Sans FireDAC.Phys.SQLiteWrapper.Stat, FireDAC charge sqlite3.dll au
demarrage. Sur ce poste ca marchait quand meme : le dossier bin
d'Embarcadero, qui en contient un, est dans le PATH. Sur une machine
sans Delphi, l'exe refusait de demarrer.
C'est le Win64 qui l'a revele — la DLL du PATH est 32 bits — mais le
Win32 livre etait casse de la meme facon chez quiconque n'a pas Delphi,
paquet prepare pour un tiers compris.
Prouve avec un PATH reduit a Windows : avec l'unite SQLite s'ouvre
(3.42.0), sans elle l'erreur exacte. Lire la table d'import ne
suffisait pas, le chargement est dynamique.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
L'ecran d'activation ne consultait qu'un branchement : le rapiecer
ouvrait tout, l'API ne demandant rien. Chaque route de PM.Router passe
maintenant par PM.LicenceGate.ApiLicenceOk avant son handler.
Aucune fonction de decision partagee avec PM.Licence — sinon un seul
rapiecage ouvrirait les deux couches. La verification est refaite de
bout en bout, appels CNG redeclares compris ; seules la cle publique et
le nom de la pref sont partages, parce que ce sont des donnees.
Branchee dans DispatchRequest : les fichiers de l'ecran d'activation
passent par TryServeEmbedded, non garde, sinon l'ecran ne se chargerait
plus. Un oui est retenu pour le processus, un non jamais.
Le risque etait qu'elle refuse un client licencie. Auto-test sur un
vecteur lie a une empreinte fixe, sens accepte en premier ; et accord
verifie en conditions reelles sur ce poste : IsLicensed = TRUE,
ApiLicenceOk = TRUE.
Documente aussi le piege du jour, sans lien avec ce code : des .dcu TMS
x64 deposes a cote des sources cassaient le build Win32, et la sortie
redirigee vers E: faute de place sur Z:.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
La premiere privee a ete supprimee sans copie : plus aucune licence
nouvelle n'etait emettable. Regeneration + reemission, indolore
aujourd'hui parce qu'une seule licence circulait.
licence.ps1 ecrit desormais la cle dans %USERPROFILE%\Documents, hors du
depot. Elle vivait dans tools\, ou elle n'attendait qu'un menage : le
.gitignore la protegeait du depot, rien ne la protegeait de son
proprietaire.
Verifie : la licence de la nouvelle paire passe, celle de la paire
perdue tombe — la regeneration a bien remplace, pas ajoute.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Le frontend etait lisible dans l'onglet Image de Process Explorer, avec
les commentaires. BuildAssets.ps1 brouille chaque fichier avant brcc32,
PM.EmbeddedAssets le debrouille au moment de servir.
C'est de l'opacite, pas du secret : la cle est dans l'exe. Ca supprime
strings et l'onglet Image ; la memoire, un debogueur et la reponse HTTP
rendent toujours le texte. D'ou XOR plutot qu'AES — a securite egale
(nulle contre qui a l'exe), dix lignes, et aucune force affichee qu'on
n'a pas.
Une magie PMA1 prefixe le clair : sans elle, deux constantes qui derivent
serviraient du binaire au navigateur, ecran blanc sans trace.
Verifie en comparant les flux des deux cotes bloc par bloc, puis par
aller-retour sur app.js, index.html et style.css. Le frontend ne se
trouve plus dans le binaire.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
BoardUUID s'appuyait sur un COM initialise par quelqu'un d'autre. Appele
apres WebView2 il lisait le vrai UUID de carte mere, appele avant il
echouait et retombait sur MachineGuid : deux empreintes pour une seule
machine, et une licence liee a l'une que l'autre refuse. Il suffisait que
licence/status parte avant l'init du navigateur pour verrouiller
l'utilisateur hors de son coffre.
CoInitializeEx est desormais appele sur place. Pas de CoUninitialize
symetrique : essaye, il rend une violation d'acces — les mandataires WMI
et le cache de type-info de ComObj se liberent plus tard, dans un
appartement qu'on vient de fermer.
Verifie avec un harnais qui n'initialise PAS COM : il rend maintenant la
meme empreinte que l'application.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Une cle flottante etait sortie. Un schema hors ligne n'a pas de
revocation — rien ne peut aller dire a un exe installe qu'une cle ne vaut
plus — donc la seule prise est de cesser d'accepter toute la classe.
MachineMatches exige desormais une empreinte, et -Floating disparait du
script : il ne peut plus fabriquer une cle que l'exe rejetterait.
Effet de bord assume : plus de licence de site, et un poste sans
empreinte exploitable n'est plus licenciable du tout — mieux qu'une
empreinte degeneree qui ouvrirait partout.
Verifie : liee a ce poste acceptee, liee a un autre refusee, les deux
flottantes deja emises refusees.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
C'etait la decision prise, et je l'avais mal lue : "pour le moment 1"
repondait a la question 1, pas "une seule licence".
PM.MachineId melange l'UUID de la carte mere (WMI) et la serie du volume
systeme, puis n'en publie que 16 hexa : l'empreinte se recopie dans un
courriel, l'UUID brut n'a pas a circuler. Un UUID nul ou tout a FF est
rejete — VM et images clonees l'annoncent, et deux postes tomberaient
alors sur la meme empreinte.
Le champ m est compare APRES la signature. m absent = licence flottante,
acceptee : seule la cle privee peut en emettre, et le champ ne peut pas
etre retire sans casser la signature. Cote script -Machine est
obligatoire sauf -Floating explicite.
L'ecran d'activation affiche l'Installation ID : une licence liee ne peut
plus etre emise avant l'installation.
Verifie sur ce poste : liee a lui acceptee, liee a un autre refusee,
flottante acceptee.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
L'ecran d'activation devient franchissable. Verifie en executant l'unite
contre la vraie cle : licence emise acceptee et charge utile lue, licence
signee par une AUTRE paire refusee — c'est celle-la qui prouve que la cle
tient reellement la porte — charabia et chaine vide refuses.
La cle privee reste hors du depot (gitignoree) et doit etre deplacee dans
un coffre : qui la detient fabrique des licences.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
PM.Licence verifie une chaine charge-utile.signature contre la cle
publique de l'editeur, via BCryptVerifySignature. Aucun ASN.1 : la
signature du script est deja en r||s brut et les coordonnees se rangent
telles quelles dans un BCRYPT_ECCPUBLIC_BLOB.
La cle publique est livree A ZERO — tant que la sortie de licence.ps1
-Init n'y est pas collee, aucune licence n'est valide. C'est le defaut
sur : un blob par defaut qui accepterait quelque chose serait pire.
IsLicensed REVERIFIE la licence rangee plutot que de lire un drapeau.
La signature est verifiee AVANT que la charge utile soit analysee.
SelfTestLicence a l'initialization rejoue un vecteur fige signe par une
paire de test : valable accepte, un octet retourne refuse. Verifie en
executant l'unite, pas seulement en la compilant.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Le lien n'etait garde par rien : un port pris levait, et l'application
s'ouvrait sur une boite d'erreur. Trois essais desormais - le premier
honore le port de config.txt, que l'utilisateur a peut-etre choisi
expres, les suivants prennent un ephemere.
SE LIER EST LE TEST. Sonder le port avant serait plus faible - un service
qui ecoute sans repondre passerait inapercu - et sujet a une course : le
port peut etre pris entre le sondage et le lien. Chaque deplacement est
journalise, un serveur qui change de port en silence etant pire qu'un
serveur qui refuse de demarrer.
Trouve en chemin : Randomize n'est appele NULLE PART, donc le port
"aleatoire" de port=0 etait le meme a chaque lancement - l'aleatoire cense
compliquer un balayage exterieur n'existait pas. Il vient maintenant de
RandomBytes, donc de BCryptGenRandom, deja utilise pour le jeton d'acces.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
La barrette et le raccourci se rejoignent des FireAutofillRequest, donc
tout ce qui suit etait journalise sous "hotkey" - y compris un
remplissage declenche a la souris. Un diagnostic de lenteur qui ne sait
pas par ou l'on est entre cherche au mauvais endroit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Un autofill lent apres une periode d'inactivite peut l'etre pour des
raisons sans rapport qui se ressemblent parfaitement de l'exterieur : le
moteur WebView2 qui se reveille, la prise du premier plan, ou l'attente
de la relache des modificateurs. Elles se corrigent a des endroits
opposes, donc on les separe au lieu de deviner laquelle accuser.
Trois marques : le temps que met la PAGE a repondre au raccourci, le
total, et le detail des deux attentes internes.
Premiere mesure : 13 ms de reponse apres 23 minutes en tray, total
635 ms. Le renderer ne dormait donc pas - l'hypothese la plus evidente
etait fausse - et 635 ms est le cout nominal de la sequence, faite de
pauses fixes (80 + 120 + 60 + 200 + 200). Le cas lent reste a attraper.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Le premier nomme la chose, le second decrivait l'utilisateur. Le signe
qui a tranche : l'ancien libelle avait besoin d'une parenthese pour se
faire comprendre - "Lazy mode (fill with the mouse)" - et c'etait elle
qui portait le sens.
Seul le libelle change. Les identifiants internes gardent "lazy" -
FLazyWnd, LAZY_*, et surtout les prefs lazyOn / lazyPos / lazyAlphaPct :
renommer ces cles perdrait les reglages deja poses sur la machine.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pref DPAPI lazyOn, reprise par RestoreLazyBar depuis FormCreate - apres
le theme, parce que la barrette se peint elle-meme, et apres le cablage
de OnTrayRestore, dont depend son bouton Open.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Quand la barrette est allumee, elle est souvent la seule chose de
l'application a l'ecran : revenir au coffre ne doit pas demander de
retrouver l'icone du tray.
Le menu d'opacite passait derriere : le battement qui repose la barrette
en tete de la bande TOPMOST cinq fois par seconde recouvrait le menu que
la barrette venait d'ouvrir - lui aussi dans cette bande. TrackPopupMenu
est modal, donc le minuteur battait pendant tout ce temps. Il s'arrete
maintenant le temps du menu.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
TOPMOST n'est pas un rang, c'est une bande. La barre des taches y est
aussi, et cliquer son fond vide la remonte en tete de cette bande : une
barrette posee dessus passe DERRIERE et semble avoir disparu. Le meme
battement de 200 ms la repose.
Traine depuis un bouton : les boutons sont subclasses, et DragDetect
tranche entre clic et glissement en respectant le seuil regle par
l'utilisateur. Sur un clic, on envoie BN_CLICKED a la place du bouton -
DragDetect a avale le WM_LBUTTONUP, lui rendre la main le laisserait
enfonce pour toujours.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pref DPAPI lazyPos, ecrite sur WM_EXITSIZEMOVE - une ecriture par
deplacement, pas une par pixel parcouru.
La position n'est reprise que si MonitorFromRect la trouve encore sur un
ecran. La barrette n'a pas d'entree dans la barre des taches : une
position heritee d'un ecran debranche, ou d'une resolution changee, la
laisserait hors du bureau et definitivement irrecuperable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Clic droit dessus -> 100/75/50/25 %, persiste en pref DPAPI. Pas dans
settings_json : la barrette peut etre allumee coffre ferme, ou aucune
session n'existe pour lire les reglages du serveur.
Le survol est sonde toutes les 200 ms plutot que suivi par
TrackMouseEvent : avec quatre fenetres enfants, passer de la barrette a
un bouton envoie deja un WM_MOUSELEAVE au parent, et il faudrait suivre
la souris sur les cinq fenetres pour recoller les morceaux.
Le menu appartient a la fenetre message-only : TrackPopupMenu reclame un
SetForegroundWindow pour se refermer seul, et le faire sur la barrette
detruirait le NOACTIVATE qui la fait fonctionner.
Transparente n'est pas traversante - les clics sont toujours avales.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Un quatrieme bouton X. Il POSTE WM_LAZY_CLOSE a la fenetre message-only
plutot que de detruire sur place : liberer une fenetre depuis son propre
gestionnaire libererait, en plein envoi, l'instance qui lui sert de
window proc.
Au passage, deux libelles sortaient en mojibake - "Quick search__|" dans
le menu du tray, un tiret cadratin dans l'infobulle. Le fichier est en
UTF-8 SANS BOM, que Delphi lit en ANSI, un octet par caractere : les
litteraux restent donc en ASCII.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
SetPreferredAppMode(ForceDark) ne themise que ce que le systeme dessine
lui-meme - menus, ascenseurs, infobulles. Un bouton Win32 standard reste
donc gris clair sur fond sombre, ce que la capture montrait.
BS_OWNERDRAW + WM_DRAWITEM : le systeme garde le suivi de la souris et
l'etat enfonce, nous ne reprenons que les couleurs, recopiees des
variables de css/style.css. Le cadre passe dans WM_ERASEBKGND plutot que
par WS_BORDER, qui se sert dans les couleurs systeme.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Menu du tray -> "Lazy mode". Deux boutons Full / Password qui flottent
au-dessus de tout et font ce que font Ctrl+Shift+L / Ctrl+Shift+P.
WS_EX_NOACTIVATE est toute l'astuce : une fenetre ordinaire prendrait le
premier plan au clic, le champ que l'utilisateur vient de cliquer perdrait
son curseur, et GetForegroundWindow nous designerait nous-memes comme cible.
Le clic ne deplace donc jamais le premier plan.
Le corps du handler WM_HOTKEY est extrait dans FireAutofillRequest : hotkey
et barrette partagent la capture de la cible, le matching de titre et le
picker. Un seul chemin, deux gestes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The autofill helper attaches the caller to the TARGET thread, but the dialog runs on our own thread, so the attach was a no-op and Windows kept refusing the foreground change - the dialog rose without focus. Attaching to the current foreground thread instead is what grants the call.
SetForegroundWindow is refused when the hotkey comes from another application, so the dialog was raised but never focused; ForceForegroundWindow is now exposed by PM.Bridge for that. The explanation goes through a tray balloon because our window is disabled and behind the dialog, where no toast would be seen.
Tabbing into an input already selects its whole content, so the second clear was redundant. Where Tab lands on something that is not editable - a link, a reveal button, the document, as on the Reddit login form - Ctrl+A selected the entire page instead and the password went nowhere.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ForceForegroundWindow restores an iconic target, so closing to the tray would pop back open a window the user had deliberately minimised. Skip the handback in that case and leave the choice to Windows. A -tray launch also skips it: it never took the foreground from anyone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Hiding a window hands the foreground to nobody, so we stayed foreground while
invisible. The next hotkey then captured our own window as its target and typed
the password into it, and Ctrl+Shift+A opened a New entry with an empty title —
the guard correctly refusing to prefill with "Password Manager".
Track foreground changes with an EVENT_SYSTEM_FOREGROUND hook and keep the last
window that is neither ours nor a shell one, then reactivate it just before
hiding. Sampling on demand cannot work: by then the user's application is two
foreground changes behind (their app, the taskbar they clicked, us), and
reactivating after the hide is silently refused since only the current
foreground owner may give focus away.
Autofill paths pass AReturnForeground=False — they already placed the
foreground on the window they filled.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The window still showed, taskbar entry included. Two things.
The taskbar button belongs to the FMX proxy window (TFMAppClass), a separate
top-level window from the form — hiding the form alone never removed it.
FindFMXAppWindow moves from private to public so the activate-hide can drop
both. (Guarded on Assigned(FBridge): the handler is installed at line 222,
the bridge is created at 268.)
And rather than guess a fourth time at FMX's startup order — the same trap as
the window-state saga, which only fell to instrumentation — this logs a
millisecond timeline: FormCreate begin/end, the activate-hide, and the
deferred MinimizeToTray. Whichever gap holds the second will be obvious.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Win+D / Show desktop minimises the window without any tray transition, so the
restore took the non-tray branch. FMainForm.Show runs first there and
re-applies FMX's own WindowState (normal), discarding the
WPF_RESTORETOMAXIMIZED flag Windows had recorded — the app returned in normal
size.
Read the placement flag BEFORE showing, and re-apply SW_SHOWMAXIMIZED when it
was set. The tray path is untouched: it already replays its own captured
placement.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two gaps when Windows refuses a RegisterHotKey:
- the toast said "one or both hotkeys" — wrong since there are three autofill
combos, and it never said which one lost.
- Quick search and New entry failures were only written to the debug panel,
which is off in production: the hotkey silently did nothing with no clue why.
Delphi now exposes the per-slot registration outcome and the host accumulates
the failed combo names across all five hotkeys, emitting one notification for
the batch. A combo the user deliberately cleared (vk=0) is never listed.
JS names them: "Already used by another app: 2FA code — pick another combo in
Settings".
Registration stays best-effort per slot: one clash doesn't disarm the others.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Same workflow as the password-only hotkey: click the 2FA field, press the
combo, the current 6-digit code is typed in.
Reuses the whole autofill pipeline — no new native typing path. The code is
sent in the password slot with an empty username, so Delphi's existing
password-only branch types just the code.
- Bridge: akTotp kind, hotkey id 47, registered/torn down alongside the other
two autofill combos; an unconfigured combo (vk=0) doesn't count as a
registration failure.
- Matching filters on totp_secret BEFORE scoring, so the multi-match picker
never offers an entry that has no code.
- Configurable + synced like the others (autofillHotkeyTotp), collision-checked
against the other four slots, covered by Reset to defaults and the cheatsheet.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Auto-fetched favicons and user uploads shared one icon_b64 column, so
"Clear cache" wiped custom uploads and "Re-fetch all" overwrote them — data
loss the user can't undo (a custom icon exists precisely because DDG can't
fetch that site).
New icon_custom flag (0=fetched cache, 1=user upload). Invariant: an
automatic action never touches a custom icon; only the user changes/removes
it per entry.
- Schema: icon_custom INTEGER on vault_entries.
- POST /entries/{id}/icon takes {custom}; GET /entries echoes icon_custom;
DELETE /entries/icons/all filters COALESCE(icon_custom,0)=0.
- saveEntryIcon(id, uri, isCustom): upload→true, fetch/remove→false.
- backfillFavicons skips icon_custom even on force (Re-fetch all).
- clearAllFavicons clears only non-custom locally + refreshes the open editor
(the icon preview didn't update on clear before).
- duplicate re-stamps a copied custom icon; import-overwrite carries the flag.
Not propagated across bulk-import / sync (the bulk INSERT doesn't read it) —
a restored icon counts as fetched. Documented; custom icons are typically
self-hosted sites DDG won't re-fetch anyway.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Part A — Ctrl+Shift+A (new entry from window title) is now editable like the
other 3 global hotkeys, following the same pattern end to end: state.newEntryHotkey
(synced), a capture button in Settings > Autofill, dup cross-check, reset-to-
defaults, autofillPushHotkeys, and Bridge.setAutofillHotkeys carrying ne_mods/
ne_vk. Delphi gets SetNewEntryHotkey + stores the live combo (FNewEntryHotkeyMods/
Vk) so ExecuteAutofill's post-fill re-register uses the user's combo instead of
resetting to Ctrl+Shift+A. UMainForm parses ne_* unconditionally (works with
autofill off).
Part B — new-entry from the tray pre-filled the title with "Password Manager":
GetForegroundWindow resolved to our own hidden window. Only capture the title
when the foreground window belongs to another process; otherwise pass '' (blank).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ctrl+Shift+P (password autofill) with Ctrl+Shift physically held triggered
new-entry: the clear-field step injects a real Ctrl+A, and the held modifiers
(re-pressed by keyboard auto-repeat faster than ForceReleaseModifiers clears
them) make the system see Ctrl+Shift+A = our new-entry global hotkey. Injected
keys are only Ctrl+A + Tab, and only 'A' matches a hotkey, so unregister
NEW_ENTRY around the SendInput (try/finally re-register) — no WM_HOTKEY is
generated. Password/username go as VK_PACKET unicode, can't chord.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
SQLite defaults foreign_keys OFF per connection, and the schema relies on
ON DELETE CASCADE (entry_attachments, entries_password_history, recovery_keys).
FireDAC is said to default it On but that's undocumented here — set
ForeignKeys=On so a permanent delete never orphans encrypted blobs. Correct
in both cases (no-op if already On). Base check before fix: 0 orphans.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
RequireCSRF looked up csrf_token by user_id ORDER BY created_at DESC LIMIT 1
— the newest session, not the one whose Bearer token was presented. Two
consequences: created_at's 1-second resolution made two same-second sessions
a coin-flip (phantom 403s), and conceptually it accepted "any valid CSRF for
this user" rather than the one bound to this session. Now queried by
token_hash = SHA256(bearer) AND user_id. Authenticate already located this
row + checked expiry, so the expiry clause is dropped. Signature unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
CreateSession wrote FormatDateTime(IncHour(Now,24)) = local, but RequireCSRF
and CleanupExpired compare against SQLite datetime('now') = UTC. On negative
UTC offsets the CSRF window fell short of the 24h auth window: between the two,
GETs authed fine but every write got 403 "Invalid CSRF token". Now
IncHour(NowUTC,24) on write + LExpires < NowUTC in Authenticate, so all three
readers are UTC-consistent. Last FormatDateTime(...,Now) in the backend —
closes §2.2 for real. NowUTC + DateUtils already in uses.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Hunted the classes behind the recent quick-search fixes across the codebase:
- Esc priority: the command palette is .cmd-palette (not .modal), so the
slideover's capture-phase Esc handler didn't see it and popped the
discard-confirm UNDER the open palette. Capture handler now also yields
to the palette.
- Esc fall-through: the fallback Esc branch closed palette + slideover +
entry modal + generator ALL on one keystroke. Now closes exactly one
surface per keystroke, topmost first.
- Modifier chord: the real-VK Tab between username and password becomes
Shift+Tab if the user holds Shift mid-fill -> focus moves backward and
the password lands in the username field. ForceReleaseModifiers before
SendVKey(VK_TAB).
- Stale placement: RestoreFromTray replayed the MinimizeToTray snapshot for
a merely taskbar-minimised window, teleporting it to the last tray-hide
position. Snapshot now replays only on a genuine tray return (captured
before Show flips Visible); iconic windows use Windows' own placement.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Copy a password (30s auto-clear armed), then copy something else from
another app before the timer fires: the tick emptied the clipboard anyway,
destroying the user's newer content. Guard with the Win32 clipboard sequence
number: SetText snapshots GetClipboardSequenceNumber, ClearIfOurs only
empties when it hasn't moved. Applied to the auto-clear timer AND the
clear-on-minimize path (same bug class); the explicit JS clipboard/clear
command stays unconditional (user-initiated).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Residual new-entry trigger (1 in 6): password chars go out as
KEYEVENTF_UNICODE (VK_PACKET, can't match a hotkey) — the real chord risk is
the Ctrl+A clear-field, which sends a real VK_A. If the user re-presses
Ctrl+Shift mid-sequence, that VK_A becomes physical Ctrl+Shift+A = our own
new-entry hotkey. ForceReleaseModifiers now runs inside
SendSelectAllAndDelete, at the risky instant, not just once up front.
Balloon: was gated by "Show tray notifications" (OFF for this user) — now
gated by its own synced setting "Tray alert when autofill is blocked"
(autofillFailBalloon, Settings > Autofill, default ON), carried as notify=0
on cmd://autofill/execute. ShowBalloon no longer gates internally; each
caller applies its own setting.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Root cause of "Ctrl+Shift+P opened the new-entry modal": if the user still
holds Ctrl+Shift when WaitForModifierRelease times out (1s), every password
letter is typed as a Ctrl+Shift+<letter> chord — garbage in the field AND it
fires our own global hotkeys (a password containing 'a' triggers Ctrl+Shift+A
= new entry). ForceReleaseModifiers now injects KEYUP for any still-held
modifier before typing.
Also: when the fill is blocked (elevated target) while the window is hidden
in the tray, the in-app toast is invisible — show a tray balloon instead.
ShowFirstTimeBalloon generalized into ShowBalloon(title, text, warning),
gated by the existing "Show tray notifications" setting.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Ctrl+A + Del sent before each field misbehaves on targets where Ctrl+A
isn't select-all (terminals, some remote desktops). New synced setting
(autofillClearField, Settings > Autofill) gates it: JS appends clear=0 to
cmd://autofill/execute when off; ExecuteAutofill wraps the three
SendSelectAllAndDelete calls behind AClearFirst. Absent param = ON, so
existing behavior is unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bug 1: a maximized window trayed via the quick-search fill flow came back
"normal" on the next restore. ExecuteAutofill minimizes the window BEFORE
MinimizeToTray snapshots the placement, so the snapshot said SHOWMINIMIZED
and the never-restore-minimized guard forced SHOWNORMAL. Now honours
WPF_RESTORETOMAXIMIZED (Windows keeps the pre-minimize state in flags).
Bug 2: filling into an elevated app (admin Notepad) showed "password sent"
while UIPI silently discarded the keystrokes (SendInput even reports
success). ExecuteAutofill is now a function: it checks the target process
elevation up front (can't-open counts as elevated) and returns False without
typing. UMainForm feeds the result to JS via Bridge.onAutofillResult; the
quick-search success toast is deferred until Delphi confirms, and a failure
shows "Autofill blocked - the target window runs as administrator".
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The minimize-then-SW_SHOWNOACTIVATE approach flickered and sometimes lost the
restore race (window stayed minimized). Root fix: when the window was open
before the hotkey (ARestoreAfter), skip the minimize entirely — being the
foreground process is exactly what allows handing focus to the target, so the
window simply stays in place beside it. Tray-origin flow keeps the old
minimize (the window is a temporary overlay, trayed after the fill anyway).
Safety: if the target refuses the foreground (elevated / UIPI) and our window
is still foreground, bail before typing — otherwise the password would be
typed into the vault's own visible UI.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two bugs when Ctrl+Shift+Q fires while the app window is open beside the
target app:
- RestoreFromTray re-applied FSavedPlacement (captured at the LAST
MinimizeToTray) to an already-visible window -> it jumped to a stale
position. Now: visible and not iconic -> just SetForegroundWindow.
- ExecuteAutofill minimizes our window when it is foreground (the user just
clicked the entry) and never brought it back when hide_after was false.
New ARestoreAfter param (= not HideAfter): restore with SW_SHOWNOACTIVATE
after the fill, so the window returns to its position without stealing
focus from the freshly-filled target.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Migration verified (0 unmigrated cleartext templates across both accounts)
before contracting:
- GET/POST/PUT/bulk no longer read or write the cleartext template column
(only template_enc/iv); the PUT partial-update gate stays keyed on the
'template' JSON key presence.
- Removed AddColumnIfMissing for template AND tags/title — those two had
silently re-added the dropped columns as empty ghosts at every start.
- CREATE TABLE: removed site/username NOT NULL cleartext columns — a FRESH
database rejected the very first INSERT (which no longer ships them).
User can now DROP COLUMN template (and re-drop the ghost tags/title).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Authenticate/RequireCSRF write a 401 then raise ESessionRejected; when it
reached the dispatcher catch-all, the generic `on E: Exception` overwrote it
with a 500. Added `on ESessionRejected do Exit` before the generic clause in
both dispatchers (GET + Other) — one place, covers every handler whether or
not it wraps Authenticate. Root cause, not per-handler patch.
ponytail: runtime check only (expired token → 401) — no Delphi unit harness.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Extends the username-at-rest scheme to site, title and tags — the last
searchable metadata still stored cleartext. Same design: dedicated
<f>_enc/<f>_iv columns (AES-GCM under the vault key), decrypted at load into
e.<f>, so client-side search/sort/render/favicon/autofill-match are unchanged.
Full-strength random-IV AES-GCM (no searchable encryption) because search is
client-side.
Generalized the helpers over ENCRYPTED_META_FIELDS = [username, site, title,
tags]:
- withEncryptedUsername → withEncryptedMeta (encrypts all four, blanks
cleartext) — wraps every POST/PUT body.
- decryptEntryUsernames → decryptEntryMeta (decrypts all four at load).
- migrateUsernamesAtRest → migrateMetadataAtRest (sweeps any field still
cleartext, live + trash).
- doChangeMasterPassword re-encrypts all four under the new key.
Server (Entries + Auth + Database):
- Columns site_enc/iv, title_enc/iv, tags_enc/iv; GET emits them (new
AddNullableField helper); POST/PUT/bulk read+persist (BindNullable helper);
rotation UPDATE re-encrypts them.
- Removed the server "Site required" validation (site='' when encrypted — the
client enforces it) at POST/PUT/bulk.
- ?q= server search neutralized (site+username ciphertext → LIKE useless; the
frontend never sends ?search=).
Tests: merge assertions updated to decrypt site (encrypted on import). 65/65.
username was runtime-validated earlier; site/title/tags NOT yet compiled/
runtime-tested (Delphi) — large multi-handler change. Rebuild BuildAssets +
PMServer, then create/edit/dup/move/tag/import/rotate and verify the DB shows
no cleartext site/title/tags (and the app still renders/searches).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
r-zakarya