Qodana's full analysis of main flagged `Text ?? string.Empty` as never null. The text box can set null when it is
cleared, so the property is declared nullable instead of dropping the guard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Qodana gate (no critical or high findings) caught these in the changed code of the previous pull request: the
expected-hash grid on Verify gained a second row, so its buttons and the computed-hash box now set Grid.Row and
Grid.Column explicitly, and Shortcuts no longer imports System.Linq (C# 14 binds the array Contains to the span
overload).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From 1.3.0 Hash Tool is in maintenance mode. What it accepts is written down once, in CONTRIBUTING.md ("Project
status"): security fixes (high/critical vulnerabilities in .NET or a dependency get a prompt release), dependency and
.NET updates, bug and UI/UX fixes, accessibility, translation fixes, and hash algorithms on real demand; no new
features unless truly necessary.
- README: maintenance, Qodana and .NET 10 LTS badges and a "Project status" section.
- CONTRIBUTING: the policy, the issue forms, and the hash algorithm candidates (BLAKE2b-512, SHA3-256/512, CRC-32C)
with what each would take; ARCHITECTURE: a checklist for adding an algorithm.
- Issue forms: the feature request form is replaced by "Hash algorithm request" and "Translation problem"; feature
ideas go to Discussions (Ideas). The accessibility form gives the email address for people who cannot use GitHub.
- Pull request template: the kinds of change that are accepted.
- SECURITY: 1.3.x supported, the security commitment during maintenance, Qodana among the checks.
- CLAUDE.md: the project status for agents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
"Expected hash", "Computed hash" and the announced "{0}: {1}" result in the 66 translations (with AI assistance, in
the terms each language already uses).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ShellIntegration kept its own list of extensions, which had drifted from the one the app reads: .sha384sum and .b3sum
had no "Verify with Hash Tool" entry. It now uses ChecksumFileNames.KnownExtensions, and since the entries are written
again at every start, existing installations pick them up. Tests keep the Core list, the algorithms' extensions and
the macOS document types in build/macos/Info.plist in step.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fixable known limitations of ACCESSIBILITY.md:
- Hashes, file names and paths keep left-to-right order in right-to-left languages (FlowDirection on each control,
not on the rows, so buttons keep their mirrored order).
- Verify shows the computed hash of a mismatched file under the expected one, in a read-only box that is focusable and
named "Computed hash"; it used to be only in the status tooltip, which bypassed hash masking (now masked too). The
expected-hash box is named "Expected hash" instead of its placeholder.
- A run of one file ends with its result in a live text at the top of the page ("photo.jpg: VALID"), which screen
readers announce; several files still end with the summary dialog.
- Shortcut tooltips use the platform's notation (⌘O on macOS) and the buttons and search box expose them as
AcceleratorKey.
ACCESSIBILITY.md is reviewed for 1.3.0: the maintainer's keyboard-only and Narrator test, the email address for people
who cannot use GitHub, the remaining limitations, and two new rules for contributors. New strings: Lbl_ExpectedHash,
Lbl_ComputedHash, Msg_RowResult.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
FsCheck runs each property on 500 random inputs and reports the smallest one that fails:
- the checksum parser accepts any text (random and checksum-shaped lines) without throwing, and whatever it accepts
is a hexadecimal digest of the algorithm's length;
- GNU (with backslashes and line breaks in names), BSD and SFV files read back exactly what ChecksumWriter wrote;
- ChecksumEntry.ResolvePath never returns a network or device path and refuses names starting with // or \;
- HashMatcher ignores letter case, spaces and a 0x prefix, rejects wrong lengths and non-hex digits, and accepts the
reversed CRC32 that Hash Tool 1.1 wrote;
- TextHashing writes every line break the chosen way.
OpenSSF Scorecard counts FsCheck as fuzzing. FsCheck (BSD-3-Clause) is a test-only dependency, updated by Dependabot
with the other test packages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
High-severity findings in the shipped code:
- HashEngine disposes exactly the hash functions it created (a list instead of an array with null checks).
- SingleInstance allocates the pipe header once, not on the stack in every retry, and AllowSetForegroundWindow's
unused result is no longer declared.
- Lambdas whose first parameter was named "_" assigned tasks to that parameter instead of discarding them
(UnhandledException handler, settings save timer).
- ChecksumWriter sets the writer's NewLine after the using declaration; LocalizationService raises the indexer change
without a caller-info argument; VerifyCommand's Outcome drops a field nobody read.
- Nullability: converters return non-null values, the language combo box may set null, FocusManager is never null.
- Redundant usings (global usings of the app), qualifiers and default arguments; C# 14 lambda parameters.
- FocusHandOff, RowTabNavigation and HighlightTextBlock move to Views/Controls, matching their namespace.
qodana.yaml analyses the shipped code (tests and benchmarks are covered by the build, analyzers and SonarCloud),
leaves Roslyn analyzer suggestions to the build, documents each remaining exclusion, and fails on critical or high
findings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The last new feature before maintenance mode, suggested in Discussions #23.
- Hash Text page (after Verify): type or paste text and see its hash in every algorithm, updated shortly after each
change, with a copy button per algorithm. UTF-8, UTF-16 LE or UTF-16 BE without a byte order mark; LF or CRLF line
breaks (every line break in the text is written that way); nothing is appended, so the SHA-256 matches
printf '%s' text | sha256sum. Follows the hash letter case and hash masking settings; usable while a list is busy.
- Core: TextHashing turns text into those bytes; the page hashes them in memory with HashEngine.
- hashtool hash - reads standard input once per command, like sha256sum -.
- Tests for the conversion, the view model, navigation, accessibility names and headings of the new page, the smoke
test and the CLI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- SDK 10.0.404 with the 10.0.15 runtime, a security release (CVE-2026-109475, CVE-2026-105544, CVE-2026-114155,
CVE-2026-118747) that the self-contained downloads include.
- Avalonia 12.1.6, Microsoft.Extensions and System.IO.Hashing 10.0.15, System.CommandLine 2.0.15,
Meziantou.Analyzer 3.0.341, dotnet-sonarscanner 12.1.0.
- DisableTransitiveFrameworkReferenceDownloads: projects with a runtime identifier (the NativeAOT CLI) no longer
restore the ASP.NET Core and Windows Desktop runtime packs, whose advisories raised Dependabot alerts for runtimes
Hash Tool never ships.
- Dependabot keeps Microsoft.Extensions, System.IO.Hashing, System.CommandLine and the SDK on .NET 10 (LTS): moving
to a new major version is done on purpose.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
JetBrains Qodana (qodana-dotnet, Ultimate licence) in native mode, with the SDK installed by setup-dotnet and the
settings in qodana.yaml. Results go to Qodana Cloud, to annotations and a pull request comment, and to the
qodana-report artifact. Pull requests analyse the files they change; pushes to main, a weekly run and manual runs
analyse everything. Skipped without QODANA_TOKEN (forks, Dependabot); not a required check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ACCESSIBILITY.md (goals, what works, supported environments, known
limitations, reporting and severity, contributor checklist), an
"Accessibility issue" form, a PR checklist item, keyboard shortcuts in
docs/USAGE.md, links from README and CONTRIBUTING, and the changelog.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Every focusable control has a translated automation name: list rows by
file name, buttons with an icon and a label, filters, number boxes (and the
text box inside them), progress bars, switches and lists. Avalonia 12.1
otherwise announces "Avalonia.Controls.StackPanel" or the row's view model.
- Page titles are level-1 headings, section titles level 2; navigation and
page are landmarks; update/notice banners and the update status are polite
live regions; decorative logos are hidden from assistive technology.
- RowTabNavigation: Tab moves through the controls of the focused list row,
then leaves the list (ListBox is a single tab stop).
- FocusHandOff: focus moves to Cancel when COMPUTE ALL / VERIFY ALL hides
itself, and back; same for a row's run/stop buttons.
- Number box spin buttons are no tab stops and hidden from screen readers;
Up/Down change the value.
- Colour-blind light palette: error, warning and info statuses reach 4.5:1.
- Dialogs follow the UI scale and stay within the screen.
- AccessibilityTests: names on every page in two languages, headings, Tab
in rows, focus hand-off, WCAG contrast of every palette.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(release): 1.2.2
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: 1.2.2 changelog names the System.CommandLine version
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Release and CI logs then show which runtime a self-contained build
bundles, such as 10.0.14 after a security update.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
.NET 10.0.14 is an out-of-band security release (Brotli decoder
CVE-2026-111497, SocketAsyncEventArgs CVE-2026-70197). global.json now
requires SDK 10.0.403, so self-contained builds bundle the patched runtime.
- Microsoft.Extensions.*, System.IO.Hashing 10.0.13 -> 10.0.14
- System.CommandLine 2.0.13 -> 2.0.14
- Meziantou.Analyzer 3.0.303 -> 3.0.317 (no new findings)
- github/codeql-action v4.39.0 -> v4.39.1, zizmor-action v0.7.1 -> v0.8.0
(zizmor 1.31.1 is still the latest)
xunit.v3 stays on 3.2.2: Avalonia.Headless.XUnit 12.1.4 still fails with
xunit.v3 4.x (AvaloniaUI/Avalonia#22072 is open).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CodeQL default setup never analyses Dependabot pull requests, so the
Analyze (actions) and Analyze (csharp) checks that the main ruleset
requires never report there and those pull requests cannot be merged.
The new CodeQL workflow runs the same analysis itself: actions, csharp
and python with build mode none, the default query suite, and default
setup's job names and /language: categories, so existing alerts carry
over. It runs on pushes to main, on pull requests to main (Dependabot's
included: code scanning accepts uploads from pull_request runs despite
their read-only token), weekly and on demand.
Default setup has to be switched off in the repository settings first;
while it is enabled, code scanning rejects uploads from workflows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fail() exits, but the taint analysis cannot tell, so S8707 stayed open for assert-release.py; the check now raises SystemExit itself, like the other scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Avalonia 12 creates an application per test by default, so the app's
startup ran again for every test, and its asynchronous part (system font
loading, then applying the settings) could land on the dispatcher of a
later test: "The calling thread cannot access this object" on cleanup,
seen on macOS CI and now and then on Windows. The app itself has one
application per process, and so do the tests now.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MainWindowAccessor gets an internal constructor so tests can hand the
services a headless window of their own (the test application has no
desktop lifetime): dialogs are clicked and dismissed for real, and the
pickers, clipboard and launcher run against the headless platform.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
UpdateService gets an internal constructor that takes the Velopack update
manager and the HTTP client for release notes, so tests run checks and
downloads against Velopack's own SimpleFileSource and TestVelopackLocator
in a temporary folder: no network, and nothing is ever applied.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Tests that switch the UI language run in a collection of their own so they never race other tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The common folder of the hashed files was found with a plain prefix test,
so "photos" and "photos-2024" counted as one inside the other and every
file of the second folder was stored with its absolute path. Tests cover
the Generate page's exports, ZIP, copy and error paths.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The row handler was registered for the tunnel route, but drop events only
bubble, so it never ran: the file was added to the list as a new row
instead. It now sits on the list ahead of the handler that adds files.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AppPaths resolves the data folder once, and only the first Avalonia test
used to point it at a temporary folder: a plain test running first would
have used the real settings folder. A module initializer now sets it
before any test. The dialog, storage and clipboard fakes move to
Fakes.cs, can be scripted, and never accept the restart as administrator.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The CLI awaits its console writes (S6966) and a running row is stopped
with CancelAsync. Where an overload takes a cancellation token, the call
now says which token applies (S8949): the pipe listener stops with its
own token, while its retry delay, file import, the batch's Task.Run (the
processor watches the batch token itself) and the update check opt out
explicitly. No behaviour change.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SonarCloud flags paths taken from the command line (S8707: an agent or a
mistyped argument could make the scripts read anything). The release
folder for assert-release.py and the changelog for release-notes.py must
be inside the repository; loc.py set also accepts JSON files in the
temporary folder, where they are usually written.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
zizmor-action 0.6.4 refuses versions it does not know ("Unknown version:
1.31.1"); 0.7.1 ships 1.31.1 as its default.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Automatic Analysis is off in SonarCloud; the new SonarCloud workflow runs
SonarScanner for .NET 12.0 around a real Release build, with Cobertura
coverage from the unit tests and from the smoke test run under
dotnet-coverage, and waits for the quality gate. Fork and Dependabot pull
requests get no SONAR_TOKEN, so their analysis is skipped instead of
failing.
Both tools join .config/dotnet-tools.json. Benchmarks are analysed as test
code, eng/ scripts stay out of coverage, and gate failures on main are
reported like other CI failures.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dependabot only reads a tool manifest from .config/dotnet-tools.json, so
vpk, nuget-license and CycloneDX were never proposed for updates. vpk is
grouped with the Velopack library it packs for, and manifest changes now
get the release dry run with the upgrade gate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On Unix, .NET named pipes are one Unix socket shared by all server
instances; disposing the only instance before creating the next one closed
the socket and dropped a launch that had already connected, while that
launch reported success. The next instance is now created before the
connected one is handled.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Large batches: the final drain applied at most 2,000 events, so batches
of more than ~1,000 files left rows "Queued"/"Hashing" and corrupted the
next run; it now drains until the queue is empty.
- Second launches: PipeMessage had two constructors, so the generated
serializer threw and the pipe listener stopped after the first message;
the listener now survives any bad connection. Requests that arrive
together (one Explorer process per selected file) are merged and handled
one at a time, waiting while the target page is busy.
- Verify summary counts missing files and rows without a hash as failures
instead of reporting success; no "folder is empty" after a checksum file
fills in existing rows; hash files cannot be loaded into queued rows.
- Changing a row's algorithm clears a result computed with another one;
exports always use the algorithm that produced the hash.
- .sfv files read "name CRC32" first (names starting with eight hex digits
were taken for GNU lines); SFV lines are parsed without a backtracking
regex; only LF separates lines (U+2028/U+0085/form feed are valid in names).
- Logging no longer reopens the file for every line after a 10 MB rollover.
- Explorer "Generate hash" on a drive root works (C:\ arrives as C:").
- A timed-out update check reports a network error instead of staying on
"Checking"; installing an update saves settings and window placement
first, and the update banner waits until no batch is running.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Run All / Verify All and Cancel now come before the Clear buttons, so the
primary action no longer wraps to a second row at the default window
size. Screenshots updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>