The web interface could create a vault and never remove one, so the only
way to undo a mistake was the CLI. There is now a Delete link on every
vault you administer, leading to a page that says what is about to go —
owner, versions stored, grants pointing at it — and asks for your
account password before it goes.
The gate is re-authentication, not a confirmation checkbox. A session
token is enough to read and write a vault; it should not be enough to
erase every version of one, because a session is what an unlocked
laptop leaves lying around. The one-time code is required too when the
account has one enrolled.
It lives in confirmDeletion, so the rule holds wherever deletion is
reachable: DELETE /api/v1/databases/{id} now takes {password, totp?}
and answers 401 without it, and keeporg db -delete prompts. That is a
breaking change to the endpoint, which is why the server goes to 0.7.0
and the desktop clients to 0.4.0.
An account that signs in only through an identity provider has no
password to type. Rather than lock those deployments out of deleting
anything, they confirm by typing the vault's exact name — the weaker
gate, and deliberately the fallback.
Deleting takes the record, its grants, every stored version and the key
a merge vault left with the server, and tells the watchers over SSE.
Copies already in people's mirror directories stay where they are: the
server stops tracking a vault, it does not reach into anybody's home
directory.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
location / covered /invite/{token}, /oidc/callback, /device and
/healthz along with the web interface, so Vouch stood in front of all
of them. An invite link was the worst of it: it goes to somebody who has
no account anywhere yet, and Vouch bounced them to a login they cannot
pass. The OIDC callback was fragile for a different reason — the
authorization code it carries is single-use and short-lived, and a
second login round trip is a poor place to spend it.
Those four paths now proxy straight through. KeepOrg authenticates each
of them itself: an invite is a capability token, the callback is checked
against its state, a device code expires in ten minutes, and /healthz is
unauthenticated by design. /web keeps Vouch in front of it, and /api was
already exempt.
The proxy settings move to the server block, where every location
inherits them, and the upstream becomes a named block. location / has to
repeat the headers in full: a location that sets one proxy_set_header
replaces the inherited set rather than adding to it.
Checked with nginx -t and against a running keeporgd with a 401 standing
in for Vouch: /, /web and /web/* redirect to auth.tiredsysadmin.cc,
while /healthz, /invite/, /device and /oidc/* serve KeepOrg's own pages
and /api returns KeepOrg's 401.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Running the server meant running the API by hand: invites, accounts,
groups and grants existed only as endpoints and CLI commands, so the
browser could create a vault and nothing else.
/web/groups walks the group tree and opens onto one group's members,
roles and vault access, including the grants it inherits from above and
where they come from. /web/users lists the accounts you may see, and
lets a server administrator disable, enable, clear a lost second
factor, promote, demote or delete one. /web/invites shows the
outstanding links with their use counts, issues new ones and revokes
them.
Every administrative change is now a function of the state and the
caller in admin.go, which both the JSON handlers and the forms call. The
alternative was a second copy of the permission rules behind the pages,
which would have drifted on the first change. The same move gave the API
the two verbs it was missing: DELETE /api/v1/invites/{token} and
PATCH /api/v1/users/{id}.
Account changes are server-administrator only. A group administrator
decides who reaches which vault; disabling an account or clearing
somebody's second factor reaches past the group they share. Disable and
delete revoke that account's sessions, and the three changes that would
lock the caller out are refused on their own account.
Pages go through webPage and forms through webForm, so the session and
CSRF checks cannot be forgotten on a new form. No page carries any
JavaScript.
Server 0.6.0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two things in the deploy did not do what they looked like:
build-server passed --build-arg APP_VERSION, but the Dockerfile declares
ARG VERSION. An unknown build arg is ignored with a warning, so every
deployed image was stamped "dev": `keeporgd -version` and the image's
version label both said so, whatever the tag claimed.
The build pipeline's path filter listed 'VERSION', a file this repository
does not have, and '*.go', which under doublestar matches only top-level
Go files, of which there are none. So a push that changed only Go code or
only the version did not trigger a build. Now versions.mk and '**/*.go'.
Proved: built with --build-arg VERSION=0.5.0 and ran -version in the
container, which now reports 0.5.0.
Why: two questions about operating this turned up two gaps. "Where do I get
the admin password" had no answer once the line scrolled out of the log,
and "how do I set up OIDC" ran straight into a sign-in link that could not
work.
What:
- /oidc/login with no parameters now signs a browser in and ends in a
session cookie. It previously required the user code of a waiting client,
so the link on the web sign-in page always failed. With a code it still
approves a device, and the two paths no longer overlap: an empty user
code matches no pending device.
- POST /api/v1/auth/password, `keeporg passwd`, and a form on /web/vaults.
Changing a password requires the current one, and the one-time code when
a second factor is enrolled, and revokes that account's other sessions.
Until now nothing could change a password, which made the advice printed
at first start impossible to follow.
- keeporgd -reset-admin, for when that first password was not kept: a new
one printed once, any second factor cleared, sessions revoked. It writes
the state document directly and says to stop the server first.
- The client read stdin through a new buffered reader per prompt, so the
first prompt swallowed the rest and anything asking twice could not be
scripted. One shared reader now. A prompt for a password being set no
longer reads $KEEPORG_PASSWORD either; that variable holds the password
you sign in with, and taking it here would have set the new password to
the old one.
- HANDOVER gains a first-sign-in section and an OIDC setup section covering
the redirect URI, the config block, both entry points and how an identity
becomes an account.
Proved: OIDC now has tests, against a stub provider that serves discovery,
authorization, token and userinfo. They cover the web sign-in that was
broken, device approval, linking an existing account by username, the
refusal when auto_create is off, and a forged callback state. Eight more
cover the password change. Checked by hand with the real binaries: the
first-start password logs once and works, -reset-admin replaces it while
the old one stops working, and `keeporg passwd` changes it from a pipe.
Server 0.5.0, desktop clients 0.3.0.
Why: the usual way to hand a secret to a container is to mount it as a file
and name the path in the environment. Compose, swarm and Kubernetes all do
it that way, and it keeps the key out of the process environment and off any
disk the container can write.
What: the master key is now looked for in KEEPORG_MASTER_KEY, then
KEEPORG_MASTER_KEY_FILE, then the configured master_key_file. A secret file
may hold base64, with or without padding or a trailing newline, or exactly
32 raw bytes, because secret stores produce both.
Only the configured path is ever created, which is what keeps a plain
package install working with no setup. The two environment variables name a
key something else owns, so a missing one is reported and vault creation
refused. Minting a replacement there would be worse than failing: it opens
none of the existing vaults, and the mistake would otherwise surface one
failed merge at a time.
The compose example now provisions the key as a real secret rather than an
environment variable, and the Dockerfile and README show the same. Neither
names a version any more, so they stop drifting as the server is bumped.
Proved: unit tests for every source and precedence, all four accepted
encodings, and that a missing secret writes nothing anywhere. In real
containers: a podman secret carrying base64 survives a stop and start with
"1 vault keys available for merging" and merges a stale push afterwards; a
secret holding 32 raw bytes creates a vault; and a secret that failed to
mount is named in the startup log, with vault creation answering with the
three places a key may come from.
Why: the server is a single static binary with no dependencies, so its
container should be nothing but that binary. Shipping one also makes the
master key easy to lose, which needed handling rather than documenting.
What:
- Dockerfile: two stages, runtime is scratch plus the binary and a CA bundle
for reaching an OIDC provider, running as uid 65532. About 9 MB. The build
stage copies only go.mod, cmd/keeporgd and internal, so client or docs
changes do not invalidate it.
- The image has no shell, so HEALTHCHECK runs `keeporgd -healthcheck`, a new
flag that probes /healthz on the configured address and exits non-zero
with a reason. podman needs --format docker to keep the healthcheck at
all, which the Makefile passes when it picks podman.
- New -master-key-file flag, so a mounted secret works without a config file.
- The server now checks at startup whether the master key still opens the
stored vault keys, and says so. Without it, a container restarted without
its key discovers the problem one failed merge at a time.
- make docker, docker-run and docker-save; a compose example with read_only,
cap_drop ALL and no-new-privileges.
- The client and the server now version independently, so their Debian
changelogs are split per package and `make changelog` takes PKG=.
Server 0.3.0; the clients stay at 0.2.0.
Proved: built with podman and run. The container serves /healthz and /web,
its own healthcheck reports healthy, and a vault created through it was
edited from the same base by keepassxc-cli twice, merged by the server into
version 3, and read back with both entries intact. Restarted without the
master key it warns "1 of 1 stored vault keys do not open with this master
key" and answers a merging push with 409 merge_failed, leaving the vault
untouched; restored, it logs "1 vault keys available for merging" and merges
again. The same key supplied through -master-key-file from a mounted file
works identically. Both .deb packages build with their own changelog, and
make check is clean.
Why: the point of a sync server is that two people can work at once. Picking
a winner was all this one could do while it could not read a vault.
What, end to end:
- A vault is created at /web or with `keeporg vault -create`: the server
builds an empty KDBX 4 file, encrypts it with the password from the form,
and stores it as version 1. An existing database can adopt a stored key
with `keeporg db -id <id> -store-password`, which is verified against the
stored file before it is accepted.
- A stale push to a `merge` vault is no longer refused. The server opens the
stored version and the pushed one in memory, folds them together, and
stores the re-encrypted result. Clients push as usual and pull what comes
back. Any failure answers 409 merge_failed and the client keeps its copy
aside, so a failed merge still loses nothing.
- At rest: vault contents stay KDBX-encrypted as always. A merge vault's
password is kept AES-256-GCM-wrapped under a master key that lives outside
the data directory, bound to the vault id, never returned by the API. A
stolen data directory is ciphertext; a stolen data directory plus the
master key is not. The server runs without a master key and simply refuses
to create or merge vaults.
This moves the server inside the trust boundary for vaults it can merge,
which is inherent to merging and is stated plainly in the README, the
handover notes and the web form itself. Key-less strategies are unchanged and
remain the default for a database created without a password.
Proved: two clients editing the same version are merged rather than refused,
the newer edit of one entry wins, a merge failure leaves the stored vault
untouched, `merge` is refused without a stored key, the wrapped key never
appears in state.json or the API, and the web interface rejects a form
without its CSRF token, withholds its cookie over plain http, and only offers
groups the user administers. Checked by hand against the real thing: a vault
created in a browser, edited by keepassxc-cli on two sides from the same
version, merged by the server, and both entries read back out with
keepassxc-cli. make check, make test-race, android lint and both .deb builds
are clean.
Why: picking a winner is what a sync server does when it cannot read the
file. Once it can, the right answer is the union of both edits.
What: kdbx.Merge folds one database into another. Objects are matched by
UUID; the most recently modified version of each wins and the loser is filed
in the entry's history, so nothing is dropped. Deletions apply only when the
deletion is newer than the other side's edit, otherwise the edit resurrects
the object and the stale record is discarded. Groups keep their nesting,
objects move when the other side moved them later, database settings follow
Meta/SettingsChanged, and custom data and icons are unioned by key.
Attachments are remapped from the incoming pool into the target's, since an
entry that kept its old index would silently gain the wrong file.
Proved: twelve tests over independent additions, both conflict directions
with history, deletion, resurrection, nested groups, relocation, attachment
remapping, idempotence, order independence and settings. The last one
marshals a merged database and has keepassxc-cli decrypt an entry that only
ever existed in the incoming copy.
Why: merging two edits of a vault means decrypting it, so the server needs a
real KeePass implementation. Nothing else in this repository can be built on
guesswork about the format.
What: internal/kdbx reads and writes KDBX 4.0/4.1 with AES-256-CBC or
ChaCha20 payloads, AES-KDF or Argon2d/Argon2id derivation, and the ChaCha20
inner stream for protected values. ChaCha20, BLAKE2b and Argon2 are
implemented here, since the standard library has none of them and this
repository carries no third-party dependencies.
The XML payload is kept as a generic node tree rather than Go structs on
purpose: a merge server that dropped elements it did not model would silently
destroy data written by a newer KeePass. Anything unrecognised round-trips
untouched. Unsupported ciphers, KDFs and KDBX 3 files are reported as such
rather than guessed at, so a caller can fall back instead of losing data.
Proved: ChaCha20 against RFC 8439, BLAKE2b against RFC 7693, all three
Argon2 variants against RFC 9106. Round trips preserve unknown elements and
multi-line values; a wrong password, a flipped payload byte and a flipped
header byte are each rejected. Entry passwords appear nowhere in the file,
encrypted or merely compressed. And the part that actually matters:
keepassxc-cli 2.7.10 lists, reads and decrypts the protected values of a
vault this code wrote, including at the production Argon2d cost, and this
code reads back the file KeePassXC saves after editing it.
Why: the module was written but never compiled, so nothing about it was
proved. Building it found a name clash and a theme that would have crashed
every device below Android 10.
What:
- KeepOrgProvider.requireContext -> ctx(): ContentProvider gained
requireContext() in API 30, and Kotlin refused the silent hiding.
- The application theme was @android:style/Theme.DeviceDefault.DayNight,
which is API 29 while minSdk is 26. Replaced with @style/AppTheme, plain
DeviceDefault in values/ and DayNight in values-v29/.
- Backup exclusions: data_extraction_rules.xml and backup_rules.xml exclude
every domain for both cloud backup and device transfer. The session token
authenticates as the user and the mirrored files are the organization's
password databases; neither belongs in a backup.
- User-facing strings moved from code into strings.xml, so the app is
translatable. Layouts stay in code, as before.
- mipmap-anydpi-v26 -> mipmap-anydpi (the qualifier is redundant at minSdk
26), and the redundant activity label is gone.
- Lint is configured to write a text report, and `make android-lint` runs it;
`make android` now fails with a clear message when ANDROID_HOME has no SDK.
Proved: the unsigned release APK builds (656 KB), lint reports "No issues
found", and the packaged DEX contains all 16 classes including every manifest
component. Toolchain used: JDK 21, Gradle 8.10.2, SDK platform 35 and
build-tools 35.0.0; on this arm64 host Google's x86-64 aapt2 runs under
qemu-user-static with amd64 multiarch libraries, which android/README.md now
documents. The app has still not run on a device: there is none here.
Why: the set has to be installable and publishable, not just buildable, and
the next person needs the reasoning, not only the code.
What:
- android/: Kotlin client with no third-party dependencies. A
DocumentsProvider publishes the mirrored databases to every file picker, so
KeePassDX opens them as ordinary documents; closing a file it wrote
triggers a sync. A foreground service holds the event stream open.
Mirror.decide transcribes the Go merge table, so both platforms resolve a
conflict identically.
- Makefile as the single entrypoint, with file targets so only what changed
is rebuilt: host and cross builds, both .deb packages, the macOS DMG with
its /Applications symlink, the Windows zip, the APK under its F-Droid
publishing name, and manual-only android-publish.
- versions.mk holds one version per platform plus the Android versionCode;
`make bump` and `make changelog` move them. Debian changelog is tracked.
- systemd units (hardened system unit for the server, user unit for the
agent), macOS bundle and launch agent, Windows Task Scheduler job.
- docs/HANDOVER.md on architecture and every decision that shaped it,
docs/PROTOCOL.md for the API, README for the quick start.
Proved: make check and make test-race both pass; built both .deb packages and
inspected their contents, built the Windows zip, cross-compiled all five
desktop targets, and exercised make bump. The Android module is written but
not compiled: no JDK or Android SDK in this environment, which HANDOVER.md
states plainly.
Why: a KeePass application must see an ordinary .kdbx file. A mirror
directory of real files gives that on every desktop platform without FUSE,
WinFsp or macFUSE, and without cgo.
What: internal/client holds the API wrapper, an SSE reader, and a mirror
agent whose entire policy is one pure decide() table: adopt, pull, push,
force-push, server-wins, restore or manual. Saves are detected by rescan,
remote changes arrive over the event stream, and every overwrite of local
bytes leaves a conflict copy. cmd/keeporg drives it: login (password, TOTP or
browser/OIDC device flow), run, sync, status, plus the group, grant, database
and invite administration.
Proved: go test ./internal/client covers the decision table and runs a real
server in-process for adopt, follow, server-wins, client-wins, manual,
read-only restore, server-side rename and event-driven update. Also smoke
tested live: built both binaries, bootstrapped a server, logged in, created a
database and adopted a local file.
Why: collaborative KeePass use needs one arbiter of "who wrote last" that
never sees a master password, plus access control that follows an org chart.
What: stdlib-only Go server. Metadata is a single JSON document with one
writer and atomic commit; KDBX payloads are versioned blobs on disk. Access
comes from a group tree where a grant on a group reaches every member below
it. Login is password+TOTP, invite links, or OIDC through a browser-approved
device flow. Pushes carry If-Match and are arbitrated by a per-database merge
strategy (server-wins by default). Updates are pushed to clients over SSE.
Proved: go test ./internal/... covers RFC 6238 vectors, ACL inheritance and
isolation, store rollback, blob pruning, every push verdict, and an
end-to-end HTTP pass over login, conflict, force, read-only grants, invite
redemption, TOTP enrolment, device login and the event stream.
Reactorcoremeltdown
root