Skip to content

Sync main into builtin-engines (2026-11-19) - #1669

Merged
KnifeLemon merged 154 commits into
builtin-enginesfrom
p4/sync-2026-11-19
Nov 18, 2026
Merged

KnifeLemon merged 154 commits into
builtin-enginesfrom
p4/sync-2026-11-19

Conversation

@KnifeLemon

@KnifeLemon KnifeLemon commented Nov 18, 2026 •

Copy link
Copy Markdown
Owner

Merges main (32ea617, 152 commits) into builtin-engines (4df0682). Merge base: 770e982, the previous sync (#1529). No change goes to main.

What comes over from main: the 2.5.0 release with its security hardening (Setup EoP fixes, MOBI / ODS / table / path-depth limits, re-zipped copies counted against the unpack limit), RAW tones (#1560), MSIX / APPX (#1585), GMA / VPK / BSP unpacking (#1591) and GMA packing (#1608), the VTF preset (#1611), .cxlog (#1636), UDF 2.01 ISOs (#1491), the updater mirror (#1638), the PDF page editor (#1641, #1660), the Windows ARM64 release, the CI split into build.yml (#1553) and the rest.

Settings schema

Main's library steps keep 95–99 (MSIX, GMA / VPK / BSP, GMA preset, VTF preset, cxlog). The built-in engines steps move above them, in their order, so users on 2.5.0 and the next main release run them on upgrade:

step was now
unhwp → hwp-text in the engine order 95 100
LibreOffice for HWP → Filee renderer, LibreOfficeForHwp cleared 96 101
Filee renderer on, LibreOffice for slides off (once) 97 102
  • CurrentSchemaVersion = 102.
  • Tests renamed with them: Version_99_settings_put_the_built_in_hwp_text_engine_where_unhwp_was, Version_100_settings_drop_libreoffice_for_hwp…, Settings_before_102_… (InlineData 99 = main's last, 101 = a branch build).
  • Doc comments, docs/ENGINES.md and docs/RENDER-QUALITY.md say 100/101/102.
  • A preview build already at 97 skips main's library steps 95–97 (MSIX, GMA / VPK / BSP, GMA preset) and runs the branch's steps once more (all three are idempotent apart from turning the renderer on and slides off again). Harmless for unreleased builds.

Conflicts (38 files) and how they were resolved

Archives (ArchiveExtractor.cs, ArchiveConverter.cs, ArchivePacker.cs, ArchiveSafety.cs, ArchiveTests.cs)

  • MSIX / APPX, GMA, VPK and BSP join the built-in formats; MSIX goes through the in-process ZIP path like NUPKG (and 7-Zip's retry, when installed).
  • GMA and VPK are read by main's readers inside ExtractInProcess. A BSP's pakfile is copied out and unpacked through the in-process ZIP path; a PS Vita .vpk (a ZIP) too. None of them needs 7-Zip.
  • GMA is written in-process (packer, folder → GMA edge, Gma.Root for packs from a folder).
  • Both sides had added a 512-folder path depth limit: one MaxPathDepth, IsTooDeep is the branch's Depth check, and hard-link targets are checked too (main). The branch's refusal (final, no 7-Zip retry) is kept.
  • Main's package test gets broken.msix under the branch's test name.

Fixes on top of main's new readers (Gma.cs, Vpk.cs, Bsp.cs)

  • ArchiveFiles.UniquePath doesn't exist on the branch (replaced by UniquePaths, which avoids the quadratic probe): GMA and VPK use UniquePaths.
  • GMA and VPK now go through ArchiveSafety.CheckListing like every in-process reader (entry count, name lengths, the bomb ratio, room on the drive). For a VPK the ratio is taken against the directory file and its data parts together.
  • Comments and docs no longer say 7-Zip unpacks BSP pakfiles and Vita paks.
  • The SkipWithout7Zip guards in GmaTests, GmaPackTests, VpkTests and BspTests are removed (the archive engine is always available here). These four classes join the "Archive tests without 7-Zip" CI step.
  • Udf1443Tests.An_actual_sparse_file_over_four_gib_converts_and_reads_back lists the image with 7-Zip only when it is installed.

DiscUtils (Directory.Packages.props, Filee.Engines.csproj, Filee.Engines.Tests.csproj, THIRD-PARTY-NOTICES.md)

  • Main moved ISO creation to LTRData's fork (LTRData.DiscUtils.Iso9660 / .Udf 1.0.90) for UDF 2.01. The branch reads UDF, HFS+ and DMG with the original DiscUtils 0.16.13. Both families ship DiscUtils.Core.dll and the other assemblies under the same names, so they can't be mixed.
  • Every package moves to LTRData 1.0.90: .Iso9660, .Udf, .HfsPlus, .Dmg and .Fat (tests). lzfse-net is gone; LTRData's .Dmg brings the managed LzfseSharp (BSD-3-Clause, in the notices). LZFSE chunks still go to 7-Zip before DiscUtils reads them.
  • On Linux and macOS, LTRData's DiscUtils separates UDF and HFS+ paths with / (0.16.13 always used \). DiskImageReader now splits on the platform's separator too (19354c2); before, every name came out with a leading _ and a UDF folder pointing back at the root went unnoticed (first CI run: 6 DiskImageTests on Linux, 9 on macOS).
  • HfsPlusLinks reaches _wrapped and GetDirectoryEntry(string) by reflection; both are still there in 1.0.90 (checked by reflection on the built assemblies). HFS+ DMGs are only tested on macOS, so this PR carries full-ci.

Limits both sides added for the same advisories (HuffCdic.cs, MobiReader.cs, PalmDatabase.cs, HwpxReader.Body.cs, HwpxWriter.Body.cs, MarkdownWriter.cs, OdsReader.cs, UnpackLimit.cs)

  • MOBI: main's book-wide budget (TextLimit, HuffCdic(…, maxBytes), "too large") and the branch's per-record room, dictionary limit and inflate ratio ("too complex") are all kept. HuffCdic.Decompress(data) stays for main's callers.
  • Tables: the branch's PlaceCells (jump lists, DocumentLimits refusals) is kept. HwpxWriter.MaxTableColumns is now DocumentLimits.Default.MaxTableColumns (both 16,384) and TableTooLarge is internal, for main's TableSheets. From main, the HWPX reader cuts a column span to the columns left after the cell.
  • HWPX text runs: the branch's Settle join is kept and main's FinishJoin dropped; both fix the same quadratic join.
  • Markdown tables: the branch's budget and SpanCursor are kept.
  • ODS: the branch's caps (20M cells, 1,024 spaces, cell text counted once) are kept and main's duplicate counters dropped.
  • UnpackLimit: main's Copy / Read and the branch's Extract(entry, source, path) overload share one copy loop.
  • Three of main's ConversionBudgetTests expected main's error key or behaviour. They now expect DocumentLimits.ErrorKey (ODS, writers). A_cell_past_the_widest_table_goes_to_its_last_column follows the branch, which clamps the column instead of refusing it.

CI (ci.yml, build.yml, release.yml, portable.yml)

  • Main's structure is taken: the Windows builds are in build.yml, called as windows, behind the build gate job.
  • build.yml gets an arm64 input (default true). ci.yml passes needs.changes.outputs.full != 'false', so pull requests into builtin-engines without full-ci still skip Windows ARM64, as before.
  • Unhwp is removed from build.yml's ARM64 job and the release's new windows-arm64 job (fetch list and -p:UnhwpNativeLibrary).

Other

  • Cli.cs: the branch's RefusalPackages and main's same-audio name, both kept.
  • build-installer.ps1, CONTRIBUTING.md: main's ARM64 signing text, without 7-Zip and Unhwp.
  • README.md and 15 localized READMEs: both sides' rows combined (main: INSP, program icons, XLSX tables, PDF/A, split modes, NOV, WEBP, M4R, trackers, INO/PDE, GMA/VPK/BSP, packages, UDF ISOs).
  • docs/ENGINES.md: rows combined (CXLOG, GMA/VPK/BSP, WebP, ARM64 release); the DiscUtils section names LTRData 1.0.90 and LzfseSharp.
  • mn.json: main's own Mongolian texts for the 23 keys this branch had translated itself are taken. Every locale has en's 1986 keys. No key's English changed on both sides.

Tests (Windows 11 x64, local)

  • dotnet build Filee.slnx: 0 errors.
  • Filee.Core.Tests (full): 1766 passed, 11 skipped.
  • Filee.Engines.Tests, with FILEE_HIDE_7ZIP=1 (Archive*, DiskImage, Gma, GmaPack, Vpk, Bsp, FolderIso, FolderArchives, Udf1443, GamePack, SevenZip, ConversionBudget, EbookLimits, KindleWriter, WriterGrid, PackageRepacker, Markdown, *HostileInput, NestingLimit, MarkupWriterLimit): 784 tests; the 4 that failed first (3 of main's budget tests, the UDF test needing 7-Zip) were fixed and pass (44 passed, 2 skipped on rerun of those classes); the rest passed (68 skipped).
  • Filee.Engines.Tests with 7-Zip (ArchiveTests, ArchiveInProcess, Gma, GmaPack, Vpk, Bsp, Udf1443, FolderIso, DiskImage, Graphics, PortableEngine): 351 passed, 10 skipped.
  • Filee.Cli.Tests (full): 135 passed, 2 skipped.
  • Filee.App.Tests (Localization, Udf1443, Settings, Migration, EnginesPage, PresetEditor, EngineStatus): 312 passed, 11 skipped.
  • dotnet format Filee.slnx --verify-no-changes --severity warn: clean.

🤖 Generated with Claude Code

KnifeLemon and others added 30 commits November 10, 2026 16:45
The release workflow builds and tests win-arm64 on windows-11-arm and keeps it
unsigned as windows-arm64-unsigned; the draft waits for both Windows builds.
sign-release.ps1 signs and packs it next to the x64 build, writes and signs
Filee-<v>-win-arm64-SHA256SUMS.txt (one passphrase prompt) and mirrors it to
dl.filee.sh. It also removes its worktree when git trips over long paths.

make-setup takes --runtime, so an ARM64 Setup packed on an x64 PC offers what
the ARM64 app downloads. The updater of an ARM64 build takes the ARM64 Setup
and checksums, without falling back to x64; the x64 target skips -win-arm64-
files.

Refs #236

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Filee 2.4.0 and older take the last asset ending in -SHA256SUMS.txt that
isn't Linux or macOS, so a separate Filee-<v>-win-arm64-SHA256SUMS.txt could
become their pick and break every x64 update. sign-release.ps1 now appends
the ARM64 hashes to Filee-<v>-SHA256SUMS.txt (one file, one .sig) and refuses
to publish a draft with another Windows sums file; the ARM64 updater uses
that same file. A test runs 2.4.0's matching rules over the new assets.

The Explorer menu package registered for the other architecture (an install
switched between x64 and ARM64 in place) is now Outdated and is removed
before this build's is added, by Setup and by Settings.

Refs #236

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Setup runs elevated for an install for all users, and some of the trees
it deletes are the user's own: Filee 1.1's %LocalAppData%\Filee and, at
uninstall, the settings in %AppData%. Directory.EnumerateFiles with
AllDirectories enters junctions, so a junction planted there made Setup
delete files anywhere with admin rights.

FileSwap.DeleteTree now walks the tree one level at a time and removes a
link itself instead of entering it. Each entry is opened without
following a link and deleted through that handle, and only when its
real path is inside the tree, so swapping a folder for a junction while
Setup runs doesn't help either. A tree reached through a link isn't
touched. In the user's folders nothing is left for the next restart,
since that deletes by path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An uninstall for all users started without administrator rights copied
uninstall.exe into the user's %TEMP% and started that copy with "runas".
Any program of the same user could swap the copy between the copy and
the start, and the prompt and the administrator token went to it.

The unelevated uninstaller now asks for administrator rights for the
uninstall.exe in the install folder itself, which only administrators
can change. The elevated one makes the copy in a new folder in
%SystemRoot%\Temp, created with SYSTEM and Administrators as the only
ones allowed to change it and nothing inherited from the parent, and
holds the copy open without write or delete sharing while it starts it
with its own token (CreateProcess, no shell lookup). Standard users have
no delete-child right on %SystemRoot%\Temp, so they can't rename the
folder away either, as they could in their own %TEMP%. The trash for
files in use stays inside that folder.

Both uninstallers in the install folder run from the file the copy
deletes, so /FIRSTPHASE now takes a list and the copy waits for both.
An uninstall without administrator rights still copies to %TEMP%.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Setup and the uninstaller were published single-file with
IncludeNativeLibrariesForSelfExtract, so the .NET host unpacked
libSkiaSharp.dll, libHarfBuzzSharp.dll and av_libglesv2.dll into
%TEMP%\.net\<exe name>\<bundle id> and, on later starts, reused whatever
files it found there without checking them. That folder belongs to the
signed-in user, so when Setup or the uninstaller ran with administrator
rights it could load native code another program of that user had put
there.

The publish now embeds Skia and HarfBuzz as resources of Filee.Setup.dll
and leaves them out of the bundle and deps.json, so the host has nothing
to unpack. Before the window opens, Setup writes them into a folder made
new for this run (elevated: in %SystemRoot%\Temp with the admin-only
rights the uninstaller copy uses; otherwise in %TEMP%) and loads them
right away by full path through a DllImport resolver. Old folders are
removed on a later run, this run's after it ends.

Setup draws in software with the redirection surface (transparency
kept), so ANGLE is no longer carried at all. At start Setup also limits
the default DLL search to System32 and resolves every other library
named without a folder from System32, instead of next to Setup.exe or on
the PATH. A publish check fails if any DLL would land next to the exe.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A text:s element now stands for at most 1,024 spaces, all cell text of a workbook is counted against the size of an XML part, and repeated rows and columns with text expand to at most 10 million cells; past that the sheet is refused as too large.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PlaceCells cuts row spans at the table's last row and column spans at 16,384 columns, tracks covered positions per column instead of a set of every spanned cell, and refuses tables over 10 million positions. The Markdown writer and the document-to-sheet tables use the same bounds, and the HWPX reader cuts column addresses and spans when it reads them.

The HWPX reader also joins the text of a run in a StringBuilder: joining it one piece at a time cost the square of the run's length.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Unpacked folders are walked level by level (links removed, folders merged into place); entry paths and hard-link targets with more than 512 names are refused in the listing check and when an entry path is resolved.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HUFF/CDIC phrases can be made of phrases, so a small dictionary can stand for terabytes. The decompressor counts everything it unpacks, phrases included, and the text of a book may unpack to 64 times the file's size (at least 16 MB, at most 256 MB); past that the book is refused as too large.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
EPUB/KEPUB rewriting, Office variant conversion and the workbook prepared for sheet printing copied package entries without a size limit. They now copy through UnpackLimit, which also reads a whole entry with a per-entry limit for the chapters Kepub parses.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Before ImageMagick's first use its policy registers only the modules of the formats Filee reads or writes, keeps PDF write-only (reading it runs Ghostscript) and turns off MSVG; SVG stays, rasterized by librsvg. Memory, map, area, disk, width, height and list length are limited.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every input is built in code: repeated ODS spaces and cells, table spans over billions of cells, a run of 300,000 spaces, deeply nested archive paths, nested HUFF/CDIC phrases, package copies past the unpack limit, and pictures in formats Filee does not convert.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The shared configuration folder was made in the temp folder, which on Linux is /tmp:
another user could create it first and change the files between the check and
ImageMagick reading them. It now lives in WorkFolders.Root (%TEMP%\Filee on Windows,
the mode-700 cache folder on Linux and macOS). The app also sets the policy as it
starts, since it calls ImageMagick itself for crops and previews.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PDFsharp's ContentReader parses arrays and dictionaries recursively without a limit,
so a page of a million "[" overflowed the stack in Compress PDF and Remove blank
pages. PdfContentNesting scans the content first and treats anything nested deeper
than 64 as content that can't be read, as other unreadable content already is.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Add Yazi plugin in installer/linux/yazi/filee.yazi/ supporting quick conversion and Donut GUI trigger
- Bundle yazi integration in build/build-portable.ps1
- Document setup instructions and keybinding in installer/linux/INSTALL.txt
- Remove @sync entry to keep execution asynchronous
- Update keymap to modern [[mgr.prepend_keymap]] (Yazi 25.5+) in docs
- Drop CLI fallbacks and notify on error if GUI launcher fails
# Conflicts:
#	src/Filee.App/ViewModels/PresetEditorViewModel.cs
…onverting

Adds OutputRule.TrashOriginalAfterConvert (off by default) and a checkbox in the preset editor's output card. After a conversion to another format completes, the original goes to the Recycle Bin / Trash. If the Trash refuses, the original stays with a note saying why.

Never applied for watch folders, 'Ask where to save', files from a browser, the command line, merges, ZIP results, 'Only if smaller' copies, or results in the same format as the source (that is 'Replace the original').

Adds tests in TrashOriginalTests and a CLI test.

Refs #1387
Adds presets.trash_original, presets.trash_original_hint and notice.trash_original_kept to all 23 language files.

Refs #1387
trash?.Move does not compile (CS8978). Use trash is null ? null : trash.Move instead.

Refs #1387
Mongolian landed while those PRs were open, so main's LocalizationTests
failed on mn. Adds the 23 missing strings in Mongolian and removes
engines.state.system_copy, which en.json no longer has.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part B of the preset gallery (#1530).

- Filee.Core/Gallery: an explicit allow-list of the fields a shared preset.json
  may have (unknown fields fail, fields that carry the sharer's folders,
  passwords, picture paths, clip music, metadata edits or the AI model may only
  be empty), the bundle and folder checks (pictures by header, sizes, no
  animation), index.json reading with relative paths and the counted download
  link, and an installer that verifies SHA-256 and the rules, then adds new
  presets and a donut under new ids without replacing anything.
- filee presets check <folder> [--json]: what the gallery's CI runs, with
  SkiaSharp decoding every picture; codes, messages and derived meta fields.
- App: "Preset gallery" window from the Presets page (Hot/Best/New, filters,
  search, details, Install, Comments), contacting filee.sh only while open,
  through the proxy setting with time and size limits and a picture cache.
  Texts in all 23 languages, labels matching filee.sh.
- Export writes settingsSchema into bundles.

Refs #1530

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…entre

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ures per language

Maintainer decisions for the preset gallery (#1530):

- The allow-list now takes every preset kind and tool (Copy to folder and
  Metadata included), replacing/moving originals, overwriting, metadata edits,
  every save location, a translate model, annotations, hidden areas, backdrops
  and clips. Only secrets, folders of one's own, picture and sound paths, and
  metadata naming people, places or copyright owners stay left out.
- The check derives warnings (replaces_original, moves_original,
  overwrites_existing, copies_to_folder, edits_metadata, fixed_edits,
  strips_metadata, uses_translate_model) with a severity; the CLI prints them,
  meta.json may carry them and must match.
- The gallery window works them out from the downloaded file, shows them above
  Install, and asks for a tick before installing a preset with a high one.
  Installed presets remember the gallery entry; the preset editor says so.
- Export for the gallery: presets (and a donut) to a checked preset.json, with
  the submission form a click away.
- Pictures per language: images/<lang>/ in the check (app languages, 40 MB
  folder limit) and imagesByLang/thumbByLang in the window, which switches
  pictures when the app's language changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ared presets

A document's title, name or subject may be written; author, artist, creator,
copyright and rights holder, by-line, owner, credit, GPS and location fields
and contact details may still only be removed (#1530).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
i18n(mn): add the 23 keys main gained after Mongolian landed
KnifeLemon and others added 24 commits November 17, 2026 22:49
fix(clip): exact seeks, so the frame shown is the frame cut
feat(pdf): drop PDFs from the file manager into the page editor
fix(whisper): transcribe with the largest installed model
fix(video): offer 4:5 and 5:4 in Crop video
fix(i18n): say when the video codec setting rewraps a video
feat(archive): unpack Garry's Mod addons, Valve paks and Source map pakfiles (GMA, VPK, BSP)
…ngs schema 98

Main took 95 (Windows app packages), and the Source PRs took 96 and 97.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#1591 took 96 for the Source engine archives, and 97 and 98 are kept for
the Source engine presets in #1608 and #1611.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix(portable): tell packagers to drop filee.portable, make the error readable
feat(archive): pack a folder into a Garry's Mod addon (GMA)
fix(update): take release files from dl.filee.sh when GitHub fails
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
feat(presets): a "VTF" preset to make Valve textures from any picture
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix(formats): read CrossOver logs (.cxlog) as text
The move up/down buttons on the Engines page, Toolbar page, Merge order
window, Group editor and PDF page editor used filled triangles
(Icon.ArrowUp / Icon.ArrowDown), which looked heavier than the thin chevrons
of the number boxes (the DPI field and similar).

Add Icon.ChevronUp and Icon.ChevronDown and use them in those five places.
The unused triangle icons are removed.
The thin chevrons look smaller than the filled triangles they replaced at
the same size. Raise the icon size by 2px on the move up/down buttons
(Engines, Toolbar, Merge order, PDF page editor: 14 → 16, Group editor:
12 → 14).
…saves

After pages were deleted, reordered, duplicated or added from another PDF,
Save and Extract copied the pages one by one. That dropped the bookmarks
and the form field tree, and links could end up pointing outside the saved
pages. They now reuse the merge code that keeps this structure, limited to
the saved pages.

- Map each source page to its first place in the saved order, so bookmarks,
  links and named destinations go to the right page; a repeated page keeps
  its links as separate copies and its form fields only on the first copy.
- Remove links and named destinations whose target was deleted; keep a
  bookmark to a deleted page only as a title when it still has children.
- Keep only the form fields whose widgets stay on a saved page, and point
  each widget's /P at the page it is on. Signature fields are left out, as
  in the merge.
- Bound reading and copying annotations with three limits; when one runs
  out, the save leaves the rest out instead of failing.

Merge PDF and the original-order copy path are unchanged.

Closes #1379
fix(pdf): keep bookmarks, links and form fields when the page editor saves
feat(ui): use thin chevrons for the move up/down buttons
Brings main's 2.5.0 release and the work after it (MSIX / APPX, GMA / VPK /
BSP unpacking and GMA packing, the VTF preset, .cxlog, UDF 2.01 ISOs, the
updater mirror, the PDF page editor, the security hardening of 2.5.0 and
the Windows builds in build.yml) into the built-in engines branch.

- Settings schema: main's library steps keep 95-99; the branch's three
  steps move to 100 (unhwp -> hwp-text), 101 (LibreOffice for HWP ->
  renderer) and 102 (renderer on, LibreOffice for slides off).
  CurrentSchemaVersion 102.
- Archives: MSIX / APPX, GMA and VPK are read in-process; a BSP's pakfile
  and a PS Vita VPK go through the in-process ZIP path, so none needs
  7-Zip. GMA and VPK go through the shared listing checks.
- DiscUtils: every package moves to LTRData's 1.0.90 (main's ISO and UDF
  writer); the original 0.16.13 packages ship the same assembly names.
- Limits that both sides added (path depth, HUFF/CDIC, table spans, ODS
  cells, the ZIP entry limit) are kept once, with the branch's
  DocumentLimits refusals and main's extra checks.
- CI: build.yml takes an arm64 input, so pull requests into
  builtin-engines without full-ci still skip Windows ARM64.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@KnifeLemon KnifeLemon added built-in engines Filee: built-in engines (Project 4), on the builtin-engines branch, not released yet full-ci Run the whole CI matrix on a pull request into builtin-engines labels Nov 18, 2026
LTRData's DiscUtils separates the paths of UDF and HFS+ volumes with "/" on
Linux and macOS, where the original 0.16.13 always used "\". The disk image
reader took the whole path as the entry's name, so every name came out with
a leading "_" and a folder that pointed back at the root went unnoticed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@KnifeLemon
KnifeLemon merged commit 3ba2e73 into builtin-engines Nov 18, 2026
23 of 26 checks passed
adnan-bhaldar pushed a commit to adnan-bhaldar/Filee that referenced this pull request Nov 19, 2026
… and retries

The metadata job hit its 30-minute limit on KnifeLemon#1663, KnifeLemon#1669, KnifeLemon#1671, KnifeLemon#1672 and KnifeLemon#1679: apt took 23-30 minutes to
install Ubuntu's FFmpeg. Skip the runner's man-db index rebuild, skip recommends, give apt network timeouts and
three attempts, and limit the step to 12 minutes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

built-in engines Filee: built-in engines (Project 4), on the builtin-engines branch, not released yet full-ci Run the whole CI matrix on a pull request into builtin-engines

Projects

None yet

Development

Successfully merging this pull request may close these issues.

9 participants