Delivery status
Required features: delivery inventory
The maintainer's shared app requirements, checked against this app on 2026-11-04. Status words: done (built and rendered from the real app by --screenshots), partial (works, with the gaps listed), missing, n/a (does not apply, with the reason). Nothing is marked done from reading source alone. MonitoringTools.exe --selftest <folder> drives 151 checks inside the built app in a throwaway data folder and writes a pass or fail report.
#Language, voice and personalization
| Feature | Status | Evidence and gaps |
|---|---|---|
| Language modes (English, Hong Kong Cantonese, bilingual) | done | Language and voice page; about 1,100 table entries plus patterns; captures lang-cantonese-* and lang-bilingual-* for overview, language, ollama, safety, modes, tools and records. The self-test check language: every visible string on every page and window has a Cantonese entry switches to Cantonese, opens every control-centre page, the toasts, offline dialog, setup wizard, welcome, file tools, commit template, confirmations, appearance editor, colour picker, regex builder, bulk selection and guided forms, and reports any untranslated string (zero at the last run). Accessible names, help text and tooltips are translated and restored like visible text. Reviewer sheet: cantonese-review.md. Late text is covered by the window sweep (a short timer re-applies the language to every window whose layout changed): the self-test checks language: late text on every page is translated by the window sweep alone (Cantonese and bilingual) open all control-centre pages without calling Language.Refresh and report any untranslated string. language: every page shows Cantonese beside the English in bilingual mode checks that each translatable string on every page, Ollama and Safety included, shows its Cantonese beside the English. Notification titles are stored with their headline source and written again at the language and humour level in force when read (history list, export, and open toasts); the check language: notification titles are written again when the language or humour level changes switches English, Cantonese, bilingual and a level change. Every explanation line the regex builder can write (tree rows brief and detailed, the flat token list and the replacement explanation, valid and invalid patterns) has a Cantonese entry or pattern, checked by language: every regex explanation line is translated, in Cantonese and in bilingual mode over sample patterns covering every token kind; the quoted remainder of .NET engine error messages stays as the engine wrote it. Gaps: no native speaker has reviewed the wording, and the maintainer accepted the Cantonese as it is for now (2026-11-04), so it stays marked as not reviewed; the audit reaches only the states it drives. Labels added after a window opens are translated live by the sweep. Reviewer sheet: cantonese-review.md. |
| Independent English and Cantonese humour levels (1 to 5) | done | Sliders on the Language page; about 45 headline keys with five variants each; every key now has a call site: toasts, the offline dialog and start-up (earlier), plus welcome (window title), game mode and snooze (Overview title and tray snooze row), history empty (History page), no results (search empty message) and update available (About page status). The check humour: each new headline call site shows its own text at all five levels in both languages switches each level 1 to 5 for English and Cantonese in the built app and reads the rendered text of the welcome window title, the Overview snooze heading, the empty History page and the no-match search message plus game mode (the Overview heading, with a simulated running game) and update available (the About page, with a simulated newer release tag) for 60 rendered reads. Test hooks GameDetector.SimulateForTest and ControlCentreWindow.SimulatedLatestTagForTest exist only for this check. No gaps recorded for the headline call sites. |
| Emoji in dialogs switch | done | Language page; toast titles go through Language.Title. |
| Personal wording file (upload, drop and copy) | done | Service page, drop on any window, or copy the file or its path (File Explorer's Copy as path in quotes, a plain path or a file:/// address); unrelated JSON files are ignored, and a copied file meant as wording that fails shows why. Files up to 4 MiB; extra top-level fields are ignored; keys that differ only in case are separate entries; an entry may be an object with a "replacement"; schema versions 1 and 2. Phrases are matched by a linear longest-match scan (no pattern timeout), checked with 6,004 phrases (MT_ONLY=wordingcopy). Self-test applies a copied file, ignores a package.json and respects the switch. The wording reaches every label, tooltip, accessible name and window title, keeps across tab switches and comes back after a restart; a self-test check visits six pages, switches away and back, and simulates a restart. |
| School mode | done | Modes page: renamable, PIN with PBKDF2-SHA256, back-off, shared record. PIN or password (separate PBKDF2 hashes, 8 or more characters), the Language page hides the language choices, humour sliders and personal wording while it is on, and English is forced. Windows Hello is an optional third way to turn the mode off, set up on the Modes page (self-test checks drive setup, refusals, unlock and the unavailable states; capture school-hello). It is Windows Hello user verification, not a FIDO2 credential. Turning it on and off with a PIN is driven through the page (English forced, Language page choices hidden, wrong PIN refused; capture school-on-language), and a throwaway-data run keeps the shared record inside its own folder. Not tested, by the maintainer's choice (2026-11-05): a FIDO2 key and answering a real Hello prompt need hardware and a person at this computer (the checks stub the Hello answer; the real availability check ran). |
| Narrator | done | Windows voices, separate English and Cantonese pickers, queue, test button. The live check synthesises speech with a real installed English voice. Gap: Cantonese speech is untested because this computer has no zh-HK voice installed. The maintainer chose (2026-11-04) to leave Cantonese speech untested on this computer, because Windows refuses the Cantonese voice package here (error 0x800f0950) and working around that would need a Windows update policy change. |
| Action sound effects | done | Existing; now governed by Modes. |
| Scheduled and external settings | done | Weekday and time rules with optional start and end dates, in-place rule editing, external JSON file with validation and watching. The Modes page form is driven by a self-test check (rule created with a date range, a reversed range refused, Edit loads the rule, Save changes updates it in place, stored fields asserted) and captured as modes-schedule-edit. Following the external file is now driven through the page too: the check external settings file: following it through the Modes page ... types the path, presses Follow this file, asserts the sound, toast and game-mode values in effect and the status line, rewrites the file and waits for the watcher to follow the edit, writes broken JSON and a wrong-typed value and asserts the last good values stay with a reason, checks the saved settings were never overwritten, and presses Stop following. Times are now chosen with a time picker (hour and minute number fields with steppers, built from the app's number fields because WPF has none) and dates with the system date picker and its calendar; the checks type into the hour and minute fields and set the picker dates, and show that an hour of 25 or a minute of 99 is refused. The calendar popup is now opened and clicked through (MT_ONLY=dialogs, step date): the check fills the Modes form, opens the Start date picker's popup (March 2027), and a real click posted with the lowlevel mouse_click tool to the popup window on day 17 selects it; the check asserts the picker holds 2027-03-17, the popup closed, and after Add rule the stored rule's start date is 2027-03-17. Limit: no physical mouse. |
| App rename | done | Appearance → display name drives the drawer title. |
| ADHD modes | done | Five switches and reminders, a Focus spotlight (Modes page and Ctrl+Shift+Space) and palette entries for every switch. Driving the effects found that Focus, One thing at a time, the elapsed-time display and Momentum had switches but no consumer; each now has one (the toast manager, the task toast and the control centre). Self-test checks read the effect in running windows: adhd focus (a non-urgent toast goes to the notification history, a warning still shows), adhd one thing at a time (the second and third toast wait, in order, and appear as each goes), adhd low stimulation (a 1.2 s fade finishes at once with the mode on and runs with it off; a toast is at full opacity straight away; ordinary sounds are refused while warnings still play), adhd time awareness (a running task toast reads N s elapsed, then Took N s; the reminder timer starts at the chosen interval and stops with the mode) and adhd momentum (the window reopens on the page left open with the mode on, and not with it off); the spotlight is driven by its own earlier check and captured as adhd-spotlight. Momentum now names the next step: a finished task toast offers Next: open Activity (checked on and off). The Time check reminder toast is waited for with a test-only one-second interval hook and shows even under Focus. The sound service itself is now checked (adhd low stimulation: the sound service really starts playback ...): with a test hook that answers the quiet-time question, an ordinary cue reaches the media player and starts playback (the Windows Notify System Generic sound, no media error reported) with the mode off, is refused before the player with Low stimulation on while a warning still starts, and nothing reaches the player with sounds switched off, the volume at zero or during quiet time. Limit: nobody listens, so whether audio actually comes out of a speaker is not heard; the check proves playback was started and the media engine reported no failure. |
| Dim sum surprise | done | Modes page; reads the public photo catalogue, with an honest unavailable state. The live check loads the live catalogue and picks a dish. The Modes page switch looks like an off switch and springs back on with a different line each time, announced politely, at once under Reduce motion (ModesPage.DimSumSwitch). Gap: the spring-back was not driven by a self-test. |
#Material 3, appearance and layout
| Feature | Status | Evidence and gaps |
|---|---|---|
| Material 3 conformance | done | Colour roles, Roboto (pinned and embedded), pill buttons with state layers, 24 to 28 corner containers, Material 3 switch, 56 high search bar, tab strip; toasts, the offline dialog, setup wizard, welcome tour, file tools, commit dialog, tray menu and confirmation all read theme roles live. Audited concretely on the built app in both themes against the named checklist in material3-audit.md: colour roles, type scale, typeface, shape scale, state layers, elevation and component anatomy. The self-test checks material 3: ... fail on a literal colour, an off-scale font size or radius, or a style without hover, pressed, focus and disabled states. Found and fixed: light-theme warning, accent and error colours, about 80 files of off-scale font sizes and corner radii, switch and text field state layers. Elevation is recorded, not judged (one shared shadow plus tonal surfaces). Captures before and after were compared by eye on 8 pages; the other pages were not compared by eye. The two gaps found by the audit were closed and looked at: a capture scene (states-dark, states-light) shows every interactive style (switch off and on, tonal, filled, chip, row, text field and the date field) at rest, hovered, pressed, focused and disabled in both themes, with hover, pressed and focus forced on the real templates through their state properties because a capture cannot hold a pointer; and the date field now has a themed template (container, text part, calendar button) and a themed calendar (calendar-dark, calendar-light), where before it drew the system's light box and a dark-on-dark calendar. The drop-down popup itself is not part of a window picture, so the calendar is captured on its own. |
| Motion with a full reduced-motion path | done | Spring and decelerate curves; Reduce motion and low-stimulation mode skip entrance motion, toast slides, the builder container transform and menus. The every-surface sweep runs with motion off and found two paths that still moved (the setup step slide and the command palette opening); both now follow the setting. |
| Appearance editor (theme, seed, text size, density, contrast) | done | Appearance page; per-element editor from any element menu; per-tab and per-group looks (Shift+right-click); density applied to rows, cards, tabs and buttons. The editor is driven by a self-test check: every tab opens, a Hover and a Pressed text colour are typed through the picker and stored per state (the normal look is untouched), Undo and Redo step through them, a preset is saved and applied to a second element, and an export then import keeps the per-state colours. Captured as appearance-editor-font, -colour, -shape, -library, -states and -presets. Export all looks and Import looks are now pressed as real buttons with the file dialogs answered by a test hook (AppearanceEditorWindow.FilePickForTest): a cancelled dialog changes nothing, the exported file holds the look, import brings it back, a file with no format name or the wrong one is refused (this found that a file without a format name was accepted as an empty set), and an oversized file is not read. The native Windows file dialogs are now driven too (MT_ONLY=dialogs, step appearance, no test hook): the check presses Export all looks, the real Save dialog ("Export element looks") opens on the hidden desktop, the lowlevel tools set its file name box and press Enter, and the check asserts the saved file holds the look and the editor says it exported; then it removes the look, presses Import looks, the real Open dialog is answered the same way, and the look comes back with the "Imported" message. Limit: the dialog is answered by setting the file name box and sending Enter to it, not by clicking its Save or Open button, and no physical mouse or keyboard was used. |
| Colour picker | partial | HSV square, hue and alpha, hex, RGB, HSL, HSV, HWB, CMYK, OKLCH, CIELAB and LCH fields with gamut clipping notes, theme swatches, recent colours, eyedropper and a contrast readout. A self-test check types the value each of the nine fields shows into that field (after changing the colour) and asserts the colour and the HEX field come back, for an exact colour and a rounded one; an out-of-range OKLCH value reports clipping, a short hex expands and an unreadable value changes nothing. Captured as colour-picker. The screen eyedropper is now driven on a hidden desktop: the check shows a window of a known colour (#12C84D), presses Pick from screen, clicks the overlay at a point inside that window and asserts the picker and its HEX field take the colour, then opens it again and cancels. The click now uses the position of the click itself instead of the global cursor. A hidden desktop has no display, so the screen read there returns black; the pixel read then asks the window under the point to draw itself (this app's own windows from their visuals, others through PrintWindow into a 32-bit picture). The overlay now reads the click's position from the window message, and a real click posted to the overlay's window handle with the lowlevel mouse_click tool on a hidden desktop (MT_ONLY=lane4clicks) picked the known colour. Gap: no click from a physical mouse, and on a desktop with a display only the first read path was used, not the fallbacks. |
| Logo customization | done | Logo colour from the seed or a custom hex, plus twelve glyphs from the icon set or one custom PNG or ICO (256 KB, 16 to 512 px, signature and decode checked, copied into the data folder) for the drawer logo and the tray icon, with reset. Self-test checks validation, import, glyph choice, reset, the drawer header and the tray bitmap; capture logo-custom. The tray icon itself was not photographed (it lives outside the app windows); its bitmap is shown in the preview and compared in the check. |
| Four looks (Graphite, Instrument, Blueprint, Pop) | done | Typefaces, corner shapes, card outlines, shadows, the rail marker, the page label and neutral colours change together; picked on the Appearance page, in setup, in the Make a look guide and in the command palette. Designs in design/looks. Captures look-*-light, look-*-dark, look-*-appearance, look-*-toast, look-*-tray and setup-look come from the real app, and the every-surface sweep runs every surface in all four looks. Found by the first capture run: corners were not reshaped, because a class handler for Loaded never reaches most elements; corner mapping now runs on each border's first layout. Gap: the looks were judged from captures, not on a screen by a person. |
| Accessibility and sizing | done | Numeric settings are typed number fields with steppers, keys and inline validation (see app/number-fields.md). Text size 90 to 150 %, keyboard paths for search, menus and the confirmation gate, live regions. Captures scale200-* render overview, history, appearance, safety and Ollama at 960 by 540 (1920 by 1080 at 200 %) in English and bilingual with no clipped content. Bilingual text is audited on every page, Ollama and Safety included (language: every page shows Cantonese beside the English in bilingual mode). Audited concretely on the built app in both themes against the named checklist in accessibility-audit.md: accessible names on every control, tab stops plus a real Tab traversal of every window (326 traversals per run), 24 by 24 target size with the spacing exception, text contrast on the real backdrop and every palette pair in all five seeds. Found and fixed: 431 buttons and 21 inputs with no name, light-theme and muted text contrast, and a real keyboard fault (leaving the Display name field rebuilt the page and dropped focus). Not tested: a screen reader (Narrator, NVDA). |
| Every surface fits the screen and the text | done | UI/SurfaceSweep.cs, a build step on every run: every control-centre page, four toasts, the tray menu, welcome, offline dialog, confirmation, all eight setup steps, the command palette and six clipboard dialogs, in all four looks, in English, Cantonese and bilingual, on a desktop window and on a 1080p screen at 200 percent (960 by 500), and in English at 150 percent text (1176 scans in 7.6 minutes in run 42399945303). It fails the build on text cut off or cut by its own slot, text taller than its box, content wider than a box that cannot scroll, parts outside their window, windows bigger than the screen, overlapping text, controls with no accessible name or no focus ring, and keyboard focus on something that cannot be seen; trimmed text with an ellipsis is listed for review. The first run found 1,110 such findings (some from the check itself measuring text differently from the window); they were fixed at the root: windows of a set size grow with the text and stay on the screen with scrolling parts, fields narrow and their buttons wrap, dialogs give way in their lists. Limit: the sweep sees what the layout reports, not pixels, and covers the surfaces it opens, not every state of every dialog. |
#Navigation, search and commands
| Feature | Status | Evidence and gaps |
|---|---|---|
| Onboarding wizard | done | Existing setup wizard. |
| Regex builder on every search | done | Every page, palette, tray, version picker, lists and menus; Build (with guided editors including conditionals, and reordering by buttons or a mouse drag), Test (scope, line breaks, clipboard, file, hex, expected outcomes), Debug (look-around, atomic groups, Unicode categories, class subtraction, conditionals, inline options and \G, traced and checked against .NET including captures), Compare (.NET modes plus JScript, Windows PowerShell 5.1, Python and Node.js where installed), Benchmark (growth, two patterns, per-part timing, saved history), Replace, Files (name pattern, archives, binary search with hex, per-extension settings, saved searches, history, any backup restorable, open at line), Code (five languages, match/replace/split, editable templates), Library (descriptions, parameters, search, opt-in history) and Help with builder preferences views; --regex command line with --lib, --in-place, several patterns and a JSON automation entry. Parity ledger after the second 2026-11-05 pass: 13 families, 13 done, 0 partial, 0 missing. Genuine limits: no PCRE, Perl, Ruby, Go or Java runtime offline; .NET exposes no internal tree or trace; no COM server; archives inside archives are not opened. |
| Command palette | done | Existing; now on the shared search field. |
| Context menu on every element | partial | Right-click, Shift+F10 or Menu key in the control centre and tray: press, copy, search the page, edit appearance, lock; each menu has its own filter. "Edit appearance…" opens the per-element appearance editor. Windows now carrying the menu: control centre, tray, drawing board, toasts, offline dialog, setup, welcome, file tools, commit message, dim sum, the appearance editor, confirmation and lock cards, and the command palette overlay (Clear the search, Use the full window or card size, Close the palette; the palette stays open while its menu does) (see element menu). A self-test check opens the menu on a toast, the offline dialog, the setup wizard, the confirmation window (asserted to offer only Copy text, on its button and on a key), the command palette and file tools; captured as menu-toast, menu-offline, menu-setup, menu-confirm, menu-palette and menu-filetools. The check opens the menu through the same routine the right-click handler calls, because the synthetic mouse event is not delivered to the handler on a hidden desktop. A real right click posted to the palette's window handle with the lowlevel tools on a hidden desktop opened its menu (MT_ONLY=lane4clicks). That needed a fix: WPF finds a mouse event's target from the cursor, which a posted click does not move, so every window with the menu now also reads the position inside the click message. Gap: no click from a physical mouse. |
| Menus show working shortcuts | done | Shortcut hints in element menus; Ctrl+C and Ctrl+F work. |
| Browser-style tabs | done | Tab strip docked left by default and movable to any edge, New tab catalog, pinning, groups with a picker, overflow, four tab searches, bulk close by text, per-tab and per-group appearance, shortcuts, Cantonese, persistence. |
| Every page open in the side rail | done | A fresh install opens all nineteen pages in the rail; an update opens only pages the rail has not offered before (KnownTabs), so a tab the person closed stays closed. Close buttons show on the open tab and on hover. Unit checks in tests/SmartTests/AppStartChecks.cs; captures show the full rail in every look. Gap: an update of an existing install was not watched. |
| Settings explain themselves | done | Every switch and choice on every control-centre page and every setup step shows a visible sentence saying what it does, usually with its default, beside the information button. A self-test check walks all pages and the seven setup steps and fails for any switch or choice without one (a caption directly under the control counts). It found and the same task fixed the missing ones on the Minecraft world generator, Tools, Ollama and Modes pages, the schedule form, the EULA switches and four setup switches. The walk now also visits the other windows: the appearance editor (all four tabs), file tools, the commit message, the welcome tour, the tray menu, a lock wizard, the drawing board and the Convert to dialog. It found four more and the same task fixed them: the Italic and Underline switches in the appearance editor, and the Clipboard toasts and Network alerts switches in the tray menu. The windows in the walk hold few switches (the report counts 6 switches and 1 choice outside the control centre); a window state that is opened by a later action, such as the file tools conversion options, is covered only when the walk reaches it. |
| Guided forms, info icons, blank-editor presets | done | Every switch and choice has an info button; New shortcut, Add a DNS profile, New Minecraft server, Add a schedule rule, Make a look (theme, seed, contrast, density, text size, review) and New chat profile are guided forms with Back, Next, validation and a review. The single forms remain beside them. Self-test checks drive the schedule, look and profile forms through their real buttons; capture guided-look. A blank script offers presets. The per-element appearance editor window stays one form by design (it edits many properties at once). See guided forms. |
#Safety and records
| Feature | Status | Evidence and gaps |
|---|---|---|
| Two-key super confirmation | done | Every destructive dialog; captures super-confirm-start, super-confirm-armed. |
| Toy locks on every element | done | Persistent locks with a wizard and veil, respected by the command palette; self-test locks, refuses removal while locked, unlocks with the PIN and removes. Appearance values can be locked from the editor (Library tab). A second check drives the real lock wizard window: two different PINs and a 2 digit PIN are refused, matching PINs make the lock through the Lock it button, a closed lock refuses removal, and the right PIN opens it. A real left click posted to a locked button's window on a hidden desktop was stopped by the veil and opened the unlock prompt without clicking the button; making that pass found a bug: the veil's own tooltip text was read into a locked element's key so a locked button lost its key, now fixed. Every element type found on the real pages is now walked (check locks on every element type, MT_ONLY=lockwalk): the check visits seventeen pages, locks one example of each type it meets and shows, for the element and for each interactive control inside it, that a left press is stopped with the unlock prompt opened (through the same routine the window's press handler calls, ElementMenu.BlockPress; a real posted click on a locked button was shown earlier), that Space on a button, switch or toggle is stopped, that typing and Backspace in a text field inside are stopped, that the control's value never changes while locked, that the PIN opens it, that the same press then goes through, and that an open lock can be removed. Types walked: button (Button), toggle button (ToggleButton), switch (CheckBox), number field (NumberField, with its text box and steppers), choice group (StackPanel with its buttons), card (Border), and list rows or cards built as Border, StackPanel and Grid (including a slider row with its slider, steppers and text box). Walking found two defects, both fixed: a locked text field could still be typed into from the keyboard (typing, Backspace, Delete, paste, Enter and Space are now stopped, and so are the arrow and paging keys on a locked choice or slider), and a button that carries a tooltip (New tab, Match case) lost its key the moment it was locked because the veil replaced the tooltip that is part of the key (the tooltip is now kept). Keys per element: a lock key is the element's words plus its position among the lockable elements of its window that say the same words (<words> #2 for the second); the first keeps the plain words, so locks saved before still hold, on the first such element. Check locks keyed per element locks the second of two buttons with the same words and shows the first stays free, shows a words-only lock holds on the first and not the second, and on the real Records page locks the second of three "Match case" buttons alone. More types: radio buttons, drop-down lists (the app's own and WPF's ComboBox, keyed by name, not by the value shown), WPF tabs, control centre tab strip tabs (kept under the tab strip so the lock holds on every page) and element menu items (the menu is its own window, so its items are locked from the menu itself: right-click an item or press the Menu key on it, Lock this menu item…; keyed by the element and the item) are lockable. The walk's second check (MT_ONLY=lockwalk) locks each and sends a press and a key through the window's real input route: press stopped with the prompt, the key (Space, Down, Enter) stopped with the prompt, the value unchanged, the PIN opens it, the same press then changes the value (radio checked, list index or choice moved, tab selected, page switched, the menu's Press pressed the button), and the open lock is removed. Walking these found three defects, all fixed: the release of a stopped press still reached the control (the app's drop-down list opened its list on the release), keyboard focus moving into a locked WPF tab selected it, and F4, Alt+Down, Delete and a middle click were not stopped. Real clicks: MT_ONLY=lane4clicks, step locktypes: left clicks posted with the lowlevel mouse_click tool on a hidden desktop to a locked radio button, ComboBox, the app's drop-down list, WPF tab, tab strip tab and element menu item were each stopped with the unlock prompt opened and the value unchanged (the menu item through the menu window's own hook). Capture lock-menu-item: an element menu with its Press item dimmed and carrying the lock chip. Limits named, not fixed: a locked radio button or switch can still be changed through UI Automation's select or toggle pattern called directly (the check records it); a position key follows the position, so it moves if an earlier element with the same words is added or removed; the file tool dialogs, which hold the app's radio buttons and drop-down lists, offer Copy text only from their menu by design, so the types are proven in windows that carry locks. See element locks. |
| Unlock ladder, support tickets, authenticator | done | Four-rung ladder, Support Tickets desk, TOTP authenticator (a copied picture of a setup QR code offers Register for codes on its toast, checked by the self-test check authenticator: a copied otpauth QR picture offers Register for codes, adds the account once and keeps the secret out of the toast; RFC 6238 vectors pass in the self-test) with Windows-protected secrets, and Windows Hello as an extra unlock route where available. |
| QR pairing | done | A built-in QR decoder and encoder, versions 1 to 40 and levels L, M, Q and H: read a code from the clipboard or a file, show an account as a code behind the two-key confirmation. The block layout and alignment spacing for versions 11 to 40 are derived from the standard's per-block tables; the derivation reproduces the explicit tables for versions 1 to 10 (a check), matches a byte-capacity table at versions 15, 20 and 40, and was compared with another encoder library's table for all 160 version and level pairs when it was written. Self-test checks round-trip every version from 11 to 40 at all four levels, a 2,331 byte code at version 40, a mirrored code at small and large versions (the reader transposes a picture it cannot read), pictures made by an independent encoder at versions 3, 13, 20, 27, 34 and 38 (6 of 6 decoded), and 16 simulated phone photos of a screen (perspective tilt, turn, blur, uneven light, noise): all 16 decode (16 of 16 in the full self-test; the check now fails unless every one decodes). The photo test first used a warp that was not a true perspective and the decoder was improved to refit on every alignment pattern with a finer search (14 of 16); the last two, version 24 photos with tilts of 6.5% and 8%, were two decoder faults found by logging the sampled grid against the true modules: a tidy triple of stray ring patterns was picked as the three corner squares (finders are now weighed by how often each was seen), and a mismatched alignment pattern dragged the fitted map off (patterns more than a third of a module off the fit are dropped and the map refitted). Simulated tilts up to 15% with an 8 degree turn decode on a version 9 code. Earlier: 320 of 320 rotated and noisy test images decoded. Not tested, by the maintainer's choice (2026-11-05): a real phone camera. Note: a second independent reader (OpenCV) could not read some of our big codes, but it could not read the other encoder's version 34 and 38 codes either, so that says nothing about ours. |
| Local history | done | Settings history with restore and diff (Records page). Shortcut scripts and the commit message template also keep up to 50 versions each by default, with a version list, line comparison and a two-key Restore (see local history). A self-test check saves three versions of a script through the editor, opens the History panel, picks the oldest and asserts the comparison, and that Restore asks for the two-key confirmation; captured as history-panel. Saved script and template versions can now be encrypted, either for this Windows account or with a password (PBKDF2-SHA256, 600,000 rounds, AES-256-GCM), and locked; a Local history card on the Records page is the password-protected manager, with a retention setting by count (5 to 500) and by age (days). Checks: history: Windows-account encryption ... (text not on disk, a flipped bit is refused), history: password encryption ... (short password refused, locked history unreadable and refuses saves, wrong password refused, right password unlocks, converts back), history: retention ... and history: the Records page manager ... (real buttons). Captured as history-manager. Settings snapshots on the Records page are encrypted with it too (check history: settings snapshots are encrypted ...: not readable on disk, listed only while unlocked, not kept while locked, converted back). Limits by design: a forgotten password cannot be recovered; a Windows-account key does not stop malware running as the same user. |
| Actions and discarded changes | partial | Core/LocalHistory.cs: an append-only JSON-lines record; discarded unsaved work from guided forms, the copied text editor and the Minecraft settings is kept (sealed with the clipboard history key when that is protected) with Copy it back; Base64 files with Show and Undo; virtual disk actions and Quick packs with their route. Shown on the Records page (see actions and discarded changes). Gap: no self-test discards work and copies it back yet. |
| Base64 beside the file | partial | A Tools page card writes name.ext.base64.txt beside a dropped file and decodes it back beside it, streaming, never overwriting, each result checked by a SHA-256 round trip, with Show, Undo and Cancel (see Base64 beside the file). The engine is checked on any machine by tests/SmartTests/Base64Checks.cs. Gap: the drop on the card was not driven in the built app. |
| Markdown rendering | partial | One renderer (UI/Markdown.cs, Core/MarkdownText.cs) for release notes, toasts and file previews, with :shortcode: emoji, relative links resolved against the repository, a named region and an empty state (see Markdown in the app); tests/SmartTests/MarkdownChecks.cs. Gap: not driven in the built app by a self-test. |
| Folding search bars, filters and statistics | partial | UI/Collapsible.cs on every page search, the Activity, Virtual disks and conversion filters and the overview statistics: one keyboard stop, announced, kept per panel, a badge when a closed panel hides a filter, Ctrl+F opens a folded search (see panels). The every-surface sweep opens every page with them. Gap: folding was not driven by a self-test. |
| Resizable and movable panels | partial | UI/PanelSizer.cs on the regex builder and explanations: edges and corner, a header to move a floating one, Alt+Shift and Alt with the arrow keys, kept per kind of panel, kept on the screen, and a reset (see panels). Gap: resizing was not driven by a self-test. |
| Live controls in menus and a filter in every dropdown | partial | The element menu shows a live switch, stepper or choice for a switch, number field, slider or choice control, which keeps the menu open and also changes with Left and Right; every dropdown has a filter with the regex builder and full keyboard use. Gap: no self-test drives a live control yet. |
| Live response-time chart | partial | UI/LiveChart.cs on the overview: an easing line with a live dot, a fading fill, an axis and value badge that ease, and a crosshair under the pointer or the arrow keys; it says when it waits for its first probes, is paused or is stale, never invents a value, and snaps without pulsing under Reduce motion or Low stimulation. Captured on the overview. Gap: not driven by a self-test. |
| The same features on the project site | partial | The site's dim sum switch springs back (an old stored "off" is dropped, one visit in ten draws a dish), its right-click menu has live settings and its own regex builder, a Base64 page works as downloads with the SHA-256 round trip, status, task and picture filters fold, the search builder and dialogs resize, and articles get emoji, resolved links and an empty state (see the project site). Driven in a headless browser before release, not in the build. Gaps by nature: no live chart (it needs the app's probes), no virtual disk manager (a browser cannot touch disks; the article documents it) and no record of actions. |
| Export everything | done | One Export everything action writes a zip holding settings, activity, notifications, settings history, shortcuts, the regex library, appearance and looks, schedule rules, lock records without secrets, and DNS profiles in JSON, CSV (formula-safe), Markdown, HTML and text; clipboard history is opt-in. Self-test opens the zip and checks files, formats, no lock hash or secret, and the Records page action; capture export-everything. The save dialog itself was not driven. |
| Hash and verify dialog | done | Own window opened from the notice for any single file: SHA-256, SHA-1, MD5, SHA-512 read off the UI thread with progress, rate and time left, Cancel that keeps nothing, hash in groups of eight, Copy hash, Copy as file line, a pasted value that ignores spaces, dashes and capitals and reads whole checksum lines, a length that selects the matching algorithm, match and mismatch banners with the first differing character. Self-test checks (MT_ONLY=datasec) cover the four digests against known text and the framework, every paste form, the match, mismatch and algorithm-switch paths, a 300 MB file with progress, and the Cancel button; captures datasec-hash-computing, datasec-hash-ready, datasec-hash-match, datasec-hash-mismatch. Not built: the design's "Download it again" button, because the dialog has no source address to download from. See data and security tools. |
| Who signed this dialog | done | WinVerifyTrust verdict (embedded and catalog-signed files), publisher, chain with a badge per level, timestamp, signed, not signed, broken and untrusted states, Open certificate into the inspector, Delete file behind the usual confirmation. Self-test checks use a signed system file, an unsigned copy, a copy signed with a throwaway self-signed certificate and a copy changed after signing, and open the inspector from the chain; captures datasec-signed-valid, datasec-signed-none, datasec-signed-broken, datasec-signed-untrusted. Revocation is not checked (stated in the dialog). MSI, CAB and MSIX files get the verdict and publisher but no timestamp, which is read only from PE files and catalogs. |
| Certificate inspector dialog | done | .cer .crt .pem .pfx .p12 .der: subject, issuer, validity bar and countdown, alternative names, key, serial, fingerprints, allowed uses, expiring and expired states, a PKCS#12 password asked once and cleared, conversion between PEM, DER and PKCS#12 with a private key only on request with a new password. Self-test checks read a certificate made in code field by field and drive the password, every conversion and the key rules; captures datasec-cert-valid, datasec-cert-expiring, datasec-cert-expired, datasec-cert-password, datasec-cert-convert. |
| Data viewer and converter dialog | done | .csv .tsv .json .yaml .yml: first 500 rows as a table with search, text view, separator chips, validity check with line and column, Go to the error, conversion between CSV, JSON and YAML using the whole file, Format in place with a backup. YAML uses YamlDotNet 16.3.0 (MIT) pinned in packages.lock.json. Self-test checks cover quotes, embedded newlines, separators, error positions, numeric typing, the 1,200 row conversion, all conversion pairs, a refused shape and the dialog flows; captures datasec-data-table, datasec-data-done, datasec-data-error, datasec-data-text. The confirmation window of Format in place is not clicked by a check; the rewrite itself is. |
| Encoding and line endings dialog | done | Detects UTF-8, UTF-8 BOM, UTF-16 LE and BE, a legacy code page (as a guess) and line endings; batch list, preview with visible line-end marks, mixed-endings view, strict decoding that stops a file and writes nothing, hand-picked source encoding, .bak backups and Undo from backups, binary files never touched. Self-test checks create files in each encoding and drive a batch, the error state and undo byte for byte; captures datasec-encoding-batch, datasec-encoding-mixed, datasec-encoding-error, datasec-encoding-done. |
| Bulk actions | done | One shared select mode (bulk actions) on clipboard history, activity, notifications, shortcuts, watched games, the regex Library, installed models, settings history, the changelog, the outage list and schedule rules. Self-test checks settings history delete (newest kept), release export, outage removal (ongoing kept) and select mode on the rules and activity pages. Lists without bulk actions are named in the document with reasons. The regex Library still has no search of its own. |
| Cut or copied, said on the toast | done | Files carry Explorer's own mark (Preferred DropEffect); text and pictures go by the shortcut held when the clipboard changed (Ctrl+X or Ctrl+C). The toast shows a Cut or Copied chip and history keeps it; when it cannot be told, nothing is claimed. Unit checks in tests/SmartTests; captures show the chip. Gap: a cut by a menu click with no shortcut held and no file mark reads as unknown, by design. |
| Background updates | done | The service checks GitHub a little after start and every hour, downloads a newer release in the background, keeps it only when its SHA-256 matches the published digest and its own version matches the release, and installs it at a quiet moment (no conversion, no Minecraft server started here as a local process, no game being played; a game left open with no input for 10 minutes does not count) by running it with --update-install, which replaces the exe and starts the service again in the same mode; a failed copy starts the previous version again. On by default, a switch and Check now on the About page, a notice after the swap. The self-test check updates: a newer release is downloaded, checked against its digest and version, waits while busy and installs when quiet refuses a release with no checksum, a wrong checksum and a file whose own version differs, keeps a good one waiting while busy, installs it without downloading again once quiet, and shows the notice once. Seen in the real installed app on 2026-11-04: v1.0.36 downloaded and checked v1.0.37 at 22:03, waited every 10 minutes while the maintainer was playing, installed it at 23:23 once there had been no input for 10 minutes (Roblox still open) and was back in 7 s; that version then found, downloaded and installed v1.0.40 by itself and reported up to date at 23:27. See service and tray. |
| No welcome after an update | done | Core/InstallNote.cs decides what a copy started by hand leaves behind: the welcome only on a genuine first install, a quiet "Updated from X to Y." toast after an upgrade by hand, and just the control centre when the same version starts again. Unit checks in tests/SmartTests/AppStartChecks.cs. Gap: an update of a real install was not watched. |
| Game started and stopped notices | done | "Roblox detected" (or any listed game, or a fullscreen app) when a game starts and "Roblox closed" when it stops, saying game mode is on or off; never takes focus; a switch on the Game mode page and Stop these notices on the notice; nothing while game mode is off or notifications are snoozed. The self-test check game mode: a game starting shows "detected" without taking focus, closing shows "closed", and the switch turns both off drives all three. Gap: shown by calling the notice the game watcher calls, not by launching a real game. See game mode. |
| Changelog viewer | done | Records page, cached releases rendered as Markdown. |
| Offline documentation browser | done | Help and docs page: task-led landing, article reader, search with the regex builder. |
| Notification centre | done | Records page; toasts and offline alerts are logged. |
#Built-in tools
| Feature | Status | Evidence and gaps |
|---|---|---|
| File converter, Base64 toolkit | done | 46 conversion formats from the bundled FFmpeg, every one made and read back by a self-test; a Convert to dialog lists every target with what it is for (captures convert-dialog, convert-dialog-picked, convert-dialog-batch); file tools. |
| Archive, subtitle, font and disc image dialogs | partial | Four separate windows (Extract an archive, Subtitle timing, Font preview and install, Disc image) opened from the clipboard notice, built from design/file-tools/ boards 7, 13, 14 and 15. Self-test group docsys drives each on files it creates (zip with zip-slip entries, password zip, tar, tar.gz, gz, 7z state, SRT and VTT shift, stretch, convert and problem cues, a copied system font with the install step replaced by a recorder, an ISO built by the check that is read, extracted and really mounted and ejected through Windows) plus a Cantonese audit of every state; captures docsys-*. 7z, xz, bz2, tar.xz, tar.bz2 and AES-encrypted zips are listed and extracted through the bundled 7-Zip (password, wrong password, encrypted names, zip-slip refusal and a cancel that leaves nothing behind are checked on fixtures made with it); RAR4 and RAR5 are listed and extracted through the bundled full 7-Zip (7z.exe with 7z.dll), checked on RAR archives the self-test builds byte by byte from the documented formats (stored RAR4 and RAR5 with a subfolder and SHA-256 compared, a RAR4 ..\evil.txt entry refused, a RAR5 with AES-256 encrypted data and one with encrypted names tried with a wrong and the right password, and a cancel that leaves nothing behind); the overwrite rule has Ask with a per-file question and an apply-to-all switch; a .ttc lists, previews and installs every face; the tray menu lists each image this app mounted with Eject (checked on a real mount). Gap: only ISO 9660 and Joliet images can be listed without mounting. See dialogs. |
| Virtual disk manager | partial | The Virtual disks page (UI/Pages/VirtualDisksPage.cs) over the Windows virtual disk service (Core/Disks/VirtDisk.cs), so nothing needs Hyper-V: the routes this computer has, a library (attached disks, added, scanned and dropped files, search, filters), every figure of one disk, its partitions, attach, detach, grow with the volume, shrink with the volume, compact (Quick, Full scan), convert, merge, reconnect a parent, a new identifier, and New virtual disk with presets; each action is checked first, shows PowerShell and DiskPart, runs elevated once (--vdisk-job) and is recorded. Real disks are made, changed and read back by MT_ONLY=vdisk, a build step on every push, which also reads every state of the page in Cantonese and in both languages; captures vdisks-*. Gaps: VHD Set checkpoints, attach at sign-in, physical disk copy, block size change, the other Optimize-VHD modes, chain test and tree, metadata, mirror, change tracking ranges, Dev Drive trust and filters, native boot (see Virtual disks). |
| Quick pack | done | Drop files and folders on the Quick pack card: one VHDX beside them, named after them and numbered on a clash, sized from the content, cluster rounding, records, tables and headroom, grown by itself when it runs short (detach, grow, attach, extend, no second prompt), every file copied with streams, times and attributes (permissions as an option) and checked by SHA-256, pause, resume and cancel, partial results listed, and an optional tight fit. Proved by MT_ONLY=vdisk on the build machine: a pack that has to grow, a clash, the mark of the web carried over, sources unchanged, and a cancel that leaves no file; unit checks in tests/SmartTests/QuickPackChecks.cs. Gap: a ReFS pack has not been run on a real disk. |
| Archive and virtual disk windows | done | Archive (zip, 7z, tar.gz with every board option) and Put in a virtual disk (VHD, VHDX through one elevated helper) open from the file toast. 29 self-test checks (MT_ONLY=archivedisk) drive the real windows and read the results back independently: zip and tar.gz entries compared by SHA-256 with the sources, 7z with AES-256, hidden names and 1 MB volumes opened with the right and wrong password by the 7-Zip command line, encrypted zip, damaged archives rejected by the test step, Cancel leaving no files (zip, tar.gz, 7z), the two-key confirmation before originals go to the Recycle Bin (full slide deletes, Emergency exit keeps), the disk window against a fake diskpart and a fake volume (scripts, request hash, status and command files, copy checked per file, Move confirmation, Detach, a failing step cleaning up, a declined prompt showing 1223, Cancel). The language audit visits every view of both windows with Cantonese on and reports none of theirs. Captures archive-* and disk-*; progress, done and error captures are the real views with fixed numbers. Real disk, live check MT_LIVE_ONLY=vhd (2026-11-05): with the maintainer approving the Windows consent prompt, a 64 MB VHDX was created, attached, formatted NTFS (label LIVECHECK), mounted as Z:, filled with 3 files (3,215,752 bytes) by the unelevated app, each identical to its source, then detached and deleted, in 6.4 s. The first real run found that diskpart rejects UTF-16 scripts (it printed its help and exited 0x80070057, so no real disk could ever be made); scripts are now written in the console code page and a name, folder or label that code page cannot carry is refused before the consent prompt (self-test check). See archive and virtual disk. |
| Minecraft server manager: list, header and state | done | Server list with the shared search and regex builder, state filters with counts and target badges; header with Start, Stop, Restart, Reconnect and Copy join address. State is read from the target: an unreachable SSH host (ssh exit 255) is Host offline, never Stopped. The live parts of each section redraw only when something shown changed and then put keyboard focus back on the same control and keep the scroll position. Self-test checks drive the filters and a regex search, probe a real unreachable host (203.0.113.9), load an older servers.json, and focus a button in each of the eight sections through two real 4-second polls (with a new console line between them) and find focus and the scroll position unchanged. The live check ran real servers on all four targets in throwaway folders and containers: a local process (46 s), local Docker (create, start, list, stop, remove by label; 142.5 s), SSH to a throwaway OpenSSH container after approving its host key (56.5 s), and Docker over SSH on a throwaway Docker-in-Docker container (72.5 s). Capture mc-overview. See servers. |
| Minecraft server manager: overview | done | Tiles for players (from list), ticks per second (Paper and Purpur tps, otherwise Unavailable), memory, uptime and port; recent console; players online; attention items for Host offline, a failed start, a port in use, a newer build and backup age. A Players over the last 7 days card keeps the peak per server per day in player-peaks.json (30 days, upward within a day, a day with no reading shown as a dash) and shows Peak today, Peak in 7 days and a bar per day. Capture mc-overview uses sample values and mc-overview-peaks shows the card. Checked by self-test (peak kept per server and day, saved and read back after a reload, pruned after 30 days, shown on a real server's Overview and redrawn when a new peak arrives) and by the live check MT_LIVE_ONLY=minecraft-lane on a real Paper 1.21.4 build 232 server: tps read through the Overview's own poll (20.0, 535 ms), and the update check against the real Paper API (build 232 current; a server recorded as build 1 told build 232 is newer). The update check was also run against the real Purpur (58 builds), Fabric (255), Forge (53) and NeoForge (158) APIs: for each the newest build is current and the oldest is told the newest is newer. Capture mc-overview now shows the peaks card whole. See overview. |
| Minecraft server manager: console and command builder | done | Live log with level filters, follow and the shared search, with favourites and recent commands that run again; no free shell field. The command builder (its own Commands section, opened from the console and the header's Commands button) has a searchable gallery of 24 tasks in Players, World, Rules and Server, each a form walking the server's own reports/commands.json: online players and @p @a @r @s with explanations and a selector filter builder (distance, limit, sort, game mode); items, blocks, effects, enchantments and entities as searchable lists from the version's reports/registries.json and the jar's data folder; typed number fields with the parser's real bounds; positions with a ~ toggle per axis and a player's position read with data get entity; choices and true or false. Every command walks the whole tree (redirects followed) with a control per parser type and a text box naming any other parser. Coloured preview whose parts focus their field, a status line naming exactly what is missing or out of range, Run (two-key confirmation for stop, ban and op), Copy, favourites per server in servers.json, and Schedule it (new Run a command schedule kind). Self-test checks build exact lines for nine tasks from a trimmed copy of the real 1.21.4 reports, refuse out-of-range numbers, unknown ids and wrong targets, write ~ and ~5, read ten commands of eight argument types back through the walker, read enchantments from inside a bundler jar, drive the form until Run turns on, confirm stop, ban and op, persist and reopen a favourite and open a command schedule; the language audit and settings walk visit every task. The live check (MT_LIVE_ONLY=minecraft-commands) built and ran time set day, give @a minecraft:diamond 2 and gamerule keepInventory true on a throwaway Vanilla 1.21.4 server and read each reply (66.7 s in all). Item components (Java 1.20.5 and newer) are built with controls under the item list: custom name, lore, enchantments with levels, damage, unbreakable, custom model data and a generic row for any component id of the version's minecraft:data_component_type list with an SNBT value, written in each version's form (1.20.5, 1.21.4 and 1.21.5 differ). NBT arguments get a structured tree editor (add key, byte, short, int, long, float, double, string, list or compound, typed values with ranges, nesting) that renders SNBT; older versions use it for item data. Plugin and mod commands are read from the running server's help and minecraft:help, listed under Plugins and mods labelled as from the live server, and turned into typed forms from their usage text. Self-test checks (SNBT for every tag type read back, components per version, the item editor driven to an exact give line with a refused bad value and a removed chip, the NBT tree driven to an exact summon line and a refused byte of 300, plugin help read into forms, Use a player's position pressed with a stand-in reply) and the live check on a real Paper 1.21.4 server with a headless mineflayer bot player: the server's data get entity position equalled the bot's own (-25.50 72.00 -39.50), a give with seven components answered Gave 1 [Big sword] and the inventory read back the name, lore, enchantment and model data, a summon with tree-built NBT answered Summoned new Bessie, and help gave 15 commands the jar does not list (about, callback, icanhasbukkit, mspt, paper, pl and others), all with usage text. Command replies were also run through the builder over the Docker target (time, give with components, gamerule, which exposed and fixed Windows quoting that stripped the quotes of a Docker rcon-cli command) and over the SSH target (time, gamerule, read from the host's console log). Captures mc-commands-give, mc-commands-tp, mc-commands-tree, mc-commands-favourites, mc-commands-item-components, mc-commands-item-components-more, mc-commands-nbt, mc-commands-plugins, mc-console. A real Fabric server in Docker with Fabric API and the spark mod had help read: /spark was found as a mod command and became a form (Fabric prints its usage as [<args>], so one optional text field); RCON returns that reply glued into one run of text, which the reader now splits apart. The builder's time, give with components and gamerule lines were answered over the SSH Docker target too (throwaway Docker-in-Docker over SSH). An invalid NBT tree now puts the exact reason in the status line. A real Forge 1.21.4 server in Docker, created and started by the app on the Java 21 image (ready after 2,781 s), answered list and help (mod commands config and forge; /forge became a choice form) and the builder's /time set day (live check MT_LIVE_LANE=forge); replies are stripped of the terminal colour codes Forge sends. A real NeoForge 1.21.4 server the same way (ready after 1,529 s) answered list, help (mod commands config and neoforge; /neoforge became a choice form) and /time set day with clean replies (live check MT_LIVE_LANE=neoforge). Limits: a plugin command missing from help or printing no usage gets one text field; item predicates keep an id list and a typed value; spark tps over Docker RCON answered only "Sent." because the mod prints to the console. See command builder and console. |
| Minecraft server manager: players | done | ops.json, whitelist.json and banned-players.json read and written while stopped (UUIDs kept), console commands while running, two-key confirmation for bans and operator changes, kick and game mode from the menu. SSH servers now read and change those files over SSH while stopped (one round trip to read, only the changed file written back, text sent as base64 so no character is interpreted by a shell). Bedrock has its own Players section for allowlist.json and permissions.json (Xbox user ids, visitor/member/operator, id-only entries shown, console allowlist/op/deop while running). The list is a real multi-select ListBox (Ctrl/Shift, Ctrl+A, arrow keys) with the shared search field and its regex builder, a shown/total count and a typed selection line, and bulk actions on the selection: kick, ban, pardon, operator, allow list or whitelist, game mode, teleport to a player or coordinates, and give an item (a plain multi-target form, plus Give in the command builder which opens the builder's give form with the item, the count and the first selected player filled in, since the builder takes one target at a time); ban, operator and kick ask for the two-key confirmation naming the count and the players; while running one console line per player with a per-player result line, while stopped the file-backed actions edit the files (this computer, SSH, Bedrock) and the rest say they need a running server. Self-test checks: the file store, SSH players through a stand-in host, Bedrock players (ids, visitor, id-only entries, refusals), the bulk list (regex and plain search, selection line, exact commands for every action, confirmations, invalid input, per-player failures) and the bulk file edits; the live check minecraft-ssh-files ran the players and a two-player bulk change against a real OpenSSH host and minecraft-bedrock-files read an allow-list entry on a real Bedrock server. Captures mc-players, mc-players-bulk, mc-ssh-players, mc-bedrock-players. The live check minecraft-bulk-bot (72.4 s) ran the bulk teleport, kick and ban against a real online player, a headless mineflayer bot on a throwaway Vanilla 1.21.4 server with online-mode=false: the server logged Teleported LiveBot to 100.5, 150.0, -79.5 and the bot's own position became the same; the bot received the kick reason see the rules and the server logged Kicked LiveBot: see the rules; the ban kicked the bot and a new login was refused with the ban reason, and after Pardon it joined again. A self-test check drives the hand-off. Gaps: the bulk game mode and give were not run against the bot (their exact commands are asserted, not their effect); a Bedrock operator change from the files needs the Xbox user id typed or recorded; the hand-off takes one target because the builder's Who control does. See players. |
| Minecraft server manager: settings | done | Groups with counts and the shared search; rows with label, help, key, default, an applies badge and a switch, typed number field with stepper and bounds, or choice list; unknown keys and comments kept byte for byte; unsaved bar with Discard, Save and Save and restart; live keys sent as commands while running; game rules from the jar's gamerule command. Self-test checks round-trip a file and save a change through the section; capture mc-settings. See settings. |
| Minecraft server manager: worlds | done | Dimensions with size and the seed read from level.dat, Download ZIP, import a folder or ZIP (checked and copied), reset and replace (stop, back up first, two-key), and the world generator moved here. Capture mc-worlds. A self-test check drives the real buttons on a throwaway local server with real folders: Import refuses a folder without level.dat, copies a world folder, asks before a ZIP replaces it (declining changes nothing, confirming swaps the content after a backup), and Reset removes the world after a second backup (the file pickers and the confirmation window are test hooks there; those dialogs were not driven). Pre-generation has Pause and Resume (see the world generator). For a server on an SSH host the world folders are listed from the host with sizes and the seed, and Download ZIP copies a folder from the host and zips it (checked by a self-test check and by the live check minecraft-ssh-files against a real OpenSSH host; capture mc-ssh-worlds). Import, reset and replace work for an SSH server too: the new world is staged here, checked for level.dat, uploaded beside the live one, and swapped in after a dated copy on the host, with the old folders moved back if any step fails; a self-test check drives the real buttons on a sample SSH server (including an injected mid-swap failure), and the minecraft-ssh-files live check ran folder import, ZIP replace, a failed swap that restored the world, and reset on a real OpenSSH host with level.dat and a 2 MB region file read back byte for byte (synthetic worlds, not ones a server generated). See worlds. |
| Minecraft server manager: plugins and mods | done | Installed jars with names from their own metadata, on and off by renaming to .jar.disabled, install from a file, remove with two keys, a folder backup before every change and Undo the last change; a compatibility badge per jar (Compatible, Update, May not work, Not in a catalogue) from Modrinth by SHA-512 and from Hangar for files installed from it; Update per jar and Update all, which rolls the folder back if any update fails; Modrinth and (for Paper and Purpur) Hangar search, each installing only after the catalogue's SHA-512 or SHA-256 matches. Self-test checks drive the rename and backup, every badge rule, Update all with a failing and a succeeding update plus Undo, and the section on a sample server (badges, Update all, Undo, the Hangar chip only for Paper). The live check searched both catalogues, installed a Hangar plugin and an older Modrinth file with their hashes checked, found the badges Update and Compatible and ran Update all. Captures mc-plugins, mc-plugins-badges, mc-plugins-hangar. SSH servers list their plugins or mods folder over SSH (names from file names), switch jars on and off by renaming, remove, and install from a file by scp, each after a dated copy of the folder on the host. Bedrock add-ons (behavior_packs and resource_packs with the world's world_behavior_packs.json and world_resource_packs.json): install from .mcpack, .mcaddon or a ZIP (zip-slip safe, nested packs, skin packs and templates refused), a switch per pack that also switches on the packs it depends on, remove with the two-key confirmation, a dated copy of what each change touches and Undo the last change; the live check minecraft-bedrock-files installed an add-on on a real Bedrock server and read it back as a Pack Stack line in the server log (34.9 s). Self-test checks drive the SSH folder, every add-on path above and the Bedrock section. Captures mc-ssh-mods, mc-bedrock-addons. Catalogues over SSH: Modrinth and Hangar search and install, the compatibility badges (SHA-512 computed on the host, so no jar is downloaded to check it), Update and Update all (rolling the host's folder back if one fails) and Undo the last change work for a server on an SSH host: the file is downloaded here, checked against the catalogue's hash, copied with scp after a dated copy on the host, and Undo puts that copy back. The live check minecraft-ssh-catalogue (12.8 s, a throwaway OpenSSH host) installed a Hangar plugin and an older Modrinth file with both hashes matched on the host, read the badges Update and Compatible, ran Update all and Undo. Not covered: plugin names and versions on a host are file names only; add-ons of a Bedrock server whose world does not exist yet are installed but switched on only after the first start. See plugins. |
| Minecraft server manager: backups | done | Back up now with save-off, save-all, copy and save-on, a list with reason, size and contents, restore that stops the server and keeps the current world as a backup, keep-N as a typed number and a chosen folder. A self-test check backs up, restores and prunes a test world; capture mc-backups. SSH servers are backed up to this computer (saves paused while running, the world folders pulled with scp and written by the same code, so the backup lands in the same list) and restored to the host (current folders moved aside, the backup's copied up, the old ones removed only when all have arrived, moved back on a failure, a safety backup first); self-test checks cover both and the section, and the live check minecraft-ssh-files backed a world up from a real OpenSSH host and restored it (31.8 s for the whole group). A running SSH server is backed up with save-off, save-all and save-on (Bedrock save hold and save resume) over the console pipe and its folders copied here, and a switch Keep backups on the SSH host archives the world into .mt-backups on the host as tar.gz with a description beside it, lists them from the host, restores them there (safety backup, moved aside, extracted, rolled back on failure) and prunes to the number to keep; the live check minecraft-bedrock-ssh (174.9 s, a Debian host) backed up a running Linux Bedrock server both ways, read the on-host backup back from the host's own folder, and restored it with a safety backup. Gaps: a running Java server over SSH was backed up only by the same code path (the live run was Bedrock); one call to the host hung for 60 s once under load and the app now retries a hung call once. See backups. |
| Minecraft server manager: schedules | done | Restart with a chat warning, backup, message and update check; daily, weekly or every N minutes; stored in servers.json, run by the background app once a minute, a missed run happens once; on switch and a guided New schedule form. A self-test check computes next and missed runs and runs a due schedule through the runner; capture mc-schedules. See schedules. |
| Minecraft server manager: create flow | done | Six steps on the shared guided form with a step rail: the full Mojang catalogue with snapshot and old alpha/beta toggles, software builds from the Paper (Fill), Purpur, Fabric, Forge and NeoForge APIs (read live in the capture run), the four targets with live readiness checks, typed numbers with a free-port and memory check, world options, review with the EULA. Self-test checks drive the refusals and the full walk; captures mc-create-1 to mc-create-6. Bedrock edition offers Mojang's current Windows Bedrock Dedicated Server and is driven to the review by a self-test check; captures mc-create-bedrock, mc-create-bedrock-review. See create a server. |
| Minecraft server manager: Bedrock | done | The official Bedrock Dedicated Server for Windows from Mojang's download links service (minecraft.net over HTTPS only; Mojang publishes no checksum, so bedrock_server.exe must carry a valid Microsoft Corporation signature); create, start, stop, console with a typed command line, its own server.properties settings, backups of worlds with save hold and save resume. Self-test checks read its log line and list answer and walk the create flow; the live check created 1.26.61.1, saw "Server started." after 15.1 s, read 0 of 7, took a backup while it ran and stopped it (44.6 s). Added: the Players section edits allowlist.json and permissions.json (stopped: files, here and over SSH; running: console), add-ons are managed (see plugins and mods), an existing Bedrock world folder or .mcworld is imported (checked for level.dat and db, Java worlds refused, replace backs up and stops first, nothing left behind on a failure), worlds list with size and Download ZIP, and Bedrock now runs on Docker (the community image itzg/minecraft-bedrock-server, port published over UDP, console through its send-command; live check minecraft-bedrock-docker: created in 2 s, "Server started." after 286 s because the image downloads the server on its first start, list read, an allow-list entry written through send-command, stopped, container removed) and over SSH (Mojang's Linux zip downloaded and unpacked on the host over HTTPS from minecraft.net, settings written, started through the same console pipe as the Java process). The live check minecraft-bedrock-files started a real Windows Bedrock server twice: it made its world, then read the add-on and the allow-list entry the app had written (Pack Stack line and allowlist list reply) and an app-built allowlist add for a name with a space reached allowlist.json (34.9 s). Linux over SSH, started for real: the live check minecraft-bedrock-ssh made a Debian bookworm container with openssh-server, installed the server through the app's SSH route (8 s), started it ("Server started." after 3 s), read list (There are 0/7 players online), backed up the running server, and stopped it. Gaps: the Linux zip has no checksum or signature from Mojang (a genuine limit), so only HTTPS on minecraft.net is checked (the install also forces HTTP/1.1 because Debian's curl got an HTTP/2 error from that CDN); the Docker image is a community image, not Mojang's; Docker Bedrock settings changed in server.properties are kept except those the image's environment sets (none are set beyond EULA and VERSION); Bedrock world import is now driven through the real Open dialog (MT_ONLY=dialogs, step bedrock): on a sample Bedrock server the check presses Import a world, the native Open dialog ("Choose a .mcworld file, or cancel to choose a folder") opens on the hidden desktop, the lowlevel tools set its file name box to a prepared .mcworld and press Enter, and the check asserts worlds/<level>/level.dat and db/000001.ldb landed in the server folder and the page says it imported. Download ZIP is driven through its native Save dialog (step bedrock-zip): the dialog opens with Zip level.zip filled in, the lowlevel tools set its file name box and press Enter, and the check asserts the ZIP holds Zip level/level.dat and Zip level/db/000001.ldb and the page says it saved. The folder-chooser fallback of the import is driven too (step bedrock-folder): the Open dialog is cancelled with Escape, the native folder dialog ("Choose a Bedrock world folder (the one with level.dat and db)") opens, the lowlevel tools set its Folder box and press Select Folder, and the check asserts worlds/Folder level/level.dat and db landed, the source folder is untouched and the page says it imported. Genuine limits that stay: Mojang publishes no checksum for the Windows or Linux download, and the Docker route uses a community image. See servers. |
| Minecraft server manager: discover and adopt | done | Found on this computer lists Java processes running a server jar (command line and working folder read from the process), Bedrock server processes, and Docker containers whose image is a Minecraft server, leaving out servers the app manages and other Java programs; nothing is touched until Adopt, which records an entry pointing at the existing folder or container. An adopted container runs through Docker and is never removed by the app; an adopted process started elsewhere is shown Running and takes commands and a graceful stop only over RCON, and is never killed. Self-test checks classify a sample process and container list and drive Adopt (the folder is unchanged; stop and commands without RCON are refused). The live check discovered and adopted a real Java server started outside the app and stopped it over RCON (38.8 s), and a real itzg/minecraft-server container without the app's label, reading its state and server.properties and refusing to remove it (18.4 s). Captures mc-found, mc-adopted. See discovery. |
| Ollama suite manager | done | Detect, install with consent, start and stop, a dated model catalogue with a hardware-fit verdict from real RAM, VRAM and disk, a pull cart, model details, loaded models, a chat test and a troubleshooter. Chat takes up to six image or text attachments. The live check ran it against Ollama 0.33.3 with gemma4:12b: model list and capabilities, a text attachment read correctly and an image attachment described correctly. |
| Minecraft world generator | done | In the server manager's Worlds section. Seed, level type, structures, single biome, game-rule editor and superflat layers from the real jar, bounded pre-generation in the overworld, Nether and End, folder or ZIP output. Noise-settings and dimension-type editors. The live check downloads the real 1.21.4 jar and Java, reads game rules, blocks and biomes from it, and writes noise and dimension edits from its default files. Pause pre-generation and Resume hold the build at a batch boundary; batches are now cut to near-equal widths, the world is saved after each, and a batch that stops loading still restarts the staging server as a fallback (up to three times per batch). Live checks (MT_LIVE_ONLY=minecraft-lane) on the real 1.21.4 jar, read back from the files: pre-generation radius 16 paused after two batches with no progress for eight seconds, resumed and finished with all four regions; radius 17 and radius 23 (nine batches each) finished with no restarts at all (170 s and 221 s) once the staging server stopped pausing itself when empty; a normal world pre-generated the Nether and the End (4 non-empty region files each beside 4 in the overworld) and its level.dat holds keepInventory=true and randomTickSpeed=7; a custom superflat world (bedrock, emerald block, gold block) written as a ZIP holds those blocks in its finished chunks; a normal world written as a ZIP holds overworld, Nether and End regions (4 each) and level.dat; the superflat ZIP and the Single biome world, each with the Nether and the End switched on, hold 4 Nether and 4 End region files; a Single biome world of cherry grove holds only minecraft:cherry_grove in its finished chunks. Cancel while paused is checked at unit level, and the batch tiling by a self-test check. Stall root cause proven: since 1.21.2 a server with no players pauses ticking after 60 s (pause-when-empty-seconds), and force-loaded chunks only generate while it ticks; with the restart recovery off, 60 stalled radius 17 at batch 4 and 0 finished radius 17 and 23 cleanly. The staging properties set 0; the restart stays as a fallback. Other versions run live (MT_LIVE_MC, group gen, read back from the files): 1.20.6 passed in 362 s (4 region files each in the overworld, Nether and End, game rules in level.dat, normal and superflat ZIPs, a cherry grove single biome), and 26.3, the newest release in Mojang's manifest, passed in 337 s after two version differences were handled in the check: 26.x names game rules in snake case (keep_inventory; the engine refuses an unknown name and says so) and stores them in data/minecraft/game_rules.dat, and keeps dimensions under dimensions/minecraft/ rather than DIM-1 and DIM1. Limit, not a gap: Bedrock has no equivalent route. |
| Drawing and annotation | partial | Native Windows ink board (the named JavaScript library needs a web view this app does not have); pens, shapes, text, undo, PNG and SVG export. A self-test check opens a real board window and drives stroke, undo, redo, SVG and PNG render. The width now applies live: a stroke in progress is drawn by a renderer that reads the pen attributes at each moment, the keys [ and ] (Shift for 5) change the width while the pen is down, and the finished stroke takes the latest width. A check renders segments of one stroke before and after a width change and measures them (4, 20 and 10 px). Not tested with a real pen or touch device, by the maintainer's choice (2026-11-05) (the segments are fed to the renderer directly); the library itself is deliberately not used. |
| Publishing to a forge | done | Publish a folder to GitHub with bundled Git and an optional GitHub CLI, explicit consent, no stored tokens, recovery steps. Also GitLab (gitlab.com or self-hosted) and Gitea/Forgejo, with glab and tea preferred when signed in and a one-time pasted token otherwise (see publishing). The live check published to a real Gitea 1.22.6 server and to a real GitLab CE 19.5.3 server (a throwaway container, a one-time token) with the repository created, the commit pushed and the README read back through the server's API (MT_LIVE_ONLY=forge). The tea route ran live (2026-11-05, with the maintainer's approval to download it): the official tea 0.16.0, checked against its published checksum and signed in only to a throwaway Gitea 1.22.6 with its configuration in a temporary folder, created the repository ("with tea" in the steps, required by MT_LIVE_FORGE_EXPECT_CLI=1), and the pushed README was read back (8.8 s). The glab route ran the same way (official glab 1.128.0, checksum-checked, signed in only to a throwaway GitLab CE): it created the repository ("with glab" required) and the README was read back (6.7 s); glab kept that throwaway token in the Windows credential store during the run, and glab auth logout removed it afterwards. GitHub ran live too (2026-11-05, with the maintainer's approval): the signed-in GitHub CLI created a private repository on the maintainer's account, main was pushed and the README read back through gh (11.9 s, MT_LIVE_FORGE_KIND=github), and the repository was deleted afterwards. |
| External editor hand-off | done | VS Code, Notepad++, the default app or a custom program, for scripts and copied file paths. |
#Media file tool dialogs
| Feature | Status | Evidence and gaps |
|---|---|---|
| Trim and cut video (lossless or exact, keep or remove) | done | VideoTrimDialog; self-test checks video trim: ... cover lossless copy, exact cut, remove (both ways), existing name, Esc cancel with no leftover files, damaged file, keyboard route; captures media-trim-*. Details in files/media-dialogs.md. |
| Video to GIF, animated WebP or short MP4 with a target size | done | VideoGifDialog; checks video to GIF: ... cover a real GIF (codec, width, length, fps), WebP and MP4, a target that needs step-down, and a target that cannot be met (nothing saved); captures media-gif-*. |
| Audio trim, fade and loudness | done | AudioEditDialog; checks audio: ... cover length, loudness within 1 LU of the target measured with ebur128, a real fade in and out, FLAC and MP3 output, undecodable file; captures media-audio-*. The limiter path is not exercised by a clipping source. |
| Images resize, crop and compress (batch) | done | ImagesResizeDialog; checks images: ... cover exact pixel sizes with a 4:3 crop, originals untouched, a byte-target quality search, one damaged picture in a batch, the empty state; captures media-images-*. HEIC depends on a Windows decoder and is not tested; animated GIF is refused. |
| Image privacy clean-up | done | ImagePrivacyDialog, ExifData, ImagePrivacyEngine; checks privacy: ... read hand-built JPEG, PNG and WebP metadata, remove it (raw byte search, WPF metadata reader, pixel comparison, orientation kept), remove only some groups, replace in place, nothing to clean; captures media-privacy-*. TIFF is saved again losslessly but has no dedicated check; HEIC is unsupported. |
| Make an icon | done | MakeIconDialog, IconMaker, SvgLite; checks icon: ... read the ICO directory and decode every frame, check transparency, padding, circle shape, opaque background, a small source, an SVG source, an unreadable file; captures media-icon-*. |
| Media dialog notice buttons | done | MediaActions; check toast buttons: ... opens each dialog from the real notice action list. |
| Media dialogs in Cantonese | done | Strings.yue.Media.cs; check media dialogs: every visible string of every dialog state has a Cantonese entry audits 26 scenes; capture media-trim-ready-cantonese. |
#Not applicable here
| Feature | Reason |
|---|---|
| Browser-extension download hand-off | The app is not a browser extension. |
| Roblox models and realism | The app is not a game. |
| Encrypted public builder | The repository is public. |
View this article's Markdown source on GitHub
Suggested articles
- Language and voice
Cantonese review sheet
This sheet lists every entry of the Hong Kong Cantonese table (Strings.Table in Core/Strings.yue.cs and Core/Strings.yue.More.cs), English on the left and Cantonese on the right, grouped by the page…
Read it - Tools
Live check
the parts that only mean something against the real thing. It uses a throwaway data folder, so the real settings are
Read it - The app itself
Material 3 audit
A named checklist for Material 3 conformance, measured on the built app rather than read from source. The self-test tours every control-centre page, every other window and every state the language…
Read it - The app itself
Accessibility audit
A named checklist for keyboard use, accessible names, target size and contrast, measured on the built app. It runs on the same tour as the Material 3 audit: every page, window and state, in the dark…
Read it