FieldEncryptionBoundary declared a full sensitive/plaintext policy for six
tables and was referenced only by its own test. Nothing in production used it.
Its policy is byte-for-byte the same as FieldCrypto.ENCRYPTED_FIELDS, which is
the copy that actually runs, so nothing is lost by deleting it. Keeping a second
list is the real risk: someone adds a field to this one, sees it classified as
sensitive, and ships something that was never encrypted.
The one apparent improvement it had — requiring an explicit recordId instead of
DataCrypto's temp-${Date.now()} fallback — turns out to guard against nothing.
decryptField derives its context from the recordId stored inside the ciphertext,
not from the argument, so a temporary id at encryption time still decrypts.
The OIDC client secret and LDAP bind password were stored behind an encoded:
prefix that is base64, not encryption. Anyone reading the database read the
secrets. A second path wrote the same thing behind an encrypted: prefix, which
was also base64 — and the reader even documented that it could not decrypt it.
These belong to the installation rather than to a user: sso_providers has no
userId, and the values must be readable during login, before anyone has
authenticated, so the per-user DEK used elsewhere does not apply. They are now
sealed with AES-256-GCM under the system encryption key, which already protects
other installation-level material.
Reading handles both legacy prefixes so an existing install is not locked out of
SSO login, and a legacy value is upgraded the next time the provider is saved.
The three scattered encode/decode sites are replaced by one module.
Retention became configurable in #1132, which only helps if entries can be moved
somewhere before they expire. Until now the only way out was two GET endpoints
built for the UI.
Adds GET /audit-logs/export, taking the same filters as the list endpoint and
streaming the whole matching set as CSV or NDJSON in batches, so an export is not
bounded by the 200-row page cap and does not buffer the result set. Reading the
entire trail is itself recorded as export_audit_logs.
CSV fields starting with =, +, - or @ are prefixed with a quote. Audit rows carry
attacker-influenced values like resource names, and spreadsheet software treats
those as formulas on open.
Adds optional live forwarding to a collector via AUDIT_LOG_FORWARD_URL, with an
optional bearer token. Delivery goes through safeOutboundFetch so a misconfigured
URL cannot be turned into an internal network probe, and it is fire-and-forget:
the local write stays the source of truth and a dead SIEM must never delay or
fail the operation being audited. Repeated failures are reported five times and
then suppressed until delivery recovers, so an outage does not bury the logs it
is supposed to appear in.
Two ways audit evidence still disappeared silently.
Deleting an account removed its audit entries and session recordings outright.
#1128 relaxed those foreign keys to ON DELETE SET NULL, but deleteUserAndRelatedData
deletes the rows explicitly, so the schema change had no effect on the path that
actually matters. Both repositories gain anonymizeByUserId, which nulls the
reference and leaves the row; username is already denormalised on both tables, so
entries stay attributable to whoever acted.
Separately, the log pruned itself at a hard-coded 10000 rows with no signal.
Entries well inside any retention window were discarded and nothing recorded it.
Retention is now configurable by age via AUDIT_LOG_RETENTION_DAYS, the row cap
via AUDIT_LOG_MAX_ENTRIES, and the two are reported differently: expiring an old
entry is routine and logged at info, while hitting the cap means the ceiling is
too low for how much this install audits and is logged at warn, naming the range
discarded and how to stop it.
* fix: load notification channels on mount in AlertsPanel
Channels only loaded when the Channels tab was visited, so opening Edit
Alert Rule before ever switching to that tab showed the channel picker
as empty even when channels existed.
(cherry picked from commit caed913ee91990a853f5a048849c67ed3f7c329e)
* fix: register login-alert route before auth middleware
Global JWT auth middleware ran before this internal
service-to-service route, rejecting it with 401 before
its own IP+token check ever ran — silently dropped every
SSH-login alert. Also surface non-OK responses instead of
swallowing them.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* test: add coverage for alert-notification fixes
Channel-load-on-mount, login-alert non-OK handling, and a
source-order guard for the route/auth-middleware regression.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* format AlertsPanel test with prettier
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Co-authored-by: ZacharyZcR <zacharyzcr1984@gmail.com>
Node's happy-eyeballs autoSelectFamily calls custom dns lookup functions
with all:true and expects the full address array back. Always replying
with a single (address, family) pair corrupted net's internal state,
surfacing as "Invalid IP address: undefined" instead of a real connect
error, breaking outbound notification delivery (webhook/ntfy).
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Co-authored-by: ZacharyZcR <zacharyzcr1984@gmail.com>
Only SSH terminal sessions were audited. Opening a file manager session, an RDP,
VNC or Telnet desktop, a Docker session or an SSH tunnel left no audit entry at
all — which covers most of the ways data leaves a host or a foothold is
established.
Each of those four now writes an entry when the session is established, matching
the existing ssh_connect: who, which host, from what address, and for tunnels
the endpoint and local port being forwarded.
Audit writes are fire-and-forget so they cannot delay or fail the connection,
consistent with logAudit already swallowing its own errors.
getAuditUsername was defined identically in two route files and is needed in
four more, so it moves next to logAudit.
audit_logs and session_recordings both referenced users with ON DELETE CASCADE,
so removing an account erased everything it had ever done. An audit trail that
disappears with the account it recorded cannot answer the question it exists for,
and a recording is evidence about a host as much as about a person.
Both foreign keys become ON DELETE SET NULL. audit_logs already denormalises
username, so an entry still names who acted once the reference is gone.
session_recordings did not, so the column is added and backfilled first —
otherwise relaxing the constraint would only trade deleted evidence for
anonymous evidence.
SQLite cannot alter a foreign key in place, so existing databases are migrated
by copy-and-swap, guarded by a PRAGMA check that makes it idempotent. Fresh
databases are created in the target shape and skip it. Recordings still cascade
from their host.
DatabaseContext handed every repository a raw better-sqlite3 handle alongside
drizzle, and three of them used it for retention queries built on datetime('now',
?) — a SQLite-only function. That handle is the one thing standing between the
repository layer and a second engine.
Drop it. The two time-based prunes compute their cutoff in JS against the
CURRENT_TIMESTAMP text format, which every engine writes the same way and which
compares correctly as a string; the health-history prune becomes a select of the
rows to keep followed by a NOT IN delete. All three turn async, so their two
callers await them.
Name the dialect rather than repeating a string literal, so adding an engine is
one edit instead of a search.
Tests built their schema through context.sqlite?.exec(). Optional chaining meant
removing the field type-checked cleanly and then silently created no tables, so
the fixture now owns exec() and a raw handle for direct assertions — schema setup
belongs to the test harness, not to the interface repositories consume.
No behaviour change, and no Postgres yet: this only removes the coupling that
would have to be undone first.
#1090 bumped typescript to 7.0.2. typescript-eslint declares
`typescript: >=4.8.4 <6.1.0`, and TypeScript 7 removed `ts.Extension`, which
@typescript-eslint/typescript-estree dereferences at import time:
node_modules/@typescript-eslint/typescript-estree/dist/create-program/shared.js:59
ts.Extension.Cjs,
TypeError: Cannot read properties of undefined (reading 'Cjs')
ESLint hits that while loading eslint.config.mjs, so `npm run lint` fails before
linting anything. Node reports it as ERR_INTERNAL_ASSERTION, which hides the
cause. Every open PR fails this check, not just new ones.
Even the latest typescript-eslint prerelease still caps at <6.1.0, so there is
nothing to upgrade to yet. Pin back to ~6.0.3 and tell dependabot to hold major
typescript bumps until the ecosystem catches up.
Also fixes biome.json pointing vcs.defaultBranch at dev-2.5.0, a branch that no
longer exists.
* fix: show remote sync account identity
* cover getRemoteSyncUserInfo and make its null contract hold
Nothing asserted the renderer-side gate: browser builds must not reach for the
IPC bridge, and a missing bridge, an unconfigured server, an expired JWT or a
failed channel all have to degrade to no identity rather than throw.
Writing that turned up a mismatch — with no preload bridge the optional chain
resolved to undefined while the signature promises null. The only caller uses
??, so nothing is broken today, but the type was not telling the truth.
The main-process half (token expiry, /users/me, the roles fallback) stays
uncovered: remote-sync.cjs requires electron at load, so exercising it means
stubbing safeStorage and the filesystem, which is a bigger change than this PR
warrants.
* fix: clarify desktop local profile
* cover the AccordionSection hidden branch
The desktop build hides the Security section because the embedded profile signs
in automatically and has no login password, so the controls there would imply a
protection that does not exist. Nothing asserted that hidden actually keeps the
children out of the DOM rather than merely collapsing them.
Export the component and cover both states, including that an expanded hidden
section still renders nothing.
Each patch bails out with a console.log and process.exit(0) when its anchor
string is missing. The write-back happens at the end of the file, so an upstream
release that moves any one anchor drops every patch, exits successfully, and
leaves postinstall reporting nothing wrong. Termix then builds and starts
normally and drops VNC/RDP sessions at runtime — with no signal pointing at the
patch.
Every patch here is required for correctness: protocol negotiation, the guacd
1.6.0 name handshake, dynamic argument answering, UTF-8 tokens, read-only joins.
A missing anchor means the patch no longer applies, so exit non-zero and say
which one and what to do.
Unchanged: a missing guacamole-lite still skips quietly, and an already-patched
tree still exits 0.
The test mocks db, logger and AuthManager, but the route module also calls
PermissionManager.getInstance() at import time and pulls in the repository
factory, which loads the drizzle schema and the better-sqlite3 native binding.
Importing that costs seconds when the full suite runs its projects
concurrently, and the test times out at 5s. On its own it passes, so it read as
flaky rather than as a missing mock.
Mock both. None of it is under test here, and the file now imports in
milliseconds regardless of load.
A control string (OSC/DCS/APC/PM) carries text that must never be displayed —
an OSC 0 title holds the user, host and path, and PROMPT_COMMAND emits one on
every prompt. Its opener and its terminator routinely land in different
websocket frames, and the continuation frame contains no escape byte at all, so
every guard in the highlighter misses it: TUI_SEQUENCE, CONTROL_STRING_SEQUENCE
and hasIncompleteAnsiSequence all only look at one chunk.
Highlighting that continuation injects an SGR sequence into the middle of the
open string, which aborts it early in xterm.js and prints the remainder as
ordinary text — the stray ~/path glued to the prompt, and the cursor arithmetic
drift behind the duplicate prompts and Ctrl+R corruption.
Track the state across chunks the way alternate-screen mode already is, and skip
any chunk that starts or ends inside a control string. A trailing lone ESC counts
as inside, since its meaning only arrives with the next chunk.
ClosesTermix-SSH/Support#1025
A tap discovers casks in a top-level Casks/ directory. The cask sat in
packaging/Casks/, so tapping the repository succeeded and every subsequent
brew install --cask termix reported that no cask with that name exists.
Move it and repoint the five workflow references. The release job still rewrites
the version and checksum in place, and the electron job still copies it into the
generated and submission trees.
ClosesTermix-SSH/Support#1044
resolveConnectionOrigin() pins RDP/VNC/Telnet to "remote" because the embedded
desktop backend does not bundle guacd, and the Guacamole websocket already
follows that. The status check and both token calls did not: they use the shared
authApi, which in Electron is hard-coded to the embedded backend.
So the desktop app asked the backend without guacd whether guacd was available,
got "disconnected", and refused to connect — while the connected server it would
actually have used reports it as connected and serves the same host fine from the
web client.
Send those three calls through a remote-origin instance in Electron, alongside
the existing file-manager, tunnel and stats ones.
ClosesTermix-SSH/Support#1043
The session recording section offers a recording path, a filename template and
four content toggles, but the backend overwrote five of the six on every
connection. A host could set none of them and get no indication why.
Location and filename genuinely are not the host's to choose — recordings are
indexed by them for playback and the backend refuses to read outside its
recordings directory — so drop those two inputs rather than keep pretending they
apply. The content flags are a host-level decision, so default them instead of
forcing them.
That still leaves the reported case, where guacd writes the file somewhere the
backend cannot see it. The warning now reports both paths and names the two env
vars that align them, which is otherwise guesswork for a split-container setup.
ClosesTermix-SSH/Support#1041
The host editor stores width and height in guacamoleConfig, and the backend
passes them to guacd in the connection token. The renderer then appends its own
width and height query parameters measured from the container, which take
precedence, so a configured resolution never reached the session — only dpi did,
because that was the one display field GuacamoleApp read back.
Pass the configured width and height alongside dpi, and skip the container-driven
sendSize on connect and on resize when a resolution is pinned. rescaleDisplay
still fits the fixed display into the available space.
ClosesTermix-SSH/Support#1039
Shared hosts hide their edit, share and delete actions based on the recipient's
permission level, but the sidebar row stays draggable regardless. Dropping one on
a folder issues a bulk folder update the server rejects, so a recipient without
edit rights gets a failure toast for an action the UI offered them.
Gate draggable on canEditHost, and skip hosts the recipient cannot edit in the
move handler so a mixed selection moves what it can instead of failing whole.
ClosesTermix-SSH/Support#1011
When the data directory holds no database, startup treats it as a first run and
silently creates an empty one. A deployment that loses DATA_DIR — an .env file
the service no longer loads, a volume that did not mount — lands in exactly that
state, so the user is asked to register an admin account again while the real
database sits untouched one directory over. It is indistinguishable from the
upgrade having deleted everything.
Check the known data locations before creating a new database and refuse to
start when one of them already holds a database, naming both directories.
ALLOW_EMPTY_DATA_DIR=true starts anyway for anyone deliberately starting over.
This matches how a failed decryption already behaves: it throws rather than
falling back to an empty database.
ClosesTermix-SSH/Support#1006
verifyOIDCToken passed the raw id_token straight to jose's jwtVerify, which
throws JWSInvalid when the token is not a three-segment compact JWS. Authentik
issues an encrypted JWE id_token when the provider has an encryption key set,
so the callback threw and every OIDC login failed with 'Invalid Compact JWS'.
2.5.0 hid this behind a catch-all that decoded the unverified payload; removing
that fallback fixed the trust bug but turned the pre-existing verification
failure into a hard login failure.
Check the segment count before verifying and raise a distinct
OIDCTokenFormatError, which the callback treats as 'no usable claims here' and
falls through to the userinfo endpoint. Signature and claim failures still
reject the login.
FixesTermix-SSH/Support#1016FixesTermix-SSH/Support#1018
ZacharyZcR