaves the other
Renaming one half keeps the other half's name, also once a later feature
joins that other half away, after both an incremental and a full
evaluation.
* claim: #4846
* MCP tools: unknown command ids suggest the closest ids (#4846)
An unknown command id in `execute` or `batch` now lists up to three closest
ids, and an id with parameters packed into it says to pass them in `params`.
* claim: #4804
* Keyboard Shortcuts page: draw only the visible rows (#4804)
Off-screen rows keep their height but skip their cells, the list is
filtered once before measuring and drawing, and each row's shortcut
comes from the command spec already in hand instead of a registry
lookup.
* claim: #4899
* Sketch: offer Break Offset Association for offsets of lines (#4899)
Trim and Extend refuse an associated offset output and point to Break
Offset Association, but the right-click menu offered it only for
free-form offsets. Offer it whenever the curve is an offset output
(analytic, topology or free-form), using the sketch's shared lookup.
* claim: #4851
* solidcraft-cli bench-assistant: the standard Assistant model test
Runs the eleven v1 tasks of #4851 against a desktop app that is already
running with a control port. Each task starts a new design and a new chat
with All permissions, runs only the prescribed setup, sends the prompt
once and waits for the turn. The result is checked on the design
(document.inspect with measurements): body count, volume within 1%
(0.3% for the fillet), bounding-box size in any orientation, the lowest
corner, the body name, live feature kinds, and changed or unchanged
content; the volume question is checked on the reply. Failures get the
issue's categories from the conversation's items. Prints a table, and
writes the issue's comment template (--markdown) and the raw results
with each conversation (--json).
Dropping a feature just behind the history marker now rolls it back, and
dropping it just before the marker makes it active again, in both the
bottom strip and the left panel. timeline.reorder and timeline.move_many
take an optional `active` param that sets the marker in the same command,
so the move is one undo step and one recompute. A position on the wrong
side of the marker for the requested `active` is refused.
Drop feedback:
- bottom strip: the drop line sits centred between two features (it
used to touch the previous one) and shows at the first position
(it used to be drawn outside the strip there)
- left panel: a drop slot below the marker when it is at the end
- no drop line when the marker is dragged back to where it is
- a dragged feature is outlined in the drop line's green
- grab cursor on the marker, grabbing cursor while dragging the marker
or a feature
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #4932
* icon_audit: compare the requested-icons manifest with CRLF checkouts ignored (#4932)
On Windows, Git for Windows checks out requested_icons.txt with CRLF line
endings (core.autocrlf; the repo has no .gitattributes), while the test
builds the manifest with \n, so the comparison fails although the lists
are the same. Read the file with \r\n normalised to \n, as #1931 did for
the icon tools. Regenerating still writes \n.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #4966
* Agent chat: serve the bundled skills' reference through an MCP tool (#4966)
The bundled skills start with "read reference.md", but SolidCraft's own
chat turns the agent's file tools off, so its agent could not follow them.
The MCP server gains a read-only `skill_reference` tool that returns a
skill's documents from copies built into the binary. It takes a skill
name and a file name from a fixed list, never a path, and refuses an
unknown skill or file with the available ones. The skills point to it,
and docs/mcp.md and skills/README.md describe it.
#5187 gave the scene its own offscreen target with the app's sample count,
which skipped the OpenGL clamp to 1x from 6dd202d. The scene's count now
goes through effective_msaa_samples as well.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* test(mcp): reproduce whitespace bypass of request-size limit
Add a source-level regression that rejects a JSON-RPC request larger than
MAX_LINE even when most of it is whitespace, plus bounded-stdio framing
coverage and invalid UTF-8 recovery.
* fix(mcp): bound stdio line buffering and validate raw request size
Consume oversize JSON-RPC lines in bounded memory while keeping the
following request available. Reject whitespace-padded lines before
trimming, preventing a request-size cap bypass.
* docs(mcp): keep server framing comment idiomatic
* claim: #4585
* Sketch drag: a linkage held by redundant parallelograms still moves (#4585)
Cursor motion looks for the sketch's freedom in the null space of the
constraint Jacobian. Redundant constraints (a parallelogram held both
parallel and equal on each pair of sides) are dependent only at an exact
solution; after a correction to the solver's tolerance their rows differ
by about 1e-9, which the 1e-9 pivot tolerance counted as rank, so the
null space came out empty and a one-DoF linkage stayed locked. Use the
rank analysis' 1e-7, so the drag sees the freedom the sketch reports.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #4991
* Browser: Shift+click above the anchor selects the whole range (#4991)
The browser rebuilds its row order while drawing, and a Shift click looked
the anchor up in that half-built list, so an anchor further down was not
there yet and only the clicked row was selected. Resolve the range against
the previous frame's complete row order instead.
* claim: #5005
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Timeline: delete the selected features, Shift+click ranges, delete after the history marker (#5005)
Delete in a feature's menu deletes the selected features when the
right-clicked one is among them ("Delete 3 Selected Features").
Shift+click selects the range from the last feature clicked (or the last
selected one), in the order taken before drawing. The menu of a feature
after the history marker deletes every feature after the marker, with
their number in the label. All of it goes through delete_features: one
timeline.delete and one undo step. The confirmation no longer lists
features that are themselves being deleted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #5015
Co-Authored-By: Claude Code <noreply@anthropic.com>
* Assistant: add Anthropic-compatible API agents (#5015)
The Add agent wizard now offers Anthropic-compatible endpoints beside OpenAI-compatible ones. The API backend speaks the Anthropic Messages protocol with the same MCP bridge, tool gate, limits, undo and keychain rules: x-api-key and anthropic-version, a top-level system note, input_schema tools, content-block answers and one user tool_result message per round. Test sends a short message because compatible endpoints (Qwen's) need not list models.
Co-Authored-By: Claude Code <noreply@anthropic.com>
---------
Co-authored-by: Claude Code <noreply@anthropic.com>
* claim: #5021
* Opening a design: range-check integer-valued coordinates too (#5021)
The file-level number check skips every integer-valued number so that ids
and counters may be large, which let a coordinate such as
9223372036854775808 through. The typed design is now checked as well:
float fields (coordinates, sizes) are serialized as floats and must be
within the same +-1e9 limit, while integer fields stay integers and are
not limited.
* claim: #5022
* Plastic features: refuse sizes under 0.01 mm (#5022)
A snap fit with a near-zero thickness (1e-9 mm) got past the size checks,
which only asked for a positive number, and panicked in the kernel's
intersection code. The boss, lip, groove and rest had the same gap.
Every length of a plastic feature that becomes geometry (snap fit length,
thickness, width, catch_depth and catch_length; boss diameter, height,
hole_diameter, hole_depth and rib sizes; lip and groove width and height;
rest width, length, height and thickness) must now be at least 0.01 mm.
The check sits where the feature is evaluated, so a parameter changed later
is refused the same way, and the message names the parameter.
* claim: #5023
* Extrude and loft: refuse repeated profile indexes (#5023)
`profiles: [0, 0, ...]` with 10 000 copies built 10 000 identical bodies in
about 35 s, and a loft section listing one profile 10 000 times took 25 s
before it was refused.
A list of profile indexes now refuses a repeated index while it is parsed
("profile 0 is listed twice"), as offset association ids already were.
Before the profiles are resolved, a selection naming more profiles than its
sketch has, an index the sketch doesn't have, or two points in the same
profile is refused too. A design file that repeats an index still opens,
using the profile once.
Catch up with en.json: the new Preferences pages, assistant settings,
3D mouse axis assignment, bug report form, history names and the tool
help texts. pl now covers every message of en.json.
Coincident with two circles (or arcs) picked was refused: "two curves must both
be straight lines; otherwise choose a point and a curve". It now lays the second
pick on the first: Concentric and Equal together, in one step, with the first
staying put. A pair already concentric or equal gets only the other half; one
already lying on the other is refused as nothing to add.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* A renamed body keeps its number (#4957)
The design's body counter starts over at every evaluation, so a body
renamed after a join (the join changed it, so the rename cannot be
followed without evaluating the design again) gave its number back: the
next body took the name the rename freed, and a script that names a body
found another one.
The design now saves the names a rename freed, and the counter skips
them.
* A renamed body keeps its number (#4957)
The design's body counter starts over at every evaluation, so a body
renamed after a join (the join changed it, so the rename cannot be
followed without evaluating the design again) gave its number back: the
next body took the name the rename freed, and a script that names a body
found another one.
The design now saves the names a rename freed, and the counter skips
them.
* Format
A favourite is listed twice, under Favourites and under User. Double-clicking its name opened a
name editor in both rows, under one widget id; the two took each other's focus, one lost it at
once, and the rename closed with the name unchanged. The rename now opens only in the table it
was started in, under an id that includes that table.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #5087
* view.apply: keep the camera's roll, and refuse an up axis (#5087)
An explicit camera was read as a named view, which has no roll or up, so
roll was silently reset to 0 and up was ignored. Read roll as a finite
number in radians, clamped to ±π, and refuse up with a message saying the
up axis comes from the design settings.
A docked dialog is placed from the previous frame's drag offset, so a fast
drag can leave the pointer outside the dialog, over the model. The viewport
then treated the pointer as hovering the model and highlighted faces and
edges under it.
drag_dock now marks the frame in egui temp data while its title band is
dragged, and navigation::handle treats the pointer as not hovering the view
while that mark is current and a button is held. This covers every dialog
docked through drag_dock.
NVIDIA's EGL answers wgpu's request for a GL 3.3 context with exactly 3.3
(GLSL 3.30) yet lists GL_ARB_compute_shader. wgpu-hal then reports compute
shaders, wgpu-core builds its indirect-call validation pipeline, naga cannot
emit compute GLSL below 4.30, and request_device fails as Device(Lost).
The test draws nothing indirect, so it opts out of that validation.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #5049
* GPU tessellation: an empty plan no longer fails buffer validation (#5049)
Empty tessellation inputs were padded to 16 bytes, smaller than the
shader's 32-byte vertex and 48-byte surface records, so wgpu refused the
dispatch. Skip the dispatch when there is nothing to evaluate, and pad
each empty buffer to one whole element of what the shader reads.
The component's sketches, features, bodies, subcomponents, the joints among
them and the parameters they use move to a new .solidcraft design, keeping
their ids and names, and the component becomes linked to it as Insert Design
makes it. The new design must build the same bodies and the assembly must
look the same afterwards, else nothing changes and no file is written. Joint
faces on the component are found again on the linked bodies, and components
that build the same bodies can become placements of the link.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A STEP file inserted into a new design sits in its own component under the
empty root, and a sketch on its face must join that part going out and cut it
going in rather than make a separate body beside it. With a distance, main
already resolves this; Through All with no distance gave `auto` no reach to
probe, so it fell back to New Body and cut nothing. Through All now reaches
past every body in its direction, and `auto` reads it as a cut (#1234).
Co-authored-by: Steven <steven2021067613@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The 14 new built-in themes still listed four logo colours, so none of them read after #5465 and main's theme test failed.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #5461
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* App icon: draw the borderless mark everywhere (#5461)
app_icon.rs drew the older mark with white lines between the faces and a
white outline around it, so the Windows, Linux, web and README icons did
not match the macOS icon, the social card or the title bar. It now draws
the borderless mark from the faces and gradients of the macOS artwork
(03-solid-blue-cutout.svg): five faces with transparent gaps between them.
The filler fills even-odd spans so the concave faces work.
The Linux hicolor PNGs and both Windows .ico files are regenerated with
`cargo xtask app-icons`. The README logo, the web app's icons and every
package pick them up from there.
The logo colour token's fourth entry (the white edge) is gone: nothing
draws it any more.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The agent cursor only acted out engine commands: pointer input sent over the control channel
reached the app as synthetic events without moving it, so an agent's clicks, right-clicks and
drags were invisible to a person watching. While the cursor is shown, each synthetic pointer event
now moves it to where the input lands, a press leaves a ripple, and a drag is followed step by
step. A request being acted out keeps the pointer; with the cursor hidden nothing changes.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #5172
* Extrude without profiles keeps nested loops as holes (#5172)
With no `profiles`, solid.extrude took every region of the sketch, so
circles drawn inside a plate outline were extruded as well and filled the
plate's holes. The default now applies even-odd nesting to the drawn
profiles: a loop inside an outline is its hole, a loop inside that hole is
material again, and separate outlines all extrude. Sketches without nesting
keep the `all` selection, and explicit `profiles` (including "all") are
unchanged. Saved designs store their selection, so they recompute as before.
* claim: #5174
* component.create refuses parameters it does not take (#5174)
A call with `bodies` (or a misspelled key) used to succeed and make an empty
component. Unknown keys are now refused before anything changes, the error
lists the accepted parameters, and `bodies` points to component.move_bodies
and component.from_bodies.
* claim: #5175
* Joints: a joint origin's name given as a face says to pass it as the origin (#5175)
When the name in `face` or `circle` matches no face but is a joint origin,
the error now says so instead of "no face is named".
* Viewport: show the picture it drew last when nothing changed (#4820)
A frame redrew the whole scene whenever egui asked for a frame, even when
the camera, scene, highlight and overlays were identical: pointer moves over
empty space and UI animations over a still view paid for the full scene.
The renderer now hashes everything the picture depends on (scene keys,
camera, size, section, images, analysis, accessibility map, occlusion) and
blits the last offscreen picture when the hash matches. The first still
frame filters a supersampled picture down once and later frames copy it as
it is. A frame with the same content costs a blit instead of a scene draw.
Co-Authored-By: Claude Code <noreply@anthropic.com>
* Viewport: draw the scene offscreen, into a 1x window pass (#4820)
The scene keeps the app's depth, stencil and MSAA, but egui's window pass now
has one sample and no depth. Whenever the two targets differ, the scene
(product previews too) is drawn offscreen and blitted, so the UI pass no
longer clears and resolves a multisampled window.
Over 8 s of pointer moves on the 450E rotor head, the app's GPU intervals fell
from 1325 ms to 1006 ms and egui_render from 892 ms to 509 ms; idle stays at
0%. A GPU test checks the split path's picture matches the direct one.
Co-Authored-By: Claude Code <noreply@anthropic.com>
* 3D mouse: a repeated report at rest asks for no frame (#4820)
A device may repeat its latest report while the cap rests. Publishing it
again asked egui for a frame although nothing changed. Only a report that
changes the reading now asks for one; while the cap moves, the view asks for
the next frame itself.
Co-Authored-By: Claude Code <noreply@anthropic.com>
---------
Co-authored-by: Claude Code <noreply@anthropic.com>
The File view's recent row already shows its path on hover; Home added
a second hover text with the raw path, so both appeared.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Add the format registry: one table of formats, extensions and options in crates/io.
The hosts' Open-dialog filters and the web accept list derive from it, and
registry tests pin it to the path predicates and the command docs.
* Fix format-registry tests that compared FormatSpec without Eq.
assert_eq on Option<&FormatSpec> does not compile; the none cases
only need is_none, matching the other rows that compare ids.
* claim: #5230
* Viewport: draw the ground shadow on the direct path (#5230)
The ground shadow went into the model's translucent triangles, so every
frame with a body counted as see-through and took the order-independent
transparency path. Give it its own batch, drawn alpha-blended after the
opaque geometry as before, so only truly see-through bodies use that path.
* claim: #5248
* feat(motion): regenerate kept traces when source inputs change
* feat(motion): add a parametric Chebyshev trace demo
* perf(joints): reuse face measurements across motion samples
Face-centred joint origins called Body::faces on every solve, bypassing the
reference-mesh reuse used for rigid placements. Share face measurements with
unchanged model geometry at the existing tolerance, while guarding replaced
bodies and applying offsets after the cached measurements.
Cover repeated named/nearest-face solves, shape edits, retained snapshots and
direct replacement with deterministic work counts. The 145-point Chebyshev
dimension and range updates retain identical sampled and interpolated curves.
* perf(naming): skip preparation for cached face names
Both face lookup paths spawned fresh workers over every body even when their name cells were ready. Prepare only cold cells, with direct handling of zero or one cold body. Preserve the existing parallel naming path and all face selection semantics.
Deterministic regressions cover repeated picked and unique-name lookups, misses, partial warmup, renames, snapshots and moved copies. Release Chebyshev medians fall from 859 to 50 ms for a dimension edit and 871 to 34 ms for a sweep edit; all 145 samples and interpolated points remain identical. Full cargo xtask ci passes with RUST_TEST_THREADS=4 and the existing local Windows validation overlays.
* claim: #5287
* web-offline-test: wait only for the files the browser can ask for at once
The fail-fast step waited for every other file of build v2 to reach the
server. Chrome sends at most six requests to one server at a time, and
the page's failure cancels the download, with it the files still waiting
for a connection. A real build keeps nine files, so only six were ever
asked for and the Pages job failed with "only 6 of 8 slow files were
asked for" whenever the installing worker didn't happen to ask again.
Wait for as many as can be in flight (less the page's connection), and
give the stand-in app more files than that, so the default run covers it.
raykholo