* claim: #6755
* Frame rate readout stays beside a right-side pane, not under it (#6755)
The readout and its details sat in the viewport's bottom-right corner even under the
floating Assistant. They now move just left of any pane over that corner and come back
when it closes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Every view attached to an API-agent message stayed in the history as a
base64 data URL, and the whole history goes with every request (up to 24
per message). Requests grew by one screenshot per message until the
endpoint refused them with HTTP 413, after which every message failed.
Keep the newest two views as pictures and swap older ones for a short note;
the text sent with them and every other message stay. HTTP 413 now says the
conversation is too large for the endpoint and to start a new chat.
Fixes#2298
Main now speaks the Anthropic protocol too, where a view is an `image`
block rather than an OpenAI `image_url` part: drop_old_views trims both,
before trim_history. ui.ai_http_too_large is in every catalog that has the
other ai_http_* messages.
Co-authored-by: Framewrk CI <ci@framewrklabs.ai>
* claim: #6396
* Sketch: a regular line on a construction line is picked before it (#6396)
Coincident sketch curves tied on pick distance and the one drawn first
won, so a construction line drawn first hid the regular line on part of
it. Construction and centre lines now get half a screen point added to
their pick distance: on a tie the regular curve is hovered and picked,
and a construction line clearly nearer the cursor still wins. Clicking
again and Select Other still reach the construction line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FN5eEMrqWWXqi65FcK8GXq
---------
Co-authored-by: sascha2101 <221609842+sascha2101@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
wgpu's default uncaptured-error handler panics ("wgpu error: Out of Memory").
The headless picture devices already report instead; the window's device had
no handler. Errors the code expects stay under error scopes.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
A line tangent to an arc at the arc's end had no sense: the tangency held
with the line on either side of its touch point, so pinning a line's far
end across the corner left the line cutting back through its fillet, and
when the solver found the mirrored corner the arc kept its stored
direction and became its three-quarter complement, inside out.
The solver now keeps each fillet an arc of its corner. A line tangent at an
arc's end keeps its sense there (it leaves the arc as it did) and the arc
its side of its chord, compared with the positions before an interactive
move or the stored ones. A corner that mirrored gets its arc on the
mirrored side: its ends swapped where the solve ran the arc round the
other way with the lines following, or the arc reflected across the line
that stayed and solved again where a line cut back through. A solve that
would loop the arc the other way round, or fold a line onto it, is refused
as a fold is. A drag's cursor-following steps keep the fillets too, so a
line dragged through its own fillet stops at the arc.
The kit slider adds a real `egui::Slider`, lets egui paint it, then paints an
opaque rectangle over it and draws the kit's track and thumb on top. That mask
was exactly `slider_width` wide, and egui places its handle so the circle's
right edge lands on `rect.right()` at the maximum, so the fill stopped flush
with the handle and its last column survived as a light sliver beside the
value box. The kit's own thumb cannot hide it either: its radius is
`t.space_8` while egui's handle radius is `input_radius`, about two pixels
larger.
Extend the mask one pixel past its right edge. The gap before the editor is
`item_spacing.x`, so nothing else is covered, and only the maximum was
affected: at the minimum the handle's left edge coincides with the mask's left
edge, which a fill does cover.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* claim: #6524
* Sketch display: each sketch samples within its own budget (#6524)
After a large DXF was inserted, lines drawn in a later sketch vanished
once finished: only their end points showed and could be picked, and
hiding the DXF sketch brought them back.
The display cache had one sampling budget for every shown sketch (4096
curves, 131 072 vertices per model state and view). Sketches are drawn
and picked in timeline order, so a large imported drawing used it all
and every later curve came back empty (BudgetLimited): not drawn, not
pickable, while the points, drawn separately, stayed.
Each sketch now has its own budget with the same limits (curves,
profiles, vertices and nodes). The cache as a whole is capped at 65 536
curves and 16 384 profiles, so memory stays bounded however many
sketches are shown.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FN5eEMrqWWXqi65FcK8GXq
---------
Co-authored-by: sascha2101 <221609842+sascha2101@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #6657
* Context menu: the Appearance flyout closes when the pointer moves to another row (#6657)
Hovering Appearance opened its quick-choice flyout, but resting on a plain
row of the same menu never closed it: only another submenu row replaced it.
Now a flyout closes once the pointer moves onto another row and rests there
for 0.3 s, so a diagonal path to the flyout across a neighbour keeps it open.
A flyout opened with the keys stays open while the pointer is still.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A feature's face points are moved onto their named faces before it is
evaluated, so a To Object extrude follows its target face when an earlier
feature moves it. A pattern's or mirror's copies were built from the
feature as stored, with the face where it was picked: after a gear's disc
was made thicker, the tooth reached the new top but its pattern copies
kept the old height. The copies now use the same moved points; the move
is one helper used in both places.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Test: sketch Intersect keeps its exact cuts, crosses curved faces on the surface, cuts on a top face
Intersect's cut of a box, a cylinder, a sphere and a tilted cylinder, as it is on main, pinned.
Two cases where it is not exact: a plane through a cylinder's rim cuts the mesh's chord (its sides
come out at 6.3159 and 6.3085, not the square root of 40; a hole's walls at 25.0122), and a plane
on a body's top face finds nothing ("projects to nothing on this sketch plane").
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PsiqyXWBFGTCX9J51bhysc
* Sketch Intersect: crossings on curved faces lie on the surface; a plane on a top face cuts it
Problem: the plane cut that sketch Intersect makes of a body's mesh (project::section) took a
crossing on an edge where a curved face meets another (a cylinder's rim on its cap) from the
merged normals, which are undefined at the crease, so it fell on the chord: a cylinder's sides came
out 0.01 mm off and uneven. A plane through a vertex or on a face counted the vertices on it as
below, so a plane on a body's top face crossed no edge and found nothing.
Solution: the crossings are found by section_segments, which section fits as before: an edge where
a curved face meets another takes the curved face's normals, so the crossing lies on its surface;
a vertex within 1e-9 x the mesh's size of the plane is on it, on one side for every edge (below
when the mesh lies on or above the plane, else above), and an edge ending there crosses at that
vertex, which is an exact point for the fit. Each segment keeps its B-rep face, for a section that
groups them by face. Cuts that were exact are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PsiqyXWBFGTCX9J51bhysc
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The first test in each GPU path shard compiles the path pipeline cold
on lavapipe (335-502 s on 16 vCPU, 665 s on 8, against the 600 s setup
budget); the later ones read Mesa's disk cache in about 30 s. Put that
cache on a sticky disk that main commits, so the first test reads it
too until the shaders change.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A change computed off the window thread ran here with only the results of the evaluation
ahead's last evaluate primed. A parameter change evaluates twice (before and after its sketch
stores the new solution), so its first evaluate here found nothing to take, and on a design whose
states pass the evaluation cache's budget it recomputed the features on the window thread: the
window froze for seconds after Computing ended. The run now also takes what the evaluation ahead
kept (KeptResults, as an agent's control requests already do).
The catalogue entry Quantity served both the pattern dialogs' number of
instances and the motion study's plotted quantity. Nine locales had
translated the second meaning, so the rectangular pattern dialog showed
e.g. "Größe" above "Anzahl 2" in German.
The motion study's row now reads motion.plot_quantity (a pending file
that keeps every locale's previous word there). Quantity takes each
catalogue's own ui.quantity text, which the "Quantity 2" row already
shows; nothing is newly translated. The glossary lists Quantity under
words with more than one meaning.
Fixes#6538
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
On 8 vCPU the first merge-queue run after #7066 failed: one gpu path
test went past the 600 s pipeline-setup budget (665 s; 204 s for the
same test on the PR run) and shard 2 took 35 minutes. lavapipe compiles
on every core, so the shards move to 16 vCPU, and four count partitions
(27 tests each) replace three hash ones (41/40/27).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The 108 gpu path:: tests run one at a time (gpu-path test group) and took
about 36 of Native Linux tests' 45 minutes in a merge-queue run. They now
run in a GPU path tests job, three hash partitions on 8 vCPU, beside the
other tests; Native Linux tests leaves exactly them out. A `gpu` output of
Select checks drives both the job's `if` and the `ci` check.
The Windows jobs now run in a merge group only when its diff reaches what
they exercise (the pull request patterns and the Cargo.lock graphics-stack
check), not on pushes to main, and still nightly and by hand.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The README and docs/downloads.md ship a universal macOS disk image, but the
contributing guide listed only Windows and Linux.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Docs: on Windows, give claude mcp add the CLI's full path (#5895)
The Windows installer puts solidcraft-cli.exe in C:\Program Files\SolidCraft but not on PATH,
so the README's and docs/mcp.md's `claude mcp add solidcraft -- solidcraft-cli mcp` registers a
server that can't start there. Say so next to the command and give the full-path form.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Docs: word the Windows full-path note so it holds once the installer sets PATH (#5895)
#5898 makes the MSI add its folder to PATH. Say "if a terminal doesn't find solidcraft-cli
(older Windows installers don't put it on PATH)" instead of stating that the installer never
does, so the note stays true whichever change lands first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Docs: the Windows CLI path is the default, not the only place (#5895)
The MSI installs to ProgramFiles6432Folder and lets people choose the folder
(WixUI_InstallDir), so C:\Program Files\SolidCraft\ is only the x64 installer's default; the
x86 MSI on 64-bit Windows lands in C:\Program Files (x86)\SolidCraft\. Say "by default" in
the README and docs/mcp.md (review on #5896).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
SolidCraft runs through XWayland whenever an X display is available. A
compositor that doesn't let X11 apps scale themselves draws the window at
100 % and stretches it. docs/nix.md said so only for fractional scaling,
but GNOME 47 (RHEL and Rocky Linux 10) does it at any scale: the report
shows 1.05 physical pixels per point at 200 %.
The troubleshooting page gets a section with the cause, the Window system
preference, SOLIDCRAFT_WAYLAND=1 for 0.21.0 and earlier, and the file drop
trade-off; docs/nix.md links to it.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The component colour tests no longer assume a design always carries a list of
pending format changes; once the release folds them, there is none.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #4214
* Viewport: the grid fades with the distance instead of cutting off (#4214)
The ground grid stopped at a fixed distance from the view, so its edge showed
as a cut across the picture. It now fades out with the distance from the
middle of the view, and it is gone before either of the two lines that used to
cut it: its own edge, and the picture's far plane.
Each grid line is drawn in pieces, and each piece is as faint as the fade
makes it at the distance where it lies. The 'under' batch is drawn by the
overlay pipeline with alpha blending, so the colour's alpha reaches the
frame.
* The headless grid fades with the distance too (#4214)
The headless path (solidcraft-cli snapshot, the home-page thumbnail, the MCP
screenshot) drew the grid from the engine's palette, where a line had no
alpha: the grid still ended in a cut, and at a low camera angle the cut was
in the picture.
A SceneLine now carries an alpha, the CPU rasterizer blends with it and the
GPU capture passes it on, so the engine's grid can fade. The fade itself
moves to the engine, next to the grid's other geometry, and both paths use
it: the headless grid now reaches as far on the ground as the view is wide
(the ground stretches as it is seen at a slant) and dissolves into the
distance instead of ending in a line.
* Fix the headless/viewport grid fade divergence (#4214)
The engine stretched the grid's size with the slant of the plane it is seen at, the viewport stretched it only on the ground, so on a sketch's plane the picture's grid reached further than the window's (153.42 mm against 120.00 mm). Both now share one grid_ext, which the picture and the window draw the same grid from.
* The grid fades out before the picture's far plane cuts the ground (#4214)
The window's grid faded against a far plane of its own: the scene's radius
clipped by the projection's camera (260 mm away at 400 mm of distance), while
the picture itself cuts the ground at the far plane of its own projection
(223.7 mm at the same distance, because the projection frames the scene's
bounds and not its radius). The grid was still a third of its strength where
the picture cut the ground, so the cut showed as a straight line across the
view.
The fade now ends at the far plane the picture's projection uses, and the
grid is gone before that plane cuts the ground.
The test that watched the fade now measures the ink of each row of the
picture, and checks that the grid is already gone in the ten rows above the
row where the far plane cuts the ground: the point it compared before lay
beyond that plane, where both renders draw nothing at all, so the comparison
was always true.
* claim: #5734
* Docs: don't offer winget install before the package is in winget-pkgs (#5734)
The README and docs/downloads.md said Windows users could run
`winget install bherbruck.SolidCraft`, but winget-pkgs has no manifests for
bherbruck.SolidCraft (manifests/b/bherbruck doesn't exist), so winget finds
no package. winget.yml only updates a package that is already listed, and
does nothing until then. Drop the command from both pages, and say in
docs/releasing.md that the first version has to be submitted once by hand
and the command added back when it's listed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Releasing: a first WinGet submission is already open; take it up (#5734)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ui): finish the open sketch when a modeling feature starts
* fix(ui): wait for the marker roll before starting a feature from a sketch
* fix(ui): start the feature once the marker is back
* test(ui): keep the sketch open in the palette clearance test
Saving into Documents fails with the Windows dialog's "File not found" when
Controlled folder access blocks the app, and the message gives no hint.
Refs #4479.
* claim: #4640
* Docs: AppImageLauncher 2.x can't start the AppImage; update to 3.0 (#4640)
AppImageLauncher 2.x intercepts every AppImage through binfmt_misc and
can't start static-runtime AppImages such as SolidCraft's: it fails with
"execv error: No such file or directory" after a FUSE message. 3.0 starts
it. One paragraph in the Linux AppImage section says so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Docs: AppImageLauncher 3.0 is so far only published as betas (#4640)
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
GitHub Pages reports the base URL as http:// for the custom domain, so the
.flatpakrepo file, the remote-add line on the repository page and the mirror
of the previous repository all used plain HTTP. Force https:// once where the
base URL is read.
* docs: two notes for agents and contributors (the gate's working tree, and linking a picture)
* docs: how to show a picture without attaching it
The reviewer asked for the pictures to be attached to the PR (or in a gist),
not hosted on a branch of a fork, so the note no longer teaches the branch
trick.
* docs: say how to attach a picture, and what a gist is for
The review of this pull request pointed at two things in the note on
evidence:
- Attaching works in the browser for anyone who can comment: drag the
file into the comment. The limit is the command-line agent, whose
upload needs write access to this repository. Say that, instead of
claiming that a contributor without write access cannot attach a file.
- A gist suits a log or another text file; it does not render a picture
inline. Say what it is for, so nobody hits that dead end.
The note in AGENTS.md about the gate reading the working tree is
unchanged.
* claim: #4851
* User guide: local models with the Assistant (#4851)
A page on choosing and setting up a local model (Ollama, LM Studio, or any
OpenAI-compatible server) for the Assistant: tool support, VRAM, permissions,
the measured results from the standard test in #4851, and troubleshooting
for the common failures.
* Guide: local models table with the full set of measured results
Point to the `<file>.bak` copy every save keeps, the design.solidcraft in crash
reports, the recovery folder for unsaved designs, and how to attach the file
to an issue.