Shift did nothing with the Pencil: the stroke stayed freehand, while the
Line tool and the Pen keep theirs to 45° steps. pencil_add, shared by the
drag and the release, now keeps one straight segment from where the
stroke started to the pointer at the nearest 45° step while Shift is
held (the Marker and Highlighter too); a Shift-click without a move still
draws nothing.
#2468 moved SVG import onto craft_svg::tree::parse, which runs the size,
entity, attribute and nesting checks itself, but left FigCraft's own
scanners (too_wide, too_deep, nesting) and import/budget.rs unused and
import.rs unformatted: main failed cargo fmt, clippy -D warnings and the
svg tests, which still passed budget::Limits to import_within.
The dead code goes; budget_tests use craft-svg's tree::Limits (bytes,
embedded_bytes, embedded_depth) and its error texts, and hostile_tests
check attribute width with craft_svg::guard::too_wide at the default
limits and the nesting refusal at craft-svg's 256 levels.
Publishing keeps the variables its styles bind and where its copies of other libraries' assets came
from; importing files such a copy under that library when the file links it.
* Dependencies: take craft-libs main at 3af2280ee3
Every craft-libs crate in the workspace now comes from one commit,
3af2280ee3434a2c860012df2f440652621c2c81. They were on five revisions:
e11f9d9a, d13ac58a, 976eb51c, e9829954 (craft-update) and fd5e2fef
(craft-web-api and craft-web-host). Cargo.lock now holds one craft-zip
instead of two.
No FigCraft code changes. The new craft-font-store adds cap-primitives
and its dependencies to Cargo.lock.
* Dependencies: drop pin comments the craft-libs move made untrue
craft-update now uses the same craft-libs rev as the other crates, and
craft-window comes from craft-libs main. Remove the comment that says
craft-update is pinned separately and the comment that says craft-window
is pinned to an unmerged branch.
* Tests: look up craft-fig with cargo metadata filtered to the host
The i18n key test finds the craft-fig sources with `cargo metadata
--offline`. Since craft-font-store depends on cap-primitives, the graph
holds winx, which only Windows uses. Without a platform filter, cargo
metadata needs every platform's sources and failed on Linux CI with
"failed to download `winx v0.36.4`". The query now passes
`--filter-platform` with the host triple from `rustc -vV`.
* Dependencies: take craft-libs main at 453b0da782
Every craft-libs crate in the workspace moves from 3af2280ee3 to
craft-libs main at 453b0da782, and craft-sketch moves from ff142b3ad5
to the same commit. Cargo.lock now holds one craft-figma-rest and one
craft-zip instead of two of each; no other package changes.
* Dependencies: drop the design-file pin comment that no longer holds
craft-penpot, listed above the design-file crates, uses the same
craft-zip as craft-fig and craft-figcraft-format, and every craft-libs
crate is now on one rev, so the comment's reason for a separate rev is
removed.
Without a device, a frame taller than the window scrolled its content and then panned whole, taking
fixed and sticky layers off the top. A frame with them now takes the window as its screen.
From review: the asset list follows the layers placed in the tree, so a
smaller tree (maxNodes, or the MCP tool's retry for maxBytes) means a smaller
list, and `truncated.assetsLeftOut` counts the rest; following aliases reports
`truncated.variables` when it stops at its budget with variables still to
read; and a font is checked with figcraft_text::has_family by its trimmed name,
as the missing-fonts list does, so a family stored with outer spaces isn't
taken for installed. A regression test for each fails without its fix.
From review: the variables a style binds and those an alias or colour
calculation reads are joined too, followed to any depth; fonts come from the
text layers placed in the tree, so a second target's font isn't lost to a
separate walk's budget; and a malformed `id` or `ids` entry is an error
instead of reading the selection or dropping the entry. A regression test for
each fails without its fix.
inspect.designContext and the MCP tool design_context read the selection (or
ids, or a scope) in one call. Each layer keeps apart the values the designer
set (explicit), those layout or the file worked out (derived) and those with no
direct CSS equivalent (unsupported, with the reason). The variables (every
mode's stored and resolved value), styles, components, fonts and assets the
tree uses are joined once, and `missing` says what the design doesn't specify.
A connected instance gives its code snippet instead of its layers. Large trees
are capped and say so, and the MCP tool asks for fewer layers until the result
fits the client.
node.get and document.inspect write `dashCap` on a layer with a dash cap, but
it was missing from NODE_KEYS, so `fields: ["dashCap"]` was refused as an
unknown key (and the MCP tools' `fields` schema left it out). A test now walks
the keys a varied layer tree writes and checks each is one `fields` accepts.
A fill-only text was left as it was: Outline stroke skipped every layer
without a visible stroke. A text now flattens in place into a vector of
its glyphs, filled as the text was and keeping its colour style,
constraints, visibility and lock; a text with no glyphs stays.
The nested frame is 100 x 100 at (180, 180) and its rectangle is 40 x 40 at (200, 200), as the code creates them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 driver, no subagent
A press on a layer inside a selected one now drags the selection, and a click
without a drag selects that layer (also in Dev Mode and for a single selection),
as in Figma. Cmd/Ctrl, Shift and layers outside the selection are unchanged.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove; Opus 5.5 briefed; Sonnet 5.5 wrote at medium
transform.nudge applied the step in the parent's local space, so Down inside a frame
rotated 90 degrees moved the layer sideways. The step is now turned into the parent's
space first (falling back to the plain step if the parent's transform is not invertible).
Auto layout children are unchanged.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove; Opus 5.5 briefed; Sonnet 5.5 wrote at medium
inserts_and_detaches_count_per_style_and_variable_in_the_period returned
as soon as detachments reached 2, then asserted layers == 0. Asset-action
reports can arrive before the usage snapshot, so CI saw layers still at 1.
The wait now requires both.
Sonar on the same-repo vehicle (#2451) evaluated the coverage gate at
79.74% of 232 new lines (one line under 80%). A unit test covers grow's
branch that drops a new key once peaks are already at MAX_REPORTED.
Signed-off-by: FredApps <153173854+FredApps@users.noreply.github.com>
CodeRabbit's majors on #1735: evicting a file from the capped usage list
dropped asset_peaks, so a later identical report was counted twice.
Retention now includes asset_updated, and the file just heard from is
kept even when timestamps tie. The document keeps up to 1000 asset
counts, but each wire message is still capped at 250, so reports are
sent in batches an older server will accept.
Tests: evicting_capped_usage_rows_keeps_asset_peaks_so_a_repeat_report_is_not_counted_twice
and asset_action_reports_cover_every_count_past_the_message_cap.
Signed-off-by: FredApps <153173854+FredApps@users.noreply.github.com>
counting_links compared library_links, which walked every node, so applying or
detaching a library style or variable on an instance or a layer in it counted,
while duplicating that instance counts only its component. library_links now
walks the pages and skips instances with what they hold, as count_inserted does.
Test: library_asset_counts_tests::applying_and_detaching_inside_an_instance_counts_nothing
(an instance, its layer, a nested instance and its layer). The server test waits
for the detachments too, not only the insertions.