Calendarr-IOS wird mit vollstaendiger Historie als Unterverzeichnis
ios/ eingebunden. Damit liegen Server, Web, Android und iOS in einem
gemeinsamen Repo.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Calendarr-Android wird mit vollstaendiger Historie als Unterverzeichnis
android/ eingebunden, damit Server, Web und die mobilen Apps in einem
Repo liegen.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Device family becomes 1,2 for the app and the widget extension, and the iPad
gets all four interface orientations. That is not a preference: App Store
Connect rejects an iPad-capable bundle that declares fewer (error 90474),
because iPad multitasking can present the app at any size. The iPhone key stays
portrait-only. Mac Catalyst still resolves to family 2 alone, which is what
84f5c00 set up and what this had to avoid breaking.
Building for Catalyst also turned up a regression from the watch work: the
Embed Watch Content phase carried watchOS content into a macOS build, which
fails outright. It went unnoticed because no Catalyst build was run after the
watch targets were added. Both the copy and its target dependency are now
filtered to iOS.
Verified on all three destinations: iOS carries Watch/CalendarrWatch.app with
its complications, Catalyst carries no Watch folder, and both report the right
device family.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Requiring both ends to sit exactly on midnight was too strict. A holiday stored
as the 16th at 00:00 through the 21st at 01:00 slipped through and rendered as
a one-hour slot. Starting on a day boundary and running into a later day is
enough to call it all-day.
The agenda row now names the weekday on both sides when an event crosses a day.
Showing only "00:00 – 01:00" for something that runs from Wednesday to Monday
is true and useless.
The timezone handling above this is deliberately untouched: timed events show
the right time, so the UTC assumption for timestamps without an offset is not
what produced the odd hour.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The previous rule only caught date-only strings. A multi-day event sent with
full timestamps and no allDay flag still arrived as a timed event, so a holiday
running from the 16th to the 21st rendered as a one-hour slot at midnight.
Now an event that starts and ends on a day boundary and covers at least one
whole day is all-day regardless. That is a statement about what the data means,
not about how any one client formats it, which is why it lives in the parser
and not in a view.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three things, one of which is not a watch problem at all.
parseDate already treated a date-only string as all-day, but isAllDay came from
the JSON flag alone. An event sent as "2026-09-12" with no flag therefore
landed at midnight with isAllDay false, and every surface that formats a time
put "00:00" in front of it — the widgets as much as the watch. Now derived from
both.
The date complication showed the day number and nothing else, because
.accessoryCircularCapacity renders only currentValueLabel and the month sat in
the label closure, where it is never drawn. It now stacks weekday over day
number, which is what makes it read as a date rather than a number.
The ring only appears while an appointment is running, and then shows how far
through it we are. An empty track around an idle day said nothing, so there is
no Gauge at all in that case. All-day events do not count as running: one would
hold the ring at an arbitrary fraction for twenty-four hours.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The position asked for is the bezel slot: a circular complication plus a label
that arcs across the top, where the tick marks otherwise are. ClockKit called
that family graphicBezel; in WidgetKit it is an accessoryCircular widget with
.widgetLabel attached. Nothing was attached, so that arc stayed empty while
Apple's Calendar fills it — which is the difference that was visible on the
watch and invisible in every check made here.
All three circular kinds now carry it. Faces without a bezel slot ignore the
label, so there is no cost to always providing it.
cornerLabel becomes eventLabel: the same text serves the corner edge and the
bezel arc, and the old name claimed otherwise.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Requested shape: the event title as the text, the time small alongside. This
inverts how accessoryCorner is built — the curved bezel is where a corner has
room for long text and the inner area holds three or four characters — so the
title shrinks and truncates. Kept alongside WatchCornerTime, which carries the
same information with the roles the other way round, so both can be compared on
the face rather than argued about.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Date and "events today" filled their slot with AccessoryWidgetBackground and
centred text, which reads as a small element with space around it. Both are now
Gauges in .accessoryCircularCapacity — the WidgetKit equivalent of ClockKit's
ClosedGauge templates, which no longer exist and cannot be reached from a
WidgetKit complication at all.
The rings carry meaning rather than decorating: the date's drains through the
day, and the count's drains as today's appointments are worked through. Both
drain, matching the countdown rather than inventing a second direction. An
empty day leaves the ring empty instead of full, so it does not read as an
untouched one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Text(style: .timer) renders 1:23:45 once more than an hour out. An
accessoryCorner has room for three or four characters beside its curved label,
and a circular gauge for about the same, so the live timer overlapped what it
sat next to.
Replaced with a compact value — "23m", "2h", "3d" — recomputed per timeline
entry, which is why the timeline now carries a mark every minute for the first
hour. Both corner kinds and the gauge also get lineLimit(1) and a minimum scale
factor, so an unexpectedly wide value shrinks instead of colliding.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A companion watch app has to carry the same CFBundleShortVersionString and
CFBundleVersion as the app it ships inside. The template left the watch at
1.0 (1) against the app's 3.5 (7), and watchOS answers that by refusing the
install without reporting anything: the Install button simply comes back.
Same class of drift as c9bd42a, which unified the app and widget versions.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The target template left AppIcon.appiconset with an entry carrying no filename
— an empty slot. watchOS refuses to install an app without a valid icon, which
is why the Watch app's Install button kept coming back after ten seconds with
no error anywhere. Reuses the iOS 1024 icon, which is already alpha-free.
Both watch targets also called themselves "CalendarrWatch", which is what would
have appeared under the icon and in the complication gallery.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The watch receives the phone's payload over WatchConnectivity and writes it
into its own App Group container through the same SnapshotStore, so there is
one wire format and one reader. The agenda, the detail screen and the filter
all read that; the complications read what the app last wrote, since an
extension cannot run a WCSession of its own.
Complication timelines land on event boundaries rather than a schedule, so the
next appointment flips in the exact minute it changes without spending refresh
budget.
CalendarrWatchShared holds what both watch targets need, attached to each the
way Shared/ is attached to the app and the widgets.
recommendations() is implemented because watchOS requires it — iOS has a
default, the watch does not, since the complication gallery has to be
populated. Found by building, not by reading.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
-sdk overrides SDKROOT for the whole dependency graph, so the watch app was
built against the iOS SDK. It still succeeded, and produced a watch app with no
WKApplication key — the kind of failure that only shows up on the device.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The companion dropdown in Xcode's target template came up empty, which left
both bundle identifiers starting with a bare dot and
WKCompanionAppBundleIdentifier as an empty string. Both are set explicitly
here, along with the App Group entitlements and watchOS 26 as the floor.
The iOS app had no Embed Watch Content phase, so the watch app would have built
and never shipped. Added, with the matching target dependency: the product now
lands at Calendarr iOS.app/Watch/CalendarrWatch.app with the complications
inside it.
Folders follow what they contain: CalendarrWatch is the app,
CalendarrWatchComplications the extension. Xcode had them the other way round.
Package.resolved is now tracked. It pins AppAuth, whose version drift is what
produced the two compile errors fixed in 75a6a5b.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Xcode exposes neither UIBackgroundModes nor BGTaskSchedulerPermittedIdentifiers
as an INFOPLIST_KEY_ build setting, and a partial plist inside a synchronized
group collides with the processed one. Both found by building, not by reading.
Also records that synchronized groups pick up new source files on their own.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
App Groups are per-device, so the watch cannot read the file the phone writes
and has to be handed the data. Three delivery paths: application context for
the latest state, a complication push only when the next appointment actually
changed (the budget is small), and a reply to the watch asking directly.
Sign-out pushes an empty signed-out snapshot. Clearing our own container is not
enough — the watch holds its own copy and would keep rendering the previous
user's calendar.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
publishWidgetSnapshot() only ever ran from a user action, so an event created
on the web or on Android did not reach the widgets until the app was opened
again. A BGAppRefreshTask now republishes without the foreground.
The two keys this needs are arrays, which Xcode does not expose as
INFOPLIST_KEY_ build settings — verified by building and inspecting the
result. Hence a partial Calendarr-Info.plist that generation merges into. It
sits at the project root rather than inside the target folder: that folder is
a synchronized group, which would also copy the file as a resource and collide
with the processed one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
UIAccessibilityIsGuidedAccessEnabled() is gone in iOS 26; use the property.
AppAuth declares lastAuthorizationResponse and its request as nonnull, so the
optional chain was rejected. Behaviour is unchanged: nonce is optional either
way, which is the tolerance the comment there already describes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Seventeen tasks. Every piece of testable logic lands in CalendarrCore first,
under TDD, because that is the only target this machine can actually run tests
for; the watch targets hold view and WCSession glue that is build-verified only.
The iOS background refresh is task 7 and commits on its own: it fixes the
home-screen widgets' staleness whether or not the watch app follows.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The watch is a read-only consumer of the phone's snapshot: WatchConnectivity
carries a trimmed CalendarrSnapshot to the watch, which writes it into its own
App Group container through the same SnapshotStore. One format, one reader.
Also specifies the iOS background refresh the watch depends on, which fixes a
pre-existing staleness in the home-screen widgets as a side effect.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Die Seite trug zweimal autocomplete="one-time-code": im Login und im Feld
zum Einrichten von 2FA in den Einstellungen. Fuer einen Passwortmanager ist
damit nicht eindeutig, zu welchem Feld sein Inline-Symbol gehoert.
Das Einrichtungsfeld bekommt data-bwignore (plus die Entsprechungen fuer
1Password und LastPass) — dort tippt man den Code aus der Authenticator-App
beim Anlegen eines neuen Geheimnisses ab, ein Ausfuellen ergibt da ohnehin
keinen Sinn. Die Passwortfelder bleiben unangetastet, die sollen weiter
befuellbar sein.
Ob das die fehlplatzierte Bitwarden-Anzeige behebt, ist offen: hier laesst
sich keine Erweiterung installieren. Der vorherige Versuch ueber das
Fokus-Timing hat es nicht geloest.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Benutzername und Passwort liegen ab Seitenaufbau im Layout, das 2FA-Feld
taucht erst aus display:none auf — und der Fokus wurde im selben Tick
gesetzt. Passwortmanager oeffnen ihr Inline-Overlay beim Fokus und messen die
Feldposition in genau dem Moment; Bitwarden platzierte sein Symbol dadurch
weit neben dem Feld.
Der Fokus laeuft jetzt ueber setTimeout(0), sodass das Layout vorher steht,
plus ein resize-Event als Anstoss zum Neupositionieren. Bewusst kein
requestAnimationFrame: das feuert in einem nicht sichtbaren Tab nicht und
haette den Fokus ganz verschluckt.
Nicht gegen Bitwarden verifizierbar — hier ist keine Erweiterung installiert.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Die Endpunkte gab es schon, die Oberflaeche fehlte — es war schlicht nicht
sichtbar, ob und womit ein Konto verknuepft ist.
Unter Einstellungen > Profil, zwischen App-Passwoertern und Privatsphaere,
steht jetzt "Single Sign-On" mit den verknuepften Konten (Anbieter, E-Mail,
letzte Nutzung) und je einem Knopf zum Aufheben. Ohne Verknuepfung erscheint
ein Hinweis, wie man eine anlegt; ist auf dem Server ueberhaupt kein Provider
konfiguriert, bleibt der ganze Abschnitt unsichtbar.
Bei einem per SSO angelegten Konto ist die letzte Verknuepfung nicht
loesbar — es hat kein brauchbares Passwort und waere danach dauerhaft
unerreichbar. Der Knopf ist in dem Fall deaktiviert und erklaert sich per
Tooltip; der Server lehnt es zusaetzlich ab.
Im Browser durchgespielt: verknuepfen, Eintrag erscheint mit Anbieter und
Datum, aufheben, Bestaetigung, leerer Zustand.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Das Sitzungs-Token trug den Benutzernamen als Subjekt, und get_current_user
suchte danach. Eine Umbenennung machte damit jedes Token auf jedem anderen
Geraet stillschweigend ungueltig — der Nutzer flog ohne Hinweis raus.
Tokens tragen jetzt zusaetzlich uid, die unveraenderliche Zeilen-ID, und die
Aufloesung geht darueber. sub bleibt als Anzeigename erhalten. Tokens ohne
uid werden weiter ueber den Namen aufgeloest, damit vor der Aenderung
ausgestellte Sitzungen gueltig bleiben — mit "Angemeldet bleiben" laufen die
180 Tage.
Fuer SSO war die Zuordnung schon vorher richtig: OIDCIdentity haengt an
(provider_key, sub) des Providers, und bei bekanntem sub wird der lokale
Benutzername nicht angefasst. Ein umbenannter Nutzer in Authentik landet also
unveraendert auf demselben Konto. Vier Tests halten beide Richtungen fest:
Umbenennung beim Provider, lokale Umbenennung mit weiterlaufender Sitzung,
lokale Umbenennung mit erhaltener SSO-Verknuepfung, und die Kompatibilitaet
alter Tokens.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bisher war eine SSO-Anmeldung ohne passendes Konto eine Sackgasse: rote
Fehlermeldung, kein Weg weiter. Die Verknuepfung ging nur ueber
LINK_BY_EMAIL, das der E-Mail-Verwaltung des Providers vertrauen muss.
Jetzt endet der Callback bei einer besitzerlosen Identitaet nicht im Fehler,
sondern parkt sie in einem kurzlebigen signierten HttpOnly-Cookie und leitet
auf /?sso_link=1. Der Login-Screen bittet daraufhin einmalig um das
Calendarr-Passwort; nach erfolgreichem Login loest der Client
POST /api/auth/oidc/link-pending ein und die Identitaet haengt an genau dem
Konto, das sich gerade ausgewiesen hat. Beide Haelften sind damit bewiesen —
das ID-Token haben wir selbst geprueft, das Konto durch das Passwort — und
die Vertrauensannahme ueber die E-Mail entfaellt.
Das greift fuer alle drei Faelle, in denen eine gueltige Identitaet keinen
Besitzer hat: kein Konto vorhanden, Selbstregistrierung aus, oder E-Mail
bereits vergeben.
Ausserdem: unbekannte Fehler-Slugs werden jetzt mit angezeigt statt hinter
"Anmeldung fehlgeschlagen" zu verschwinden — genau die Information, die zur
Diagnose fehlte. Und oidc_token_exchange_failed hat eine eigene Meldung, die
auf Client-Typ und Redirect-URI hinweist, weil das der haeufigste
Konfigurationsfehler ist.
Sechs Tests dazu; Ablauf zusaetzlich im Browser gegen einen echten
OIDC-Stub-Provider durchgespielt: Hinweis statt Fehler, Verknuepfung nach
Passwort-Login, danach Anmeldung allein per SSO.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Zusaetzlicher Anmeldeweg neben Benutzername/Passwort. Der Login-Screen holt
die konfigurierten Provider von GET /api/auth/oidc/providers; ohne Provider
oder gegen einen aelteren Server bleibt der Screen unveraendert.
Die App laeuft als Public Client: Authorization Code Flow mit PKCE in einer
ASWebAuthenticationSession (nie ein WKWebView), ohne Client Secret. Das
ID-Token wird an POST /api/auth/oidc/exchange gegen ein Calendarr-Token
getauscht; die Antwort hat dieselbe Form wie der Passwort-Login, sodass der
bestehende Session-Pfad unveraendert weiterlaeuft.
- OIDCAuthenticator kapselt Discovery, PKCE-Request, Code-Tausch und die
stille Erneuerung. Der AuthState mit dem Refresh-Token liegt in der
Keychain (kSecAttrAccessibleAfterFirstUnlock) und wird beim Logout aus
beiden Zugriffsgruppen entfernt.
- OIDCSessionRefresher erneuert die Sitzung beim Start, wenn das
Calendarr-Token binnen 24 h ablaeuft — ohne Browser, ueber den
Refresh-Token des Providers.
- Die angefragten Scopes kommen vom Server (mobile_scopes); ein Scope, den
der Provider nicht kennt, wuerde die ganze Anfrage scheitern lassen.
- Bei aktiviertem gefuehrten Zugriff weicht AppAuth auf SFSafariViewController
aus, dessen Redirect nur ueber ein registriertes URL-Schema zurueckkaeme.
Da das App-Target seine Info.plist generiert und CFBundleURLTypes dort
nicht setzbar ist, wird der Fall vorab sauber abgelehnt statt endlos auf
einen Callback zu warten, der nicht kommen kann.
- Abbrueche werden mit Fehlerdomaene geprueft (user- und programCanceled) und
still behandelt; alle uebrigen Fehler tragen Domain, Code und OAuth-Fehler
ins DEBUG-Log, statt spurlos zu verschwinden.
project.pbxproj: AppAuth-iOS 1.7.5 als SPM-Abhaengigkeit, nur am App-Target.
OTHER_LDFLAGS bekommt -ObjC, weil OIDAuthState+IOS.m ausschliesslich eine
Kategorie enthaelt und ohne Symbol sonst wegoptimiert werden koennte.
Achtung: mangels macOS/Xcode ist dieser Stand nicht kompiliert. Beim ersten
Oeffnen loest Xcode das Paket auf und legt Package.resolved an; die bitte
mit committen.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Drei Fehler, die den Flow gebrochen haetten:
- Ein Abbruch im Browser liefert bei AppAuth ein regulaeres Ergebnis mit
USER_CANCELED_AUTH_FLOW, nicht null. Der bisherige Null-Check lief ins
Leere und "Zurueck" endete in einer roten Fehlermeldung.
- launch(intent) lag ausserhalb von runCatching. Verlaesst der Nutzer den
Screen waehrend der Discovery, ist der Result-Launcher abgemeldet und die
IllegalStateException haette die App beendet.
- Nonce und Provider lagen nur in ViewModel-Feldern. Der Custom Tab laeuft
in eigener Task, dieser Prozess wird dabei routinemaessig getoetet — die
Anmeldung waere nach erfolgreichem Login beim IdP gescheitert. Der Nonce
kommt jetzt aus AuthorizationResponse.request, Provider und Client-ID
werden vor dem Start persistiert.
Weiter: AuthState.update verlangt genau eines von (Response, Exception) und
haette bei null/null geworfen; onCleared() verwarf den @Singleton-Service
fuer die zweite AuthViewModel-Instanz mit.
Neu ist der geforderte Hintergrund-Refresh: laeuft das Calendarr-Token
binnen 24 h ab, wird beim Start ueber den Provider-Refresh-Token ein frisches
ID-Token geholt und getauscht — ohne Browser. Best effort; schlaegt es fehl,
bleibt das vorhandene Token in Kraft.
Ausserdem: Scopes kommen ausschliesslich vom Server (mobile_scopes), Spinner
pro Provider statt global, fehlender Browser als eigener Fehlerfall, Issuer
wird validiert, keine namenlosen Identitaeten mehr.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Authlib wertet "Aufrufer hat eine Nonce uebergeben, das Token traegt keine"
als Mismatch. Das ist fuer den erneuerten Token der mobilen Clients falsch:
ein per Refresh-Token geholtes ID-Token laesst die nonce-Claim regulaer weg,
womit der Hintergrund-Refresh immer an einem 401 gescheitert waere.
Die Nonce wird daher nicht mehr an Authlib durchgereicht, sondern hier
geprueft: vorhandene Nonce muss passen, und der Browser-Flow verlangt sie
zusaetzlich zwingend (require_nonce) — dort erzeugen wir sie selbst, also
muss der Provider sie zurueckspiegeln. Signatur und uebrige Claims bleiben
unveraendert bei Authlib.
Zwei Tests dazu: erneuertes Token ohne Nonce wird angenommen, Browser-Flow
ohne Nonce abgelehnt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Zusaetzlicher Anmeldeweg neben Benutzername/Passwort. Der Login-Screen holt
die konfigurierten Provider von GET /api/auth/oidc/providers; ohne Provider
oder gegen einen aelteren Server bleibt der Screen unveraendert.
Die App laeuft als Public Client: Authorization Code Flow mit PKCE in einem
Chrome Custom Tab (nie ein WebView), ohne Client Secret. Das ID-Token wird
anschliessend an POST /api/auth/oidc/exchange gegen ein Calendarr-Token
getauscht; die Antwort hat dieselbe Form wie der Passwort-Login, sodass der
bestehende Session-Pfad unveraendert weiterlaeuft.
- OidcAuthenticator kapselt Discovery, PKCE-Request und Code-Tausch.
- Die angefragten Scopes kommen vom Server (mobile_scopes) — ein Scope, den
der Provider nicht kennt, wuerde die ganze Anfrage scheitern lassen.
- AuthState mit dem Refresh-Token liegt in den EncryptedSharedPreferences,
nie im Klartext, und wird beim Logout mitgeloescht.
- Provider und Client-ID werden vor dem Browser-Start persistiert, weil der
Custom Tab in einer eigenen Task laeuft und dieser Prozess dabei getoetet
werden kann; der Nonce kommt aus AuthorizationResponse.request, das AppAuth
ueber den Redirect zurueckreicht.
- Abbruch im Browser meldet AppAuth als regulaeres Ergebnis mit
USER_CANCELED_AUTH_FLOW und wird still behandelt, nicht als Fehler.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Die Apps duerfen keine Scopes erfinden: ein Scope, den der Provider nicht
kennt, laesst ihn die gesamte Authorization-Anfrage ablehnen. Daher legt
jetzt der Server fest, was die nativen Clients anfragen.
OIDC_<KEY>_MOBILE_SCOPES ist frei setzbar; ohne Angabe sind es die
Web-Scopes plus offline_access, damit Authentik ab 2024.2 ueberhaupt ein
Refresh-Token ausstellt. Der Browser-Flow bleibt unveraendert, der braucht
kein offline_access. Ausgeliefert wird der Wert als mobile_scopes in
/api/auth/oidc/providers, aber nur wenn eine MOBILE_CLIENT_ID konfiguriert
ist.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Die Platzhalter-URI calendarr://oauth/callback durch die tatsaechlichen
Reverse-DNS-Schemata der beiden Apps ersetzt und vermerkt, dass sie als
Public Clients ohne Secret laufen und ohne MOBILE_CLIENT_ID keinen
SSO-Button zeigen.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Zusaetzlicher Anmeldeweg per OIDC neben dem bestehenden Passwort-Login. Ohne
gesetztes OIDC_PROVIDERS aendert sich nichts: /api/auth/oidc/providers meldet
enabled:false und der Login-Screen bleibt unveraendert.
Backend
- oidc_config.py: Provider-Registry rein aus ENV. Ein zweiter Provider ist
reine Konfiguration (OIDC_PROVIDERS + OIDC_<KEY>_* Block), keine Codeaenderung.
- oidc_client.py: Discovery, JWKS, Token-Tausch, ID-Token-Pruefung ueber
authlib.jose mit fest gepinnter Signatur-Allowlist (nur asymmetrisch, damit
alg:none und Algorithmus-Confusion ausgeschlossen sind).
- oidc_identity.py: Zuordnung ueber (provider_key, sub). Ein E-Mail-Treffer auf
ein bestehendes Konto wird standardmaessig abgelehnt statt verknuepft — sonst
erbt jeder, der im IdP eine beliebige E-Mail setzen kann, das passende Konto.
Verknuepfen ist eine explizite, authentifizierte Aktion. Auto-Provisioning
nur per ENV-Flag, standardmaessig aus, optional mit Domain-Allowlist.
- routers/oidc_router.py: /providers, /start, /callback, /complete, /exchange,
/link, /identities. PKCE S256; Flow-State als kurzlebiges signiertes JWT in
einem HttpOnly-Cookie mit jti==state statt eines prozesslokalen Dicts, damit
der Flow Neustarts und mehrere Worker uebersteht. Das Calendarr-JWT wird ueber
ein Einmal-Cookie uebergeben, nie ueber URL oder Fragment.
- /exchange als Kontrakt fuer die AppAuth-Clients: die App fuehrt den Flow als
Public Client selbst aus und tauscht das ID-Token; kein Secret im Request.
- models.py: neue Tabelle oidc_identities, users.auth_source; Migration in
main.py im dortigen try/except-Stil.
Nebenbefund, mitbehoben: auth.verify_password liess ValueError aus
bcrypt.checkpw durch. Ein nicht wohlgeformter password_hash wurde damit zu
HTTP 500 auf /api/auth/login und bei jedem CalDAV-Basic-Auth. SSO-Nutzer
bekommen einen echten bcrypt-Hash ueber ein zufaelliges, nirgends gespeichertes
Geheimnis; fuer CalDAV nutzen sie App-Passwoerter wie 2FA-Nutzer.
Web
- SSO-Block auf dem Login-Screen, aus /providers befuellt; Buttons sind Links,
weil /start eine Top-Level-Navigation braucht.
- consumeSsoHandoff() laeuft vor den Setup-/Token-Pruefungen in boot() und
raeumt die Query per replaceState ab, bevor es awaited.
- 401-Auto-Logout um die SSO-Pfade erweitert, sonst dreht ein fehlgeschlagener
Handoff eine Reload-Schleife.
- i18n-Keys in de + en (Fallback-Kette deckt die uebrigen acht Sprachen).
- Service Worker laesst /api/auth/oidc/ unangetastet durch.
37 neue Tests (Gesamtsuite 60 gruen): Signaturpruefung mit echten RSA-Tokens,
alg:none, falscher iss/aud/nonce, unbekannte client_id, Provisioning-Policies,
Username-Kollision, Einmaligkeit des Handoff-Cookies, unerreichbarer Provider.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Statt eines einzelnen Punkts in --primary zeigt der Mini-Kalender jetzt fuer
jeden Kalender, der an dem Tag mindestens einen Termin hat, einen Punkt in
dessen Farbe (max. 4, stabil nach erstem Auftreten sortiert). Die Farbquelle
ist dieselbe wie in Monats-/Wochenansicht: ev.color || ev.calendarColor.
Ausserdem laeuft der Mini-Kalender jetzt ueber filterEvents(), damit
ausgeblendete Gruppenmitglieder keinen Punkt mehr erzeugen. Auf dem
"heute"-Kreis bekommen die Punkte einen duennen weissen Ring, damit sie auf
der farbigen Flaeche lesbar bleiben.
Cherry-pick von 7a2085b (master) auf beta; die selected-Day-Regel aus beta
(color-mix) bleibt erhalten.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The snapshot format was an internal detail shared between the app and its own
widget extension. A second Mac app needs the same data, and the way to give it
that is a documented contract rather than a format it reverse-engineers and then
drifts from. Both targets now link CalendarrCore and go through SnapshotStore.
Shared/WidgetData.swift becomes a thin facade. The typealiases and the flat
colour accessors exist so the ~70 existing call sites across the app and the
widget views compile unchanged; they are a migration convenience, not a design.
The snapshot now carries what a reader outside this app actually needs:
coverageStart / coverageEnd, because the published window is ~7 days back and
~42 ahead. Outside it the snapshot holds no information, which is not the same
as holding no events — and only the writer knows where that edge is. Without it
a consumer renders a convincingly empty March and is simply wrong.
isLoggedIn plus a session record, so a reader can say "sign in to Calendarr"
rather than "open Calendarr once". The events are deleted on sign-out, so the
absence of a cache alone cannot tell those two apart.
writerVersion, purely so a mismatch between the two apps is diagnosable.
The coverage constants move to SnapshotCoverage, so the writer and the code that
reconstructs the window for older files can no longer disagree about it.
CalendarStore needs an explicit `import CalendarrCore` because the app target
builds with SWIFT_UPCOMING_FEATURE_MEMBER_IMPORT_VISIBILITY, which requires the
defining module to be imported directly rather than picked up transitively.
Verified: builds for iOS and Mac Catalyst; CalendarrKit's 11 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Enabling Mac Catalyst pulled in the iPad device family, which made App Store
Connect reject the iOS build with error 90474: an iPad-capable bundle must
declare all four interface orientations to support iPad multitasking, and this
app declared two.
Catalyst genuinely needs the iPad idiom, but only for its own build. A per-SDK
condition gives macOS family 2 while iOS keeps family 1 — verified to resolve
correctly for both destinations, and Catalyst still builds with family 2 alone.
The shipped iOS bundle is back to UIDeviceFamily [1], exactly as before the
port started.
The iPad orientation key is removed rather than extended. With an iPhone-only
iOS build it is ignored, and macOS ignores orientation keys entirely, so it was
dead configuration that only served to trip the validator.
Supporting iPad properly is a deliberate piece of work — the layout is
portrait-only with no size-class handling — and is better done on purpose than
inherited as a side effect of the Mac port.
Verified: builds for both Mac Catalyst and iOS Simulator; built bundle reports
UIDeviceFamily [1] and portrait-only with no iPad orientation key.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The App Group identifier registered in the portal does not change, but the
string the runtime expects does: macOS and Mac Catalyst require the Team ID
prefix, iOS forbids it. CalendarrAppGroup now resolves the right one per
platform. Getting this wrong is the worst failure mode in the whole port —
containerURL() returns nil, every snapshot read and write quietly no-ops, and
the widgets show placeholder content forever with no error anywhere. A DEBUG
assertion now makes that loud during development.
Because the two platforms need different values, they need different
entitlements files — listing both strings in one file breaks iOS provisioning
on the unregistered prefixed value. Selected via CODE_SIGN_ENTITLEMENTS[sdk=macosx*],
verified to resolve correctly for both destinations.
The Catalyst entitlements are written App Store grade from the start, so one
configuration serves both the Mac App Store and a notarized DMG: sandbox,
network client, Contacts (birthday import), user-selected files (.ics import
and export), the prefixed App Group, and keychain sharing. Deliberately absent:
files.downloads, network.server, device.*, temporary-exception.* — nothing needs
them and each is App Review friction. ENABLE_HARDENED_RUNTIME is required for
notarization and ignored by the App Store, so it is safe to set unconditionally.
Verified: builds for both Mac Catalyst and iOS Simulator.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Supported Destinations now includes Mac (Mac Catalyst) on the app and the
widget extension. Catalyst requires the iPad device family, so both targets
move to TARGETED_DEVICE_FAMILY "1,2".
The first Catalyst build failed on CalendarSelectionIntent: WidgetConfigurationIntent
requires every @Parameter to be optional, and the macOS SDK enforces that where
the iOS one lets a bare array through. selectedCalendars is now optional.
Behaviour is unchanged — the timeline provider already treated an empty
selection as "show all calendars", so nil folds into the same path.
Verified: builds for both Mac Catalyst and iOS Simulator.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The app shipped 3.5 (6) while the widget shipped 1.0.1 (2). Apple requires an
embedded appex to match its host's CFBundleShortVersionString and CFBundleVersion,
so this was a latent App Store validation failure that only surfaces at upload —
and it would surface on the Mac submission too. Both keys now live at project
level, where they cannot drift apart again.
The widget also targeted iOS 17 while the container app requires 26.0, so the
extension could never have been installed on 17 anyway. Raising it to 26.0
removes the availability skew that Mac Catalyst would otherwise inherit as
macOS 14 vs macOS 26.
Verified: app and appex Info.plists now both report 3.5 / 7 / 26.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two failures here are silent rather than loud, which is why they have gone
unnoticed on iOS and would have been much harder to diagnose on a Mac.
The old `enum Keychain` discarded all four OSStatus results. On Mac Catalyst a
missing `keychain-access-groups` entitlement makes SecItem calls fail with
errSecMissingEntitlement, and because nothing checked, that was indistinguishable
from "no token stored" — the user would be signed out on every launch, with no
error anywhere. KeychainStore now checks every status, distinguishes
errSecItemNotFound from real failures, and sets kSecUseDataProtectionKeychain so
macOS selects the modern entitlement-gated keychain instead of the legacy login
keychain.
Adding the entitlement moves the default access group to the first array entry,
so that entry is deliberately the app's own group: existing tokens keep
resolving with an unqualified query and nobody is signed out. loadToken() then
migrates forward in three steps — shared group, own default group, and the
pre-Keychain UserDefaults copy.
If the entitlement is not provisioned yet, KeychainStore falls back to the
default group rather than throwing. A hard failure would make the app unusable
for everyone whose provisioning lags; the fallback asserts in DEBUG instead, so
a misconfiguration is loud in development and survivable in production.
Separately, logout() left widget-cache.json in the App Group container, so
widgets kept rendering the signed-out user's events indefinitely. That is an
existing iOS bug, and it would have leaked the same data to any other app
reading the container. WidgetStore.clear() now removes both cache files and
reloads the timelines.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds Swiss German, French, Italian, Dutch, Danish, Norwegian, Swedish and
Finnish — ten selectable languages, each complete at 418/418 keys with
matching placeholders and date arrays.
A language used to be registered in four unrelated places (dictionary,
settings dropdown, date-picker locale ternary, resolver fallback), so
adding one meant touching all of them. LANGUAGES in i18n.js is now the
single source of truth: adding a language is one entry plus one
dictionary, and the dropdown derives from it.
The registry also separates the stored code from the BCP-47 tag used for
formatting. That distinction is what makes 'ch' safe — "ch" is a country,
not a language, and would throw if handed to Intl; it maps to de-CH.
Fixes locale handling along the way: fmtTime/fmtDatetime in calendar.js,
month.js and week.js hardcoded 'de', so English users saw German-formatted
times, and two toLocaleDateString() calls silently used the browser locale
instead of the app language. All formatting now goes through getLocale().
Missing keys fall back language → English → German rather than straight to
German, and the dead per-dictionary `locale` key is gone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Replace the stale blue calendar icon with the real green Calendarr app
icon (matching iOS/Android) for the installable web app:
- icon-192/512.png regenerated from the branded 1024 app icon
- add dedicated maskable icons (192/512) with brand-green safe-zone
padding so Android/Chrome render a proper adaptive icon (no white box)
- manifest: maskable purpose entries, id, description, lang/dir,
categories; theme_color -> brand green (#16713d)
- index.html: theme-color -> green, add modern mobile-web-app-capable
- favicon.svg + icon.svg recolored to brand green
Bump APP_VERSION v89 -> v90.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Equal rights for all calendar types: the sidebar remove button is now a uniform
"hide" (eye) for every calendar instead of delete-for-some / nothing-for-shared.
- Owned local (incl. birthday) and iCal now hide via the server sidebar_hidden
flag (delete stays available in Settings' calendar table).
- Calendars shared with me get a per-device sidebar hide (they can't carry a
server flag); their events are filtered out while hidden, and the Settings
calendar table's "shared with me" rows gain an un-hide eye toggle.
- External caldav/google/ha unchanged.
Group calendars intentionally left as-is for now.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The version line in the Impressum footer ("Calendarr vNN") is now a
subtle link to the Calendarr Gitea repo (git.scarriffle.com), opening
in a new tab. Styled to stay muted at rest and reveal as a link on
hover. Bump APP_VERSION v87 -> v88.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Admin default-theme editor: "Theme importieren" button loads a .theme.json
into the editor (colours only, live preview) so an imported theme can be set
as the instance default (still confirmed via "Standard speichern").
- Branding: drop the "Standard" placeholder; show pixel-dimension hints instead
(logo shown at max 150×40, favicon 128×128). Logo preview box now mirrors its
real topbar footprint.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replaces the blue defaults with the provided green theme (primary #58B900,
accent #45A148, today #6FB669, greens for month/selected/today-bg, etc.). These
apply when no admin instance theme is set. THEME.md defaults table updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Admin settings tab (admin-only) now holds user management plus a server-wide
default theme editor (live preview + discard) and logo/favicon branding.
- New singleton InstanceSettings model + /api/instance router (public GET for
the login screen; admin-gated theme PUT and logo/favicon upload/delete,
reusing the avatar PIL/validation pattern).
- Instance default theme is the base users inherit and the target a per-colour
"Reset" returns to (baseColor: instance default -> built-in).
- Custom favicon overrides the primary-colour tinting; custom logo replaces the
top-left glyph+text, hard-capped so it only scales down and never breaks layout.
Branding + defaults load before login (public endpoint).
- Theme export now uses the recognised .theme.json extension (old .theme still
imports); import stays partial/unknown-key aware.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Theme:
- Export/import themes via UI as <date>_<time>.theme (JSON, readable keys,
link to THEME.md). Import is partial-aware: only present params are written
(server accepts partial); prompts before importing files with unknown params.
- Dynamic favicon + theme-color tinted to the primary colour on load/save.
- New themable colours: general hover-highlight, day hover/selected/bg,
today background, plus two unified sidebar action-icon colours
(inactive/active) covering bell, hide, delete and read-only icons.
- All new colours are per-setting syncable; documented in THEME.md.
UX:
- Styled confirm dialog (#modal-confirm) replaces window.confirm() for
calendar delete and account disconnect.
- Birthday/local calendars and iCal subscriptions can now be hidden from the
sidebar via Settings (new sidebar_hidden column + hide toggle).
Backend: additive nullable columns + idempotent migrations for user_settings
colours and local_calendars/ical_subscriptions.sidebar_hidden.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
