* claim: #4515
Co-authored-by: Cursor <cursoragent@cursor.com>
* Inspect: keep analysis legends clear of a left History (#4515).
Draft, Accessibility and Curvature legends were painted at the viewport's left edge under a left History sidebar. Place them in the free workspace content instead.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
offscreen_device() asked wgpu for its default adapter and ignored the
SOLIDCRAFT_GPU policy that headless pictures follow, so with
SOLIDCRAFT_GPU=software the ui-egui offscreen tests still opened the
hardware GPU. Under hardware or software the adapter now comes from
solidcraft_gpu::select (new pick_among, which pick uses too); under cpu
there is none; unset (auto) keeps wgpu's default adapter as before.
Part of #9040.
Claude-Session: https://claude.ai/code/session_013WUFq6QFBqmZjxwaKy2J1M
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #9147
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Sketch: the Sketch Grid box in the Sketch panel turns the sketch's grid off (#9147)
The box sets the open sketch's hide_grid, but the viewport drew the grid
whenever a sketch was open, so hide_grid only reached the ground grid and
the box did nothing. An open sketch now shows its grid whatever the Grid
switch says, unless its own Sketch Grid option is off; the window guides
that renders and the control channel use follow the same rule.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* claim: #7521
* fix: headless pictures return their draw counts with the picture
`the_renderer_counts_what_it_draws` drew with `Capturer::rgba` and then
read `Capturer::last_stats()`. The capturer is shared with the binary's
other tests, which draw on it from parallel threads, so another picture
could land between the two calls and the test read its counts (e.g. 10224
triangles for the empty design).
`Capturer::rgba_with_stats` returns the picture and its counts, read while
the call still holds the device. `last_stats()` is removed; this test was
its only caller (the performance overlay reads the viewport renderer's own
stats handle).
New test `pictures_drawn_at_once_each_count_their_own` draws empty and box
designs from eight threads on the shared capturer. Against the old pair it
failed 25 of 40 full runs of the `capture` binary; it passes with the new
call.
* fix: a headless picture waits for its read-back's answer
A picture maps its read-back buffer, waits for its submission with
`device.poll(PollType::Wait)`, and then checked once with `try_recv`
whether the map callback had run. wgpu calls a finished mapping's
callback from whichever thread next maintains the device, in a poll or
in any `queue.submit`, after it lets go of the device. When another
thread got there first, the wait returned with the mapping done but its
callback not yet called, and the picture failed with "not supported yet:
reading a picture back without waiting for the browser", the error meant
for a browser.
On the shared capturer, the other thread is another picture's device
check (`probe_lost` submits nothing before it takes the capturer's lock),
so the `capture` tests failed now and then: `bad_picture_sizes_are_errors`
(its last 8 x 8 picture), `pictures_show_a_plain_cut_without_the_hatch`,
`pictures_blend_a_line_that_is_not_opaque`. In the app, the window's
frames are submitted on the device its thumbnails and `ui.render`
pictures are drawn on.
A native picture now waits for the answer, up to the picture timeout; a
browser still answers later, on its own.
New test `pictures_read_back_while_another_thread_uses_the_device`: 2,000
small pictures while another thread polls their device. Without the fix
it failed 19 of 20 runs (1 to 7 pictures each); with it, 150 of 150 runs
pass. `bad_picture_sizes_are_errors` now prints the error it gets.
* Release 0.27.1
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Render job test: wait for the cancelled job's reservation to be released
The worker drops a job's reservation just after the job turns terminal, so
checking admitted bytes right after wait_timeout raced (10 of 40 local runs
failed). Poll with a deadline, as dielectric_tests does.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Sketch: snap where two recent centre and midpoint guides meet (#8694)
The alignment guide remembered one reference and did not move the cursor. Keep the previous centre or midpoint as well, and snap a drawing point to where their guides cross.
* Sketch: keep the recent-guide crossing snap off until asked (#8694)
The crossing snap was on whenever alignment guides were. Give it its own Sketch preference, off by default, so the guides stay visual until someone turns the snap on.
* Sketch recent snap: ar/he/vi strings and clippy borrow fixes
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #9072
* MCP: --param-check checks calls to the running app (#9072)
`mcp --connect` (and `mcp` once it bridges to an app) took
`--param-check` and forwarded every call unchecked, while the app's own
session is always off. `Remote` now checks each `engine.execute` and
`engine.batch` against this server's command contracts before sending:
`strict` refuses a call, or a batch with such a step, without sending
anything, and `warn` adds `param_warnings` to the app's reply (each
step's result in a batch). A command with no contract here goes to the
app unchecked. `Auto` carries the headless session's level over when it
bridges to a late app. `batch_preflight::contracts` is the contract half
of the existing preflight, for a runner in another process.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
theme::apply always sets egui's theme to light or dark, and egui's own window
sync then told the native window to be light or dark. A macOS window told so
keeps that appearance and no longer reports the system appearance changing,
so with Theme set to System the app stayed light after macOS turned dark. The
app now syncs the window itself: the system's appearance while the theme
follows it, the theme's own otherwise (for the title bar).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #9025
Co-Authored-By: Claude Code <noreply@anthropic.com>
* Sketch: an old edge-on projection keeps its line when it is re-resolved (#9025)
A face seen edge-on was saved as two lines on the same two points, one each way, and is one line now. Re-resolving the link dropped both lines and built a new one under a new id, so an offset that names the first line as its source failed with "an offset source vanished or changed shape". Keep the first line and drop only the second.
Co-Authored-By: Claude Code <noreply@anthropic.com>
---------
Co-authored-by: Claude Code <noreply@anthropic.com>
Fold the pending design format changes (sketch-equation-curves, timeline-joint-folders) into format 44.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The check for a constraint that shrinks a curve to a point took under 1e-6
for a point. One solve could squeeze a 49 mm side to about 1.2e-6 mm and
the next take it the rest of the way, so neither solve tripped it (two
rectangles' corners made coincident, found by a local sketch fuzzer). Under
1e-5 now counts: in 4000 random sketches the constraints that still
collapsed a line went from 9 to 0, and every refusal rate stayed the same.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #9015
Co-authored-by: Cursor <cursoragent@cursor.com>
* Evaluate bare length offsets in the design's units (#9015)
Feature expressions like baseLength - 1 were subtracting millimetres because val() parsed with mm defaults and only applied design units to a final unit-less result.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* claim: #6921
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* CLI: run refuses unknown options and takes --param-check (#6921)
`solidcraft-cli run` now checks its options through `known_args`, so a
mistyped option fails instead of being ignored, and takes
`--param-check off|warn|strict` with the MCP server's words and default
`off`. `warn` prints each unknown key with a suggestion on stderr and
runs the step; `strict` stops at that step and writes nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* CLI: run's option check takes --param-check; warn on timed runs as steps complete
`run` keeps #6919's (#6918) unknown-option check, with `--param-check`
among the options that take a value. `Session::run_script_timed` now
takes the same per-step callback as `run_script_each`, so `run --timings`
or `--profile` at the warn level reports a step's warnings even when a
later step fails, instead of after the script (and not at all on a
failure).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #8644
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Web: open tabs share settings changes, so one tab no longer undoes another's (#8644)
Every open tab of the web app saved its whole copy of the settings to the
same browser storage, every 30 s and on focus loss, so a tab still holding
old settings put them back over a toolbar change made in another tab.
A tab now writes its settings in the frame they change, and the other open
tabs load them from the browser's storage event, so their own saves keep
them. Settings loaded that way aren't written back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* claim: #9016
Co-authored-by: Cursor <cursoragent@cursor.com>
* UI: keep Distance entry under Preferences and Report a bug (#9016)
Canvas value boxes used Foreground order, so they covered system windows. Draw them as Background chrome instead.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* Sketch Trim: a cut at a curve's own end leaves no piece of no length
A curve that ends where another one ends crosses it there only within the
intersection's tolerance (1e-6), but bracket dropped cuts at the curve's
ends only within 1e-9 of its parameter range. A line ending on an arc's end,
trimmed from its other end, was cut a hair inside its own end and kept a
piece about 1e-7 mm long (found by a local trim fuzzer: 78 such trims in
4000 sketches, and 8 breaks). bracket now also drops a cut within that
tolerance of the curve's ends, as add_cut merges cuts that close together.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Sketch Extend: a touch at an arc's growing end is that end, not a stop
An arc whose end is where a line ends that touches the arc's circle there
found the touch a hair beyond the end (within the intersection's tolerance),
stopped on it, and added an arc from that point to itself. A hit within that
tolerance of the growing end (1e-6, as for the far end) is now the end.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A constraint whose only solution squeezes curves to a point (a rectangle's
corner made the midpoint of its opposite side, or two opposite corners made
coincident) collapsed the whole rectangle while the sketch said solved. A
solve that shrinks a line or circle from over 1e-3 to under 1e-6 now counts
as one that fails, as one that throws points far away does: a strict edit is
refused, another keeps the sketch from before its solve, and the free solve
after the held ones in apply_in_order refuses the constraint.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
- Skip drawing and interaction for history chips outside the strip's visible part, after their slot is recorded
- Keep a dragged or focused chip, and every chip while a popup is open
- Frame on a 4,880-body design: 20.3 ms -> 14.6 ms (M4 Max, release, headless)
- Test a_long_timeline_paints_only_the_chips_in_view
Co-authored-by: P'tite Tête <asso@ptitetete.com>
On a Windows account whose %TEMP% is a short (8.3) path, such as
C:\Users\MATHEU~1\AppData\Local\Temp, three tests failed because they built
their folders from std::env::temp_dir(), and the agent file scope refuses short
names ("not allowed for this agent: short (8.3) names such as `PROGRA~1`"):
- engine: cmd::render::settings_tests::embedded_hdr_capture_rejects_parameters_scope_and_cpu_primaries_without_mutation
- ui-egui: agent_chat::tests::file_actions_wait_for_the_user
- ui-egui: agent_chat::tests::permission_modes_decide_what_asks
They now use vfs::temp_dir(), which spells the folder with long names, as
libraries_tests and motion_tests already do (#1385, #1211).
Checked on Windows 11 x64 with the 8.3 TEMP: the agent_chat and
render::settings tests (113) pass, and clippy -D warnings is clean for both
crates.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>