Legt je ein layout.tsx fuer cert-manager, dkv-fleet, domaincheck und
tender-radar an, das ModuleAccessGate mit dem fest eingetragenen
Verzeichnis-Slug umschliesst. Ein Layout im App Router deckt alle
verschachtelten Unterrouten automatisch mit ab — my-sources und
settings unter tender-radar sowie settings und vehicles unter
dkv-fleet schliessen sich ohne eigene Datei (WINDOWS #10, PERM-04).
Die generische Route [category]/[moduleSlug]/page.tsx nutzt jetzt
ebenfalls ModuleAccessGate statt des bisherigen Inline-403-Markups —
das 403-Markup existiert damit nur noch einmal im Code
(module-access-denied.tsx).
module-layouts.test.tsx deckt zwei Threats ab: falscher Slug in einem
der vier Layouts (T-e8k-03) und ein kuenftig hinzugefuegtes
Modulverzeichnis ohne Layout (T-e8k-04, liest das Verzeichnis per
node:fs aus). module-access.test.tsx ist auf die Weitergabe an das
Gate umgeschrieben, die 403-vs-Rendern-Entscheidung ist bereits durch
module-access-gate.test.tsx abgedeckt.
Volle Web-Testsuite (213 Tests), Typpruefung und Produktionsbau sind
gruen. Browser-Gegenprobe folgt durch den Orchestrator.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5jtbGzC5Sf9npJ3JCjKhq
Zieht das bisher inline in [category]/[moduleSlug]/page.tsx stehende
403-Markup in ModuleAccessDenied (uebersetzungsfrei, nimmt fertige Texte
als Props) und legt mit ModuleAccessGate eine wiederverwendbare
Server-Component-Pruefung an, die checkModuleAccess aufruft und bei
jeder Ausnahme ebenfalls als "kein Zugriff" wertet (zweite
Verteidigungslinie ueber das bereits geschlossen ausfallende
checkModuleAccess, T-15-29). Vier Testfaelle decken Durchlassen,
Verweigern, Ausnahme und Slug-Weitergabe ab.
Bereitet Task 2 vor: die vier Modul-Layouts und die generische Route
werden auf dieses Gate umgestellt (WINDOWS #10, PERM-04).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5jtbGzC5Sf9npJ3JCjKhq
WINDOWS #10: der Zugriffs-Guard sitzt nur in der generischen Route
modules/[category]/[moduleSlug]/page.tsx. Die vier fest verdrahteten
Modulverzeichnisse laufen daran vorbei.
Plan: gemeinsame 403-Komponente + ModuleAccessGate (Server Component),
je ein layout.tsx pro Modulverzeichnis (deckt Unterrouten mit ab),
generische Route auf dieselben Bausteine umgestellt. Zwei Tasks,
Browser-Gegenprobe als human-check.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5jtbGzC5Sf9npJ3JCjKhq
Die Phasen 11 bis 14 standen in der Roadmap-Tabelle auf "In Progress", obwohl an
keiner davon noch gearbeitet wird. Der Status wurde gegen die Verifikationsberichte
gehalten und dabei zeigte sich, dass es kein reiner Buchhaltungsrueckstand war:
11 und 12 stehen auf passed und sind jetzt Complete.
13 traegt einen offenen Befund. tenderFingerprint() haengt buyerName, title,
cpvDivisionKey, deadlineKey und valueBucket zu einem String und hasht ihn. Die
Scraper liefern cpvDivisions immer leer, DOE meist gefuellt — dieselbe
Ausschreibung ergibt aus beiden Quellen verschiedene Fingerabdruecke. Die
Fingerprint-Stufe greift also gerade dort nicht, wo sie greifen soll. Das stand
seit dem 2026-07-23 in 13-VERIFICATION.md, aber nicht im Ledger und war damit
praktisch unsichtbar. Jetzt WINDOWS.md #11.
14 wartet auf den Live-Test gegen ein echtes Exchange-Postfach; der
NTLM/SOAP-Weg ist nur gegen Mocks geprueft. Vom User am 2026-07-24 bewusst
zurueckgestellt, stand ebenfalls nur im Verifikationsbericht. Jetzt WINDOWS.md #12.
Beide Phasen tragen den wahren Status statt "In Progress", mit Verweis auf den
jeweiligen Ledger-Eintrag.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5jtbGzC5Sf9npJ3JCjKhq
Die sechs seit Anfang August offenen Browser-Durchlaeufe aus Phase 15 und 16
(WINDOWS.md #1-#6) waren liegengeblieben, weil in den damaligen Sitzungen kein
Browser-Tool verfuegbar war. Der Stack lief fuer die Phase-17-Gegenproben ohnehin,
also wurden alle nachgeholt, die ohne echtes Active Directory pruefbar sind.
Geschlossen:
#1 (15-06) Gruppenverwaltung: anlegen, umbenennen, Standardmarkierung exklusiv
setzen und zuruecknehmen, Mitglied hinzufuegen und entfernen, Loeschdialog nennt
beide Zahlen konkret. Bei gestoppter API meldet der Loeschvorgang deutschen
Klartext und laesst die Gruppe stehen. Die AD-Bindung ist an dieser Stelle
gegenstandslos geworden — Phase 16 hat sie in den LDAP-Bereich verlagert.
#2 (15-07) Freigaben-Matrix: setzen und entziehen, Aktivierungsdialog auf beiden
Wegen (Sofort-Freigabe legt den Grant an, "Spaeter konfigurieren" nicht),
Direkt-Grant neben Gruppen-Grant mit sichtbarer Gruppenspalte, langer Gruppenname
bleibt in seinen Massen. Bei gestoppter API springt die Checkbox zurueck und die
Datenbank bleibt unveraendert — kein optimistischer Zustand ueberlebt.
#5 (16-04) Gruppen-Dialog in allen drei Zustaenden, inklusive importierter Gruppe
mit gesperrtem AD-Namen und freiem internen Namen; die Liste zeigt danach den
internen Namen und traegt den AD-Namen im Tooltip. Die AD-Bindung wurde dafuer in
der Datenbank gesetzt, nicht importiert — im Bericht ausdruecklich vermerkt.
#3 (15-08) wurde durchgefuehrt und hat einen Defekt aufgedeckt:
WINDOWS #10 (neu, offen): PERM-04 greift nicht auf den modul-eigenen Routen. Der
Zugriffs-Guard sitzt allein in modules/[category]/[moduleSlug]/page.tsx. Die vier
fest verdrahteten Modulrouten (tender-radar, dkv-fleet, cert-manager, domaincheck)
samt Unterseiten laufen daran vorbei. Ein USER ohne Freigabe bekommt unter
/modules/procurement/tender-radar korrekt die 403-Seite, unter /modules/tender-radar,
/modules/tender-radar/my-sources und /modules/dkv-fleet dagegen die volle Seite mit
bedienbaren Knoepfen. Keine Datenpreisgabe — die API antwortet durchgehend 403 —
aber der Nutzer sieht Bedienelemente, die er nicht benutzen darf, und rohe
englische Techniktexte statt einer verstaendlichen Meldung. Sidebar, Marketplace
und API verhalten sich dagegen korrekt.
Offen bleiben #4 und #6: beide messen den Sync-Vorgang selbst und brauchen ein
erreichbares Active Directory.
Berichte mit Screenshots unter 15-UAT-2026-09-07.md und 16-UAT-2026-09-07.md.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5jtbGzC5Sf9npJ3JCjKhq
Die drei human-check-Punkte aus Phase 17 waren am 2026-08-12 offen geblieben,
weil in der Ausfuehrungssitzung kein Browser-Tool verfuegbar war. Sie wurden
jetzt gegen frisch gebaute Images aus main (79015dd) mit echtem Chrome und
leerer Datenbank durchgeklickt.
#7 (17-01 Task 2): nutzer1 speichert sein Postfach, die Werte ueberleben den
Reload; nutzer2 sieht ein leeres Formular statt der Werte von nutzer1. Danach
zwei TenderEmailConfig-Zeilen mit je eigenem userId.
#8 (17-03 Task 1): Meine Quellen zeigt alle drei Abschnitte, ein eigener
RSS-Feed erscheint sofort mit Entfernen-Knopf, der plattformweite service-bund
steht darunter ohne, und das Digest-Intervall "Woechentlich" ueberlebt den
Reload.
#9 (17-03 Task 2): USER sieht auf /settings keine Bedienelemente, nur Hinweis
und Verweis; SUPER_ADMIN sieht Abrufintervall und plattformweite Feeds, aber
kein Postfach und keine Benachrichtigung mehr. Das Zahnrad fuehrt in beiden
Faellen nach Meine Quellen.
Zusaetzlich am laufenden Server gemessen, weil eine ausgeblendete Schaltflaeche
kein Beweis fuer eine serverseitige Sperre ist: als nutzer2 liefert GET
/rss-feeds nur den plattformweiten Feed, DELETE auf den plattformweiten wie auf
den fremden Feed antwortet 404 ohne die Zeile anzufassen, und POST mit
scope=platform wird mit 403 abgewiesen.
WINDOWS.md #7/#8/#9 auf fixed (open_count 9 -> 6), Verifikationsbericht mit
Nachtrag und Screenshots auf passed, ROADMAP auf Complete, STATE nachgezogen.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K5jtbGzC5Sf9npJ3JCjKhq
Every mechanism checked against live schema, migrations and code rather
than the summaries. Three browser click-throughs remain unrun (no browser
tool this session), so the phase lands human_needed, not passed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
- RssFeedListForm.test.tsx auf scope umgestellt: personal zeigt eigene Feeds
editierbar + plattformweite als schlichte, nicht bedienbare Liste darunter;
platform zeigt nur plattformweite Feeds, eigene Feeds tauchen dort gar
nicht auf; beide Faelle pruefen den scope-Parameter von createRssFeed
- settings-roles.test.tsx (neu): USER sieht Hinweis+Verweis statt
Bedienelemente, ADMIN/SUPER_ADMIN sehen Abrufintervall+RSS-Feeds, unbekannte
Rolle zeigt weder-noch, entfernte Abschnittsueberschriften kommen nirgends
mehr vor
- my-sources.test.tsx (neu): alle drei Abschnittsueberschriften vorhanden,
Verweis auf die Administrationsseite nur bei ADMIN/SUPER_ADMIN
- Erwartete Texte in allen drei Testdateien von Hand geschrieben, nicht aus
der gleichen next-intl-Zuordnung abgeleitet, die die Komponenten benutzen
- Backlog-Punkt 2026-08-11-tender-radar-einstellungen-mischen-rollen.md nach
todos/completed/ verschoben, mit Resolution-Abschnitt: beide Nutzer-
Entscheidungen (eigene Quellen je Nutzer, Trefferliste bleibt
plattform-global) und die Antwort auf offenen Punkt 4 (eigene
nutzerseitige Modulseite als Vorbild fuer kuenftige Module) dokumentiert
Verifikation: Web 205/205, API src/tenders 362/362, beide Typpruefungen
fehlerfrei.
- settings/page.tsx zeigt nur noch Abrufintervall + plattformweite Feeds;
Postfach- und Benachrichtigungsabschnitt entfernt (ziehen auf my-sources um)
- Anzeigepruefung der Rolle aus dem Anmelde-Speicher: unbekannt -> Platzhalter,
ADMIN/SUPER_ADMIN -> Inhalt, sonst Hinweistext + Verweis auf "Meine Quellen"
(verbindliche Pruefung bleibt serverseitig, T-17-08/@UseModule, siehe
Dateikommentar)
- Zahnrad auf der Modulseite fuehrt jetzt nach /my-sources statt /settings;
neuer Schluessel page.mySourcesTitle, alter page.settingsTitle bleibt als
Verweistext auf der Nutzerseite (Task 1) in Gebrauch
- settings.emailSectionTitle/emailSectionBody/notificationsSectionTitle aus
beiden Sprachdateien entfernt (gegengeprueft: nirgends mehr referenziert);
neue Schluessel settings.accessDeniedText, settings.rssSectionUserNote
- RssFeedSource bekommt isPlatformWide (server-derived), createRssFeed nimmt
einen scope-Parameter (personal/platform, Vorgabe personal)
- RssFeedListForm bekommt scope-Prop: personal zeigt eigene Feeds editierbar +
plattformweite als schlichte Aufzaehlung ohne Knoepfe darunter; platform
zeigt nur plattformweite Feeds editierbar
- DigestIntervalForm aus settings/page.tsx unveraendert herausgeloest (keine
neuen Beschriftungen, gleiche settings.*-Schluessel)
- my-sources/page.tsx um "Meine Feeds" und "Benachrichtigung" erweitert,
Verweis auf die Administrationsseite nur fuer ADMIN/SUPER_ADMIN
- settings/page.tsx vorgezogen auf RssFeedListForm scope="platform" (Rule 3,
eigener Type-Check-Verify sonst rot) — volle Rollenpruesung folgt Task 2
Rule 1: createRssFeed's Antwort traegt kein isPlatformWide (nur GET mappt es
serverseitig) — RssFeedListForm setzt es nach dem Anlegen lokal aus dem
verwendeten scope, sonst wuerde ein frisch angelegter plattformweiter Feed
bis zum naechsten Neuladen aus seiner eigenen Liste verschwinden.
- RssAdapter.fetchTenders tags every record from a feed with a tenantId
with the same D-13 ownerTenantId origin marking email-alert records
carry since Phase 14; platform-wide feeds (no tenantId) stay unmarked.
The pure parseFeed mapping is untouched — tagging happens in the
fan-out loop that knows which row a batch came from
- Extracted the service.bund.de seed out of TendersModule.onModuleInit
into seedServiceBundRssFeed() (tenders.seed.ts, same pattern as the
existing seedTendersModule), so the find-then-create idempotency added
in Task 1 is unit-tested directly instead of only via a Nest bootstrap
- New rss-feed-migration-sql.spec.ts: text-only check of the Task 1
migration file (nullable columns, dropped/created indexes, no
existing-row mutation, correct ordering)
- Files modified: apps/api/src/tenders/adapters/rss.adapter.ts, apps/api/src/tenders/tenders.module.ts, apps/api/src/tenders/tenders.seed.ts, apps/api/src/tenders/adapters/rss.adapter.spec.ts, apps/api/src/tenders/tenders.seed.spec.ts, apps/api/src/tenders/rss-feed-migration-sql.spec.ts
- remove(id, {userId, isAdmin}) replaces remove(id): single conditional
deleteMany (id AND (owned-by-caller OR admin-on-platform-feed)) — no
TOCTOU window, ownership check lives in the DB condition. Deletes
nothing -> NotFoundException (never Forbidden, no existence leak)
- createForUser rejects a caller's 21st personal feed with a clear
German message (T-17-10); platform-wide feeds are not counted
- DELETE /rss-feeds/:feedId moves from @Roles(ADMIN,SUPER_ADMIN) to
@UseModule('tender-radar') — ownership check does the gating now
- Tests use a Prisma double that actually evaluates the where condition
(not a double that always "succeeds") for both deleteMany and count
- Files modified: apps/api/src/tenders/tender-rss-feed.service.ts, apps/api/src/tenders/tenders.controller.ts, apps/api/src/tenders/tender-rss-feed.service.spec.ts, apps/api/src/tenders/tenders.controller.spec.ts
- TenderRssFeedSource.userId/tenantId (nullable): null = platform-wide
(admin-managed, includes the existing service.bund.de default),
set = personal feed owned by exactly one user
- Migration replaces url @unique with @@unique([userId, url]) — two
users can now follow the same address independently; existing rows
keep an empty owner (platform-wide, unchanged behavior)
- Service: listForUser/createForUser/createPlatform replace list/create
- Controller: GET/POST /rss-feeds move from @Roles(ADMIN,SUPER_ADMIN) to
@UseModule('tender-radar'); POST with scope:'platform' still requires
ADMIN/SUPER_ADMIN, checked inline (T-17-08)
- tenders.module.ts seed switched from upsert-on-url to find-then-create
(Rule 3, pulled forward from Task 3): the new compound unique index
requires a non-null userId in Prisma's generated type, so a
platform-wide row can no longer be addressed via upsert
- Files modified: apps/api/prisma/schema.prisma, apps/api/prisma/migrations/20260812110000_tender_rss_feed_owner/migration.sql, apps/api/src/tenders/tender-rss-feed.service.ts, apps/api/src/tenders/dto/tender-rss-feed.dto.ts, apps/api/src/tenders/tenders.controller.ts, apps/api/src/tenders/tenders.module.ts, apps/api/src/tenders/tender-rss-feed.service.spec.ts, apps/api/src/tenders/tenders.controller.spec.ts
Der mandantenuebergreifende Sammelabruf in email-alert.adapter.ts bleibt
mechanisch unveraendert (findMany({isActive:true}) in einem Zug,
Fehlerbehandlung je Zeile) — geaendert wird nur die Warnmeldung (Zeilen-id
+ Besitzer statt Mandant, T-17-03) und die Klassendoku.
- Neue Tests: zwei aktive Postfaecher DESSELBEN Mandanten werden beide mit
ihren jeweils eigenen Zugangsdaten abgeholt; ein kaputtes Postfach
blockiert das andere nicht und protokolliert eine Warnung ohne
Zugangsdaten/Adresse; die Herkunftsmarkierung folgt dem Mandantenfeld
der jeweiligen Zeile (zwei Mandanten -> zwei Werte). Erwartungswerte von
Hand geschrieben, nicht ueber die Produktivfunktion erzeugt.
- tender-email-config.service.spec.ts (bereits in der Task-2-Migration
mitgeliefert) deckt zusaetzlich: tenantId wird beim Anlegen mitgeschrieben,
zwei Nutzer desselben Mandanten erzeugen zwei Zeilen statt eine zu
ueberschreiben.
- email-config-migration-sql.spec.ts (neu, Vorbild
doe-url-migration-sql.spec.ts): prueft die Reihenfolge der
Hand-Migration textuell — Zuordnung vor Loeschung, Pflicht erst nach
Befuellung, alte Eindeutigkeit runter/neue rauf, gewoehnlicher
tenantId-Index bleibt stehen.
src/tenders: 335/335 gruen. API gesamt: 603/603. Web gesamt: 192/192.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Alert-Postfach gehoert jetzt dem einzelnen Nutzer (userId @unique) statt
dem Mandanten (D-01) — ein zweiter Kollege desselben Mandanten kann sein
eigenes Postfach anbinden. tenantId bleibt denormalisiert (SMTP-Aufloesung,
Herkunftsmarkierung), wird auf create UND update mitgeschrieben.
- Handgeschriebene Migration (prisma migrate dev verweigert die
nicht-interaktive Shell): befuellt Bestandszeilen mit dem aeltesten
aktiven Administrator ihres Mandanten, entfernt verwaiste Zeilen ohne
Administrator, ersetzt die tenantId-Eindeutigkeit durch userId.
Lokal getestet (0 Bestandszeilen lokal und auf alpha — Zaehlung im
Task-1-Checkpoint), Index-Ergebnis verifiziert.
- TenderEmailConfigService.getConfigForApi/saveConfig auf userId als
Schluessel umgestellt; saveConfig nimmt {userId, tenantId}.
- TendersController: email-config-Routen von @Roles(ADMIN,SUPER_ADMIN)
auf @UseModule('tender-radar') umgestellt (Postfach ist jetzt
Nutzereinstellung); Route-Reihenfolge vor @Get(':id') unveraendert.
- Neue Seite /modules/tender-radar/my-sources ("Meine Quellen") mit dem
unveraenderten EmailAlertConfigForm; Hinweistext benennt D-05 (Tender
bleibt plattform-global — nur wer Quellen einspeist aendert sich).
- tenders.controller.spec.ts an neue Service-Signatur angepasst (Rule 3,
nicht im Plan gelistet, aber zum Kompilieren/Bestehen erforderlich).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
/opt/tessera keeps its directory (compose derives the project name from it,
and a rename would have orphaned tessera_pgdata) and now tracks main.
COMPOSE_FILE pins it to the production file so the checkout does not switch
the server onto the dev defaults.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
NEXT_PUBLIC_API_URL is read by the browser, not by the web container, so
http://api:3001 could never work outside Docker. The test server had been
corrected by hand long ago; the fix never came back here, so the file we
would ship to a customer was the broken one.
Also adopts the server's TESSERA_FORCE_CHANGE default of true, so a fresh
install requires the initial admin password to be changed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The note predated 7bda56d and f574884, so two of its three points were
already shipped when it was picked up. Records what each commit actually
closed, and that the .env deny rules block the two committed template
files that hold no secrets.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both example files pointed at the old CALENDAR_ENCRYPTION_KEY name, and
.env.example did not mention the key at all. Since compose now aborts
startup when it is unset, anyone setting up a fresh install from these
templates hit a failure the templates never explained.
Adds the current variable name, the generation command, and a note that
losing the value makes stored credentials unrecoverable.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
No phase work in flight: Phase 15 (8/8) and Phase 16 (5/5) are both closed and
the roadmap reflects it. The handoff therefore sits at project level rather
than in a phase directory.
Records four anti-patterns discovered through actual failure this session, two
of them marked blocking: the tautological test that let the objectGUID defect
through review, and the fact that /opt/tessera is not a checkout, so compose
changes in this repository never reach the test server.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The page carries platform config, tenant config and one per-user setting. The
per-user one -- the digest interval, which is what an ordinary user actually
comes for -- sits last, below three blocks they may not change.
Checked before writing it up: this is not a permission hole. The admin
endpoints are @Roles-guarded server side, so a normal user cannot change
anything; the page simply has no role check of its own, so they see controls
that fail on save.
The second half is a product question deliberately left open, because it is
not specific to this module: DKV-Fleet has the same shape, and every future
module with a personal setting will ask it again.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
/opt/tessera is not a working copy -- the compose file there is maintained by
hand, and the deploy path only pulls images. Two consequences showed up on the
same day: the renamed encryption key never reached the container although it
was in the server .env, and the "refuse to start without a key" guard does not
apply on alpha at all, because it only exists in the repository file.
The item deliberately stops short of proposing a fix to apply: it first asks
why the file is hand-maintained, since it may carry host-specific settings the
repository lacks, and moving to a checkout blindly would break the running
system.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
CALENDAR_ENCRYPTION_KEY was named after the calendar module because that
module needed encryption first, in Phase 5. Every feature since has shared the
same key -- SMTP, the DKV and tender mailboxes, and as of today the LDAP bind
password -- so the name has been describing one of five users rather than the
thing itself, and each new feature inherited the confusion.
TESSERA_ENCRYPTION_KEY is the name now. The old one is still read, because
renaming outright would stop every existing installation at the next start:
their .env carries the old name, and compose was just made to fail hard on a
missing key. When only the old name is present the API logs a deprecation
warning naming both, and when both are set the new one wins -- otherwise a
half-migrated .env would encrypt with one key and decrypt with the other.
CalendarCryptoService becomes CryptoService in its own global CryptoModule.
Four modules used to import CalendarModule purely to reach the provider, which
read as a dependency on calendars where there was none; that import is gone.
Compose keeps the hard failure: without either name the stack refuses to
start. Verified in both files for all three cases -- neither name set (abort),
only the old name (starts), only the new name (starts).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The base compose file carried a hardcoded fallback key, so a stack whose .env
never set the variable started anyway and encrypted every stored credential
(LDAP bind, calendar, SMTP, DKV and tender mailboxes) with a value that is
public in this repository. That is encryption which looks present and protects
nothing.
Both compose files now use the ${VAR:?message} form, so an unset or empty key
fails at compose level with a message naming the variable and how to generate
one, instead of starting with a known key or dying later inside the API with a
stack trace.
Verified both ways: without the variable `docker compose config` exits 1 and
prints the hint; with a key present it exits 0.
Consequence for a fresh clone: the local stack no longer comes up until
CALENDAR_ENCRYPTION_KEY is set in .env. That is the point.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Fallout from encrypting the LDAP bind password: the base compose file still
carries a hardcoded fallback key, so an install that never sets the variable
starts anyway and encrypts everything with a value that is in the repository.
The prod compose already requires it, which is the behaviour the base file
should have too.
Whether the example env files explain the key could not be checked in that
session, so the item says to look first rather than asserting it is missing.
Also notes, as an optional follow-up, why the variable is called
CALENDAR_ENCRYPTION_KEY and what a rename would have to handle.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Notes what was deliberately left out: extracting and renaming the crypto
service out of calendar/ touches five modules and belongs in its own change,
so the existing provider is reused as-is and the naming smell is recorded in
LdapModule instead.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The LDAP bind password was the only credential still stored in clear text.
CalendarSource, SmtpConfig, DkvModuleConfig and TenderEmailConfig have been
AES-256-GCM encrypted for a while; LDAP simply predated the encryption service
and was never brought along.
Hashing is not an option here: Tessera has to replay this password to bind
against the directory, so it must stay recoverable. Encryption at rest covers
the case a hash cannot help with either way -- a database dump or backup
leaving the host without the key, which lives in the application environment.
It does not protect against a compromised host, and does not pretend to.
Reuses CalendarCryptoService, the same provider SettingsModule, DkvModule and
TendersModule already inject, rather than introducing a second crypto path.
The name is a historical accident and is noted as such in LdapModule; renaming
it touches five modules and belongs in its own change.
Decryption sits in getConfig()/getAllActiveConfigs(), the two methods every
consumer already goes through, so callers keep reading a plain `bindPassword`
and the controller keeps masking it to '********' in responses.
The migration only renames the column -- SQL cannot encrypt, since the key is
not in the database. An idempotent bootstrap backfill encrypts rows written
before this change, and until it has run the read path passes a legacy
plaintext value through unchanged so the sync does not break in that window.
A failed decrypt throws rather than returning null: a wrong key must not read
as "no password configured" and silently turn an authenticated bind into an
anonymous one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The AD group membership sync shipped on 2026-08-04 as 614de28; only its
summary was never written, which left Phase 15 sitting at 7/8 as if work were
outstanding. Nothing was.
Each promise the plan made is checked against today's source rather than
against the commit message: bound-only selection, no second sync job, deletion
restricted to source LDAP, manual memberships preserved, memberOf reverse query
instead of attribute reads, RFC-4515 escaping of the group DN, no nested-group
resolution, order independence, per-group error isolation. The 2026-08-11 sync
run on alpha additionally exercised this path against a real directory.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Records the user's choice between the two candidate link targets (the notice
page on oeffentlichevergabe.de, not the awarding portal's own page), what was
checked before touching the adapter, and the trap in verifying it: the target
is a single-page app that answers 200 with an identical shell for any id, so
only rendered content proves the link works.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The doe-opendata adapter stored the OCDS document's own `uri` as sourceUrl.
That is the API address of the record and serves OCDS JSON by design, so
anyone following the link from the results list, the detail view, or an alert
mail landed on raw JSON instead of the notice.
Build the human-readable page from the notice id the row already carries
instead. `/ui/de/search/details?noticeId=...` is the redirect target of
`/ui/de/notices/...`, so it needs no redirect. Verified in a browser for both
id shapes the feed uses -- numeric (25673764 -> "Feuerwehr-Geraetehaus Miehlen
Fliesenarbeiten") and UUID (7085ba12-... -> "Holzfassade"). The page is a
single-page app that answers 200 with an identical shell for any id, so this
had to be checked on rendered content; a status code proves nothing.
The adapter alone only fixes new ingests, so a backfill migration rewrites the
rows already stored -- in Tender and in TenderSource, since the detail view
lists per-source links separately. It touches only rows still pointing at
/api/notices/ and only ids of a shape that was actually verified, which makes
it idempotent and keeps an unexpected id from being pasted into a URL. Counted
read-only against the live database beforehand: 2846 DOE rows affected, none
skipped.
Closes the 2026-08-05 backlog item, which was deliberately held until Phase 16
was done.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The user released the full run once it was clear nothing there is productive
yet. All three report lines appeared (401 created / 9 memberships added / 0
groups adopted, renamed or deleted), and the amber default-marker line stayed
absent as it should when nothing was deleted.
The run doubles as the end-to-end proof for the 260811-f9i fix: the imported
group survived the sweep with its memberships filled, where the old filter
would have deleted it in exactly this run.
Roadmap marks Phase 16 complete; Phase 15 keeps its 7/8 with the reason named.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Neither is a Phase 16 defect; both surfaced while testing it and would
otherwise have been lost with the session.
1. Module activation has no licence check — a tenant admin can activate any
catalogue module for their own tenant. The grants matrix is NOT the hole: it
only distributes what is already active. Carries open product questions
(who issues licences, what expiry does), so it is written up as a draft, not
a decision.
2. The LDAP bind password is stored in clear text although an AES-256-GCM
service already exists and is used for calendar, DKV and tender inbox
credentials. Hashing is not an option here — the password must be replayable
to bind against the directory — so encryption at rest is the fix. No open
questions, just work.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The user has no write access to the company AD, so tests 1, 3, 4 and 8 cannot
be run there, and building a throwaway domain controller for them was judged
disproportionate now that the one substantive defect at this spot is found and
fixed (UAT test 2 -> quick task 260811-f9i).
Recorded rather than hidden: A1 (objectGUID survives a rename) now rests on
Microsoft's documentation, not on our own measurement. The Tessera-side rename
and delete logic stays covered by unit tests against fixtures. If a rename ever
fails to propagate in production, 16-VERIFICATION.md names that as the starting
point.
Phase 16 marked complete in STATE.md.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Quick task 260811-f9i: plan and summary of the fix, STATE.md row, and the UAT
test 2 result. Test 2 was the read-only A2 check against the real directory --
it turned assumption A2 from "unverified" into "false as implemented" and
surfaced a defect that would have deleted every AD-bound group on the first
real sync.
Also records why the defect survived review: the spec mocks built their
expected filter with the same escape helper the production code used, so the
test asserted self-consistency rather than directory behaviour.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The existence sweep in syncBoundGroupsForTenant() built its filter by
interpolating a byte-wise \xx escape of the stored objectGUID into a filter
string. ldapts parses that string before encoding it and does not turn the
escape sequences back into the bytes they stand for, so the assertion value
that reached the directory was a different value and matched nothing.
Measured read-only against a real Active Directory on 2026-08-11, probing a
group whose GUID had just been read from that same directory:
(objectGUID=\1e\4b...) 0 hits
(objectGUID=\1E\4B...) 0 hits
EqualityFilter{attribute, value: <16 bytes>} 1 hit, correct DN
(cn=Domain Admins) [control] 1 hit
Both the narrow base-DN sweep and the wider WR-03 move-detection sweep shared
that filter, so neither could ever hit: every AD-bound group looked deleted and
would have been removed together with its GroupMembership and ModuleGrant rows
on the first real sync, after handing off the default-group marker.
Build the filter as an EqualityFilter over the raw Buffer instead, and drop
escapeLdapFilterBuffer() -- it has no remaining caller and is the trap the code
walked into. escapeLdapFilterValue() is untouched: escaping STRING values into
a filter is correct and still in use.
The existing spec mocks matched on the escaped string, which is how the broken
shape passed review. They now match on the filter object's Buffer value, and
two added tests fail if a stringly-typed objectGUID filter ever comes back.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Ran the Phase 16 browser walkthrough against alpha.tessera.ctl.de with the
real AD (balios.ctl.local) behind it.
- Test 5 (AD group import section): passed. 236 discovered entries, none of
them OUs; selection count, disabled-at-zero button, already-imported badge,
result block and the discovery error state all behave as specified.
- Test 6 (group dialog, three states): passed. Includes the WR-02 case — a
409 name collision stays visible in the dialog with the input preserved.
- Test 7: partial. Error path of the sync report and the grants-matrix column
search under both internal and AD name pass; the number rows of a successful
sync run are still open because a real sync was skipped by request (no
group/OU filter set, so it would import every person under DC=ctl,DC=local).
Tests 1-4 and 8 remain open — they need AD write access or are backstop
assertions. CN=Claude_VT stays imported on alpha because tests 3 and 4 need it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Marks all four Warning findings from 16-REVIEW.md as resolved with their
fix commit hashes; the two Info findings (IN-01/IN-02) remain open and
out of scope for this fix pass. Appends a note to 16-03-SUMMARY.md
recording the WR-02/WR-03/WR-04 behavior change in
syncBoundGroupsForTenant(), since that summary still described the
pre-fix behavior.
syncBoundGroupsForTenant()'s existence sweep only searched the configured
base DNs, so an AD group MOVED to an OU outside that subtree (still
present in the directory) was indistinguishable from a genuine
disappearance and got deleted along with its memberships/module grants —
a silent access loss from a non-destructive AD operation, and a bigger
blast radius than D-05 ("group genuinely gone") was accepted for.
Before concluding disappearance, a second (objectGUID=...) sweep now runs
against each base DN's own domain root (skipped when a base DN already IS
its domain root — the common case, nothing wider to search). A hit there
is reported as an error line and the group is left untouched; only when
the wide sweep also finds nothing is deletion (SC-4/D-05/D-06) actually
established — mirroring the existing conservative stance already taken
for a legacy binding whose DN no longer resolves. Deleting on uncertainty
was the failure mode; this closes it without widening it.
syncBoundGroupsForTenant() compared cn/dn for byte equality, so any
casing difference AD returns between two runs (e.g. after a
domain-controller switch) would look like a rename and re-write
name/ldapDn every single sync — violating the 'sync twice over an
unchanged AD state = no-op' idempotency guarantee. The comparison used
to DECIDE 'is this a rename' is now case-insensitive; the value written
on an actual rename is still stored byte-for-byte as the directory
reports it, per D-03.
syncBoundGroupsForTenant()'s rename write updates name and ldapDn in one
call, so a P2002 there can come from either @@unique([tenantId, name])
or @@unique([tenantId, ldapDn]). The catch previously reported every
P2002 as a name collision unconditionally; it now inspects
err.meta.target the same way importGroupsByDn() already does for its
own create() call, so a non-name unique violation is no longer
mislabelled and sent the admin down the wrong troubleshooting path.
A legacy binding from plan 15-06 has ldapDn set but ldapObjectGuid stays
null until the first syncBoundGroupsForTenant() run backfills it. Until
then, GroupsService.update() let a direct PATCH rename through even
though GroupFormModal.tsx already treats the same group as AD-bound
(isImported = ldapDn != null) — the D-03 name lock was only a UI
convention for that window, not the backend invariant the 16-02 summary
claimed.
schalli