With image mips on (the canvas always sets them), drawing an image
fetched its full decoded pixels first, decoding them again when the
shared image store no longer held them, and only then looked up the mip
it draws. The full pixels are only peeked, so they are the least
recently used entries and the first the store drops over its budget
(1 GiB on macOS and Windows). Once a document's full-size images exceed
it (20 photos of 6000 x 4000 px are 1.9 GB as RGBA), every frame decoded
every image again, though only their small mips were drawn: moving one
image redrew seconds later.
The store now remembers the size each blob's pixels decode to, so the
mip level is found without the pixels, and a mip the store holds is
drawn (and marked used) without fetching or decoding them. Without the
mip, at full size, or with mips off (exports) drawing is unchanged.
Measured on an Apple M4 (load average 15-24 from other work), 20 noise
PNGs of 6000 x 4000 px fitted in a 2880 x 1800 view, one moved 5 pt per
frame: before, 20 decodes and 1.1-3.0 s every frame; after, the first
frame decodes 20 and the later ones none, in 3-5 ms.
The readout was a moving average of 1/dt over any two consecutive frames.
An idle window wakes now and then (a timer, a control request, a repaint
after input) and each wake-up draws two frames about 4 ms apart, the second
one not waiting for vsync because the swapchain has a free image. Every
wake-up pushed the average up by 0.1 x 240 and the idle gap pulled it down
by only 0.1 x 2, so at rest it climbed past the display's refresh rate
(105 fps on a 30 Hz monitor that was drawing 4 frames a second). As soon as
the pointer moved and frames became vsync-paced, it fell back to the real
rate, which looked like a slowdown.
- frame_rate: frames counted over the last second of continuous drawing.
A gap more than 100 ms longer than the previous frame's own work means
the window was idle and starts a new count. The readout changes only
after half a second of drawing, so a wake-up's frames keep the last
rate. Slow frames still read slow, because their gap is their own work.
- chrome: "– fps" until the window has drawn continuously, instead of 0.
Measured on an Apple M4 driving a 30 Hz 4K display: drawing continuously
reads 30; at rest with a wake-up every 0.5 s it stays 30 (before: 45 -> 105).
In a clip group or a clipping layer the bottom child is the clipping path.
Object › Arrange moved the selected children with no regard for that slot:
Send to Back (and Send Backward from just above it) put the art under the
clipping path, so the art became the clipping path, lost its role as art,
and the group drew nothing. Bring Forward / to Front on the clipping path
itself would likewise have made art the clip. Layers-panel moves already
keep the clipping path at the bottom (layer.rs, `index.max(1)`); arrange
didn't.
- object.rs `arrange`: in a parent that clips, the clipping path is left
where it is and the art is arranged within the slots above it (Back goes
to just above the clipping path, Backward stops there).
- tests: Back, Backward, several objects, and Forward/Front on the
clipping path in a clip group (with a render check); Send to Back in a
clipping layer.
Include guide visibility in effect raster cache inputs, skip hidden guide paths in outline members, and clear selected guide paths when hiding guides. Add regression coverage for cached effects and selection overlays.\n\nThe cached-effect regression test is adapted from the overlapping #3264 implementation.\n\nCo-authored-by: Mahtab <smmalam9@gmail.com>
Address CodeRabbit review on #3307: the new `Seq::Erase` arm discarded
the `tool_key` Result, dropping an erase error silently, whereas the
old `route_type_input`/`text_key` path reported it via `app.status`.
Mirror that behaviour so a failed Backspace/Delete still surfaces in
the status bar. Behaviour otherwise unchanged; fmt/clippy clean and the
three IME tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Hook IMEs recompose by injecting Backspace key events and VK_PACKET
characters, which egui surfaces as interleaved Event::Text and
Event::Key(Backspace) in one frame. type_text handled text/IME in one
pass and left Backspace to a later pass, so within a frame all text was
inserted before any delete and the diacritics were eaten (tiếng -> tieng).
Take Text, IME and plain Backspace/Delete in a single event-ordered pass,
removing handled keys so route_type_input does not repeat them. Modified
keys and composing frames are unchanged. Adds two canvas IME tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
craft-docpath follows a stored path only on this computer or on the
share the document is on, but its check is lexical and cannot see
mapped drives. fileio::document_path therefore followed a link stored
as Y:\img\logo.png, where Y: is mapped to \\server\other, from any
document: opening the document, links.check, Link Info and the
linked-file watch read or stamped the file on that server.
On Windows, document_path now reads the final path of the root folder
of a link's drive (std::fs::canonicalize). Windows opens that root
folder on the drive's server to answer, so the check still contacts
the server once per lookup; no file on it is read, stamped or watched.
When the final path names a share, the link is checked against the
document's path on its share, and a link on another share is refused
as on another computer, as a link stored with that share's own path
is. Links on this computer's drives and on the document's share are
followed as before, and a link on the document's own drive is not
looked up. A document opened as \\?\Z:\... is looked up on its share
as Z:\... is. Other systems are unchanged.
Docs: docs/mcp.md, docs/target-app-parity.md.
Preview in Browser wrote each preview into `VectorCraft Save for Web`,
and before each preview it removed every file and folder there that
was an hour old or older. Where the temporary folder is the shared
/tmp, another user could make that folder first, writable by everyone,
and the app wrote its previews inside it, in folders that user could
read and replace. Where the system follows other users' links in /tmp,
the folder could also be a link, and the app removed the old entries
of the folder it named.
A preview now uses that folder only when it is a folder and not a
link, and on Linux, macOS and FreeBSD only when the current user owns
it and no other user can write to it. On those systems the folder and
each preview's folder are made with mode 0700. The cleanup removes
only preview folders, on those systems only those the current user
owns. When the shared folder can't be used, the preview goes to a new
folder in the temporary folder. A later preview removes such folders
once they are an hour old.
Docs: docs/target-app-parity.md.
The version 7 .ai writer wrote the %_Bs color stops of every gradient
from ramp point 100 down to 0. The Adobe Illustrator File Format
Specification version 7.0 lists a linear gradient's stops in that order
(section 1.4.3). Its radial definition, "White & Purple Radial", lists
them from 10 to 100 (section 3.5.2). As reported on #2807, the
reference app reads both stops of #2687's radial gradient, written from
100 down to 0, at ramp point 100 and draws the circle in #1b2a6b. It
reads a radial gradient with three stops written that way in reverse
order.
The writer now lists a radial gradient's stops from the first to the
last. Each stop is written with its own midpoint to the next stop, and
section 3.5.2 says the last stop's midpoint is ignored. Linear
gradients are unchanged.
The Offset Path dialog remembers the values it was last confirmed
with (#1152), but the Preview checkbox was not among the kept
fields, so every open started unchecked and Preview had to be
ticked manually each time. Preview now rides along with offset,
joins and miterLimit in the fields kept on OK, as Simplify's
retained fields already keep theirs.
Unchanged: a very first open (nothing kept yet) still starts
unchecked, and Cancel keeps nothing.
Adds a regression test beside offset_path_remembers_its_last_values:
open, tick Preview, confirm, reopen - the checkbox is still on.
Verified by reading the flow on current main (menus.rs opens on
app.ui.offset_path when it is an object, d.bool("preview") defaults
to false when the field is absent, form::preview_result ticks the
checkbox into d.fields) and by tracing the existing #1152 test;
no Rust toolchain on this machine, so the test was not executed
here - saying so plainly for the reviewer.
Co-authored-by: rahul05ranjan <11793455+rahul05ranjan@users.noreply.github.com>
The desktop app reads the User Defined swatch and graphic style
libraries while it creates its window: set_user_dir runs
Libraries::rescan, which read and parsed every file in the folder
outside the panic guard that commands have. A parser bug that panicked
on one file stopped the app before its window appeared, at every
launch while the file stayed in the folder.
rescan now reads and parses each file inside guard::catch_panic and
skips a file whose reading panics, as it skips one that can't be read.
The other libraries are listed.
Found while testing crates/color/src/palette_io.rs:373
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Found while testing crates/tools/src/xform/measure.rs:20 (shown at :66, reported at :80)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Found while testing crates/doc/src/xml_id.rs:80 (strip_unique), triggered by from_object_name at :32
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
One pattern fill tiles at most 250,000 times (PatternDef::offsets_covering),
but Scene::fill writes each tile's own pattern fills again, down to the
nesting limit of 8, with nothing bounding the total. A pattern whose tile
art is filled with another pattern (or with itself) multiplies the tile
count at every level: a small document makes EMF export (the engine's
"emf" format, also reachable over MCP and the control channel) run out
of memory or never finish, and the unsaved document is lost with it.
The scene now keeps one tile budget for the whole walk. A fill whose
tiles don't fit in what is left, and every pattern fill after it, is
written as one colour like WMF writes patterns, with a warning.
The test tiles squares filled with dots (over half a million fills before)
and a pattern filled with itself, and checks both stay under the budget.
Reported-by: BS.ehlas.cz
One Repeat makes at most 250,000 instances (a 500 x 500 grid; radial
counts stop at 1,000), but its source may hold another Repeat and every
instance draws that one again: expand() clones the inner Repeat once per
instance, and the renderer, the exporters (expand_deep) and hit testing
evaluate each copy. Nothing bounds the product, so a document under 1 KB
with a 500 x 500 grid of a 500 x 500 grid asks for 62.5 billion
instances: opening it freezes the app or runs it out of memory, and the
unsaved work in other documents is lost with it.
A Repeat now counts the instances the Repeats nested in its source make
in all and keeps only as many of its own as fit in 250,000, the budget
pattern fills already keep (at least one, so the source still shows).
A Repeat with nothing nested keeps all its instances, as before.
The test nests a 500 x 500 grid in a 500 x 500 grid and in a mirror,
counts the instances from the transforms (62.5 billion before) and
checks they stay within the budget, while one grid and a small nest keep
all of theirs.
Reported-by: BS.ehlas.cz
GradientMesh::remove_point_lines (the Mesh tool's Alt-click,
object.mesh.deletePoint) only checked that the clicked point exists,
then indexed the points above and below or left and right of it by the
mesh's rows and columns. A gradient mesh read from a document is not
checked against its rows and columns, so one with fewer points than
they need (rows 2, columns 1 and 3 points instead of 6) indexes past
the end and panics: the engine guard rolls the command back with an
internal error, and the web app, which has no guard, aborts.
It now returns false for a mesh that isn't valid, as locate (and with
it adding mesh lines), the outline and the mesh lines already do, and
the command reports that nothing could be deleted.
The test Alt-clicks an interior row's point and an interior column's
point of two such meshes and checks nothing is deleted and the mesh is
unchanged.
Reported-by: BS.ehlas.cz
GradientMesh::is_valid and the two mesh envelope checks (envelope_grid
and the envelope's surface) compare the number of points with
(rows + 1) × (cols + 1), multiplying the document's u32 rows and
columns unchecked. With u32::MAX rows and columns the product is 2^64:
debug builds panic on the overflow and release builds wrap it to 0, so
a mesh or a mesh envelope without points passes as valid. Selecting
it, Outline view, hit testing and export then index points that aren't
there (the mesh lines, outline and edges), tessellating it asks for an
impossible capacity, and a mesh envelope with handles to fill in loops
2^64 times making them until memory runs out. On wasm32, where usize
has 32 bits, 65,535 rows and columns are enough, and a panic there
aborts the web app.
The three checks now share grid_len, which counts the points with
checked arithmetic and gives None when the count doesn't fit in a
usize: a grid that can't be counted matches no point list, so it is
malformed like any other wrong count.
The test checks that a u32::MAX × u32::MAX mesh without points isn't
valid and has no lines, outline or quads, that envelope_grid refuses
those counts, and that a mesh envelope with them leaves art in place.
Reported-by: BS.ehlas.cz
A live-effect plug-in was given only its object's geometry. An effect
that draws as wide as the stroke it sits on, or tells its fills apart,
had to ask for them again as parameters, which then went stale when
the paint changed.
A live effect's input object now holds the object's fills and strokes
too, read only and in the form a filter gets them (one encoder,
objects::paint_lists, for both), and _context.item says where the
effect sits: "object", or {"fill": i} / {"stroke": i} for an effect in
a fill's or a stroke's own effects. effects::GeomContext carries the
object's appearance and the item, and the plug-in cache keys on the
paints, so a change of paint runs the plug-in again. Hosts from before
this give geometry only; nothing in the output changes.
Expand worked a fill's geometry effects out in the object's context, so
a plug-in effect there would be told "object" and a gradient fill was
expanded into a shape the canvas doesn't draw: fill_shape now gives them
the fill's context, as drawing does (expand.rs).
Siblings named with a declared prefix, the other half of such context,
are left out: the effected geometry is worked out from the object alone
wherever it is drawn, measured, hit-tested or expanded.
wat::when_given is a plug-in whose output depends on whether its input
or parameters hold a text. tests_plugins.rs uses it to check that an
effect sees the stroke's width and the fill, runs again for a thinner
stroke, and on a second stroke is told so; effect.rs checks the item
counting; a gradient fill under such an effect is drawn and expanded
into the same triangle. CRAFT_PLUGIN_PAINTS_WASM=<folder> saves a test
effect to try in the app.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HhvnnnD5fTFupK8WWq35un
A plug-in's dialog labelled each field with its parameter's name spaced
out ("Line_spacing_mm:"), showed numbers without their unit and listed
at most 64 parameters in one column. A plug-in with many settings, such
as a generator's or an exporter's, could neither explain nor sort them,
and its dialog showed settings that don't apply to the choices made.
Each parameter may now say how dialogs show it, in keys that hosts
from before them ignore (plugins::dialog):
- label; unit, after the value ("0.25 mm"); help, on the label; and
helpUrl, an https:// page opened from a button after the field.
- group: the parameters outside any group come first, then each group
as a section that folds, in the order the groups first appear, the
fields lined up down the dialog.
- visibleWhen: {"pattern": ["lines"]} shows a parameter only
while each checkbox or choice it names, declared before it, has one
of the values listed and is shown itself. Manifest::visible decides
every parameter in one pass, in declaration order. A hidden parameter
still goes to the plug-in with its value, so a run never depends on
what the dialog showed.
Two types join number, int, bool and choice: string, a text field cut
to maxLength, and numberList, numbers typed separated by commas or
spaces, each clamped, minItems to maxItems of them. A plug-in declares
up to 256 parameters, at most 64 outside groups. The filter, live-effect
and exporter dialogs use every key. The Appearance panel's editor of a
plug-in effect uses the labels, units, help and conditions, and reads a
typed number list as the dialog does.
dialog.rs tests parsing, the round trip of a manifest written out,
sections, visibility (also of 256 overlapping conditions, which a
recursive check would take forever on) and the keys refused;
manifest.rs the new types and limits. tests_plugin_fields.rs draws a
dialog with kittest: labels and units, sections in order with their
fields lined up, visibleWhen hiding and showing rows, and the help
button opening its page.
appearance.rs checks the editor's labels, units, conditions and a
typed number list, and form.rs units after numbers (a unit starting
with a digit is spaced from it) and typed lists.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HhvnnnD5fTFupK8WWq35un
# Conflicts:
# docs/gaps.md
# docs/target-app-parity.md
A plug-in could filter objects or reshape them as a live effect, but
nothing it made could leave the app: pen plotters, cutting machines and
laser cutters read formats VectorCraft doesn't write. An exporter is a
third kind of plug-in ("kind": "exporter", with the file "extensions"
it writes), in the same sandbox and with the same ABI v1 exports.
- It adds a write-only format, plugin.<id>. document.formats lists it
with its plug-in id, and document.export writes it by that id, or by
a path ending in one of its extensions that no built-in format has;
"params" carries its parameters, checked against its manifest as a
filter's are.
- File › Export › Export As lists it after the built-in formats, with
Use Artboards greyed out, since the plug-in writes the whole
document. One with parameters asks for them in a dialog generated
from its schema (pluginExport) once the file is picked.
- The plug-in is given the native file's JSON (format::json_part:
images of 8 KiB or more without their data) and _context {mode:
"export", extension, name}. It returns {"dataBase64", "warnings"?}
or {"error"}. The host writes the file, and nothing when the plug-in
declines or returns bad data.
- An export gets 10 G instructions plus 4,000 per byte and 300 s.
Background Export's Cancel stops it between two slices of that
budget (Error::Cancelled), and a cancelled live-effect run is not
cached as a failure.
- plugin.run and Object › Plug-ins say where an exporter is instead of
calling it a live effect.
engine/tests/plugin_export.rs reads back what an exporter is handed
with wat::echo_exporter, an exporter whose file is its input or its
parameters (its base64 encoder is written in WAT). The document it is
given is document.save's JSON, without the large image's data, and its
parameters arrive defaulted, clamped and with the file's name and
extension. It also covers listing and removal, built-in extensions
winning, selectedOnly, declines, bad base64, missing data, warnings,
and Cancel stopping a plug-in that never returns. It is a test binary
of its own: installed plug-ins are process-wide, and the format-list
unit tests check exact lists. dialogs/tests_plugin_export.rs drives
Export As headlessly; plugins/tests/abi.rs covers the budget, the
parameters and the echo exporter; format/tests/native_v4.rs covers
json_part.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HhvnnnD5fTFupK8WWq35un
With Include SVG Code on, Copy offered the SVG markup as plain text with every placed image
embedded as base64: a 13 MB photo made 17 MB of text. A paste reads the text first (eframe on
every paste key; on Linux and macOS it is the one format written, and holds_ours reads it again),
so each paste moved it through the system clipboard on the UI thread, and on GNOME the app
stopped responding.
The SVG markup now goes out as text only while the images it embeds total at most 1 MiB; above
that it is offered as image/svg+xml and PNG. Type keeps its text, and vector art keeps its SVG
text at any size (only embedded image bytes count).
Windows 11, Ryzen 5 2600X, 13 MB PNG, real Ctrl+C: the longest UI stall fell from 823-1021 ms
to 133 ms. Not run on Linux.
Match the Selection tool: Shift on an already selected hit removes it on
release without a drag, including one path of a compound path.
Co-authored-by: Cursor <cursoragent@cursor.com>
Export Selection… and the Asset Export panel open Export for Screens on the
Assets tab with the new assets checked, but new assets land at the end of
the list, so in a document with many assets the checked one sat below the
list's visible part and had to be scrolled to by hand. The list now scrolls
to the first checked asset when the dialog opens on them. It keeps asking
until the row is in view (the dialog's first frame only measures it), then
stops, so the list scrolls freely afterwards.
Squashed from #2533 onto #2525 (Smudge Stick, Sponge, Watercolor): both sets in PIXEL_EFFECTS (45), the catalogs and the docs, whose counts now read 45/57.
File › Export Selection… collected the selected objects one asset per
object, so a few ungrouped shapes picked to export together came out as
separate images unless they were grouped first. It now collects them as one
asset of them all (`assets.add {multiple: false}`), the selection exported
as one image; `file.exportSelection {multiple: true}` still makes one per
object, and Object › Collect for Export › As Multiple Assets and the Asset
Export panel's Multiple Assets item are unchanged.
Store the click as __anchorX/__anchorY and recompute x/y when the locator opens or changes so the chosen point stays on the click like Illustrator.
Co-authored-by: Cursor <cursoragent@cursor.com>
Taha-Mahmoodi