NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
pub.dev
pub.dev, #2759 most downloaded on pub.dev
COS object layer for PDF: tokenizer, parser, filters (incl. CCITT, JBIG2, JPEG 2000), xref machinery, encryption, and serializer. Pure Dart, web-ready.
Latest stable
15.0.0
Last release 17 days ago
05 Nov 2026
Ships on a steady schedule
a new release about every 1 weeks
Nearly every release is documented
notes for 50 of 51 stable releases
Nothing withdrawn
no release was ever pulled
5 months old
51 releases · first in 2026
Drag across the ruler to compare, or tap a release to open it.
Add GitHub and email feedback links to Settings > About
Add GitHub and email feedback links to Settings > About (#1109)
Keep an edited annotation painted when it is deselected before the page catches up (#1110)
Fix image crop chrome: constant-size chips, crop-style handles (#1108)
Fix image crop chrome: constant-size chips, crop-style handles
The crop confirm/cancel chips lived inside the viewer's zoom transform
without counter-scaling, so they grew with zoom (~130px at 386%) and sat
over the image. They now counter-scale by chromeScale and ride below the
crop (above, or inside, when the page runs out).
The normal selection chrome (tint, square knobs, rotate knob) kept
painting under the crop frame; it is now suppressed while cropping, and
the crop frame draws corner brackets and edge bars instead of square
knobs.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01Mor1BnC9MigXVgh9Ja3Dn9
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01Mor1BnC9MigXVgh9Ja3Dn9
Co-authored-by: Claude noreply@anthropic.com
Callouts get the nodes-only selection chrome (#1102)
Callouts get the nodes-only selection chrome
A lone selected callout now drops its box outline and rotate knob like a
selected line does, keeping its leader nodes and the text box's resize
handles. The painter gains a selectionOutline flag so the handles can sit on
the selection rect without the box.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01JvbRYftUbFgYACV1cc6RMm
_selectedViewRects listed a callout by its full /Rect while the primary
selection uses the text box, so the mismatch was painted as a second
selection box around the leader. With the primary outline gone it was the
only box left. Both now use calloutBox.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01JvbRYftUbFgYACV1cc6RMm
Co-authored-by: Claude noreply@anthropic.com
CAD sets that tile a sheet into thousands of small /Indexed Flate images
spent most of their render time expanding palette samples. _indexedToRgba
now expands through a 2^bits RGBA word table (one store per pixel), and a
display-sized decode of an opaque unmasked Indexed image averages palette
entries straight into the target cells instead of building and then
downsampling the native RGBA surface - byte-identical to the old
full-decode + downsample.
AOT: 503 tiles at display size 854 ms -> 173 ms; full decode 802 -> 263 ms.
Claude-Session: https://claude.ai/code/session_01KwVwtEGKBKXzmjjLg5Tf9C
Co-authored-by: Claude noreply@anthropic.com
The start card from #1084 stretched to the recents column's width (900px
in grid mode), so with no recents it became a wide grey slab around a
small icon and button. Cap it at 420px, centred, independent of the
recents column, and drop the "DartPDF" heading - the app bar and window
title already name the app.
Claude-Session: https://claude.ai/code/session_0162CZ53LBd5wb9wyiZufAg6
Co-authored-by: Claude noreply@anthropic.com
Take the pen dot down when a touch/stylus ink stroke ends (#1093)
Take the pen dot down when a touch/stylus ink stroke ends
A finger never hovers or exits the page, so nothing cleared the painted
pen cursor after a touch-drawn stroke - it stayed parked where the stroke
ended. The raw-pointer stroke end (and the two-finger bail) now clears
it; a hovering stylus re-places it on its next hover event. Mouse
strokes keep the dot at the stroke end as before.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01V7eQc9EqyZztCuaFWpB4zo
Flutter web follows a finger's pointerup with a synthesized hover of kind
touch at the lift point. MouseRegion delivers it through hit testing
(MouseTracker skips touch, RenderMouseRegion.handleEvent does not), so
_onHover put the pen dot straight back where the stroke ended - the
previous commit's clear on lift never survived on web. A finger can't
hover, so _onHover now returns for touch; that also keeps the eraser
ring, guides and placement previews from parking at a finger's lift.
Reproduced and verified in headless Chromium with touch emulation
against a release web build of the example.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01V7eQc9EqyZztCuaFWpB4zo
Co-authored-by: Claude noreply@anthropic.com
Open documents in Hand mode on mobile (#1113)
Link badge, word-level text edit, and a long-press page menu (#1112)
Fold-aware shell and iPhone Duo screenshot slots (#1111)
Fix "Cannot clone a disposed image" when OCR decodes a large scan (#1115)
Fix "Cannot clone a disposed image" in PdfBudgetedCache.putAndClone
putAndClone stored the value and only then cloned it. put() enforces the
PdfCacheRegistry process ceiling, whose hard trim may evict the entry it
just stored, so a decode big enough to exceed the ceiling on its own
(OCR rasterizes scans at 216 dpi, and iOS runs a small ceiling) was
disposed before the clone and decodeImages threw, failing the OCR job.
Clone before the store, and hand a value that rejectOversize would turn
away back uncached instead of cloning it and leaking the master.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01WnasbPSpLEkwbKJr6ds1Rf
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01WnasbPSpLEkwbKJr6ds1Rf
Co-authored-by: Claude noreply@anthropic.com
The first time automatic OCR adds a text layer, a one-time dialog asks
whether to keep it on for every scan or turn it off. Dismissing keeps it
on; the offer is persisted (and counts as made once the user picks the
Settings switch), so it is asked only once.
Claude-Session: https://claude.ai/code/session_01TfxQC46LgFdM6Eqavt6icV
Co-authored-by: Claude noreply@anthropic.com
PdfEditTool.measureCustom drags a dimension line (closed arrows both
ends) and prompts for the value it shows - pre-filled with the scaled
length when a scale exists, needing none otherwise. The value is kept in
/Takeoff /Value, becomes the caption and /Contents, survives reshapes,
and still totals (pdfParseMeasurementValue reads number + unit, incl.
feet-and-inches). Existing measurements get "Edit measurement value" in
the annotation context menu (PdfEditor.setMeasurementValue; blank goes
back to the computed value). The compact tools sheet now scrolls when a
group outgrows it.
Claude-Session: https://claude.ai/code/session_0183K9JEnHZwYgm1z417D8cK
Co-authored-by: Claude noreply@anthropic.com
Prepare the 10.0.0 package suite and app build 47, with companion dependency bumps, store notes and migration guidance. Correct nested Android CI signing, the CLI version report, and Linux/Flatpak/Microsoft release metadata.
Local validation: all 14 suite groups passed, including Chrome; release metadata, dependency floors and static analysis passed. Independent AI review was unavailable, not approved.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Correct the public demo version manifest after the 10.0.0 release. No native app or package code changes.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Reuse the existing portable Android compatibility preparation in Storeflight CI. Application sources, dependency constraints and release version remain unchanged.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Document dart_pdf_printing 0.7.0 for publication (#1122)
docs: add markdown pdfjs render gallery
Release dart_pdf_editor 3.7.1 with safe annotation disposal (#712)
Update AUR package for 3.6.0
Fix Linux store metadata for 3.6.0
Fix Linux launch crashes and add GUI smoke tests to CI (#693)
Linux: disable Flutter's experimental windowing (both the Dart feature
gate and the native headless bootstrap). The headless engine bootstrap
leaves GTK plugins such as desktop_drop registering against a null
FlView, which fails gtk_drag_dest_set and prevents window creation -
the snap (and any strict-confined launch) exited at startup.
Snap: add the dartpdf-dbus session-bus slot so strict confinement lets
the GApplication own dev.milanko.dartpdf instead of dying in AppArmor;
also check in the from-source snapcraft.yaml alongside the release
repack recipe.
CI: replace the non-blocking 'survive 8s' launches with
tool/smoke_test_linux_gui.sh, which fails the job on GTK-CRITICAL /
D-Bus AccessDenied output and verifies the DartPDF window actually maps
under an isolated Xvfb display. Wired into the Linux bundle, AppImage,
snap, and flatpak pipelines, plus new macOS and Windows GUI launch
checks in release-app.yml.
Co-authored-by: Claw (Bot) ben-claw@res.app
Prepare 3.6.1 release
Pin Linux packages to the 3.6.1 release
Fix snap release build: keep cli insertion point inside the apps section
The dartpdf-dbus slot block moved the CUPS comment away from the end of
the apps: mapping, so build-release-snap.sh's cli-app insertion landed
inside the top-level slots: block (snap warning: unknown interface
"cli") and the built snap lost dartpdf.cli.
New interface attributes (the dartpdf-dbus slot) send the uploaded
revision to human Store review; snapcraft exits non-zero even though
the upload succeeded and the revision releases on approval. Skip the
clean-install verification until the store actually serves the new
version.
Omit standalone Dart CLI from Mac App Store archives (#695)
Sanitize Mach-O private symbols from Dart AOT binary for App Store compliance
Omit Dart CLI from Mac App Store archives
Co-authored-by: dart-pdf release automation probe actions@github.com
Fix Windows open/save dialog crashes under the multi-window runner (#696)
Fix Windows open/save dialog crashes under the multi-window runner
file_selector_windows takes the dialog's owner HWND from the plugin
registrar's implicit FlutterView. DartPDF's Windows runner starts a bare
FlutterEngine so Dart owns every native window, so there is no implicit
view: registrar->GetView() returns null, the plugin dereferences it, and
the process dies at the first "Open PDF in new tab" or "Save as". It is an
access violation inside the method-call handler, so nothing on the Dart
side can catch or report it.
The runner already knows its active window without a view - it supplies one
to the print dialog and the image clipboard - so drive the common-item
dialogs from there instead:
Directory picks route through the channel too - the app does not use them
yet, but they were the plugin's other two null-view crash sites.
Drag-and-drop stays inert on Windows for the same underlying reason
(desktop_drop bails out when the registrar has no view); that needs a
per-window IDropTarget in the runner and is not part of this change.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_019pif2PSbZYWmho5beuSEu9
The MissingPluginException path for getDirectoryPaths was the one line
the new file left untested.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_019pif2PSbZYWmho5beuSEu9
Co-authored-by: Claude noreply@anthropic.com
Fix simple-font character mapping in content editing; add search-panel replace (#697)
Read and write simple-font text through its own encoding
A user editing a form's date field saw the content tool report the text
as -=>-/>?-?@ where the page reads 05/08/2026, and edits that
"changed a lot of content on the page". Two independent bugs.
A simple font's byte is not its character. PdfPageElements.of
decoded a non-/Type0 show string with latin1.decode, and
_SimpleRunCodec mirrored that: matching on String.fromCharCode(code)
and writing replacements as replace.codeUnits. That only holds when the
font's codes happen to be Latin-1. Section 9.6.6 lets a document remap
every code through /Encoding /Differences, and a subsetted font
routinely renumbers its codes from scratch - the reporter's font drew
05/08/2026 from the bytes 2D 3D 3E 2D 2F 3E 3F 2D 3F 40, which is
exactly -=>-/>?-?@ in Latin-1. Their "B = 8" note is the same table read
backwards: typing B wrote 0x42, and 0x42 is whatever glyph the
subset put there.
The renderer never had this problem (PdfFontInfo.charFor resolves
/ToUnicode, /Differences, the built-in Symbol/ZapfDingbats encodings, then
the base encoding), which is why selection and search read correctly the
whole time. Only the content editor had its own shortcut.
Adds SimpleFont, the simple-font sibling of Type0Font: one place that
resolves a code to text the way the renderer does, plus the reverse table
for re-encoding replacements and the /Widths lookup (which absorbs
content_editor's _widthsFor). Used by PdfPageElements, both run
codecs, the '/" operator path, and reflow's _SimpleReflowFont.
Encodability is deliberately narrower than decodability: decoding keeps
the lenient Latin-1 fallback, but a replacement may only use codes the
font declares (a glyph name from the base encoding or /Differences, or a
/ToUnicode entry), because a subset that dropped 9 has no glyph at
0x39. An undrawable replacement leaves the run untouched - the contract
Type0Font already had. A font that declares no encoding at all (base-14,
and everything this library writes) keeps the identity in both
directions, so existing round-trips are unchanged.
Fixing the decode also created a latent bug in paragraph reflow, fixed
here: content_reflow reads element.text, so once that decoded
correctly a remapped font would start matching, and showOp would have
written text.codeUnits as bytes.
"It changes it everywhere." Separately, the element strip's "Replace
text" went through the page-wide PdfEditor.replaceText. The targeted
replaceElementText already existed (it backs the selection-menu edit);
this path just never used it. Both element-strip entry points now share
one helper that uses the targeted API.
SimpleFont needs the Adobe glyph-name tables, which lived one layer up
in pdf_graphics. They are pure data, so they move down to
pdf_document/src/fonts/encodings.dart; the old path is now a show-list
re-export so the font engine's imports are untouched.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
PdfSearchResultsPanel(editing:) turns the find panel into find and
replace: a replacement field under the options bar with Replace (the
current hit alone) and Replace all (every hit, one undo step). Null
editing - the default, and what a read-only shell passes - leaves it a
pure find panel; the editor shell wires the session wherever the editing
toolbar is available.
Two controller methods back it:
replaceMatchText(page, rects, find, replace) resolves the hit throughtextElementForSelection - the same conservative gate the selectionreplaceTextInElement. It returns 0 rather thanreplaceTextOnPages(pages, find, replace) is the page-wide form forapply() so it is a single undo step.This only became possible with the preceding encoding fix: search hits
come from the text layer while elements come from PdfPageElements, and
before it those two disagreed about what a remapped run said, so no
search-driven replace could ever have resolved.
The re-search after an edit has an ordering trap worth noting. The viewer
calls clearSearch() when it swaps in the new revision, which happens on
the next frame - so a re-search issued straight after the edit is thrown
away moments later and the panel goes blank. _ReplaceBar._refresh awaits
endOfFrame first.
New strings land in the template ARB; the other locales fall back to
English until translated.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
CI's tool/check_arb_coverage.dart gate is stricter than gen-l10n: every
non-English ARB must carry the exact same keys as the template, precisely
so a half-translated locale can't ship looking finished while random
strings render in English. Leaving the five new keys English-only failed
that gate.
Adds them to all 19 locales, following each one's existing search
vocabulary (its own word for "match" from searchMatchCount, its own verb
for "replace" from tbReplaceText) and its own ICU plural categories -
few/many for pl/ru/uk, two/few/many for ar, other-only for the
CJK and other categoryless locales. Inserted at the template's position
rather than re-serializing, so each file's existing layout is untouched.
dart run tool/check_arb_coverage.dart now reports OK across all 57
locale files, and gen-l10n regenerates with no untranslated-message
warnings.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
Codecov flagged 11 uncovered lines in the new SimpleFont, all of them
branches where a mistake is invisible: /Encoding given as a name rather
than a /Differences dictionary, the built-in Symbol and ZapfDingbats
encodings, and the width fallbacks. Covering them found a real bug.
standardGlyphName deliberately reports only StandardEncoding's high
bytes - the ASCII range falls through to the Latin-1 identity, which is
the same answer for decoding. But declared was built from the glyph-name
table alone, so under /StandardEncoding or /MacRomanEncoding no ASCII code
was ever encodable: such a font read correctly and then silently refused
every ordinary replacement. Annex D defines all of printable ASCII for
every base encoding, so a named base encoding now declares 0x20-0x7E.
This deliberately does not loosen the subset case that motivated the
original gate: for an /Encoding dictionary, ASCII is declared only when the
document states a /BaseEncoding. A /Differences-only dictionary - the shape
in the bug report - still restricts writing to the codes it names, so a
character the subset dropped is still refused rather than drawn as notdef.
Tests added for all of it: /WinAnsiEncoding round-tripping a high-byte
glyph (eacute), /MacRomanEncoding, Symbol and ZapfDingbats (including a
subset-tagged one), a /ToUnicode that is junk, a /ToUnicode whose stream
cannot even be inflated, /MissingWidth for codes outside /FirstChar../
LastChar, and the Helvetica metric fallback. simple_font.dart is now at
100% line coverage; pdf_document is 835 tests green.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
The last 8 uncovered lines of the diff were the ' and " branch of
_replaceText - the one path this PR rewrote that no test reached. Those
operators carry a line break, so they stand alone and their string is
rewritten in place rather than through the run rewriter; the change made
both ends go through the font's own encoding, and nothing exercised it.
Three cases against the remapped subset font: ' rewriting its string into
the font's codes, " doing the same while keeping its word/char-spacing
operands, and ' declining a character the subset dropped.
Every line this PR touches in content_editor.dart is now covered (the file
as a whole goes 84.4% -> 91.0%), alongside simple_font.dart and
content_reflow.dart at 100%. pdf_document is 838 tests green.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
The replacement field and its two buttons sat permanently under the options
bar. That is a lot of vertical space to spend on a panel whose job is
listing hits, so they now hide behind a find_replace disclosure that
lives in the options row itself - right-aligned, so it reads as a
disclosure rather than a fifth match option, and styled by the extracted
PdfPanelToggleButton so it is indistinguishable from the ones beside it.
Collapsed by default; the state persists via
PdfEditingPreferences.searchReplaceExpanded, which the app shell already
passes to the panel. The replacement TextEditingController moves up to
the panel state, because the bar is removed from the tree when collapsed
and would otherwise take the typed text with it.
The padding was wrong in two ways, both worth spelling out.
The panel's content edge is 16 - the title, the count header, and the
result tiles all sit there. The options bar uses 8 and still looks right,
because its toggles are IconButtons with zero padding in a 30pt box, so
the glyph lands ~8pt in: 8 + 8 = 16. Copying that 8 for the replacement
field, which has a real border, put it 8pt left of everything else.
Content carrying its own inset takes 8; content without it takes 16.
Separately, both bars hard-coded their horizontal padding and ignored
geometry.contentStartInset/contentEndInset - the resize-grip clearance
the dividers already honoured - so the field ran under the grip on the
docked edge. Measured after the fix, on a 280pt panel with an 8pt grip:
title 16, field 16..256, buttons 16..256, count 16; and 16..264 as a
bottom sheet, where there is no grip.
The two dividers collapse into one below the whole control block, and the
field drops its find_replace prefix icon now that the disclosure carries
that glyph - it only cost width on a narrow panel. Button height stays 40:
VisualDensity.compact reaches the Material touch-target floor, and going
below it would trade tablet accessibility for a few pixels.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
Co-authored-by: Claude noreply@anthropic.com
Render pages as you scroll past them (#698)
Render ordinary pages through the scroll, and keep them
On a long text document the reader waited ~600 ms per page arrival while
the render hold sat out the whole scroll animation, then paid the same
cost again scrolling back because the retained-scene LRU kept four pages
for the entire document.
The hold is priced for a dense sheet's 100-400 ms interpret walk. With a
render worker attached that walk is not on the UI thread at all, so
STARTING a page during a scroll costs the frame nothing, and what does run
here - the replay of the returned buffer - has a size that is known by the
time it matters. PdfPageRenderScheduler gains a motion-safe lane on that
basis: a request may declare itself motionSafe and re-declare it at
paceUiWork once the record lands. PdfPageView decides in two halves - a
live worker and no page-level /XObject up front, then <=4000 commands and
no image draws when the buffer arrives - and revokes the verdict the moment
a pass falls through to a local walk. Grants during a hold stay near the
scroll's destination so a fling does not record every page it passes.
A page that finishes mid-scroll is usually not on screen yet, where the
direct picture presentation declines and the raster branch would burn a
full-page readback for pixels it replaces moments later; it now defers to
the existing off-screen visibility hook instead.
Retention pairs with it. The retained-scene LRU's byte estimate priced a
text page at ~0.5 MB against ~9 MB of measured browser heap, so the budget
governed nothing and misinformed PdfCacheRegistry. priceRetainedScene now
floors an entry at the raster it stands in for, the entry cap goes back to
being a backstop, and the budget becomes platform-aware alongside its
neighbours in performance_policy.dart.
Measured with a new read harness scenario (each page arrival, and the
wait after the scroll lands). read-text, 3 interleaved runs vs the base:
settle p50 207 -> 0.05 ms first visit and 196 -> 0.03 ms on the way back,
100% of revisits sharp on arrival, whole-arrival p50 614 -> 477 ms. Costs
~34 MB of retained scenes on a text document and a median frame build of
2.8 ms against 2.2. read-plan, the CAD arm, moves the same way for +5.6%
memory. bench.mjs now copies the Dart harness into the baseline worktree,
without which a new scenario silently compared two different workloads.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
The follow-up report was that a continuous scroll still waits: scroll two
pages, then ~0.5 s before anything is readable. The first round measured the
wrong journey - read navigates with jumpToPage, while a reader drags the
document past the viewport - so a new wheel scenario streams pointer scroll
signals at frame cadence and samples every frame which pages are visible and
which of those are sharp. It reproduced the report at once: settle 716 ms
after the scroll stopped, 15% of visible-page frames readable while it moved,
3 of 5 pages never sharp at all.
The trace showed the motion-safe lane granting correctly and the page the
reader actually landed on being granted only after the 500 ms scroll-quiet
window expired - because nothing had ever queued it. Nothing in a page's
render intent changes as it scrolls into view (onScreen is deliberately not
part of it) and the image-ratio promotion cannot fire for a page that has
never rendered, so the settle generation bump at the end of the scroll was
the only thing that queued a mounted cache-window neighbour. The lane cannot
grant what was never queued.
A page on screen with nothing of its own to paint now asks on every rebuild,
guarded by PdfPageRenderScheduler.isQueued so the ask cannot pile a second
pass behind a render already in flight. Level, not edge: the on-screen span
flips before the page is geometrically visible.
Also drops the off-screen raster deferral added earlier in this branch - it
was firing for pages scrolling into view and costing them an extra round
trip - and sizes the retained-scene budget to the speculative warm window,
with the measurement showing why a smaller one is wasted work rather than a
saving.
wheel-text vs the base, 3 interleaved runs: settle 716 -> 0.23 ms, sharp
while scrolling 15% -> 96%, a page's soft time p50 1322 -> 0 ms, no page left
unsharp, median frame build +3% with p95 better. wheel-plan (dense vector)
settle 864 -> 0.23 ms and tab memory down 32%. Attribution: with retention
off the scrolling fix still lands settle 0.2 ms at 68 MB, below the 75 MB
baseline - the memory cost belongs to retention, not to scrolling.
Harness fixes found on the way: bench.mjs now copies the Dart harness into
the baseline worktree (without it a new scenario silently compared against
scroll), and non-warm scenarios no longer inherit the warm scenarios' 1 GB
page-raster budget, which flattered revisits and inflated memory.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
A field trace from the desktop app showed the previous rounds had barely
moved it, and named two causes the browser harness could not reproduce.
The thumbnail strip owned the wait. At each settle it ran 400-530 ms of
tile interpret+raster on the platform thread ahead of the page the reader
landed on, whose own walk was 12 ms. The queue does stand down while the
viewer is busy, but it read busy once before the first tile and then ran
them back to back - and at that instant nothing was queued, so the frame
that would have queued the page could not run. The drain now yields a frame
between tiles and re-reads the gate (and the queue) after it.
That session also had no render worker at all - every interpret on the UI
thread - which made the motion-safe lane inert, since it required an
off-thread walk. So the lane gains a third class, and the boolean becomes
PdfRenderMotionClass evaluated at GRANT time rather than baked into the
request: a request made mid-fling is granted after the hand has stopped, and
the answer differs between those moments. free (worker-backed) runs during
any motion; quiet (no worker, small page) runs in the scroll-quiet window
but never inside a live gesture; held is unchanged. A live gesture needed
its own signal - _motionRenderHoldActive includes the 500 ms quiet window,
which is the thing the lane exists to bypass - so the viewer drives
activeGesture from a 120 ms timer restarted on every scroll event. A local
walk that overruns its budget marks that page held for good.
The interpret line now distinguishes recorded(no-worker) from
recorded(declined), so a field trace answers "did this session have a
worker" in one word.
Measured: wheel-text settle 0.2 ms and 96% sharp while scrolling hold; the
no-worker arm goes 612 -> ~410 ms, and that residue is three UI-thread walks
rather than waiting. Full suite 2338 passing, counter gate clean.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
Chasing a field trace that showed every page interpreting on the UI thread
turned up why that state is so hard to diagnose: a failed isolate spawn is
caught with catch (_), sets _spawnFailed, and from then on every record()
resolves to null and every page walks on the UI thread - permanently, for
that document, with nothing in any log. A trace of that session reads exactly
like one where the worker simply declined a page.
It now logs to PdfPerfLog and, in debug, prints a plain-language warning
naming the cause. With the recorded(no-worker) / recorded(declined) split
already on the interpret line, a trace answers "did this session have a
worker, and if not why" on its own.
No behaviour change: the fallback path is untouched.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
The motion-safe lane refused any page that declared an /XObject, on the
reasoning that a three-operator content stream can hide a hundred
megapixels. It can - but the same rule also refuses the 42x10 letterhead
mark that every page of a corporate report carries. A field trace of one
such document (113 pages, a mark on each) shows the lane switched off for
the whole file: every scheduler grant reads cost=held.
An image's size is a dictionary entry, so measure it. The request-time gate
now sums the declared /Width x /Height of a page's image XObjects against
motionSafeMaxImagePixels (1 MP ~ 4 MB of RGBA), and the record-time gate
weighs the buffer's actual image draws the same way
(PdfPageRenderer.imageDrawPixels), which also covers inline images and Form
XObject contents nothing at page level advertises. Anything that will not
say how big it is - a Form XObject, an image with no declared dimensions, an
unresolvable resource - still answers "held", exactly as before.
Second half of the same trace: scene-cache reject reason=image-lod have=1.79 need=2.00. Scrolling back onto a page threw away its retained
display list over a 10% difference in embedded-image resolution - blank
paper plus a re-interpret for something no one can see. What a cached scene
has to clear is display sharpness at the current zoom, not the 2x decode
headroom a fresh render aims for, so it is compared against _effectiveRatio
now; a genuinely soft prefetch buffer is still rejected, and the headroom
arrives as the existing soft-to-sharp refinement behind the painted page.
The corpus had no document of this shape - the ordinary middle between
text-only and full-page raster - so letterhead-report-40p.pdf joins it, with
wheel-letterhead / wheel-letterhead-noworker scenarios.
Measured (tool/perf.sh webdiff, 3 interleaved iterations, wheel-letterhead):
settle 644ms -> 0.22ms, sharp-while-scrolling 15% -> 96%, pages never sharp
3 -> 0, soft p50 1300ms -> 0, frame build p95 -24%. Cost: 72 MB -> 173 MB of
tab memory, the retention trade this branch already priced - the baseline
held nothing because it drew nothing. With the worker off, the settle halves
(621ms -> 360ms) and sharpness during the fling is unchanged, which is that
path's ceiling.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
Co-authored-by: Claude noreply@anthropic.com
Lockstep minor for the eight core packages (3.6.0 -> 3.7.0), with
dart_pdf_cli 0.1.0 -> 0.1.1 and dart_pdf_editor_flutter_gpu 0.1.4 -> 0.1.5
for their constraint bumps, and the app at 3.7.0+31.
No breaking public API changes since 3.6.0: PdfPageRenderScheduler.request
kept its positional parameters and only gained an optional named one, and
the glyph-name tables that moved from pdf_graphics to pdf_document were
never part of pdf_graphics' public surface (src/fonts/encodings.dart is now
a show-list re-export).
Changelogs authored from the commit bodies of #695-#698, plus the store
"What's New" text for 3.7.0 in all 20 store locales.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Fix release-mode annotation picture disposal
chore(release): prepare dart_pdf_editor 3.7.1
chore(release): align editor dependency constraints
Co-authored-by: dart-pdf release automation probe actions@github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: Claude noreply@anthropic.com
Release DartPDF 4.3.0
Fix StoreFlight isolated checkout inputs
Add 4.3.0 Linux release metadata
Release DartPDF 7.0.0 and versioned companions
Add DartPDF 7.0.0 AppStream release notes
Release DartPDF 8.0.0 across packages and apps
Enforce release package order and serialize local Flutter builds
Release DartPDF 8.1.0 across packages and apps
Keep canonical Linux release metadata in sync for 8.1.0
Promote dartPDF 10.0.0 for Storeflight publishing [skip ci]
Promote the tested dartPDF 10.0.0 release snapshot for Storeflight dependency-ordered package publishing. The promoted tree is identical to main 9d4b26c; the release and CI-repair PR checks passed. Ben explicitly authorised this administrative merge on 19 October 2026. Skip the separate automatic publisher to preserve the existing Storeflight release owner.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Identical file tree to deploy cf453de and tested main 9d4b26c. A fresh metadata-only commit permits tag-triggered publication. No application source, package contents, version or workflow changes.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claude Opus 4.8 (1M context) noreply@anthropic.com
Co-authored-by: dart-pdf release automation probe actions@github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: Ben Milanko 20643387+ben-milanko@users.noreply.github.com
Select installed App Store distribution profiles and explicit macOS export options for hosted Apple builds. Wait for GPU submission completion before attachment-reuse assertions.
Exact-head CI, preview and native Patrol checks passed. Preserve the paused iOS release and its existing product-page test.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Select the installed Xcode 26.3 toolchain on macos-15 and fail before build/upload when the iPhoneOS SDK is too old.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
build(deps): bump actions/setup-java from 4 to 6 in the actions group (#1141)
fix(editor): autosize text boxes in rotated page coordinates (#1127)
Keep the visible top-left corner and paragraph width when fitting free text on rotated pages. Cover all four page rotations, callouts and Alt+Z.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
fix: preserve snapshot orientation on rotated pages (#1128)
fix: preserve snapshot orientation on rotated pages
Counter-rotate vector snapshot appearances for the destination page and place natural-size pastes in display coordinates. Keep explicit paste points and make repeated pastes cascade visibly down-right.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Use R for clockwise quarter-turns and Shift+R for reverse turns. Keep rectangle R without a rotatable selection, custom bindings, and text-editor ownership.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat: vertically align text inside text boxes (#1129)
feat: vertically align text inside text boxes
Add top, middle and bottom placement for plain and rich free text. Preserve placement through rendering, edits, resize, callouts, scoped defaults, inline editing and save/reopen.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Defer the synchronous page parse until the initial preview frame and skip stale or ended drags. Add dense-page regression coverage and a reproducible Chrome resize-start benchmark.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(editor): find and replace colours in snapshots and annotations (#1132)
feat(editor): selectively replace snapshot and annotation colours
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Fixes #1124
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Fixes #1125
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(app): resumable OCR with saved-page checkpoints (#1139)
feat(app): persist OCR jobs and continue background processing
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Remove detached workers and mobile background hosting. Checkpoint completed pages and resume when the app reopens or returns to the foreground.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(editor): duplicate library stamps and select imports/exports (#1135)
feat(editor): duplicate library and placed stamps independently
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(editor): editable tables and Excel clipboard transfer (#1140)
feat(editor): add editable tables and Excel cell copying
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Use a Table grid with full-content rows, shared column widths and cell keyboard navigation. Rebuild table appearances at constant font size during resizing and retain reflowed drag previews and undo.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Preserve table colours, standard font variants and grid settings across editing, resizing and save/reopen. Reuse existing font and colour controls and cover saved appearance pixels, mobile layout and plain hosts.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Use existing viewer input paths to edit selected or unselected tables in default and Select modes. Preserve protected-table guards and undoable Apply behavior.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Setup ran app-processes.ps1 under 32-bit PowerShell (SYSDIR is SysWOW64
for a 32-bit process), where Process.Path is empty for a 64-bit process,
so the running-app count read 0 and the locked-exe retry box came back.
Use Sysnative PowerShell, and fall back to CIM's ExecutablePath.
Claude-Session: https://claude.ai/code/session_014nwq5uZzoq6CgE3iHtfRyS
Co-authored-by: Claude noreply@anthropic.com
Floor thin rectangle fills at one device pixel (#1146)
Floor thin rectangle fills at one device pixel
Word and similar producers draw table and box borders as filled 0.5 pt
rectangles (re f), not strokes. Strokes already had a one-device-pixel
floor; fills did not, so zoomed out those borders covered a fraction of a
pixel and faded to nothing (peak ink 0.10-0.30 before, 0.58-0.92 after).
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_014TUcAN7jDYoL2LfGkKHNn5
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_014TUcAN7jDYoL2LfGkKHNn5
Co-authored-by: Claude noreply@anthropic.com
The pen-preview cursor read bigger than the line it draws: the radius was
floored at 1.5 logical px, the white hairline straddled the disc's edge,
and a translucent black halo disc sat underneath (also darkening a
translucent pen colour). With a pressure-sensitive stylus the dot also
stayed at the nominal width while the stroke drew 0.4x-1.6x of it.
The coloured disc is now exactly the stroke width at the current zoom,
the hairline and halo are rings drawn wholly outside it, and mid-stroke
the diameter follows the latest pressure sample through
pdfInkStrokeWidth, the same mapping the live stroke and the committed
appearance use.
Claude-Session: https://claude.ai/code/session_01YWuHqTtehKMybkjJQBisXC
Co-authored-by: Claude noreply@anthropic.com
feat: temporary pan with Space-drag and middle mouse (#1143)
feat: add focused Space and middle-button temporary hand panning
Route temporary mouse gestures ahead of editing overlays, reuse viewport pan math, and preserve tools and selections. Reset on key release, focus loss and cancellation; retain normal typing and context menus.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Expose original targets outside event dispatch and retain the original entries for ordinary input. Cover cached paths, transformed pointer events, form hit-testing and pan suppres
Note truncated.
Your coding agent can read these notes before it upgrades. Set up the MCP server →
One column per month.
Prepared by Alder, Ben's AI assistant.
Prepared by Alder, Ben's AI assistant.
Add GitHub and email feedback links to Settings > About
Add GitHub and email feedback links to Settings > About (#1109)
Keep an edited annotation painted when it is deselected before the page catches up (#1110)
Fix image crop chrome: constant-size chips, crop-style handles (#1108)
Fix image crop chrome: constant-size chips, crop-style handles
The crop confirm/cancel chips lived inside the viewer's zoom transform
without counter-scaling, so they grew with zoom (~130px at 386%) and sat
over the image. They now counter-scale by chromeScale and ride below the
crop (above, or inside, when the page runs out).
The normal selection chrome (tint, square knobs, rotate knob) kept
painting under the crop frame; it is now suppressed while cropping, and
the crop frame draws corner brackets and edge bars instead of square
knobs.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01Mor1BnC9MigXVgh9Ja3Dn9
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01Mor1BnC9MigXVgh9Ja3Dn9
Co-authored-by: Claude noreply@anthropic.com
Callouts get the nodes-only selection chrome (#1102)
Callouts get the nodes-only selection chrome
A lone selected callout now drops its box outline and rotate knob like a
selected line does, keeping its leader nodes and the text box's resize
handles. The painter gains a selectionOutline flag so the handles can sit on
the selection rect without the box.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01JvbRYftUbFgYACV1cc6RMm
_selectedViewRects listed a callout by its full /Rect while the primary
selection uses the text box, so the mismatch was painted as a second
selection box around the leader. With the primary outline gone it was the
only box left. Both now use calloutBox.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01JvbRYftUbFgYACV1cc6RMm
Co-authored-by: Claude noreply@anthropic.com
CAD sets that tile a sheet into thousands of small /Indexed Flate images
spent most of their render time expanding palette samples. _indexedToRgba
now expands through a 2^bits RGBA word table (one store per pixel), and a
display-sized decode of an opaque unmasked Indexed image averages palette
entries straight into the target cells instead of building and then
downsampling the native RGBA surface - byte-identical to the old
full-decode + downsample.
AOT: 503 tiles at display size 854 ms -> 173 ms; full decode 802 -> 263 ms.
Claude-Session: https://claude.ai/code/session_01KwVwtEGKBKXzmjjLg5Tf9C
Co-authored-by: Claude noreply@anthropic.com
The start card from #1084 stretched to the recents column's width (900px
in grid mode), so with no recents it became a wide grey slab around a
small icon and button. Cap it at 420px, centred, independent of the
recents column, and drop the "DartPDF" heading - the app bar and window
title already name the app.
Claude-Session: https://claude.ai/code/session_0162CZ53LBd5wb9wyiZufAg6
Co-authored-by: Claude noreply@anthropic.com
Take the pen dot down when a touch/stylus ink stroke ends (#1093)
Take the pen dot down when a touch/stylus ink stroke ends
A finger never hovers or exits the page, so nothing cleared the painted
pen cursor after a touch-drawn stroke - it stayed parked where the stroke
ended. The raw-pointer stroke end (and the two-finger bail) now clears
it; a hovering stylus re-places it on its next hover event. Mouse
strokes keep the dot at the stroke end as before.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01V7eQc9EqyZztCuaFWpB4zo
Flutter web follows a finger's pointerup with a synthesized hover of kind
touch at the lift point. MouseRegion delivers it through hit testing
(MouseTracker skips touch, RenderMouseRegion.handleEvent does not), so
_onHover put the pen dot straight back where the stroke ended - the
previous commit's clear on lift never survived on web. A finger can't
hover, so _onHover now returns for touch; that also keeps the eraser
ring, guides and placement previews from parking at a finger's lift.
Reproduced and verified in headless Chromium with touch emulation
against a release web build of the example.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01V7eQc9EqyZztCuaFWpB4zo
Co-authored-by: Claude noreply@anthropic.com
Open documents in Hand mode on mobile (#1113)
Link badge, word-level text edit, and a long-press page menu (#1112)
Fold-aware shell and iPhone Duo screenshot slots (#1111)
Fix "Cannot clone a disposed image" when OCR decodes a large scan (#1115)
Fix "Cannot clone a disposed image" in PdfBudgetedCache.putAndClone
putAndClone stored the value and only then cloned it. put() enforces the
PdfCacheRegistry process ceiling, whose hard trim may evict the entry it
just stored, so a decode big enough to exceed the ceiling on its own
(OCR rasterizes scans at 216 dpi, and iOS runs a small ceiling) was
disposed before the clone and decodeImages threw, failing the OCR job.
Clone before the store, and hand a value that rejectOversize would turn
away back uncached instead of cloning it and leaking the master.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01WnasbPSpLEkwbKJr6ds1Rf
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01WnasbPSpLEkwbKJr6ds1Rf
Co-authored-by: Claude noreply@anthropic.com
The first time automatic OCR adds a text layer, a one-time dialog asks
whether to keep it on for every scan or turn it off. Dismissing keeps it
on; the offer is persisted (and counts as made once the user picks the
Settings switch), so it is asked only once.
Claude-Session: https://claude.ai/code/session_01TfxQC46LgFdM6Eqavt6icV
Co-authored-by: Claude noreply@anthropic.com
PdfEditTool.measureCustom drags a dimension line (closed arrows both
ends) and prompts for the value it shows - pre-filled with the scaled
length when a scale exists, needing none otherwise. The value is kept in
/Takeoff /Value, becomes the caption and /Contents, survives reshapes,
and still totals (pdfParseMeasurementValue reads number + unit, incl.
feet-and-inches). Existing measurements get "Edit measurement value" in
the annotation context menu (PdfEditor.setMeasurementValue; blank goes
back to the computed value). The compact tools sheet now scrolls when a
group outgrows it.
Claude-Session: https://claude.ai/code/session_0183K9JEnHZwYgm1z417D8cK
Co-authored-by: Claude noreply@anthropic.com
Prepare the 10.0.0 package suite and app build 47, with companion dependency bumps, store notes and migration guidance. Correct nested Android CI signing, the CLI version report, and Linux/Flatpak/Microsoft release metadata.
Local validation: all 14 suite groups passed, including Chrome; release metadata, dependency floors and static analysis passed. Independent AI review was unavailable, not approved.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Correct the public demo version manifest after the 10.0.0 release. No native app or package code changes.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Reuse the existing portable Android compatibility preparation in Storeflight CI. Application sources, dependency constraints and release version remain unchanged.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Document dart_pdf_printing 0.7.0 for publication (#1122)
docs: add markdown pdfjs render gallery
Release dart_pdf_editor 3.7.1 with safe annotation disposal (#712)
Update AUR package for 3.6.0
Fix Linux store metadata for 3.6.0
Fix Linux launch crashes and add GUI smoke tests to CI (#693)
Linux: disable Flutter's experimental windowing (both the Dart feature
gate and the native headless bootstrap). The headless engine bootstrap
leaves GTK plugins such as desktop_drop registering against a null
FlView, which fails gtk_drag_dest_set and prevents window creation -
the snap (and any strict-confined launch) exited at startup.
Snap: add the dartpdf-dbus session-bus slot so strict confinement lets
the GApplication own dev.milanko.dartpdf instead of dying in AppArmor;
also check in the from-source snapcraft.yaml alongside the release
repack recipe.
CI: replace the non-blocking 'survive 8s' launches with
tool/smoke_test_linux_gui.sh, which fails the job on GTK-CRITICAL /
D-Bus AccessDenied output and verifies the DartPDF window actually maps
under an isolated Xvfb display. Wired into the Linux bundle, AppImage,
snap, and flatpak pipelines, plus new macOS and Windows GUI launch
checks in release-app.yml.
Co-authored-by: Claw (Bot) ben-claw@res.app
Prepare 3.6.1 release
Pin Linux packages to the 3.6.1 release
Fix snap release build: keep cli insertion point inside the apps section
The dartpdf-dbus slot block moved the CUPS comment away from the end of
the apps: mapping, so build-release-snap.sh's cli-app insertion landed
inside the top-level slots: block (snap warning: unknown interface
"cli") and the built snap lost dartpdf.cli.
New interface attributes (the dartpdf-dbus slot) send the uploaded
revision to human Store review; snapcraft exits non-zero even though
the upload succeeded and the revision releases on approval. Skip the
clean-install verification until the store actually serves the new
version.
Omit standalone Dart CLI from Mac App Store archives (#695)
Sanitize Mach-O private symbols from Dart AOT binary for App Store compliance
Omit Dart CLI from Mac App Store archives
Co-authored-by: dart-pdf release automation probe actions@github.com
Fix Windows open/save dialog crashes under the multi-window runner (#696)
Fix Windows open/save dialog crashes under the multi-window runner
file_selector_windows takes the dialog's owner HWND from the plugin
registrar's implicit FlutterView. DartPDF's Windows runner starts a bare
FlutterEngine so Dart owns every native window, so there is no implicit
view: registrar->GetView() returns null, the plugin dereferences it, and
the process dies at the first "Open PDF in new tab" or "Save as". It is an
access violation inside the method-call handler, so nothing on the Dart
side can catch or report it.
The runner already knows its active window without a view - it supplies one
to the print dialog and the image clipboard - so drive the common-item
dialogs from there instead:
Directory picks route through the channel too - the app does not use them
yet, but they were the plugin's other two null-view crash sites.
Drag-and-drop stays inert on Windows for the same underlying reason
(desktop_drop bails out when the registrar has no view); that needs a
per-window IDropTarget in the runner and is not part of this change.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_019pif2PSbZYWmho5beuSEu9
The MissingPluginException path for getDirectoryPaths was the one line
the new file left untested.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_019pif2PSbZYWmho5beuSEu9
Co-authored-by: Claude noreply@anthropic.com
Fix simple-font character mapping in content editing; add search-panel replace (#697)
Read and write simple-font text through its own encoding
A user editing a form's date field saw the content tool report the text
as -=>-/>?-?@ where the page reads 05/08/2026, and edits that
"changed a lot of content on the page". Two independent bugs.
A simple font's byte is not its character. PdfPageElements.of
decoded a non-/Type0 show string with latin1.decode, and
_SimpleRunCodec mirrored that: matching on String.fromCharCode(code)
and writing replacements as replace.codeUnits. That only holds when the
font's codes happen to be Latin-1. Section 9.6.6 lets a document remap
every code through /Encoding /Differences, and a subsetted font
routinely renumbers its codes from scratch - the reporter's font drew
05/08/2026 from the bytes 2D 3D 3E 2D 2F 3E 3F 2D 3F 40, which is
exactly -=>-/>?-?@ in Latin-1. Their "B = 8" note is the same table read
backwards: typing B wrote 0x42, and 0x42 is whatever glyph the
subset put there.
The renderer never had this problem (PdfFontInfo.charFor resolves
/ToUnicode, /Differences, the built-in Symbol/ZapfDingbats encodings, then
the base encoding), which is why selection and search read correctly the
whole time. Only the content editor had its own shortcut.
Adds SimpleFont, the simple-font sibling of Type0Font: one place that
resolves a code to text the way the renderer does, plus the reverse table
for re-encoding replacements and the /Widths lookup (which absorbs
content_editor's _widthsFor). Used by PdfPageElements, both run
codecs, the '/" operator path, and reflow's _SimpleReflowFont.
Encodability is deliberately narrower than decodability: decoding keeps
the lenient Latin-1 fallback, but a replacement may only use codes the
font declares (a glyph name from the base encoding or /Differences, or a
/ToUnicode entry), because a subset that dropped 9 has no glyph at
0x39. An undrawable replacement leaves the run untouched - the contract
Type0Font already had. A font that declares no encoding at all (base-14,
and everything this library writes) keeps the identity in both
directions, so existing round-trips are unchanged.
Fixing the decode also created a latent bug in paragraph reflow, fixed
here: content_reflow reads element.text, so once that decoded
correctly a remapped font would start matching, and showOp would have
written text.codeUnits as bytes.
"It changes it everywhere." Separately, the element strip's "Replace
text" went through the page-wide PdfEditor.replaceText. The targeted
replaceElementText already existed (it backs the selection-menu edit);
this path just never used it. Both element-strip entry points now share
one helper that uses the targeted API.
SimpleFont needs the Adobe glyph-name tables, which lived one layer up
in pdf_graphics. They are pure data, so they move down to
pdf_document/src/fonts/encodings.dart; the old path is now a show-list
re-export so the font engine's imports are untouched.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
PdfSearchResultsPanel(editing:) turns the find panel into find and
replace: a replacement field under the options bar with Replace (the
current hit alone) and Replace all (every hit, one undo step). Null
editing - the default, and what a read-only shell passes - leaves it a
pure find panel; the editor shell wires the session wherever the editing
toolbar is available.
Two controller methods back it:
replaceMatchText(page, rects, find, replace) resolves the hit throughtextElementForSelection - the same conservative gate the selectionreplaceTextInElement. It returns 0 rather thanreplaceTextOnPages(pages, find, replace) is the page-wide form forapply() so it is a single undo step.This only became possible with the preceding encoding fix: search hits
come from the text layer while elements come from PdfPageElements, and
before it those two disagreed about what a remapped run said, so no
search-driven replace could ever have resolved.
The re-search after an edit has an ordering trap worth noting. The viewer
calls clearSearch() when it swaps in the new revision, which happens on
the next frame - so a re-search issued straight after the edit is thrown
away moments later and the panel goes blank. _ReplaceBar._refresh awaits
endOfFrame first.
New strings land in the template ARB; the other locales fall back to
English until translated.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
CI's tool/check_arb_coverage.dart gate is stricter than gen-l10n: every
non-English ARB must carry the exact same keys as the template, precisely
so a half-translated locale can't ship looking finished while random
strings render in English. Leaving the five new keys English-only failed
that gate.
Adds them to all 19 locales, following each one's existing search
vocabulary (its own word for "match" from searchMatchCount, its own verb
for "replace" from tbReplaceText) and its own ICU plural categories -
few/many for pl/ru/uk, two/few/many for ar, other-only for the
CJK and other categoryless locales. Inserted at the template's position
rather than re-serializing, so each file's existing layout is untouched.
dart run tool/check_arb_coverage.dart now reports OK across all 57
locale files, and gen-l10n regenerates with no untranslated-message
warnings.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
Codecov flagged 11 uncovered lines in the new SimpleFont, all of them
branches where a mistake is invisible: /Encoding given as a name rather
than a /Differences dictionary, the built-in Symbol and ZapfDingbats
encodings, and the width fallbacks. Covering them found a real bug.
standardGlyphName deliberately reports only StandardEncoding's high
bytes - the ASCII range falls through to the Latin-1 identity, which is
the same answer for decoding. But declared was built from the glyph-name
table alone, so under /StandardEncoding or /MacRomanEncoding no ASCII code
was ever encodable: such a font read correctly and then silently refused
every ordinary replacement. Annex D defines all of printable ASCII for
every base encoding, so a named base encoding now declares 0x20-0x7E.
This deliberately does not loosen the subset case that motivated the
original gate: for an /Encoding dictionary, ASCII is declared only when the
document states a /BaseEncoding. A /Differences-only dictionary - the shape
in the bug report - still restricts writing to the codes it names, so a
character the subset dropped is still refused rather than drawn as notdef.
Tests added for all of it: /WinAnsiEncoding round-tripping a high-byte
glyph (eacute), /MacRomanEncoding, Symbol and ZapfDingbats (including a
subset-tagged one), a /ToUnicode that is junk, a /ToUnicode whose stream
cannot even be inflated, /MissingWidth for codes outside /FirstChar../
LastChar, and the Helvetica metric fallback. simple_font.dart is now at
100% line coverage; pdf_document is 835 tests green.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
The last 8 uncovered lines of the diff were the ' and " branch of
_replaceText - the one path this PR rewrote that no test reached. Those
operators carry a line break, so they stand alone and their string is
rewritten in place rather than through the run rewriter; the change made
both ends go through the font's own encoding, and nothing exercised it.
Three cases against the remapped subset font: ' rewriting its string into
the font's codes, " doing the same while keeping its word/char-spacing
operands, and ' declining a character the subset dropped.
Every line this PR touches in content_editor.dart is now covered (the file
as a whole goes 84.4% -> 91.0%), alongside simple_font.dart and
content_reflow.dart at 100%. pdf_document is 838 tests green.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
The replacement field and its two buttons sat permanently under the options
bar. That is a lot of vertical space to spend on a panel whose job is
listing hits, so they now hide behind a find_replace disclosure that
lives in the options row itself - right-aligned, so it reads as a
disclosure rather than a fifth match option, and styled by the extracted
PdfPanelToggleButton so it is indistinguishable from the ones beside it.
Collapsed by default; the state persists via
PdfEditingPreferences.searchReplaceExpanded, which the app shell already
passes to the panel. The replacement TextEditingController moves up to
the panel state, because the bar is removed from the tree when collapsed
and would otherwise take the typed text with it.
The padding was wrong in two ways, both worth spelling out.
The panel's content edge is 16 - the title, the count header, and the
result tiles all sit there. The options bar uses 8 and still looks right,
because its toggles are IconButtons with zero padding in a 30pt box, so
the glyph lands ~8pt in: 8 + 8 = 16. Copying that 8 for the replacement
field, which has a real border, put it 8pt left of everything else.
Content carrying its own inset takes 8; content without it takes 16.
Separately, both bars hard-coded their horizontal padding and ignored
geometry.contentStartInset/contentEndInset - the resize-grip clearance
the dividers already honoured - so the field ran under the grip on the
docked edge. Measured after the fix, on a 280pt panel with an 8pt grip:
title 16, field 16..256, buttons 16..256, count 16; and 16..264 as a
bottom sheet, where there is no grip.
The two dividers collapse into one below the whole control block, and the
field drops its find_replace prefix icon now that the disclosure carries
that glyph - it only cost width on a narrow panel. Button height stays 40:
VisualDensity.compact reaches the Material touch-target floor, and going
below it would trade tablet accessibility for a few pixels.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LVvqBwGCWDyVnxUMmpZ1Ry
Co-authored-by: Claude noreply@anthropic.com
Render pages as you scroll past them (#698)
Render ordinary pages through the scroll, and keep them
On a long text document the reader waited ~600 ms per page arrival while
the render hold sat out the whole scroll animation, then paid the same
cost again scrolling back because the retained-scene LRU kept four pages
for the entire document.
The hold is priced for a dense sheet's 100-400 ms interpret walk. With a
render worker attached that walk is not on the UI thread at all, so
STARTING a page during a scroll costs the frame nothing, and what does run
here - the replay of the returned buffer - has a size that is known by the
time it matters. PdfPageRenderScheduler gains a motion-safe lane on that
basis: a request may declare itself motionSafe and re-declare it at
paceUiWork once the record lands. PdfPageView decides in two halves - a
live worker and no page-level /XObject up front, then <=4000 commands and
no image draws when the buffer arrives - and revokes the verdict the moment
a pass falls through to a local walk. Grants during a hold stay near the
scroll's destination so a fling does not record every page it passes.
A page that finishes mid-scroll is usually not on screen yet, where the
direct picture presentation declines and the raster branch would burn a
full-page readback for pixels it replaces moments later; it now defers to
the existing off-screen visibility hook instead.
Retention pairs with it. The retained-scene LRU's byte estimate priced a
text page at ~0.5 MB against ~9 MB of measured browser heap, so the budget
governed nothing and misinformed PdfCacheRegistry. priceRetainedScene now
floors an entry at the raster it stands in for, the entry cap goes back to
being a backstop, and the budget becomes platform-aware alongside its
neighbours in performance_policy.dart.
Measured with a new read harness scenario (each page arrival, and the
wait after the scroll lands). read-text, 3 interleaved runs vs the base:
settle p50 207 -> 0.05 ms first visit and 196 -> 0.03 ms on the way back,
100% of revisits sharp on arrival, whole-arrival p50 614 -> 477 ms. Costs
~34 MB of retained scenes on a text document and a median frame build of
2.8 ms against 2.2. read-plan, the CAD arm, moves the same way for +5.6%
memory. bench.mjs now copies the Dart harness into the baseline worktree,
without which a new scenario silently compared two different workloads.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
The follow-up report was that a continuous scroll still waits: scroll two
pages, then ~0.5 s before anything is readable. The first round measured the
wrong journey - read navigates with jumpToPage, while a reader drags the
document past the viewport - so a new wheel scenario streams pointer scroll
signals at frame cadence and samples every frame which pages are visible and
which of those are sharp. It reproduced the report at once: settle 716 ms
after the scroll stopped, 15% of visible-page frames readable while it moved,
3 of 5 pages never sharp at all.
The trace showed the motion-safe lane granting correctly and the page the
reader actually landed on being granted only after the 500 ms scroll-quiet
window expired - because nothing had ever queued it. Nothing in a page's
render intent changes as it scrolls into view (onScreen is deliberately not
part of it) and the image-ratio promotion cannot fire for a page that has
never rendered, so the settle generation bump at the end of the scroll was
the only thing that queued a mounted cache-window neighbour. The lane cannot
grant what was never queued.
A page on screen with nothing of its own to paint now asks on every rebuild,
guarded by PdfPageRenderScheduler.isQueued so the ask cannot pile a second
pass behind a render already in flight. Level, not edge: the on-screen span
flips before the page is geometrically visible.
Also drops the off-screen raster deferral added earlier in this branch - it
was firing for pages scrolling into view and costing them an extra round
trip - and sizes the retained-scene budget to the speculative warm window,
with the measurement showing why a smaller one is wasted work rather than a
saving.
wheel-text vs the base, 3 interleaved runs: settle 716 -> 0.23 ms, sharp
while scrolling 15% -> 96%, a page's soft time p50 1322 -> 0 ms, no page left
unsharp, median frame build +3% with p95 better. wheel-plan (dense vector)
settle 864 -> 0.23 ms and tab memory down 32%. Attribution: with retention
off the scrolling fix still lands settle 0.2 ms at 68 MB, below the 75 MB
baseline - the memory cost belongs to retention, not to scrolling.
Harness fixes found on the way: bench.mjs now copies the Dart harness into
the baseline worktree (without it a new scenario silently compared against
scroll), and non-warm scenarios no longer inherit the warm scenarios' 1 GB
page-raster budget, which flattered revisits and inflated memory.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
A field trace from the desktop app showed the previous rounds had barely
moved it, and named two causes the browser harness could not reproduce.
The thumbnail strip owned the wait. At each settle it ran 400-530 ms of
tile interpret+raster on the platform thread ahead of the page the reader
landed on, whose own walk was 12 ms. The queue does stand down while the
viewer is busy, but it read busy once before the first tile and then ran
them back to back - and at that instant nothing was queued, so the frame
that would have queued the page could not run. The drain now yields a frame
between tiles and re-reads the gate (and the queue) after it.
That session also had no render worker at all - every interpret on the UI
thread - which made the motion-safe lane inert, since it required an
off-thread walk. So the lane gains a third class, and the boolean becomes
PdfRenderMotionClass evaluated at GRANT time rather than baked into the
request: a request made mid-fling is granted after the hand has stopped, and
the answer differs between those moments. free (worker-backed) runs during
any motion; quiet (no worker, small page) runs in the scroll-quiet window
but never inside a live gesture; held is unchanged. A live gesture needed
its own signal - _motionRenderHoldActive includes the 500 ms quiet window,
which is the thing the lane exists to bypass - so the viewer drives
activeGesture from a 120 ms timer restarted on every scroll event. A local
walk that overruns its budget marks that page held for good.
The interpret line now distinguishes recorded(no-worker) from
recorded(declined), so a field trace answers "did this session have a
worker" in one word.
Measured: wheel-text settle 0.2 ms and 96% sharp while scrolling hold; the
no-worker arm goes 612 -> ~410 ms, and that residue is three UI-thread walks
rather than waiting. Full suite 2338 passing, counter gate clean.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
Chasing a field trace that showed every page interpreting on the UI thread
turned up why that state is so hard to diagnose: a failed isolate spawn is
caught with catch (_), sets _spawnFailed, and from then on every record()
resolves to null and every page walks on the UI thread - permanently, for
that document, with nothing in any log. A trace of that session reads exactly
like one where the worker simply declined a page.
It now logs to PdfPerfLog and, in debug, prints a plain-language warning
naming the cause. With the recorded(no-worker) / recorded(declined) split
already on the interpret line, a trace answers "did this session have a
worker, and if not why" on its own.
No behaviour change: the fallback path is untouched.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
The motion-safe lane refused any page that declared an /XObject, on the
reasoning that a three-operator content stream can hide a hundred
megapixels. It can - but the same rule also refuses the 42x10 letterhead
mark that every page of a corporate report carries. A field trace of one
such document (113 pages, a mark on each) shows the lane switched off for
the whole file: every scheduler grant reads cost=held.
An image's size is a dictionary entry, so measure it. The request-time gate
now sums the declared /Width x /Height of a page's image XObjects against
motionSafeMaxImagePixels (1 MP ~ 4 MB of RGBA), and the record-time gate
weighs the buffer's actual image draws the same way
(PdfPageRenderer.imageDrawPixels), which also covers inline images and Form
XObject contents nothing at page level advertises. Anything that will not
say how big it is - a Form XObject, an image with no declared dimensions, an
unresolvable resource - still answers "held", exactly as before.
Second half of the same trace: scene-cache reject reason=image-lod have=1.79 need=2.00. Scrolling back onto a page threw away its retained
display list over a 10% difference in embedded-image resolution - blank
paper plus a re-interpret for something no one can see. What a cached scene
has to clear is display sharpness at the current zoom, not the 2x decode
headroom a fresh render aims for, so it is compared against _effectiveRatio
now; a genuinely soft prefetch buffer is still rejected, and the headroom
arrives as the existing soft-to-sharp refinement behind the painted page.
The corpus had no document of this shape - the ordinary middle between
text-only and full-page raster - so letterhead-report-40p.pdf joins it, with
wheel-letterhead / wheel-letterhead-noworker scenarios.
Measured (tool/perf.sh webdiff, 3 interleaved iterations, wheel-letterhead):
settle 644ms -> 0.22ms, sharp-while-scrolling 15% -> 96%, pages never sharp
3 -> 0, soft p50 1300ms -> 0, frame build p95 -24%. Cost: 72 MB -> 173 MB of
tab memory, the retention trade this branch already priced - the baseline
held nothing because it drew nothing. With the worker off, the settle halves
(621ms -> 360ms) and sharpness during the fling is unchanged, which is that
path's ceiling.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_015kkAUG6UA5bFDN2JqFqzfF
Co-authored-by: Claude noreply@anthropic.com
Lockstep minor for the eight core packages (3.6.0 -> 3.7.0), with
dart_pdf_cli 0.1.0 -> 0.1.1 and dart_pdf_editor_flutter_gpu 0.1.4 -> 0.1.5
for their constraint bumps, and the app at 3.7.0+31.
No breaking public API changes since 3.6.0: PdfPageRenderScheduler.request
kept its positional parameters and only gained an optional named one, and
the glyph-name tables that moved from pdf_graphics to pdf_document were
never part of pdf_graphics' public surface (src/fonts/encodings.dart is now
a show-list re-export).
Changelogs authored from the commit bodies of #695-#698, plus the store
"What's New" text for 3.7.0 in all 20 store locales.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Fix release-mode annotation picture disposal
chore(release): prepare dart_pdf_editor 3.7.1
chore(release): align editor dependency constraints
Co-authored-by: dart-pdf release automation probe actions@github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: Claude noreply@anthropic.com
Release DartPDF 4.3.0
Fix StoreFlight isolated checkout inputs
Add 4.3.0 Linux release metadata
Release DartPDF 7.0.0 and versioned companions
Add DartPDF 7.0.0 AppStream release notes
Release DartPDF 8.0.0 across packages and apps
Enforce release package order and serialize local Flutter builds
Release DartPDF 8.1.0 across packages and apps
Keep canonical Linux release metadata in sync for 8.1.0
Promote dartPDF 10.0.0 for Storeflight publishing [skip ci]
Promote the tested dartPDF 10.0.0 release snapshot for Storeflight dependency-ordered package publishing. The promoted tree is identical to main 9d4b26c; the release and CI-repair PR checks passed. Ben explicitly authorised this administrative merge on 19 October 2026. Skip the separate automatic publisher to preserve the existing Storeflight release owner.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Identical file tree to deploy cf453de and tested main 9d4b26c. A fresh metadata-only commit permits tag-triggered publication. No application source, package contents, version or workflow changes.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claude Opus 4.8 (1M context) noreply@anthropic.com
Co-authored-by: dart-pdf release automation probe actions@github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: Ben Milanko 20643387+ben-milanko@users.noreply.github.com
Select installed App Store distribution profiles and explicit macOS export options for hosted Apple builds. Wait for GPU submission completion before attachment-reuse assertions.
Exact-head CI, preview and native Patrol checks passed. Preserve the paused iOS release and its existing product-page test.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Select the installed Xcode 26.3 toolchain on macos-15 and fail before build/upload when the iPhoneOS SDK is too old.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
build(deps): bump actions/setup-java from 4 to 6 in the actions group (#1141)
fix(editor): autosize text boxes in rotated page coordinates (#1127)
Keep the visible top-left corner and paragraph width when fitting free text on rotated pages. Cover all four page rotations, callouts and Alt+Z.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
fix: preserve snapshot orientation on rotated pages (#1128)
fix: preserve snapshot orientation on rotated pages
Counter-rotate vector snapshot appearances for the destination page and place natural-size pastes in display coordinates. Keep explicit paste points and make repeated pastes cascade visibly down-right.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Use R for clockwise quarter-turns and Shift+R for reverse turns. Keep rectangle R without a rotatable selection, custom bindings, and text-editor ownership.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat: vertically align text inside text boxes (#1129)
feat: vertically align text inside text boxes
Add top, middle and bottom placement for plain and rich free text. Preserve placement through rendering, edits, resize, callouts, scoped defaults, inline editing and save/reopen.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Defer the synchronous page parse until the initial preview frame and skip stale or ended drags. Add dense-page regression coverage and a reproducible Chrome resize-start benchmark.
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(editor): find and replace colours in snapshots and annotations (#1132)
feat(editor): selectively replace snapshot and annotation colours
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Fixes #1124
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Fixes #1125
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(app): resumable OCR with saved-page checkpoints (#1139)
feat(app): persist OCR jobs and continue background processing
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Remove detached workers and mobile background hosting. Checkpoint completed pages and resume when the app reopens or returns to the foreground.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(editor): duplicate library stamps and select imports/exports (#1135)
feat(editor): duplicate library and placed stamps independently
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
feat(editor): editable tables and Excel clipboard transfer (#1140)
feat(editor): add editable tables and Excel cell copying
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Use a Table grid with full-content rows, shared column widths and cell keyboard navigation. Rebuild table appearances at constant font size during resizing and retain reflowed drag previews and undo.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Preserve table colours, standard font variants and grid settings across editing, resizing and save/reopen. Reuse existing font and colour controls and cover saved appearance pixels, mobile layout and plain hosts.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Use existing viewer input paths to edit selected or unselected tables in default and Select modes. Preserve protected-table guards and undoable Apply behavior.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claw (Bot) ben-claw@res.app
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Setup ran app-processes.ps1 under 32-bit PowerShell (SYSDIR is SysWOW64
for a 32-bit process), where Process.Path is empty for a 64-bit process,
so the running-app count read 0 and the locked-exe retry box came back.
Use Sysnative PowerShell, and fall back to CIM's ExecutablePath.
Claude-Session: https://claude.ai/code/session_014nwq5uZzoq6CgE3iHtfRyS
Co-authored-by: Claude noreply@anthropic.com
Floor thin rectangle fills at one device pixel (#1146)
Floor thin rectangle fills at one device pixel
Word and similar producers draw table and box borders as filled 0.5 pt
rectangles (re f), not strokes. Strokes already had a one-device-pixel
floor; fills did not, so zoomed out those borders covered a fraction of a
pixel and faded to nothing (peak ink 0.10-0.30 before, 0.58-0.92 after).
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_014TUcAN7jDYoL2LfGkKHNn5
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_014TUcAN7jDYoL2LfGkKHNn5
Co-authored-by: Claude noreply@anthropic.com
The pen-preview cursor read bigger than the line it draws: the radius was
floored at 1.5 logical px, the white hairline straddled the disc's edge,
and a translucent black halo disc sat underneath (also darkening a
translucent pen colour). With a pressure-sensitive stylus the dot also
stayed at the nominal width while the stroke drew 0.4x-1.6x of it.
The coloured disc is now exactly the stroke width at the current zoom,
the hairline and halo are rings drawn wholly outside it, and mid-stroke
the diameter follows the latest pressure sample through
pdfInkStrokeWidth, the same mapping the live stroke and the committed
appearance use.
Claude-Session: https://claude.ai/code/session_01YWuHqTtehKMybkjJQBisXC
Co-authored-by: Claude noreply@anthropic.com
feat: temporary pan with Space-drag and middle mouse (#1143)
feat: add focused Space and middle-button temporary hand panning
Route temporary mouse gestures ahead of editing overlays, reuse viewport pan math, and preserve tools and selections. Reset on key release, focus loss and cancellation; retain normal typing and context menus.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Expose original targets outside event dispatch and retain the original entries for ordinary input. Cover cached paths, transformed pointer events, form hit-testing and pan suppres
Note truncated.
Promote reviewed DartPDF 12.0.0 snapshot to deploy
Promote reviewed DartPDF 12.0.0 snapshot to deploy (#1171)
Identical file tree to deploy cf453de and tested main 9d4b26c . A fresh metadata-only commit permits tag-triggered publication. No application source,
Identical file tree to deploy cf453de and tested main 9d4b26c. A fresh metadata-only commit permits tag-triggered publication. No application source, package contents, version or workflow changes.
Co-authored-by: ben-milanko 20643387+ben-milanko@users.noreply.github.com
Bump the lockstep packages to 9.0.0 (PdfRenderWorker.record gained the decodeGate parameter, which breaks custom worker overrides - see doc/MIGRATING-
Bump the lockstep packages to 9.0.0 (PdfRenderWorker.record gained the
decodeGate parameter, which breaks custom worker overrides - see
doc/MIGRATING-9.0.0.md), dart_pdf_cli and dart_pdf_printing to 0.6.0 and
dart_pdf_editor_flutter_gpu to 0.8.0, which now needs the 9.0.0 editor and
graphics APIs. App 9.0.0+46.
Changelogs, store notes for 20 Apple and 21 Play locales, Microsoft listing
notes and AppStream releases. Upgrade locked dependencies. Regenerate the
Storeflight workflow for the current CLI, naming its own repository as the
CLI source. Allow the antialiased edge column macOS counts as ink in the
substituted-glyph width test.
Co-authored-by: Ben Milanko 20643387+ben-milanko@users.noreply.github.com
Co-authored-by: Claude Opus 5.5 noreply@anthropic.com
Keep canonical Linux release metadata in sync for 8.1.0
Keep canonical Linux release metadata in sync for 8.1.0
Enforce release package order and serialize local Flutter builds
Enforce release package order and serialize local Flutter builds
Add DartPDF 7.0.0 AppStream release notes
Add DartPDF 7.0.0 AppStream release notes
Describe 6.0.0 in Linux release metadata
Release DartPDF 6.0.0
Describe 6.0.0 in Linux release metadata
Co-authored-by: dart-pdf release automation probe actions@github.com
CosDocument.trimDecodedStreamCache() to release decoded payloads
without closing the document or changing subsequent decode results.Bump the app to 5.1.1+41 and the lockstep package suite to 5.1.1 (dart_pdf_cli 0.2.2, dart_pdf_editor_flutter_gpu 0.4.2, dart_pdf_printing 0.2.2), wit
Bump the app to 5.1.1+41 and the lockstep package suite to 5.1.1
(dart_pdf_cli 0.2.2, dart_pdf_editor_flutter_gpu 0.4.2, dart_pdf_printing
0.2.2), with every runtime inter-package constraint at ^.
This is a patch release: no public API breaks since 5.1.0, and one addition
integrators may see:
Write the 5.1.1 changelog sections from the one commit since app-v5.1.0,
the ink afterimage fix (#986): quick back-to-back ink commits each hid the
previous one on screen until the page re-rendered, because the editing
overlay kept only the latest commit as its afterimage. Commits now
accumulate and retire per page once the raster and annotation layer are
current. The other packages carry a version-bump entry. Dev dependency
ranges (<6.0.0) already admit 5.1.1.
Store notes cover the ink fix and note that 5.1.1 includes 5.1.0's
improvements, written platform-neutral per release-notes/README.md, across
20 iOS locales, 21 Play changelogs (41.txt) and both Linux metainfo copies.
The forms guide now says it applies to 5.1.1.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Bump the app to 5.1.0+40 and the lockstep package suite to 5.1.0 (dart_pdf_cli 0.2.1, dart_pdf_editor_flutter_gpu 0.4.1, dart_pdf_printing 0.2.1), wit
Bump the app to 5.1.0+40 and the lockstep package suite to 5.1.0
(dart_pdf_cli 0.2.1, dart_pdf_editor_flutter_gpu 0.4.1, dart_pdf_printing
0.2.1), with every runtime inter-package constraint at ^.
This is a minor release: no public API breaks since 5.0.0, with two things
integrators should know:
Write the 5.1.0 changelog sections from the commit bodies since
app-v5.0.0: the P-521 order fix, faster ECDSA, tolerant zlib inflate on
the web, the object-cache key and its side map, T-table AES, web SHA-2,
JPEG 2000, in-place updates on recovered documents, rollbackTo and
openAppended(password:) in pdf_cos; the page-loss fix under a short /Count,
linear page loops, reusable signature crypto cores and onFields in
pdf_document; the flagged masked-image output change, codec v10/v11,
cached luminosity masks, the revision-stable colour context, ICC, colorant
buffer and text-extraction speedups in pdf_graphics; trackpad signatures,
annotation previews, touch row menus, the moved-stamp white box, the web
worker fixes and in-place revisions in dart_pdf_editor; the new overprint
and ICC fixtures. Dev dependency ranges (<6.0.0) already admit 5.1.0.
Store notes cover trackpad signatures, annotation previews, touch row
menus, the moved-stamp fix, the page-loss fix, P-521 signatures and the
speedups, written platform-neutral per release-notes/README.md, across 20
iOS locales, 21 Play changelogs (40.txt) and both Linux metainfo copies.
The forms guide now says it applies to 5.1.0.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
u1*G + u2*Q, and curve
constants parsed once per curve. P-256 verify goes from 11.3 to 1.3 ms
(AOT). Signing uses the same arithmetic and stays byte-identical (RFC 6979).
verifyCertificateChain now verifies each certificate against its issuer
once instead of twice.ZLibDecoder read any bytes after the Adler-32 (a trailing EOL before
endstream) as a second zlib header, failed it and returned nothing for
the whole stream: pages rendered blank, fonts went missing, and
CosCompactor (Reduce file size) re-deflated the empty result into the
file. The new inflateZlib stops at the final block, as zlib does, and
FlateFilter uses it. CosCompactor inflates strictly (header, length and
Adler-32 checked) and never replaces a payload of more than 16 bytes with
an empty one. The VM path is unchanged.objectNumber * 65536 + generation) clustered in the VM's int hash, so
whole-graph walks spent most of their time probing: PdfCompressor.optimize
runs 0.39-0.69x and applyRedactions about 0.31x on 14k-37k-object files.
References the packed key cannot hold exactly (an object number of 2^32 or
more, as an edit on a file with a junk /Size allocates, or a generation
over 65535) are cached in a side map, which also stops n 65536 R
resolving to object n+1.applyIncrementalUpdate now folds edits into a document opened through
xref recovery instead of refusing it, so each edit no longer re-runs a
full-file recovery scan. The recovery scan itself is one pass for both
obj and trailer (0.56-0.61x on large files).CosDocument.rollbackTo(length), which takes a document back to an
earlier revision it folded in with applyIncrementalUpdate, in place,
keeping the caches of objects the undone updates did not touch. It returns
null, changing nothing, when that length is not journaled.CosDocument.openAppended accepts an earlier revision's prefix as well as
an append, and takes a password for revisions it cannot donate keys to.
It donates the security handler only while the revision's /Encrypt matches
the dictionary the keys were derived from, entry for entry, so a revision
that re-keys /Encrypt under the same object number authenticates its own
password.ContentOperationCursor.nextOperator and takeOperation, and
CosTokenBuffer.keywordCode, so an interpreter can dispatch short
operators on an int code and read numeric operands without allocating
operation objects. nextOperation and parse keep their contract, and
materialized operands follow the same int-or-real rule on the web as on
the VM.PdfPerf counters for glyph outline builds and the overprint colorant
buffer (glyphOutlinePaths, colorantBufferPages, colorantDraws,
colorantRasterized, colorantBackdropReads, colorantGroups).Roll the Unreleased sections over with Breaking changes subsections and migration notes, and write the ones the batch never got: AF* scripts, comb/Max…
Bump the app to 5.0.0+39 and the lockstep package suite to 5.0.0
(dart_pdf_cli 0.2.0, dart_pdf_editor_flutter_gpu 0.4.0, dart_pdf_printing
0.2.0 - their dependency majors moved under types they expose). This is a
major release because public behaviour and API changed incompatibly since
4.5.0:
Roll the Unreleased sections over with Breaking changes subsections and
migration notes, and write the ones the batch never got: AF* scripts,
comb/MaxLen/password fields, XFA detection, encrypted signing and tab
order in pdf_document; openAppended in pdf_cos; the /ImageMask box filter,
justified-text reflow spaces and multi-select field context in
pdf_graphics; the form-layer, secret-store, Tab navigation and
thumbnail drag-out work in dart_pdf_editor; multi-select values in the
CLI; the print range and copies fixes in dart_pdf_printing; the new
fixtures. Dev dependency ranges on workspace packages now admit <6.0.0 so
the workspace resolves and publishing still accepts the hosted 4.5.0.
Store notes cover form calculations and checks, Tab and multi-select,
password fields kept out of the file, field creation, signing encrypted
PDFs, revocation checks with EU/Adobe trust lists, and Calibri via
Carlito, across 20 iOS locales, 21 Play changelogs, and both Linux
metainfo copies. The forms guide now says it applies to 5.0.0.
Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
StandardSecurityHandler.decryptObjectGraph (and encryptObjectGraph)
now leave a signature dictionary's /Contents untouched, as ISO 32000 7.6.1
requires (StandardSecurityHandler.isSignatureContents: /Type /Sig or
/DocTimeStamp, or a dictionary with a /ByteRange when /Type is missing).
Code that decrypted /Contents itself after loading must stop doing so - the
loader now hands over the raw CMS bytes.OcspResponse.responderFor), and
OcspResponse.forCertificate matches the whole CertID. Responses that used
to pass on a serial match alone, or from an unauthorized responder, are now
ignored.X509Certificate gained members (listed below). Only classes that
implements X509Certificate need to add them.CosDocument.openAppended, which reopens a document after appended
updates and reuses its already-authenticated security handler (while the
/Encrypt object is unchanged), so follow-on revisions of an encrypted file
do not need the password again.OcspResponse.forCertificate matches
the whole CertID (serial, issuer name hash and issuer key hash) instead of
the serial alone, a delegated responder must be issued and signed by the CA
and carry id-kp-OCSPSigning (responderFor), and the nonce extension and
revocation reason are parsed (nonce, echoesNonce, ocspNonceValue).X509Certificate exposes extendedKeyUsages, isCa, hasOcspNoCheck and
isValidAt, and now reads the RSA key of an id-RSASSA-PSS certificate.issueCertificate can add OCSP (AIA), CRL distribution point, extended key
usage and ocsp-nocheck extensions, and can issue a subordinate CA (isCa).Bump the app to 4.5.0+38 and the lockstep package suite to 4.5.0 (dart_pdf_cli 0.1.8, dart_pdf_editor_flutter_gpu 0.3.4, dart_pdf_printing 0.1.1). Not
Bump the app to 4.5.0+38 and the lockstep package suite to 4.5.0
(dart_pdf_cli 0.1.8, dart_pdf_editor_flutter_gpu 0.3.4, dart_pdf_printing
0.1.1). Nothing public was removed or changed incompatibly since 4.4.0 -
showPdfShellViewOptionsSheet lost its reflow/pageGrid parameters, but it is
not exported - so this is a minor release.
Roll the Unreleased sections over and write the ones the 4.4.0..HEAD batch
never got: WinAnsi widths and the DocMDP P=1 signing refusal in pdf_document,
CJK selection metrics in pdf_graphics, selection continuity, the stroked-text
floor, stem darkening and the page-grid State fix in dart_pdf_editor.
Store notes cover What's new, the Linux antialiasing fix, exported pages
opening in a tab, the view-mode picker, metric-compatible standard-14
substitutes, and selection and stroked-text fixes, across 20 iOS locales,
21 Play changelogs, and both Linux metainfo copies.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Storeflight and release-app.yml both created the release for an app-v tag. Storeflight pushes that tag, so its own delivery and the workflow the tag s
Storeflight and release-app.yml both created the release for an app-v tag.
Storeflight pushes that tag, so its own delivery and the workflow the tag
starts were racing to own the same release, and would have attached two sets
of assets for the same platforms under different names.
Storeflight wins, with one exception it cannot cover. A release artifact
there is a single path per platform, and Linux ships two user-facing files -
the portable tarball and the AppImage. Dropping either is not an option:
publish-flatpak.yml downloads dartpdf-linux-x64.tar.gz from the release by
name and verifies its digest, so the Flatpak publisher fails a step later
without it.
So release-app.yml stops creating a release and contributes only the two
Linux assets. It waits for Storeflight's release rather than creating one,
because publishing the tag is what starts it and it can arrive first;
creating one there would be the same race with the roles reversed. It also
asserts the tarball is present under the name the Flatpak publisher looks
for, so a rename fails here rather than ten minutes downstream. The job now
needs only the Linux build, since waiting on the other six delayed that input
for no benefit.
publish-flatpak.yml waits for the asset for the same reason: two workflows
started by one tag have no ordering between them, and an asset still
uploading is not a missing build.
Co-Authored-By: Claude Opus 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01RNNTEpKG3vAEwMYMvghdvx
Add 4.3.0 Linux release metadata
Add 4.3.0 Linux release metadata
CosCompactor, retaining predictor
data and parameters. Preserve cyclic indirect arrays and PDF 2.0 headers.Lockstep minor release aligned with dart_pdf_editor 4.2.0. No public COS object model, filter, or parsing changes since 4.1.0.
dart_pdf_editor 4.2.0. No public COS
object model, filter, or parsing changes since 4.1.0.Treat holes in sparse progressive byte sources as cache misses tied to the source buffer, avoiding invalid scan recovery and stale reads while remote
archive to 4.2.0.Decode TIFF Predictor 2 streams with 16-bit samples, preserving carries between bytes and restoring high-bit-depth images correctly.
pdf_cos
public API remains source-compatible with 3.8.0.Add the public monotonic CosDocument.revision, incremented after every successful incremental update so higher layers can invalidate derived caches wi
CosDocument.revision, incremented after every
successful incremental update so higher layers can invalidate derived
caches without comparing object identity.Lockstep minor release to align the dart-pdf package suite at 3.7.0. No public pdf_cos API changes since 3.6.0.
pdf_cos API changes since 3.6.0.Lockstep minor release to align the dart-pdf package suite at 3.6.0. No public pdf_cos API changes since 3.5.1.
pdf_cos API changes since 3.5.1.Decode JBIG2 refined text symbols, including refinement deltas, offsets, and shared arithmetic contexts, so scanned MRC text layers render correctly.
Add PdfSourceLoadOptions.completeFirstPaintPageTree. Progressive preview callers can stop the initial page-tree walk after firstPaintPages, avoiding o
PdfSourceLoadOptions.completeFirstPaintPageTree. Progressive preview
callers can stop the initial page-tree walk after firstPaintPages, avoiding
one range request per leaf before page one can paint; the default remains the
correctness-first complete walk for existing callers.Lockstep minor release to align the dart-pdf package suite at 3.4.0. No public pdf_cos API changes since 3.3.1.
pdf_cos API changes since 3.3.1.Lockstep patch release to align the dart-pdf package suite at 3.3.1. No public pdf_cos API changes since 3.3.0.
pdf_cos API changes since 3.3.0.Lockstep minor release to align the dart-pdf package suite at 3.3.0. No public pdf_cos API changes since 3.2.0.
pdf_cos API changes since 3.2.0.Lockstep minor release to align the dart-pdf package suite at 3.2.0. No public pdf_cos API changes since 3.1.1.
pdf_cos API changes since 3.1.1.Lockstep patch release to align the dart-pdf package suite at 3.1.1. No public pdf_cos API changes since 3.1.0.
pdf_cos API changes since 3.1.0.Lossless structural compaction: new CosCompactor/CosCompactionResult rewrite a document's reachable object graph, packing non-stream objects into comp
CosCompactor/CosCompactionResult
rewrite a document's reachable object graph, packing non-stream objects
into compressed object streams behind a PDF 1.5 xref stream and
re-deflating streams stored uncompressed (kept only when smaller).
CosDocumentBuilder gains an objectStreams: mode (W [1 4 2]). The
user-facing entry point is PdfEditor.compress() in pdf_document (#368).cp_reduce
level when an image is displayed far below its native size, instead of
decoding full resolution and downscaling (#525)./JBIG2Globals decode once, not per image (#532).Lockstep major release: a breaking change in dart_pdf_editor moves the whole suite to 3.0.0. pdf_cos's own public API is unchanged.
Lockstep major release: a breaking change in dart_pdf_editor moves the whole
suite to 3.0.0. pdf_cos's own public API is unchanged.
CosDocument so a stream inflated once (xref
reconstruction, content parsing, image extraction) is not re-decoded on the
next access (#392).Append incremental saves to the existing bytes instead of rebuilding the whole file: CosUpdater.saveTail() returns only the new tail (the appended obj
Append incremental saves to the existing bytes instead of rebuilding the
whole file: CosUpdater.saveTail() returns only the new tail (the appended
objects, xref, and trailer), so an incremental revision costs the size of
the change rather than the size of the document. Save cost no longer scales
with the base file - a guard test pins that a 2.78x larger base does not
make a same-sized edit 2.78x more expensive (#413).
Memoise the per-object decryption key so a page's streams and strings derive it once per object rather than once per access, cutting repeated MD5/AES key derivation on encrypted documents (#400).
Speed up CCITT G3/G4 decoding by ~4x on dense scanned pages: a monotonic cursor for the 2-D reference-row scan (was O(transitions^2) per row), peek-once prefix tables for run and mode codes, byte-run span fills, and a byte-at-a-time bit reader. The JBIG2 MMR path, which reuses the same decoder, gains a byte-at-a-time bitmap unpack. Output is byte-identical on well-formed, malformed, and truncated input, locked by golden digests captured from the previous implementation (#398).
Major version bump for the 2.0.0 package suite. A breaking API change in dart_pdf_editor moves every package to 2.0.0 in lockstep; the COS object mode
dart_pdf_editor moves every package to 2.0.0 in lockstep; the COS object
model, syntax, and (de)serialization API is source-compatible with 1.4.7.PdfByteSource) for progressive PDF
loading: the reader pulls ranged slices on demand instead of requiring the
whole file up front, so remote and large local files can paint their first
page before the full download or read completes (#328, #359).EcPrivateKey.generate, deterministic RFC 6979 ecdsaSign, the X.509
v3 builders buildSelfSignedCertificate / buildCaCertificate /
issueCertificate, and ecSubjectPublicKeyInfo / pemEncode for
Sigstore/Fulcio (#322, #337, #355).CosXrefReader from
CosDocument, move the encryption object-graph walk behind the security
handler, share file-tail framing between the builder and updater, and dedupe
ObjStm header parsing between recovery and the stream decoder.Version bump to keep the dart-pdf package suite aligned at 1.4.7. No COS API changes since 1.4.6.
Version bump to keep the dart-pdf package suite aligned at 1.4.6. No COS API changes since 1.4.5.
Version bump to keep the dart-pdf package suite aligned at 1.4.5. No COS API changes since 1.4.4.
Version bump to keep the dart-pdf package suite aligned at 1.4.4. No COS API changes since 1.4.3.
Version bump to keep the dart-pdf package suite aligned at 1.4.3. No COS API changes since 1.4.2.
Version bump to keep the dart-pdf package suite aligned at 1.4.2. No COS API changes since 1.4.1.
Speed up dense content parsing with byte-level real-number parsing, operator interning, and streaming content-token handling.
Version bump to keep the dart-pdf package suite aligned at 1.4.0. Low-level parser, writer, filter, and crypto maintenance supports the higher-level e
Version bump to keep the dart-pdf package suite aligned at 1.3.2. No COS API changes since 1.3.1.
Add ContentStreamSerializer for writing parsed content-stream operations back to PDF syntax, including inline-image (BI/ID/EI) operations.
ContentStreamSerializer for writing parsed content-stream operations
back to PDF syntax, including inline-image (BI/ID/EI) operations.Nothing published for this version
Version bump to keep the dart-pdf package suite aligned at 1.2.3. No COS API changes since 1.2.2.
Version bump to keep the dart-pdf package suite aligned at 1.2.2. No COS API changes since 1.2.1.
Add a package example for pub.dev scoring.
Version bump to keep the dart-pdf package suite aligned at 1.2.0. No low-level COS API changes since 1.1.0.
Performance: a faster content-stream tokenizer. This is the heart of the render-speed work that puts dart-pdf ahead of PDFium on the benchmark corpus.
First stable release. Changes since 0.1.0:
First stable release. Changes since 0.1.0:
COS object model: dictionaries, arrays, names, strings, streams, references.
Initial release.
endobj, junk before the header, broken xref chains with object-scan
recovery).