Public Access

Files

T

jaap-jan d07b336868 Free the terminal from the Hosts screen, and fill the room it left

The WebView sat inside the Hosts grid, so navigating to Files or the keychain
hid every open terminal and the strip that named them. A connection you had
opened was invisible from four of the five screens. The window now has two
surfaces rather than one: a nav rail that says which page you are on, and a
terminal strip that is always there and switches the whole content area to a
shell. Screen keeps meaning "which page" and never becomes a sixth kind of
page, which is why this is two properties instead of one enum with a terminal
member in it.

Every screen lives inside one wrapper panel that collapses when a terminal is
showing. That is not tidiness — the WebView hosts a Win32 child window that
composites above everything Avalonia draws, so a screen left visible over its
rectangle is a screen sliced in half, and this window has shipped that defect
once already. One decision point, IsTerminalShowing, and a nested panel rather
than five compound bindings nobody would remember to extend.

The focus choreography is the part no test in this repo can see. Every reveal
path now focuses in the same turn the WebView appeared, so all three of them
post at DispatcherPriority.Loaded and let the native control re-push its bounds
first. Going the other way had a real bug: the screen-changed branch called a
bare Focus() where it had to release the keyboard from the native child, so
switching from a terminal to Files silently ate the first keystrokes. Rare
before this commit and the primary gesture after it.

The tab strip grew a cross inside each tab, a plus that opens the quick-connect
palette, and middle-click close. Nested buttons are correct here: Avalonia
handles a left press on the cross and deliberately does not handle other
buttons, which is exactly what lets middle-click bubble up from the cross as
well as the tab. The test is PointerUpdateKind rather than
IsMiddleButtonPressed, because the latter reports button state and is also true
for a left press made while the middle button happens to be held. The handler
is on the tab and not the strip, so the background closes nothing by
construction. Plus opens the palette rather than a flyout, since a menu
dropping into the WebView's rectangle may or may not composite above a child
HWND and this repo does not make rendering claims it has not photographed.

Everything a user reads now says keychain. The wire, the database and the
cryptographic spec still say vault, deliberately: renaming those is a migration
and a protocol change for a word. That split is written down rather than left
to be rediscovered as an inconsistency.

Four things that were squeezed into the keychain's category rail, or into
nothing at all, now have screens. Pinned host keys get one, with fingerprints
never truncated and a filter that matches them, because comparing what you have
against what the operator published is the whole workflow; the approved date is
read out of the item's UUIDv7 rather than added as a column, and says so, since
it means first approval and not last use. Keys can be generated in the client,
which needed the openssh-key-v1 container written by hand — there is no BCL or
NSec helper, and the PKCS#8 route is unverified in the SSH library this uses.
The armour carries no passphrase: encrypting it needs bcrypt_pbkdf, which is
Blowfish with a swizzle, in a project whose crypto is otherwise entirely
libsodium, for a protection the key's own remarks argue is redundant inside a
vault. Generation fills the existing editor and stops, so SAVE stays the one
thing that writes. ~/.ssh/config can be imported behind a preview that is
ticked per row and writes nothing until the button; IdentityFile records the
path and imports the key material only on an explicit opt-in, because reading
somebody's private key into a vault is precisely the act this product exists to
make deliberate. Match blocks and ProxyJump are reported rather than obeyed —
one cannot be evaluated statically and the other has nothing behind it to route
with, and a preview that implied otherwise would be worse than one that admits
it.

Files can be dragged in all four directions that are honestly available. Remote
to Explorer does not ship and is not pretended to: the shell wants the bytes
during the drop, which needs a virtual file and a native COM data object,
outside what Avalonia offers. Note for the next person that Avalonia 12
replaced the drag model outright — DataObject and DataFormats are no-op stubs
and IDataObject is not in the reference assembly, so every tutorial written for
11 does not compile here.

Hosts can be grouped, flat and never nested. A parent id merged as a scalar
lets two offline clients each re-parent A under B and B under A, producing a
cycle inside an encrypted payload that no server can police and every reader
would have to detect for ever. Membership lives in that payload rather than in
the one plaintext concession ADR 0001 allows, whose test is that the relay
cannot function without it — nothing on the server reads a group, so what
plaintext would hand over is a clustering of the estate for nothing. The
plaintext column reserved for it is dropped, provably always null, and the
server now refuses a client that sends one; it was never populated, was copied
on apply, and was not cleared on delete, so a group id would have outlived the
host it described.

Snippets insert through xterm rather than through the pump, because xterm is
the only thing that knows whether the remote has bracketed paste on, and that
is what makes a shell treat embedded newlines as text instead of as execute.
The host process moves opaque bytes and never parses output, so it would have
to guess, and guessing wrong runs every line. Running is off by default and the
copy says the text goes into whatever is there — the terminal has no notion of
being at a prompt, and may be in vi or at a password prompt with echo off, so
the Enter the user presses themselves is the entire safety property.

Connections and keychain changes are recorded as synced encrypted items, which
is what makes them auditable by a team later and costs the server knowledge of
connection rate and timing from row counts alone. ADR 0001 already concedes it
cannot hide that class of metadata; the trade is now written into it rather
than left implicit. A connection entry is written once, at close, which is what
makes a synced log tractable: nothing to merge, one outbox row, no chance of
colliding with itself. Live sessions come from memory, not from the log. The
write is void by contract and posts to a bounded channel, because putting an
encrypt-and-write on the teardown path of every session is how closing the
application comes to take four seconds. A ticket opened before a lock still
closes afterwards, since a shell outlives the vault. The activity log hooks the
one generic repository every kind writes through, so it cannot miss a caller —
which is also why the log kinds themselves declare they are not audited, or the
first entry would write an entry about writing an entry. It records the names
of the fields that changed and never their values; a log with an old password
in it would be a plaintext credential store with no vault around it. Retention
is 90 days or 5,000 entries, whichever bites first, pruned on the sync loop
rather than on a second timer.

That log traffic then broke the status line, which is worth recording because
the fix is a shape and not a patch: background sync counted its own log rows as
pushed items, so the quiet rule stopped being quiet and every action's message
was overwritten a second later by a sync report. The report now separates log
rows from user items and the rule reads the latter.

S3 buckets appear as a remote in the file browser, behind the same interface an
SFTP session implements, so the queue and both panes did not have to learn what
they are talking to. Uploads go through a pipe, because the queue wants to
write and the SDK wants to read; memory is then bounded by the part size
instead of buffering a file to disk twice.

Finally, the Windows device key store moved out of the session project, which
was the one thing keeping it from being portable — everything else in it is
platform-neutral, and a Windows CNG dependency in the middle of the vault code
meant a second head could not reference it without dragging Windows along. The
seam that made the move free was already there. docs/android-port.md is the
audit behind that: what ports, what does not, in order of cost, the four
decisions taken, and an inventory of every screen and state the interface has
to carry, written so a design can be made from it directly.

dotnet build, dotnet test and dotnet format --verify-no-changes are all clean:
1240 tests at zero warnings, including the end-to-end suite against real
containers. The manual checks that headless Avalonia cannot make — the drag
from Explorer, a generated key against a real host, twelve tabs at the minimum
window width — are listed in docs/manual-checks.md and are still outstanding.

2026-07-31 20:30:05 +02:00

7.6 KiB

Raw Permalink Blame History

ADR 0001 — End-to-end encrypted vault and its trust model

Status: accepted
Date: 2026-07-28

Context

DodoSSH stores SSH credentials — passwords, private keys, key passphrases — on a server so they can sync across a user's devices and be shared with teammates. Authentication is OIDC against a provider the operator chooses.

The product is self-hosted. Its buyers are teams who currently refuse to put infrastructure credentials into a SaaS vault. "Trust us with your production keys" is exactly the promise we cannot make.

Decision

The vault is end-to-end encrypted. The server stores ciphertext and never holds a key.

A vault passphrase separate from OIDC. OIDC authenticates but yields no secret we can derive a key from, and SSO compromise must not equal vault compromise. The passphrase goes through Argon2id (m=256 MiB, t=4, p=1) to a master key that never leaves RAM.
A per-user keypair, wrapped many ways. The master key wraps a ~200-byte UserSecretBundle (X25519 + Ed25519 private keys). The same bundle is stored under several independent wraps: passphrase, one per enrolled device, recovery code, and optionally escrow. A passphrase change therefore re-wraps 200 bytes and updates one row — no re-encryption of vault data and no coordination with other members.
Vault keys, then per-item data keys. A vault key is sealed to each member's X25519 key; each item has its own data key wrapped under the vault key. Rotating a vault key re-wraps N × 32-byte data keys and never touches content blobs.
AAD bound to row identity. Every ciphertext's AAD is recomputed from the row's plaintext columns rather than stored, over purpose, resourceType, resourceId, keyId, keyGeneration, itemVersion and schemaVersion.

The normative byte encoding is docs/crypto.md §4. It is fixed-width binary rather than delimited string concatenation, so that no field value can forge a field boundary.
Layered public-key trust — see Consequences.

Consequences

What this buys

The AAD binding is the most valuable structural property here, and it is not something ACLs can provide. A malicious server cannot paste credential A's ciphertext onto host B, cannot roll a row back to an earlier key generation, and cannot replay a revoked grant: each of those changes the AAD and fails the authentication tag on the client.

A stolen database dump, a rogue administrator, a TLS-terminating proxy and the relay operator all see ciphertext only. OIDC account takeover alone yields nothing readable.

What it costs, stated plainly

Revocation is not retroactive and cannot be. A removed member keeps whatever they already downloaded, along with the cached keys. Rotating the vault key protects only items written after the rotation. The only real remediation is rotating the SSH credentials themselves, so offboarding is built around a credential-rotation checklist rather than a "revoke access" button that implies more than it delivers.
Connect cannot be a security boundary. SSH terminates on the client, so opening a session requires the credential's plaintext on that machine. "May connect but may not view the key" is unenforceable in this architecture. The flag exists as a UI hint and must never be documented as access control.
Forgotten passphrase with no recovery code and no enrolled device means permanent loss of personal vault contents. Team vault contents survive, because a remaining member with Share can re-wrap. That asymmetry is a feature: it makes team vaults the right default even for a team of one plus a backup admin.
Public-key distribution is the real security boundary. Every guarantee is downstream of "the key I wrapped to is really Alice's". Four layers, deployed together: an IdP-signed key binding (the enrollment statement's hash is the nonce in an ID token, verified against JWKS fetched directly from the IdP and not proxied through us); TOFU fingerprint pinning with blocking warnings; an append-only key log whose head is embedded in every signed grant, so a forked view must stay consistent forever to go unnoticed; and safety numbers for out-of-band verification.

The residual is honest and must stay in the docs: this makes the IdP a key-distribution trust root, and in a self-hosted deployment the person running Keycloak is frequently the person running DodoSSH. It raises the bar from "one compromised service" to "one compromised service plus a detectable artefact in the key log" — not to zero.
No server-side session recording is possible in relay mode, since the relay forwards only SSH ciphertext. See ADR 0004.
Metadata leaks. The server sees item counts, sizes, timestamps, access patterns and the complete sharing graph regardless of settings. Host addresses are plaintext when relay is enabled for that host; see ADR 0004 for why that is a security requirement rather than a convenience.
The connection and activity logs widen that leak, deliberately. They are ordinary vault items — every field sealed, no plaintext column of any kind, not even a timestamp — but there is one row per connection and one per keychain edit, and rows have updated_at. So the operator can read a user's connection rate and timing off the change log without decrypting anything: how many machines somebody touched this morning, and at what hour they stopped. That is a real increase over what item counts alone gave away.

It is the price of the logs being auditable at all. Kept on the machine that produced them they cannot be read by an administrator, cannot survive a reinstall, and cannot be checked against anything — which makes them a diagnostic rather than an audit trail, and the point of them is the audit trail once shared vaults land. Retention bounds the exposure rather than removing it: ninety days or five thousand entries per kind, whichever bites first.

Two things keep it as narrow as it can be. The payload records the host's label and the address as dialled but not the SSH username — "who in this organisation opened a shell" is the audit question, and "which account they logged in as" is a detail of the host, whose own logs already have it. And the activity log records the names of the fields that changed and never their values — the same rule ADR 0006 imposes on the server's own audit_event.detail, arrived at independently on the other side of the encryption boundary.
Supply chain becomes the largest practical hole. An operator who wants the secrets attacks the client, not the crypto. Release signing with a key not held by the server, and eventually reproducible builds, matter more here than in a conventional product.

Rejected

Server-side envelope encryption (KMS-held master key). Far simpler and it would permit a recording bastion, but a server compromise or a rogue admin exposes every credential. That is the exact promise the product exists to avoid making.
Admin escrow of user identity keys. Turns every operator into a silent global reader and destroys the property being sold. Non-negotiable. Team-scoped break-glass escrow with Shamir M-of-N is a separate, opt-in, clearly-labelled M5 feature.
Storing an Argon2id(passphrase) verifier server-side so the server can pre-validate. It creates an offline-crackable verifier on the very server being defended against, for no gain: the AEAD tag on the bundle wrap already proves the passphrase.

7.6 KiB Raw Permalink Blame History Unescape Escape