The cross-compiled ARM64 binary can't run on the x64 packaging runner; the
windows-arm64 workflow installs and runs it on ARM64 instead. (x86 builds
still run there: x64 Windows runs them.)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- package.ps1 takes -Arch arm64 (aarch64-pc-windows-msvc, WiX -arch arm64)
and now checks each binary's PE Machine as well as its Subsystem, so an x64
build can never ship labelled arm64.
- release.yml builds, signs and packages a Windows ARM64 MSI and portable zip
next to x64 and x86, cross-compiled on the x64 runner so signing and WiX
run exactly as they do for x64.
- New windows-arm64.yml (PRs touching Cargo/packaging, main, weekly):
packages the ARM64 MSI the same way (unsigned), installs it on a Windows 11
ARM64 runner, checks both installed binaries are ARM64, opens/reads/edits/
renders/saves a PDF with the installed CLI, uninstalls; and runs the whole
test suite natively on ARM64. The MSI is kept as an artifact for 7 days.
- README: Windows builds for x64, x86 and ARM64.
Requested by @maxffarrell.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-ups on top of the table detection:
- Each block's column comes from a precomputed map and a binary search
instead of scanning every column's members, so pages with many text
blocks stay linear.
- A table has at most 64 columns. Rows are materialized to every column, so
a hostile page laid out as a staircase of blocks would otherwise make
(blocks / 2)^2 cells. Test: a_staircase_of_blocks_is_not_a_huge_table.
- RTF: a row's cell boundaries are written before its cells (RTF requires
it), and a spanning cell ends at the last column it spans. The test checks
the Totals row now spans two columns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
XFA forms aren't read yet, and opening one gave no hint: a dynamic XFA form
showed only its "Please wait…" placeholder page, and a static one looked like
an ordinary form though Acrobat shows its XFA data in preference to the
fields that get filled.
- render: DocInfo.xfa (Static / Dynamic) from /AcroForm /XFA; dynamic when
/NeedsRendering is true or there are no fields. Exported for doc_info
(automation and CLI) as "xfa".
- ui: the notice bar explains each case (above the form-fields notice); a
static form keeps its Highlight fields button.
- Tests: tests/xfa.rs (dynamic with and without /NeedsRendering, static,
ordinary forms, a malformed /XFA). Parity notes for form.xfa-static and
form.xfa-dynamic (still planned: this explains XFA, it doesn't support it).
Reported by @Sudachi-dev.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
In Edit a PDF ▸ Edit text, dragging a paragraph's box moves it, and dragging
the handle on its right edge rewraps it to a new width; a click still opens
it for typing.
- edit: BlockStyle gains offset (user space) and width. rewrite_block puts the
move inside the paragraph's own q … Q as a cm, taken back through the CTM's
linear part so a page-space drag lands right in scaled/rotated content
(the CTM is now recorded at each line's origin). Non-finite numbers are
refused; widths are clamped to one character .. 14 400 pt.
- ui: BlockDrag on DocView; the box follows the pointer (or its right edge
does) and the release applies one undoable Edit text step. The width handle
shows on upright pages only (on a quarter-turned page the screen's
horizontal is the text's vertical).
- automation: text_edit takes dx, dy and width.
- Tests: engine (move, move under a scaling cm, narrower/wider rewrap,
non-finite and zero inputs), kittest pointer drags, automation tool.
Parity: new edit.text-move-resize (shipped).
Requested by @TCB13.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On a slow Windows runner the test selected text before the page had loaded,
which selects nothing, so no highlight was made and the author check failed
(seen on PR #59's CI). Settle rendering first, as the comment tests do.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- printcraft.exe release builds use the GUI subsystem, so launching it no
longer opens a console window next to it (debug builds keep the console).
- The MSI's Icon id gains its .ico extension: Windows Installer requires it
for advertised shortcuts and ARPPRODUCTICON, so the Start Menu shortcut and
Add/Remove Programs showed a blank icon.
- Guards: package.ps1 checks the PE subsystem of both binaries (app = GUI,
CLI = console) before packaging; packaging-lint fails on an Icon id without
.ico/.exe.
- The shortcut's description said "image editor"; it's a PDF workbench.
Reported, with the exact fixes, by @MikailuReeves.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The edit applies on a later frame, so reading the annotation straight away
raced (failed about 1 run in 3 locally). Poll for up to ~100 frames.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Comments were signed with the login name (or set_option author), with no way
to change it in the app. Preferences gains an Identity section whose name
field (labelled for screen readers) sets the author of new comments; it is
saved with the settings and restored (an empty name keeps the login-name
default; restored names are capped at 200 characters). Japanese labels
added. Test: identity.rs. Parity comment.author-identity: shipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The comment tools' quick bar had colour and line thickness but no opacity,
though comments already carry it (/CA) and the Properties dialog edits it.
An Opacity slider joins them for every tool; set_opacity clamps to 10-100 %
so a comment can't be made invisible. Test: a strikeout made at 50 % carries
/CA 0.5. Parity comment.properties-bar: shipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The engine and agent tool already made squiggly underlines; the UI had no
way to. CommentTool::Squiggly joins Underline and Strikethrough in the
Highlight flyout and the Comment tool's Markup section, as the
comment.squiggly command, and existing squiggly comments edit as themselves.
Test: comments.rs::the_squiggly_tool_marks_selected_text. Parity
comment.squiggly: shipped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both features worked but had no test, so parity listed them as partial:
- core.drag-drop-open: a file dropped on the window opens in a tab
(ui.rs::dropping_a_pdf_on_the_window_opens_it, through egui's DroppedFile).
- view.annotation-hover: hovering a comment shows a popup with its author and
text (comments.rs::hovering_a_comment_shows_its_author_and_text).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
wgpu's default backends on Windows include Vulkan, and creating a Vulkan
instance loads every installed Vulkan driver into the process. On the
reporter's machine Intel's (igvk64.dll) crashed PrintCraft with an access
violation before its window appeared; WGPU_BACKEND=dx12 or gl started it
normally. PrintCraft now asks for D3D12 with OpenGL as the fallback, unless
WGPU_BACKEND says otherwise. The issue #8 low-power preference is kept in
the same configure_gpu, and a unit test (run by the Windows CI job) checks
Vulkan is left out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The UI theme include_bytes!'d ShipporiMincho-Regular.ttf, which
printcraft-fonts already embeds as SHIPPORI_MINCHO for Japanese text in
PDFs, so the app carried two copies. ui-egui now depends on printcraft-fonts
(layer 2, wasm-clean) and uses its static. The desktop app binary drops from
85.8 MB to 77.5 MB (debug build; one copy of the font left, counted).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both were marked shipped on the generic 'every registered command is
implemented' test, which the honest assessment flagged. Full screen: the
shortcut enters, Escape leaves, and the command toggles both ways. About:
the dialog shows the running version.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* i18n: add persisted English and Japanese interface language
* ROADMAP: the language entry goes at the top of the log, dated
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: midasdf <205497066+midasdf@users.noreply.github.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Builds on #45's FreeBSD workflow: it now also runs packaging/freebsd/package.sh
end to end (a quicker release profile) and triggers on packaging changes. The
release job uses the same packages, 4 vCPUs and safe.directory, and removes
target/ in the VM before copying dist/ back out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The release workflow gains a FreeBSD x86_64 job: packaging/freebsd/package.sh
builds in the same pinned FreeBSD VM and writes a /usr/local-style tar.gz
(app, CLI, desktop entry, MIME type, AppStream metadata, icons, licences).
README and CLAUDE.md list FreeBSD now that its CI job builds PrintCraft and
passes all 688 tests on FreeBSD 14.3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A new FreeBSD workflow builds the workspace and runs its tests in a FreeBSD
14.3 VM (vmactions/freebsd-vm, pinned by commit) with rustup stable, on PRs,
pushes to main, nightly and on demand. The fuzzer's address-space cap now
applies on FreeBSD too (its sh has ulimit -v).
PrintCraft's platform code was already Unix-general: everything that isn't
macOS-specific takes the same path on Linux and FreeBSD.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The nightly fuzz job found a Type 1 font that hung hayro: read-fonts 0.39
(through skrifa 0.42) looked ahead after every integer by parsing the next
token, which looked ahead again, recursing through a whole run of numbers.
A long run hung; 20,000 numbers overflowed the stack and aborted the
process. Upstream fixed the tokenizer; the vendored hayro-interpret now uses
skrifa 0.47 (read-fonts 0.44), with its one API change. The finding finishes
in 0.5 s. Synthetic regression test (it aborts with the old skrifa).
Corpus sweep: 0 regressions; text oracle unchanged (median F1 0.980, mean
0.866, every per-file score identical).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The nightly fuzz job found a JBIG2 symbol dictionary that decoded for
minutes: each bitmap side may be 65535, and up to 65535 symbols are decoded
pixel by pixel. Vendored hayro-jbig2 0.3.0 with two bounds: at most 2^28
pixels per bitmap (also the page bitmap, halftone grid and grayscale image),
and at most 2^26 symbol pixels per dictionary. The finding now finishes in
about a second. Synthetic regression test (a 65535 x 65535 page and region
behind a small /Width and /Height); the corpus sweep shows 0 regressions.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Real-file test oracles live in storytold/photocraft-corpus, explained in
craftrules standards/test-corpora.md. Never commit large binary fixtures
here; fetch them pinned and sha256-verified.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Comments: a text box's rectangle follows its edited text
Editing a text box (or typewriter or callout) rewrote /Contents but
kept the old /Rect: the appearance clipped the new text to the old box
and silently dropped lines that no longer fit, while shorter text left
an oversized box. set_contents now re-measures the text with the same
Helvetica widths as at creation and re-fits the box: the wrap width and
top edge stay, the height fits the wrapped lines. A callout's leader
line keeps its place and re-attaches to the moved box (that geometry
moved from set_rect into a shared apply_text_box). A locked box keeps
its text editable, so it re-fits too.
Covered by annot tests (grow and shrink with nothing clipped, callout
leader re-attach, locked box) and a comment_edit round trip through the
automation tools; parity entries cite them.
Co-Authored-By: GLM <noreply@z.ai>
* Polish edit text and images tool
* feat(edit-text): select source font style
* fix(edit-text): preserve detected PDF font styles
* feat(edit-text): support Japanese replacements
* fix(forms): escape non-ASCII format script strings
* Forms: unwrap_or instead of map_or with identity (clippy map_or_identity)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: GLM <noreply@z.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Every hang the nightly fuzz job found (11 findings, 8 source files) now
finishes in 0.1-2 s; each cause has a synthetic regression test that times
out without its fix.
- cos (7 findings, our own parser): a stream whose /Length pointed at
itself, with no endstream, made every parse attempt fail and each attempt
reload the same object: ~2^32 parses. Objects being loaded are a seen-set,
/Length values are memoised per load, and the shifted-header retry only
runs when the header is shifted.
- hayro-syntax (3): deciding where an inline image's data ends re-parsed the
rest of the stream, nesting through every inline image; 50 images ran for
over ten minutes. The look-ahead no longer nests and is bounded (256).
- hayro-interpret (1): a Type 3 glyph showing eight glyphs of its own font,
sixteen deep. Nested paints share a per-page budget (50,000).
The pdf.js corpus sweep matches its baseline (963 ok, same 20 known).
xtask check writes its report under CARGO_TARGET_DIR too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ROADMAP gains an honest assessment by dimension (features, effort,
foundations, robustness, fidelity, product readiness) and by area, what
'shipped' does and doesn't mean, and a ranked list of where we're lacking
and where we're going. Numbers refreshed (804 features, 49.5% shipped,
effort ≈ 30–35%). README's stale 'What's next' becomes an honest status
summary; CLAUDE.md points every agent session at both sections.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
By the owner's decision there is no update check at start and no in-place
install for the foreseeable future. The startup option, its setting and the
first-frame trigger are gone; Help ▸ Check for updates is the only request.
Settings written by a build that had the option are ignored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PrintCraft asks GitHub for the latest release and, when it is newer, offers
its release page to download from. Nothing is downloaded or installed
automatically (installing in place needs signed release artifacts).
- ui-egui: updates module (version comparison, background check, dialog);
no network code: the desktop app supplies the source, and without one
(web build, tests) the command opens the releases page.
- desktop: one HTTPS request (ureq over rustls with the OS's trusted roots,
10 s timeout, 1 MiB cap). Only a page under the releases list is offered.
No bundled root list: webpki-roots/webpki-root-certs are CDLA-licensed.
- The check at start is opt-in and silent unless there is a newer version.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- On hybrid-graphics laptops wgpu picked the discrete GPU (egui-wgpu defaults
to HighPerformance). NVIDIA's can lose or corrupt its memory across suspend
and screen lock, which matches the illegible window in #8. PrintCraft now
asks for the low-power adapter unless WGPU_POWER_PREF is set; one-GPU
machines are unaffected, and it saves battery.
- "Save changes?" had no keyboard: Enter saves, Escape cancels, and Don't save
takes Cmd/Ctrl+D (macOS) or Alt+D / Alt+N (the mnemonics the reporter tried).
Its keys are read before the canvas (which took Escape), and no other
shortcut runs underneath it (Cmd+D opened Document properties).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The second full nightly fuzz run on main found two more:
- hayro-syntax computed a Flate/LZW predictor row as columns × bits in
wrapping arithmetic: /Columns 9223372036854775807 asked for 2^61 bytes and
aborted on any machine. Row widths are now saturating usize (colors × bits
also overflowed u8), a zero width is refused (it divided by zero), and a
row longer than the data is not allocated (empty output, as before).
- hayro stroked /LW 9223372036854775807 as given, and stroke expansion
allocated 10 GB. Widths are clamped to 4 × (canvas width + height) device
pixels, which covers the same pixels.
Both findings now run in 8 MB. Synthetic regression test in render.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first full nightly fuzz run on main (now that the runner survives) found
two 4 GiB allocations, caught by the 4 GiB child cap:
- an inline stencil mask claiming /W 4294967295 was decoded in hayro-interpret
before the renderer's MAX_IMAGE_PIXELS check could skip it (locally 9.6 GB,
42 s). Vendored patch image_size_ok: images and masks over 2^28 pixels are
not decoded.
- a CCITT image whose /DecodeParms claimed /Columns 4294967295 (with a small
/Width) sized the decoder's line buffers (locally 4.3 GB). Vendored patch
ccitt_size_ok in hayro-syntax: at most 2^20 columns and 2^28 pixels.
Both findings now run in 7 MB. Synthetic regression test in render.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fuzzed inputs drove one check-one process to 2-6 GB (the 16 GB CI runner
running two of them was killed every night). Three causes, each with a test:
- lopdf (bootstrap inspector) sized PNG predictor rows from /Columns before
reading data: /Columns 4294967295 allocated two 4 GiB rows. Vendored lopdf
0.45.0 refuses rows longer than the data (and uses saturating row widths);
its load-time stream decoding also gets a 256 MiB limit (it had none).
- hayro sized a tiling pattern's cell pixmap from /XStep x scale, bounded only
by u16: a step far beyond /BBox allocated 2 GB. The vendored patch keeps
the cell within 3000 px a side, lowering the resolution instead.
- cos decoded cross-reference streams up to the general 1 GiB limit; they now
stop at what 8,388,607 objects' rows can hold (~200 MB at most), and the
section is reconstructed (logged) instead.
On the nightly's seed (41) the worst captured input fell from 6.2 GB to 6 MB,
and a 300 s fuzz run had no process above 300 MB (before: many at 2-5 GB).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- xtask looked for target/release/printcraft-cli even under CARGO_TARGET_DIR
(which parallel agents use), and a child that failed to spawn counted as a
pass: such runs fuzzed nothing and reported 0 crashes. The binary now comes
from CARGO_TARGET_DIR (gates' corpus check and text oracle too), and fuzz
stops if it is missing.
- On Linux each child runs under a 4 GiB address-space limit, so an input that
allocates without bound becomes a crash finding instead of exhausting the
machine: the nightly fuzz job's runner was killed this way every night,
losing the findings. Allocation failures are kept as the finding's message.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Windows runners check out with core.autocrlf=true, which turned the SVG icons'
LF into CRLF, so their SHA-256 no longer matched ATTRIBUTION.toml and
xtask's the_real_manifest_passes failed there (hidden until now behind the
perf-test failure, since cargo test stops at the first failing binary).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The two budget tests ran in parallel, each with its own render workers, and
a single preempted stretch on a shared CI runner pushed the average over the
40 ms budget (41-52 ms on Ubuntu and Windows). They now hold a lock while
measuring, and a round over budget is measured again (up to three rounds):
noise only adds time, so the best round is the UI's own cost and a real
regression still fails every round. The budget is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brandon Thomas
midasdf
storybored379
Justin John