- Native fetch + AbortController 15s timeout (icon-discovery idiom, no
axios); URL host hardcoded, only the internally-computed dayCursor is
interpolated (T-10-06)
- HTTP 400 treated as an expected no-op (pubDay today/future) -> []
- adm-zip extraction with a pre-extraction decompression-bomb ceiling
check (sum entry.header.size vs ~50MB) before any entry buffer is read
(T-10-07); entries are never written to disk
- D-02 open-tender filter: positive tag.includes('tender') match only —
award/planning/untagged-with-awards excluded (Pitfall C)
- Register DoeOpenDataAdapter in TendersModule.providers
- All doe-opendata.adapter.spec.ts tests green (6/6)
A favorite entered as a bare host ("ctl.de") passed @IsUrl() but had no
scheme, so `new URL()` threw inside icon discovery and it silently fell
back to a relative "/favicon.ico" — which 502'd through the icon proxy
and left the widget showing the first-letter placeholder ("C").
- add normalizeUrl() (prepend https:// when no scheme present)
- apply it in discoverFavoriteIconUrl and when storing the favorite url,
so both the link and discovery use the normalized value
- on update, re-run discovery when the icon field is cleared, so editing
a previously-broken favorite repairs its icon
- tests: normalizeUrl cases + end-to-end discovery (apple-touch extraction,
scheme-less fallback stays absolute)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a per-username denylist so individual accounts (service accounts like
administrator, krbtgt, guest, ldap$) can be excluded from LDAP sync,
independent of the group/OU include-filter which only scopes the search.
- schema: LdapConfig.userExcludeList String[] (+ migration)
- sync: skip excluded usernames (case-insensitive) before recording the DN,
so an already-imported user added to the list gets deactivated next sync
- DTO / config service / controller / scheduler: thread userExcludeList through
- web: exclude-list admin UI section (add/remove/save) + de/en translations
- tests: 3 specs covering empty list, case-insensitive skip, deactivation
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ldapts represents a missing/absent LDAP attribute as an empty array
([]), not undefined -- entry['mail'] is [] when an account has no mail
set. The field-mapping loop did Array.isArray(value) ? String(value[0])
: ..., and String(undefined) is the literal string "undefined". Every
synced entry without that attribute got mappedData['email'] = "undefined"
(a truthy string, so the `|| fallback` never kicked in), and the second
such entry onward crashed with a unique constraint violation on email
since they all shared the exact same literal string.
Found live: syncing against a real Zentyal/Samba AD directory failed
on every entry after the first (Kevin Schaller, krbtgt, Guest, the DC
computer object, etc.) with "Unique constraint failed on the fields:
(email)".
Fix: resolve array values to their first element (or use the raw
value for non-arrays) and only keep it when actually present and
non-empty, so a genuinely missing attribute falls through to the
`${username}@ldap.local` fallback instead of the string "undefined".
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Username lookups (login, admin seed, LDAP sync) compared case-sensitively
against a stored value with whatever casing it was created with, so
"Admin" and "admin" were treated as different accounts.
Normalizes at every write and read path: UserService.create/update
lowercase the username before persisting, findByUsername lowercases
the lookup input, AuthService.validateUser lowercases before the login
query, AdminSeedService lowercases the configured admin username, and
the LDAP sync loop lowercases the mapped sAMAccountName before using it
for lookup/create/update -- so AD casing differences don't create
duplicate accounts either.
Added a data migration to lowercase any existing mixed-case usernames.
It relies on the User.username unique constraint to fail loudly if two
existing accounts would collide after normalizing, rather than silently
merging them.
Verified locally: logged in with "ADMIN" (uppercase) against the
existing lowercase "admin" account after rebuilding the API image.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
bindDn and bindPassword are now optional on LdapConfig (nullable
migration) and throughout the DTOs/service/client -- an admin can
leave both blank to connect to directories that permit anonymous
read access. LdapService.bind() falls back to an RFC 4513 anonymous
bind (empty DN + empty password) whenever either field is missing,
shared across testConnection, listGroups, and syncUsersForTenant.
Frontend: removed the required attribute from Bind-DN/Bind-Passwort,
added a placeholder hint ("leer = anonymous bind"), and the
"Verbindung testen" button now only needs a Server-URL to enable
(not bindDn+bindPassword). Config responses now return bindPassword
as null (not a misleading "********") when no password is set.
Verified locally: submitted only a Server-URL with both bind fields
empty and confirmed the request reached the anonymous-bind code path
(DNS failure for the unreachable test host, not a validation error).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
"Verbindung testen" only appeared after a config was already saved,
so a fresh/never-configured tenant had no way to validate connection
details at all. Now the button is always visible; POST /ldap/test-connection
accepts optional ad-hoc serverUrl/bindDn/bindPassword and falls back
field-by-field to the saved config for anything omitted (bindPassword
in particular, since the form never re-sends the masked saved password).
Verified locally: tested with typed-but-unsaved values (base DN left
blank, nothing saved) and got a real connection attempt/error back,
not a 404 "no config" response.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
FavoriteLink existed in schema.prisma but was never captured in a
migration -- same bug class as the DashboardLayout fix in
20260629130000_add_missing_tables. It silently worked in local dev
(table created via db push) but any environment relying on
`prisma migrate deploy` never got the table, causing every
GET /favorites request to 500 with PrismaClientKnownRequestError
P2021 ("table does not exist").
Found live testing the production deploy at alpha.tessera.ctl.de:
adding a Favoriten widget triggered the 500. Confirmed via server
logs (docker logs) and by inspecting _prisma_migrations / \dt on
that database.
CREATE TABLE/INDEX IF NOT EXISTS makes this safe to apply against
environments where the table already exists untracked (verified: ran
clean against local dev's DB, which already had the table from a
prior db push, with zero errors and zero data loss).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Reproducibly confirmed (3/3 vs 3/3 direct comparison inside the API
container) that chatgpt.com's Cloudflare WAF returns 403 for the
"tessera/1.0" User-Agent regardless of Accept header, and 200 for a
real Chrome UA string. Parameterized fetchWithRedirectGuard's
User-Agent (defaulting to the existing "tessera/1.0") and override it
only for fetchIconBytes -- the HTML-discovery path (discoverFavoriteIconUrl)
keeps its original User-Agent unchanged, per the no-regression constraint
on that flow.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
A bare "image/*" Accept paired with the tessera/1.0 User-Agent tripped
Cloudflare bot mitigation on some sites -- caught live testing against
chatgpt.com/favicon.ico, which returned 403 with this combo but 200
with a realistic browser-style image Accept list. Isolated via direct
fetch comparison inside the API container before landing the fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Ownership-scoped (userId, matching update/remove) icon byte proxy.
Loads the row's stored iconUrl server-side and streams it through
IconDiscoveryService.fetchIconBytes -- never accepts a client-supplied
URL, so this can't become an open SSRF proxy.
Not-found/not-owned/no-icon -> 404. Upstream fetch failure (unreachable,
timeout, non-image, SSRF-blocked) -> 502, never a 200 with a placeholder.
Success sets Cache-Control so the browser doesn't refetch every load.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Extracts the manual-redirect/per-hop-revalidation/timeout loop from
fetchHtml into a shared fetchWithRedirectGuard, exports isPublicHttpUrl,
and adds fetchIconBytes() -- an image-content-type-gated, 1MB-capped
byte fetch reusing the same SSRF guard as the existing HTML discovery
path. discoverFavoriteIconUrl behavior is unchanged.
Prepares the fix for favicon hotlinks breaking on sites that send
Cross-Origin-Resource-Policy: same-origin (e.g. claude.ai) -- Tessera
will proxy the bytes through its own origin instead.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
CreateLdapConfigDto's optional fields (searchFilter, syncIntervalMin,
isActive, groupFilterDns) had class-field default initializers.
NestJS's ValidationPipe instantiates DTOs via plainToInstance, which
applies those defaults even when the field is absent from the request
body -- so any partial PATCH not including a given field silently
reset it to the hardcoded default instead of leaving it untouched.
Caught by testing the new groupFilterDns-only PATCH: saving the group
filter alone reset searchFilter back to "(objectClass=person)",
clobbering the configured Active Directory filter. The service layer
already has its own `?? default` fallback for create, so the DTO
initializers were redundant and unsafe. Removing them makes updateConfig's
existing "only set if dto.field !== undefined" pattern behave correctly
for every optional field, not just the ones sent together in one request.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds listGroups() to browse AD groups/OUs under base DN, and
collectSearchEntries() to restrict syncUsersForTenant to members of
selected groups or users under selected OUs. Group DNs are matched
via escaped memberOf clauses (RFC 4515); OU DNs become extra search
bases. Empty groupFilterDns keeps the original single-base-DN search
unchanged. Controller sync endpoint and the sync scheduler both pass
groupFilterDns through so manual and scheduled syncs honor it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Persists per-tenant AD group/OU DNs to restrict which directory
entries get synced. Empty array (default) preserves current
behavior — import everyone under base DN.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Column already existed in production DB — migration failed with 42701.
Hotfixed via psql UPDATE on _prisma_migrations; migration SQL updated
to prevent recurrence on fresh deploys.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Column was added to schema but migration was missing, causing
PrismaClientKnownRequestError P2022 on prod API startup.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add FileResponse interface and FORMAT_MIME map to service
- Implement convertCert: parses any input format (PEM/DER/PFX/P7B) via
same logic as parseCert; serializes to pem/der/p7b targetFormat
- DER output uses bytesToHex→Buffer.from(hex,'hex') to avoid utf-8
corruption (Pitfall 1 / T-09-06)
- P7B output: pkcs7.createSignedData + pem.encode (PEM-wrapped PKCS7)
- Wrap all forge ops in try/catch → BadRequestException (T-09-01)
- Controller: add @Body('pemText') + reject when neither file nor pemText
- Fix: re-add NotImplementedException import for mergeCerts stub
- All 23 API cert-manager tests green (including 4 new convertCert)
- splitCerts fullchain PEM: expects count 2, two single-PEM-block certs with correct CN
- splitCerts P7B PEM bundle: expects at least one cert in result
- splitCerts malformed input: expects BadRequestException
- All three tests FAIL against NotImplementedException stub (RED confirmed)
- All 16 prior tests still pass
- Added parseCert describe block with 5 failing tests
- Covers PEM input, DER input, PFX+correct-password, PFX+wrong-password (BadRequestException), malformed input (BadRequestException)
- Existing 11 helper tests still pass
- Fixtures built via node-forge (RSA-1024, DER from asn1.toDer, PFX via toPkcs12Asn1)
- SSRF check skipped for Exchange type (internal EWS servers are common)
- testConnectionFromConfig catches SSRF/validation errors, returns {success:false,error} instead of throwing 403
- updateSource reads existing.type to determine effective type for SSRF check
- Panel shows saveError/editSaveError on failed add/update
- Edit form initialValues now includes domain field
- i18n: calendar.saveError key added (de+en)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Prisma: domain String? added to CalendarSource model (db push applied)
- DTOs: domain in CreateCalendarSourceDto, UpdateCalendarSourceDto, new TestCalendarSourceConfigDto
- Service: domain in SOURCE_SAFE_SELECT, addSource, updateSource; new testConnectionFromConfig method
- Controller: POST /calendar/sources/test-config (before :id routes to avoid collision)
- ExchangeProvider: domain in all source interfaces; passed as 3rd arg to EWS WebCredentials
- Frontend: domain in CalendarSource/CreateSourcePayload/UpdateSourcePayload; testSourceConfig API fn
- Form: domain field (Exchange-only), "Test connection" button with idle/loading/success/error states
- i18n: de+en keys for formFieldDomain, formFieldDomainHint, formTestConnection, formTesting, formTestSuccess, formTestFailed
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- CR-01: fix SSRF bypass — isPrivateIpv6 now delegates ::ffff:<ipv4> to
isPrivateIpv4, covering 172.16-31.x and 169.254.x ranges
- CR-02: add ParseUUIDPipe to GET /favorites widgetId param + service guard
so missing widgetId returns 400 instead of leaking all user favorites
- WR-01: link-widget — replace raw 'link.error' key with t('link.error') (4 sites)
- WR-02: favorites-widget — fix load-path error to use t('favorites.error')
- WR-03: widget-catalog-modal — move aria-hidden from outer wrapper to backdrop
- WR-04: calculator — remove duplicate M button (MR clone); MC/MR/M+/M−/MS remain
- WR-05: schema — add FavoriteLink→WidgetInstance FK with onDelete:Cascade
- IN-01: create-widget.dto.ts — update comment from four to eight supported types
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add avatarPath String? column to User model (migration: add_user_avatar)
- POST /users/me/avatar: 2MB limit, image/png/jpeg/webp allowlist, writes to user-files/avatars/{userId}.{ext}
- GET /users/me/avatar: streams avatar with Cache-Control: no-store
- AuthService.getMe(): returns isLocalUser + hasAvatar without leaking passwordHash/ldapDn
- AuthController GET /auth/me: now returns enriched profile via getMe()
FindFolder was searching only under inbox DistinguishedFolderId, missing
folders at mailbox root level. Now searches msgfolderroot (full mailbox)
so custom folders like DKV are found regardless of placement.
Also adds HTTP status check and debug logging for FindFolder responses.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Custom folder names (e.g. "DKV" or "INBOX/DKV") now resolved by calling
EWS FindFolder deep-search under inbox. Well-known names still map to
DistinguishedFolderId directly. Falls back to inbox with a warning log
when the subfolder cannot be found.
IMAP already supported subfolder paths natively via ImapFlow.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
loadConfig used CONFIG_SAFE_SELECT which excludes encryptedInboxCreds entirely,
so username was never returned to the frontend — form always showed empty username.
Added getConfigForApi() which loads the safe config + decrypts encryptedInboxCreds
to extract username (never password) and adds hasPassword boolean. Controller
getConfig now calls getConfigForApi instead of loadConfig.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Date fix: previous regex matched payment-due date ("10 Tage nach Rechnungsdatum...
10.04.2026") instead of actual Rechnungsdatum. New approach anchors on the
invoice number line (DD/DDDDDDDDD/DDD) and takes the date on the next line,
which is always the actual Rechnungsdatum in DKV PDFs.
Exchange dedup: FindItem now filters IsRead=false (combined with sender filter
via <t:And>), so already-processed emails are skipped automatically.
After downloading attachments, UpdateItem marks the message as read
(using ItemId + ChangeKey from GetItem response), mirroring IMAP \Seen behavior.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Parser now extracts Rechnungsnummer (DD/DDDDDDDDD/DDD) and Rechnungsdatum
from PDF text, so filename doesn't rely on email subject
- Export filename changed from DKV_YYYY-MM_... to RG-DKV-{nr}-{YYMMDD}.xlsx
e.g. RG-DKV-26-650869002-002-260331.xlsx
- Subject fallback now also matches slash-separated invoice numbers (26/NNN/NNN)
- writeAndPrune simplified to accept baseName instead of separate fields
- Validation regex and prune prefix updated to match new RG-DKV- pattern
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Analyzed Invoice-4302486921-26_650869002_000.pdf text structure. Three cases:
1. FUEL (fields[4] = numeric tx-nr): km+product merged in fields[5], unit in
fields[6]. Already working; no change.
2. EV CHARGING (fields contains "DDDD KWH" or "DDDD MIN" unit): column layout
shifts — no km field, station+ort sometimes merged in fields[1]. Detected by
regex on unit field; kwhIdx drives relative offset for menge/netto/brutto.
Ort extracted from fields[2] (kwhIdx>=5) or fields[1] (kwhIdx=4, compact).
Kilometerstand = 0 (EV chargers don't record odometer).
3. SERVICE ROWS (e.g. "DKV Analytics Premiu"): appear inside a VEHICLE: block but
fields[4] is non-numeric (product description, not a transaction number). These
were being parsed as fake vehicle transactions producing wrong ort/km values.
Now filtered out (return null).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Three issues fixed:
1. Kennzeichen normalization: DKV PDF extracts plates without hyphens
("GP JL 740E" vs CSV-imported "GP-JL 740E"). Added _normalizeKennzeichen()
which strips hyphens, spaces, and dots before lookup — resolves vehicle
master match failure that caused Marke/Modell/Fahrer to appear empty.
2. Empty-string NaN: parser used ?? '0' which doesn't catch empty strings,
causing parseDE('') = NaN. Changed to || '0' for km, menge, and totals.
3. Invalid km values: EV charging rows from DKV have misaligned columns —
km position contains a decimal price (e.g. 18.64 EUR or kWh). Added
sanity check: non-integer km values are written as null (empty cell)
instead of a misleading decimal. ExportRow.kilometerstand is now number|null.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Directory must exist before nestjs user takes over — otherwise DkvExportService
cannot write xlsx export files and throws EACCES on first inbox processing run.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
schalli