manual-checks
Skill1 starsChanged 2 months agoApache-2.0View on GitHub
The manual-only verification registry for OS-bound checks and the rule for adding one; load when adding, running, or auditing an installed-build check.
- Reads credentials
- Sends data out
In this repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- userdocs/llms-full.txtllms.txt
- userdocs/llms.txtllms.txt
- briefSkill
- build-briefSkill
- drive-installed-appSkill
- run-millSkill
- mill-devSkill
- mill-useSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 91
Update now → Restart Mill, confirm the new version string.
- Very specific: line 209
Verify on an installed build: open Settings → Updates → "How updates stay trusted" → "Trust Mill's signing", confirm the authentication …
- Distinctive: line 391
Verify on an installed build on a table object and on Configure's List page: select a 2x2 range, ⌘C, paste into Numbers -- it lands as …
- Distinctive: line 447
Verify on an installed build with a multi-page draw.io file dropped on a board: select the object, two-finger scroll (the drawing moves, …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Manual-only verification registry
OS-bound checks that no headless harness can reach, listed explicitly with reasons — never silently absent (goal 0010's enforcement). Add an entry here, not a bare TODO, when a capability's interaction contract depends on real OS/hardware behavior a harness can't drive (a system dialog, a native drag, a login-item cycle) — name what's proven at another layer, what only an installed build can catch, and exactly how to verify it there.
Whether the packaged app launches at all and stays alive long enough to
render a window is NOT on this list — the desktop-launch job
(.github/workflows/ci.yml, goal 0384) already proves that on every PR
on a GitHub-hosted macOS runner, with no Accessibility/TCC grant needed.
This registry stays the ONLY list of checks that need this Mac specifically.
- Quick Panel drag/reset (goal 0377,
--wails-draggableonQuickPanel.module.css's facet-chip row and search-input header) -- a real native drag of a frameless WKWebView window is OS-bound the same way the tray panel's attachment is; the CSS drag/no-drag pairs and the position persist/clamp/reset logic are proven headless (quickPanelDragRegionConformance.test.ts,settingsservice_panelgeometry_test.go). Verify on an installed build: summon the panel, drag its header (not the search text) to a new position, dismiss and re-summon it -- it reopens where you left it; disconnect an external display the panel was dragged onto and re-summon -- it lands fully on the remaining screen, never off-screen; run "Reset Quick Panel position" from the main window's command palette and confirm it recenters and the row disappears from the palette until dragged again. - The menu-bar surface's OS half (goal 0189,
SystemTray.AttachWindow+newTrayPanelWindow) -- the tray ATTACHMENT is OS-bound end to end; the panel's CONTENT is e2e-proven at /#/traypanel (tray-panel.spec.ts). Verify on an installed build: the icon renders as a template glyph legible on BOTH light and dark menu bars; clicking it toggles the panel anchored under the icon (frontmost app stays active -- the panel is a non-activating NSPanel); clicking away dismisses it; parking an approval shows the count beside the icon (SetLabel) and clearing it removes the count; the panel's Stop actually cancels a live run; right-click still opens the Open Mill/Quit menu. - Tray panel on a secondary display (goal 0417,
traypaneldisplay.go'swireTrayPanelDisplay+CursorDisplayID) -- multi-monitor placement needs a real second display, which no headless harness has; the translation math is unit-proven (traypaneldisplay_test.go). Verify on an installed build with an external display attached: click the Mill menu-bar icon on the NON-primary display -- the panel opens under THAT display's menu bar, centred under the icon at the same 6px offset a single-display click already uses; clicking the icon on the primary display still opens it there. - Away-attention dock bounce (
dockBounceFn,settingsservice_attention.go) — the notify adapter's cgo send aborts headless; verify desktop-mode by parking an approval while unfocused. - Menu-accelerator suspension during hotkey recording
(
settingsservice_menu.go) — NSMenu's key-equivalent interception only exists in a real window; verify by arming a hotkey recorder, pressing ⌘⇧W / ⌘W / ⌘Q (each must be captured as a combo, never close/quit), then Escape/blur out and confirming accelerators work again. - The native menu bar (goal 0332,
menusvc.MenuService.Installwindowing.BuildMenu) -- server mode has no menu bar at all (InstallMenuis a build-tagged no-op there), so the projection, the accelerator translation and the enablement diff are Vitest-pinned and the tree builder is Go-tested headless, but no harness can open a real NSMenu. Verify on an installed build: every menu renders in bar order (Mill, File, Edit, View, Workflow, Atlas, Window, Help) and the app menu is titled "Mill"; About Mill opens the standard panel showing Mill and the running version; pick View > Configure and the app navigates, proving a menu click runs its command; on Home the whole Atlas menu is dimmed and on Atlas it is live; press ⌘W once and exactly ONE tab closes (never the window); press ⌘0 and it goes Home rather than resetting webview zoom, and ⌘+/⌘- still zoom the workflow canvas; Help > Report an issue... opens the issue tracker in the browser and Help
Open data folder opens Mill's data directory in Finder; finally arm a hotkey recorder (Settings > General > summon hotkey) and press ⌘W and ⌘Q -- each must be captured as a combo, including while the nested Help > Developer submenu exists.
- Dev-loop timing (
BuildIdentityBadge's amberDEV · go-stalestate,isGoSourceStale) — CI has no live file watcher; verify via.claude/skills/run-millby wedging awails3 devrebuild and confirming the badge flips. - Release-channel self-update (goal 0082) — verify after the next tagged release, on a release-installed copy: Check for updates → Update now → Restart Mill, confirm the new version string.
- Multi-select box-drag synthesis (goal 0081) — React Flow's pointermove delta sampling coalesces synthesized moves; verify by shift-dragging around 2+ cards desktop-mode.
- apply-notify's real banner (goal 0114) — an actual notification banner is OS-bound (signed-bundle handshake); verify by running the seeded Clipboard→Markdown workflow via its hotkey while another app is focused and confirming the banner.
- Close means hide, and the tray always brings the window back (goal 0276, main.go's WindowClosing hook) — the red traffic-light close, the hook's Cancel+Hide, and the tray's "Open Mill" restore are all native AppKit round trips no harness reaches. Verify on an installed build: red-close the main window, confirm Mill stays in the tray; tray → "Open Mill" restores the SAME window (state intact, no empty "No pending approvals." window appearing); then Quit from the tray and confirm the process actually exits, and ⌘Q likewise.
- Reopening Mill shows the main window only (goal 0344,
windowing.WrapAuxWindow+wireDockReopen) -- AppKit window restoration and the dock-icon reopen are both native round trips no harness reaches; the prompt's own content and close control are e2e-proven at /#/approvalprompt (approval-prompt.spec.ts). Verify on an installed build: open the Quick Panel, the run monitor and the capture window, quit Mill with all three on screen, then click the dock icon -- only the main window comes back, no floating panel and no frameless "No pending approvals." window; click the dock icon again with Mill running and every window hidden and the main window is the only thing that appears. - App archetype: closing the last window must NOT quit Mill
(goal 0188) —
Mac.ActivationPolicyis Regular andApplicationShouldTerminateAfterLastWindowClosedis false, which no headless check can observe (server mode has no AppKit delegate at all). Verify on an installed build: close the main window with ⌘W and confirm Mill is still running and reachable from the tray; reopen it from the tray; then quit deliberately from the tray's Quit item and confirm the process actually exits. The flag composes lethally with any path that empties the screen — it terminated the app on a background summon, and once before via a window-closing accelerator during hotkey recording. - Summon from a background app (goals 0151, 0182, 0188) — the real kill chain (hotkey → activation → HideOnFocusLost → the main-window restore) needs a real macOS activation dance; verify on an installed build with Accessibility granted, main window already open in the background (a different app frontmost): hotkey from the other app — confirm the panel appears, Mill's main window does NOT flash into view alongside it (goal 0035), and the process survives (no SIGTRAP); dismiss via Escape, then again via click-away, then again via pressing the hotkey a second time — for each path, confirm neither Mill nor the previously-open main window vanishes from the app switcher/dock, and focus returns to the app you summoned from; finally reactivate Mill (dock click or Cmd+Tab) and confirm the main window is back, right where it was before the summon.
- The run monitor window (goal 0294 S2,
newRunMonitorWindow+ShowRunMonitor/HideRunMonitor) -- the floating window, its close-means-hide hook and the Quick Panel / tray doors that show it are native round trips; the route's CONTENT is e2e-proven at /#/runmonitor (run-monitor.spec.ts). Verify on an installed build: ⌘⇧↩ on a Quick Panel row opens the window floating over the current app with the run stepping; the red close hides it; ⌘⇧↩ again re-shows it on the new run; a tray Recent row opens it on that run; Open in Mill hides it and lands the main window on the same run. - Bringing a hidden app back on screen (goal 0186, goal 0188 slice
2 —
bringMainToFront/bringFloatingToFront,settingssvc/settingsservice_presence.go) — every path that shows a window now un-hides the app first, which only a real app-level hide can exercise. Verify on an installed build: repeat the background- summon dismiss above until Mill is app-hidden (no window, not in the app switcher), then summon again and click the panel's "Open Mill" row — confirm the main window actually appears, where it previously did nothing silently. Separately, repeat with the tray's "Show Mill" item instead of the panel. Separately, park a guardrail/MCP-write decision while away and Mill app-hidden — confirm the floating approval prompt appears rather than staying invisible behind a still-hidden app. - Native file-drop delivery (goal 0081 A3, extended by goal 0179
S2) — verify by dragging a
.mdfile from Finder onto the running app and confirming the card lands with the real path; separately, drag a.drawioand a.mmdfile and confirm each lands as a "diagram" board object (not a card); separately, drag a.pdfand confirm it lands as a "pdf" object rendering page 1 through the vendored viewer (goal 0267 -- the routing decision is Vitest-tested, the rendered result e2e-proven via the seeded two-page document). The DROP GESTURE itself is a structural gap in this harness (Wails3'sWindowFilesDroppedneeds a real*WebviewWindow, which server-mode Playwright's connection is not) — the routing decision is Vitest-tested instead (useAtlasDiagramObjectCreate.test.ts'sisDiagramPath). The RESULT of a drop (the object's own board-face rendering, and Promote to card) IS e2e-proven despite the gesture gap:atlas-diagram-object.spec.tslands the object viafixtures/atlasNativeDropEscapeHatch.ts— a direct call to the exact sameCreateBoardObjectRPC every tray click already goes through, by its stable Go method name — then drives every assertion through the real rendered DOM. - Companion panel against a real local model (goal 0101) — verify desktop-mode with Ollama running: judge token-by-token responsiveness and whether replies parse into intended proposals often enough to be useful.
- MCP address editable/save/validate path (goal 0116) — every e2e
worker sets
MILL_MCP_ADDRfor port isolation, so an env override is structurally always active in the shared pool; verify desktop-mode (no override set) by entering an address in Settings (confirm the restart note), then a malformed one (confirm the validation message). - Signing identity survives an update (goal 0158) — real trust- settings state is machine-specific; verify by granting Accessibility once, taking two consecutive beta updates, and confirming the summon hotkey still registers with no new permission prompt.
- The real "Trust Mill's signing" authentication dialog, and the
section hiding once trust is granted (goal 0220 S3,
codesigning.TrustIdentity/IsTrusted) —security add-trusted-certagainst the live per-user Trust Settings domain needs a real Window Server session to show its SecurityAgent prompt at all (headless/non-TTY calls have been observed to return success without recording anything); e2e only reaches the server-modeErrUnsupportedPlatformfail-closed path (updates-trust-signing.spec.ts, which now asserts the section absent rather than driving its button), never the real dialog. Verify on an installed build: open Settings → Updates → "How updates stay trusted" → "Trust Mill's signing", confirm the authentication dialog appears and the button shows its trusted confirmation once you approve it; run it again and confirm it stays idempotent (no error, no duplicate identity); then revisit Settings → Updates and confirm the whole "How updates stay trusted" section is now gone. - SMAppService launch-at-login: the requires-approval state and the
legacy System-Events migration (goal 0198,
internal/adapters/launchatlogin) — real SMAppService registration state (enabled / requires-approval) and a real pre-existing System Events login item both need an actual login cycle on an installed, signed bundle; neither exists in the e2e harness (server mode has no login-item concept at all). Verify on a build from BEFORE this goal, installed and with "Launch Mill at login" turned on (a real System Events login item now exists), then upgrade to a build from this goal or later: open Settings > General, turn "Launch Mill at login" off then on again, and confirm System Settings > Login Items shows exactly ONE Mill entry, never two. Separately, on a machine where Mill has never been granted a login item before, turn it on and confirm the checkbox shows checked with the amber "Confirm in System Settings to finish turning this on." notice and its button, then approve it in System Settings and confirm the notice clears on the next Settings visit. - The real system authentication sheet gating an unlock (goal
0330,
internal/adapters/localauth) —evaluatePolicyraises out-of-process UI (LocalAuthentication's own XPC UI service) that no headless build can trigger or dismiss, and the e2e suite runs a server build where the framework is not compiled in at all, so it only reaches the "not set up on this Mac" disabled-toggle path (secrets.spec.ts). Verify on an installed build: in Secrets, unlock the vault, open Secrets > Locking, turn on the unlock requirement and confirm the toggle sticks; lock the vault, press Unlock, and confirm the system sheet appears BEFORE the vault opens; cancel it once and confirm the vault stays locked with "Unlock cancelled." on screen; authenticate and confirm it opens. The checkbox's own label is device-aware, so on a Mac with Touch ID it must read "Ask for Touch ID or your password before unlocking" and on one with only a password "Ask for your password before unlocking". On a Mac with no Touch ID enrolled and no password set, confirm the toggle is disabled with "Touch ID or a password isn't set up on this Mac." - The vault locking on sleep, screen lock and user switch (goal
0360,
internal/adapters/windowing/locktriggers_darwin.m) — the three OS notifications behind these (NSWorkspaceWillSleepNotificationvia the toolkit,com.apple.screenIsLockedon the distributed centre, andNSWorkspaceSessionDidResignActiveNotification) are only posted by a real login session; a server build subscribes none of them, and the policy that consumes them is Go-tested with synthetic triggers instead. Verify on an installed build, in Secrets > Locking with "Lock when this Mac sleeps or the screen locks" on: unlock the vault, press Control-Command-Q to lock the screen, log back in and confirm Secrets shows the locked state. Repeat with the checkbox off and confirm the vault is still unlocked. Then unlock the vault, close the lid (or Apple menu > Sleep), wake the Mac and confirm it locked. With "Lock when switching users" on, unlock the vault, use Fast User Switching to another account and switch back, and confirm it locked. With "Lock when Mill's window is minimized" on, unlock the vault, minimize Mill's window into the Dock, restore it and confirm it locked. - A vault whose key is missing shows the mismatch line, and Start a
new vault keeps a
.bak(goal 0330) — the failure only happens against a REAL OS keychain whose contents differ from the vault file beside it, which no test may touch (the keychain is machine-global; e2e servers run on an in-memory keyring instead). Verify on an installed build: quit Mill, move~/Library/Application Support/mill/secrets.kdbxaside and restore an older copy (or delete themill-secret-vault-key.<id>item in Keychain Access), relaunch, open Secrets and press Unlock. Confirm the line reads "The key on this device doesn't open this vault file." (or "There's no key for this vault on this device.") and that "Start a new vault" appears beside Unlock; take it, confirm the dialog, and confirm asecrets.kdbx.<timestamp>.baksits beside the new vault file with the old one's bytes intact. - The real sudo askpass / Touch ID escalation prompt (goal 0240
S5,
wrapArgvForAdmin/materializeAskpass) — sudo's PAM conversation (pam_tid's Touch ID sheet, or the osascript hidden-answer password dialog) is out-of-process system UI no headless harness can trigger or dismiss; unit tests pin the argv wrapping, askpass content/mode, and the secrets-refusal, and the guardrail tests pin the always-asks policy — never the real prompt. Verify on an installed build: configure the seeded "Run from clipboard" shell step with "Run with admin rights", runwhoami, approve the forced ask, confirm the Touch ID sheet appears (pam_tid configured) and the output readsroot; cancel the prompt on a second run and confirm the step fails with sudo's own error rather than hanging; then confirm an allow-listed command STILL parked for approval while the toggle was on. - The real browser-tab approval notification (goal 0132 slice A) — requires a real granted browser permission and a real OS compositor; verify via a server-mode instance reached from a real browser tab: enable notifications in Settings > Remote access, park an approval from another tab/device, switch away, confirm a system notification titled "Approval needed" appears, and clicking it lands on Review.
- The real Android ntfy delivery (goal 0132 slice B) — the ntfy wire protocol reaching a real Android device is OS/vendor-app-bound; verify by installing the ntfy Android app, subscribing to a paired device's URL, parking an approval, confirming the phone receives it backgrounded, and tapping lands on Review.
- Board ⌘V with a REAL pasteboard: screenshot bitmap and Finder
⌘C (goal 0255,
ReadPasteboardFilePaths/clipboard.ReadFileURLs) — real pasteboard file flavors and WKWebView's own ⌘V event delivery are both OS-bound (the e2e drives a synthesized files-carrying paste and the fail-closed empty-paths branch, never the real gesture). Verify on an installed build: ⌃⇧⌘4 a region (clipboard screenshot) → ⌘V on the board lands the image at the pointer; Finder ⌘C a .png → ⌘V lands an image object mirroring the REAL file path; Finder ⌘C a .md → ⌘V lands a card, same as dropping it. - "Copy as image" landing a real, pasteable picture (goal 0201,
clipboard.WritePNG/ClipboardHistoryService.CopyImagePNG) — the Go side's PNG flavor registration is Go-tested against the real pasteboard and the rasterizer's exclusion rules are unit-pinned, but whether ANOTHER app accepts what Mill put there is OS-bound. Verify on an installed build: select two cards and a shape on a board, run Copy as image (Atlas menu or the selection's right-click menu), then ⌘V into Preview (File > New from Clipboard) and into Keynote — both must show the selection at 2x with NO selection outline, no drag handles and no resize frame, on the board's own background colour. - A pdf link annotation actually reaching the system browser
(goal 0271,
openExternalUrlvia the runtime's Browser API) — launching the real default browser is OS-bound; the e2e pins the negative (the click never navigates the app/viewer) and suppresses the HTTP-transport call, but the runtime may route over its WebSocket transport, so no harness observes the real open. Verify on an installed build: click an external link inside a live pdf object — the system browser opens the page and Mill stays exactly where it was. - Canvas navigation with a REAL trackpad (goal 0257,
shared/canvasNavigation.ts) — Playwright's wheel is synthetic; a real two-finger scroll, pinch, and ⌘-scroll are OS/driver-bound gestures no harness can produce. Verify on an installed build in default Trackpad mode: two-finger scroll pans the Atlas board in both axes (no zoom), pinch zooms around the pointer, ⌘-scroll zooms; flip Settings > General > Canvas navigation to Mouse and confirm a mouse wheel zooms; confirm the workflow editor follows the same mode without a reload. Additionally (goal 0271): with a live (selected) PDF object and a draw.io object on the board, two-finger scroll in every direction with the pointer inside each viewer -- the viewer scrolls/zooms and the BOARD holds still (no simultaneous pan), including diagonal gestures over a page-width PDF whose horizontal axis has nothing to scroll; with the object UNSELECTED (shield up -- pdf AND diagram carry it, goal 0302), the same gesture pans the board only, and pinch / ⌘-scroll zoom the board. Synthetic wheel can't reproduce the real gesture stream's double-handling, so this stays manual. - The quit gate on every native quit path, and explicit mode's
close guard (goal 0295 S2b,
SettingsService.ShouldQuitwired asapplication.Options.ShouldQuit,HideMainWindowGuarded) -- ⌘Q, the Dock's Quit, the tray menu's Quit and the red close all reach AppKit's applicationShouldTerminate / the WindowClosing hook, which no server-mode harness has (the e2e drives the same handshake through RestartApp). Verify on an installed build with Settings > General > Save changes = When I choose: type into a note and click away (the dot shows); ⌘Q shows the Save all / Discard / Cancel sheet and Cancel keeps Mill running; hide the window, then the tray's Quit brings the window back with the sheet; Discard quits. Again with a held note: red-close the window shows the sheet titled "...before closing?" and Save all hides the window with the note written. Then switch back to Automatically and confirm ⌘Q quits with no sheet and the note text survives relaunch. - The adopted table grid's range and drag interactions (goal
0287, ADR-0049,
shared/ListGridGlide.tsx) -- range ⌘C/⌘V, the fill handle, header-edge resize and header drag-reorder are the library's own pointer gestures over one canvas; Playwright's forced positional click reaches a single cell but not a drag or the real clipboard. Verify on an installed build on a table object and on Configure's List page: select a 2x2 range, ⌘C, move, ⌘V (the rows write through); drag the fill handle down a column; drag a header edge (width survives a reload on this device) and drag a header onto another (the List's column order changes on every projection); then a VoiceOver pass -- the grid's accessibility DOM reads the headers and cells (ADR-0049's own verification, the library hedges its a11y). - The grid's clipboard against a REAL third-party app (goal 0349
S4,
shared/ListGridGlide.tsx) --list-grid-interactions.spec.tsgrants Chromium clipboard permission via CDP, which the desktop WKWebView engine has no equivalent for, so every e2e round trip there stays Mill-to-Mill; whether the OS pasteboard flavor Mill writes/reads actually interoperates with another app is untested. Verify on an installed build on a table object and on Configure's List page: select a 2x2 range, ⌘C, paste into Numbers -- it lands as tab/newline text across the right cells; separately, select and copy two rows in Numbers, click a cell in the grid, ⌘V -- the cells fill from that cell, rows append past the List's last row, and pasting past where appending is allowed shows the dropped-rows footer notice. - A relaunch never restores the floating windows (goal 0301,
SettingsService.HideAuxWindowson ApplicationStarted and before every approved quit / restart) -- macOS Resume re-showing windows that were on screen when the process ended is an OS behavior no harness reproduces. Verify on an installed build: open the Quick Panel, press ⌘⇧↩ on a workflow so the run monitor floats, then from the panel type "update" and restart (or ⌘Q from the panel's Open Mill) -- on relaunch only the main window appears, the panel and the monitor stay hidden until summoned; also relaunch with "Close windows when quitting an application" OFF in System Settings > Desktop & Dock (the default) and confirm the same. - The system accent colour (goal 0320,
internal/adapters/systemaccent) --NSColor.controlAccentColorreaches AppKit through the toolkit's ownEnvironmentManager, which returns "" in server mode, so e2e only exercises the keep-the-teal branch; the derivation is Vitest-pinned. Verify on an installed build: set System Settings > Appearance > Accent colour to Purple, relaunch Mill, confirm the selected sidebar row, a primary button and a focused field's ring follow it in light and dark; set it back and relaunch. The accent is read per window at mount, so a change made while Mill runs appears on the next launch. - File-promise drops: the post-screenshot floating thumbnail
(goal 0256,
MillFilePromiseDropView/AttachFilePromiseReceiver) — a promise drag needs a real AppKit drag session end to end; no harness can synthesize one. Verify on an installed build: ⌘⇧4 a region, drag the floating THUMBNAIL (before it saves) onto the board — the image object lands at the drop point, no no-entry cursor; drag an image out of a browser page — same; then drag a real file from Finder and confirm it still lands exactly as before (the promise view sits BELOW the toolkit's own drag view precisely so filename-carrying drags never reroute). - Drag regions are opt-in (goal 0333,
frontend/src/app/index.css's--wails-draggable: no-dragdefault, enforced statically byscripts/check-drag-regions.sh) — whether a pointer press actually moves the NATIVE window is a real AppKit drag session; server-mode Playwright has no native window to move. Verify on an installed build: dragging on an empty pane area (Home, a canvas, Configure) neither moves nor resizes the window; dragging the tab band moves it. - Trackpad pan and pinch zoom inside a selected diagram (goals
0340 and 0354 S2,
frontend/src/atlas/drawioInteraction.ts+frontend/src/atlas/atlasActivation.ts) — a real two-finger scroll and a pinch reach WKWebView with differentdeltaMode, delta magnitudes andctrlKeysynthesis than the Chromium-based suite dispatches, and macOS momentum scrolling has no headless equivalent: the bridge smoke's owndiagram-face-owns-wheelcheck dispatches a wheel rather than producing a gesture, so it proves the ROUTING in the real engine and never the momentum stream. Verify on an installed build with a multi-page draw.io file dropped on a board: select the object, two-finger scroll (the drawing moves, the board holds still — including at the drawing's own edges, where nothing more is left to pan — and the momentum tail does not overshoot into a runaway pan), pinch (the drawing zooms about the fingers, not the frame's centre), press and drag inside it (the drawing pans, the object stays put), two-finger scroll over its chrome band (the BOARD pans, so a selected object is never a dead zone), drag that band (the object moves), and confirm the board itself still pans when nothing is selected. - A real vendor-issued certificate against a real mTLS host (goal
0306 S1,
ConfigureService.TestClientCertificate+clientcert.StateFor) — a real TLS handshake against a host that demands a client certificate, and a real X.509NotAfterto classify, both need a real CA and a real server no harness can stand up; the e2e suite proves the resolve/status/error paths against a self-signed fixture instead. Verify on an installed build: in Configure > Certificates, add an entry pointing at a real vendor- issued PEM certificate + key pair (and separately, a real PKCS#12 export from Keychain Access or a CA portal) for a host that actually requires mTLS; press Test and confirm the handshake succeeds; make a request to that host through the request runner and confirm it presents the certificate; confirm the row's status reads "Expires in N days" matching the certificate's realnotAfterdate. - The real browser extension, loaded unpacked and paired (goal
0350 S1,
examples/browser-extension/) — an MV3 extension needs a real Chrome profile and a genuine<all_urls>grant; the suite proves Mill's half against a wire-protocol stand-in (e2e/fixtures/fakeExtension.ts), the runner's own logic under a DOM (src/shared/replayRunner.test.ts), and (goal 0418)browser-extension-mv3.spec.ts's real unpacked extension: pairing through its own popup and Mill's real Settings accept, achrome://serviceworker-internalsStop reproducing the platform's own idle teardown, and the parked replay succeeding once the worker wakes and reconnects withinbeginRun's own wait window. Verify on an installed build: loadexamples/browser-extensionunpacked atchrome://extensions, open Settings › Connections › Browsers, press Pair a browser, enter the code and the shown address in the extension popup (it reads "Connected to Mill"), then press Test the connection — a tab opens, the button is pressed, and Mill reports three steps. Then record a flow in the DevTools Recorder panel against a real signed-in site, export it as JSON, and replay it — the selector fallback chains resolve on a page whose markup nobody controls. Finally revoke the browser in Settings and confirm the popup, once reopened, reads "Not paired" (the socket closes on its next keepalive tick,KeepaliveSeconds). - The Secrets page's native pickers (goal 0306 S4,
windowing.PickFolder/PickCSVFile,SecretService.ChooseScanFolder/ChooseExportFile) — server mode falls back to a typed path, which is what the e2e suite drives instead. Verify on an installed build: in Secrets › Sources, open "Find .env files…" and press Choose — the real native folder picker opens with directories choosable and files disabled, and the chosen path fills the Folder field; in Secrets, open Import… and press Choose — the real native file picker opens and the chosen file's entry count shows in the preview. - The webhook door's real OS-banner and phone fan-out (goal 0368,
bridgesvc/bridgeservice_webhook.go+triggersvc/triggerwebhook.go) -- the door's auth/malformed-body/dispatch paths are Go-tested and the seeded workflow's title/body substitution is proven against a captured notifier, but a real OS banner and a real phone delivery are OS/vendor-bound. Verify on an installed build: add a webhook token in Settings > Connections > Webhooks (copy it at mint; it never shows again), thencurl -s -X POST http://127.0.0.1:8092/__mill/webhook -H "Authorization: Bearer <token>" -H "Content-Type: application/json" -d '{"source":"mytool","title":"Webhook fired","body":"A test event."}'-- the seeded "Notify when a webhook fires" workflow's run lands in the notification list with the posted title and body, a desktop banner fires, and a paired phone receives the notification. Then revoke the token in the same list and confirm the same curl answers 401. - A real repository install and update of an extension (goal 0349
S5,
pluginsvc/{marketplace,install,updates}.go) — server mode exercises the marketplace, install and update doors against a loopback publisher only; a live GitHub release is the one path CI never reaches (no outbound calls in tests). Verify on an installed build: in Extensions › Browse › Sources, addowner/repofor a repository carrying.mill/marketplace.json, install its entry and read its Verification tab; then publish a newer release tag on that repository with<id>-<version>.zip(andSHA256SUMS), press Check for updates, confirm the tab reads Updates (1), press Update, and confirm the version and tier the Verification tab now shows. - Bulk delete in Secrets behind a real Touch ID unlock (goal 0404
S1) -- Secrets' selection/bulk-delete is deliberately undo-less (no
journal registration, unlike every other bulk-deletable list), so
its only proof of correctness is the real no-Undo toast and the rows
actually gone; reaching it also needs the vault unlocked, which is
goal 0330's own real authentication sheet (
internal/adapters/ localauth) that no headless build can trigger, sosecrets.spec.tsdrives bulk delete only against a server build where the vault is never Touch-ID-gated at all. Verify on an installed build on a Mac with Touch ID enrolled: unlock the vault via the real sheet, select two or more secrets and press Delete, confirm the toast reads "Deleted {n} secrets." with NO Undo button (unlike every other entity's bulk-delete toast) and the rows are gone; confirm ⌘Z does nothing for this deletion; lock the vault again and confirm the deleted secrets stay gone, never recoverable. - Long-press on a real touchscreen (goal 0404 S1,
useListSelection'spointerType === 'touch'gate) --list-selection.spec.ts's companion-width long-press case proves the state machine via a syntheticpointerdown/pointeruppair (testing.md's own last-resort carve-out, since Playwright's touchscreen API offers onlytap(), no real press-and-hold), so a genuine touch-and-hold gesture's timing and cancel-on-move have never been driven by an actual touch digitizer. Verify on an installed build on a touch-capable device (or iOS/iPadOS companion once that surface exists): at companion width, press and hold a row with a finger past the threshold -- it enters selection mode; press and hold, then drag before releasing -- it does NOT enter selection mode (the same move-tolerance the pointer-based e2e case never exercises with a real finger's jitter). - The Accessibility grant survives a beta update (goal 0363 S1b,
.github/workflows/ci.yml's beta-release job +scripts/import-beta-signing-cert.sh) — whether a real beta install's designated requirement stays constant across a rebuild is a real TCC grant on this Mac's own permission database, invisible to any headless harness; CI's own "Verify beta signing identity" step only proves the SIGNED BUNDLE's own designated requirement, not that macOS actually preserves the grant across it. Verify on an installed build: install one published beta.zip(https://github.com/millhq/mill/releases), launch it, grant Accessibility when prompted (or add it in System Settings › Privacy & Security › Accessibility), and confirm the global summon hotkey works. Quit, install a LATER published beta over it the same way, relaunch, and confirm the hotkey still works with no new Accessibility prompt and no new row added under Accessibility for Mill. - Secrets surface states need the vault unlocked once by a person (goal 0408 S2b) -- the local vault is protected by Touch ID or a password; an automated driver cannot satisfy the LocalAuthentication sheet and must never reset the vault. Verify on an installed build: unlock the vault once, then walk the Secrets list at rest, a row's hover actions, a row's detail dialog, "Open source" navigation, and an unresolved source-backed row -- screenshot each.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

