A token Spelunk's token endpoint returns must begin with spelunkapp_, the prefix Spelunk reads a bearer
token as an app session's by; anything else is refused as an answer this app does not understand. Neither
the sign-in's client nor the tunnel's follows redirects, so the code, its verifier and the token go to the
address signed in to and nowhere else, and signing out sends the token only to an address that still
normalises to itself, within ten seconds of its own.
Opening the browser and waiting for it now always end in a SpelunkSignInException rather than whatever
the platform threw: a browser that will not open, a phone's browser closed without an answer (the
OidcException AndroidRedirectCallback.Abandon raises, which escaped the Security page and the editor
before), another sign-in taking over the phone's redirect, and the ten-minute timeout are each said. A
refused code says it in Spelunk's own words, and a rate limit says to wait.
Tests: a refused code's detail, a non-app token, a browser that will not open, an abandoned phone sign-in,
the timeout, the user cancelling, a phone redirect sent byte for byte both times, the real loopback
listener on 127.0.0.1 taking one answer and closing, and a sign-out that will not send the token to an
address it could not have come from.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Spelunk API key reaches one organisation, so a server elsewhere - oc-linux in Orange Climate - was
refused before it reached the agent. Each device now signs in to a Spelunk once, under Settings ->
Security -> Spelunk or from a Spelunk host's editor: the OAuth authorization code flow with PKCE for a
native app (RFC 8252), with Spelunk as the authorization server. The browser opens Spelunk's
/api/app-sign-in/authorize with a loopback redirect (the app's own scheme on Android), the person signs
in there however they normally do and allows "DodoSSH on <device>", and the code is redeemed at
/api/app-sign-in/token for an app session token. That token opens terminals in every organisation its
person is in, with the role they have on each server at the time; tunnel tickets send no organisation.
The sign-in stays on the device: sealed in the local cache under the LocalCacheKey (AAD resource type
18, bound to its row), never synced, wiped with the cache, and only ever sent to the Spelunk it came
from. A 401 from Spelunk forgets it; SIGN OUT ends it there and here, and so does signing out of
DodoSSH. Before every Spelunk connection DodoSSH asks its own identity provider whether its sign-in
still stands (a forced refresh, at most once a minute), and an account it refuses is signed out of
DodoSSH and of every Spelunk on the device.
A route names no key any more: Spelunk hosts are written at host schema 11, which an older build shows
read-only and asks to be updated for, and a payload that still names a key credential decodes with the
id skipped. The SpelunkKey item type is gone: SyncEntityType 14, ChangeEntityType 14 and AAD type 17
are retired, the server drops spelunk_key and its change-log rows, and the client cache drops its
cached items, outbox entries and conflicts of that type. A link still never starts a sign-in: from a
Spelunk this device is not signed in to it opens the editor, which offers to.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
SSH.NET 2026.0.0 runs the SFTP handshake synchronously inside ConnectAsync, with an infinite
operation timeout and without reading the cancellation token. A host whose login script prints
anything (an echo in .bashrc) puts text where the first SFTP packet belongs; SSH.NET reads its first
four bytes as a length of about a gigabyte and waits for the rest, so the file browser said
"Connecting…" for ever. The SFTP client now uses the connect timeout as its operation timeout for
the handshake and resets it to infinite once connected, so transfers on slow links are unchanged. A
timeout after authentication becomes a TimeoutException naming the likely cause.
The SSH test server gains a second port whose sessions print a banner before sftp-server, and
SftpSessionTests proves the open fails within a 3 s connect timeout with that message. Without the
fix the test runs into its 30 s guard.
Button.paneicon is a fixed 26x26 but kept Fluent's 8,5,8,6 padding, which left a 10px-wide content
box, so every 15-16px Material icon on the SFTP panes was drawn with its sides cut off. Its padding
is now 0. Rename on the host pane used U+E9AD, an italic T (text formatting); it now uses U+E9A2,
the rename glyph the phone's files screen already uses. The host pane's "Up one directory" drew the
design's arrow_downward; both panes now use arrow_upward.
SSH tests 123/123, layout tests 203/203, app tests 744/749 with the 5 Linux TPM skips.
The phone's Connections screen no longer draws the user@host + password box, and the desktop never
did, so nothing bound VaultViewModel.ManualTarget, ManualPassword, ManualStatus or
ConnectManuallyCommand any more, nor MainWindowViewModel.ConnectToRecentCommand, whose address branch
wrote into that box. They are gone, and so is VaultViewModel.AskAboutHost, whose one caller was
ConnectToRecent. TryParseManualTarget, ManualEndpoint and ConnectToEndpointAsync stay: the prompt, the
desktop's Ctrl+K palette and the phone's hosts search all use them. The reasoning the removed members
carried moved onto ConnectToEndpointAsync and ChoosePromptRecentAsync.
The tests whose subject is still live now go through the prompt instead of the box: a typed address
dialled exactly as typed, port 22 by default, the refusal sentences, the host-key retry replaying the
attempt that raised it, the host-key card staying over the surface it was typed on, the session facts
for a typed password, and no forwards for a typed address. The recent-row cases now drive
ChoosePromptRecentCommand. Two cases existing prompt tests already covered were deleted. Manual check
11.8 now describes recent rows landing on the prompt.
The phone's Connections screen with nothing open had a box that took an address alone, which left a
saved machine several screens away. It now draws the desktop's prompt over the same view-model state:
"connect to:" takes a saved host's name or user@host[:port], and a password line follows only when
one is needed. Names, addresses and refusals are read exactly as on the desktop, and nothing typed
here is saved.
What a phone changes: the hosts matching so far are thumb-sized buttons that take a host outright,
since there is no Tab; CONNECT is always drawn, since the IME's return key may only close the
keyboard; BACK, and the system back gesture, step off the password line rather than leaving the
screen; and the password's label sits on its own line. The keyboard follows the question onto the
password line, but is not raised when the screen opens, where it would cover the recent list.
Recent rows now go onto the prompt too. Covered by a new ShellFlowTests case for a tapped match on
top of the shared prompt cases; manual check 11.14.
WhenTheSessionEnds_ItsForwardsSayTheyWentWithIt failed once under a full run. Ended clears IsLive
and then tells each forward its session is over, in one call. The app's dispatcher never splits that,
but the harness's post runs it on the fake shell's thread, so the test's poll on !IsLive could return
between the two and find the forwards still listening.
The test now also waits until every forward says "Stopped with the session.", the last thing
SessionEnded sets. A 300 ms sleep injected between the two steps fails the old test and passes the new
one; the full App.Tests suite passed three runs in a row.
The terminal surface with no session used to be an empty frame, with the tab row's "+" the only
thing on it that did anything. It now holds a prompt, in the order ssh asks: "connect to:" takes a
saved host's name, matched exactly, with the hosts that match so far listed under it for Tab or a
click, or user@host[:port] for a machine that is not saved. A password line follows only when one
is needed, worded as ssh words it. Escape steps back a line; recent connections are listed below.
It is drawn by Avalonia in the WebView's rectangle, which the occlusion rule allows because the two
are exclusive: the prompt shows only while there are no tabs, the WebView only for a tab with a
session.
Saved hosts connect through a new VaultViewModel.ConnectToHostAsync, which the drawer's own connect
now goes through too, so every refusal is still made once. Each attempt now carries whether it may
act on the drawer's remember-password tick, and the prompt's may not: otherwise a tick left on in
a closed drawer would save a password typed at a prompt that never offered to keep anything.
Covered by nine ShellFlowTests cases and four ConnectPromptTests; manual check 7.10b.
The phone's only unsaved way in was the box on the terminal screen, which is drawn only while nothing
is open, so a second machine could not be reached that way while a first one's session was up. The
search well at the top of HOSTS is always there and the design already calls it "Search or
connect...", so it now reads an address the way the desktop palette does: with an @ in it, a NOT
SAVED card above the list offers a password box and CONNECT.
It connects through VaultViewModel.ConnectToEndpointAsync, the same path as the manual box and the
desktop palette, so nothing is written to the keychain. The search box and the password are cleared
once it dials, the password goes whenever the address changes, and an address the manual parser
refuses gets its sentence under the box.
Covered by six new ShellFlowTests cases; manual check 8.20 covers the card on a device.
The desktop had no way to reach a machine that was not in the keychain; the phone has had a manual
box on its terminal screen all along. The palette's query now reads as an address when it has an @
in it: a "not saved" row offers to dial it, Enter asks for the password in place, and a second Enter
connects. Escape on the password step goes back to the address.
It goes through the phone's own path, now split out as VaultViewModel.ConnectToEndpointAsync, so the
tab, the host-key question and the connection-log entry are the same on both heads and nothing is
written to the keychain. Password only, as on the phone: a key is the reason to save the host.
The typed row takes the highlight unless a saved host dials exactly that address, so typing a new
address cannot land on a saved machine whose address merely starts the same way.
Covered by eleven new QuickConnectTests; manual check 7.10a covers the real dial.
The image job pushed :main and stopped there, so nothing told the Spelunk
stack a new image existed and a commit carrying a migration never reached
the server's schema.
A deploy job now calls the stack's deploy hook with SPELUNK_DEPLOY_TOKEN and
then waits up to ten minutes for /healthz/ready. The new container applies
pending migrations as it starts, and readiness stays unhealthy while any is
pending, so a migration that did not apply fails the run instead of passing
quietly.
The Dockerfile no longer says nothing migrates from inside the container,
which had stopped being true and pointed at the setting that causes this.
The bottom bar is Vaults · Connections · Settings, drawn as Material 3 pills with icons from the embedded
Material font and a count of open sessions on Connections, which replaces the strip of shells above the
bar. Vaults is the open vault: its header carries the name as a switch to the vaults screen, the sync
light and a lock icon, and a Hosts · Keychain · Snippets segmented control sits under it. The keychain
and snippets move there from Settings (IsVaultSurface lights the tab; IsMoreSurface no longer lists
them), and back from either goes to the hosts. Settings becomes an account card and grouped sections:
security, vaults and history, files, this app.
Host rows are cards with a terminal tile and the live dot on its corner. The connect sheet names the
host and its address, labels its box, focuses it as it opens, reveals the password on request, and puts
CANCEL left and CONNECT right; every two-button row on the head now confirms on the right. While a
hosts-screen sheet is up the header and bar are dimmed too and a tap on them lowers it, and the bar
stands down while the software keyboard is up.
The keychain gains a floating + and a sheet of the desktop's five adds (generate a key, paste a key, a
password, a bucket, a tag) over the editors VaultViewModel already had, drawn as a full-screen page with
labels above the boxes; a chosen item gets EDIT and COPY PUBLIC KEY beside MOVE and DELETE. Snippets get
a floating + and cards that say RUNS or INSERTS. Files trade the seven-button block for a selection bar
(save, and a ⋮ sheet with rename, move, permissions and a red delete), folder and file tiles, an upload
button, and header actions for a terminal on the same host, refresh and close; a terminal's + sheet
opens SFTP on that terminal's own host.
The theme takes the audit's visual findings: Fluent's ClipToBounds is cleared on glowing buttons so the
glow is no longer cut to a square (the desktop's Button.accent too), prose is Montserrat with mono kept
for addresses and code, disabled secondary buttons are quieter than enabled ones, Fluent's dropdown,
checkbox, switch, focus and selection brushes come from the palette, touch no longer leaves hover fills
behind, and lists, nav entries and icon buttons carry names for TalkBack.
Builds clean; checked on the emulator end to end: unlock, the three vault views, adding and deleting a
password, connecting through the new sheet, files on the terminal's host with the selection bar and
its sheet. App tests 727 of 733 pass with 5 skipped; the one failure, TheCountsAPhoneShows_
FollowEveryForwardsState, passes three runs of three alone and is untouched here. Layout tests pass.
Moving vpk to 1.2.161 was not enough: it got further and failed the same way, one frame later.
Both versions use a generated Gitea SDK that gets null back from 'create release' on the forge's
28.0.0, although the release is created, so the step threw before uploading anything and left an
empty draft behind on every push since 2 October.
The upload is now the android job's curl pattern. The release is created as a draft prerelease,
the three packages go up, then releases.win-nightly.json, which vpk pack already wrote and which is
exactly the index vpk upload would have generated, and only then is it published. The read-back
check, the RPM and the Linux manifest after it are unchanged.
TheChoiceIsMadeOnce_OffTheCallersThread failed on CI with the code doing the right thing. It
compared the probe's managed thread id with the test's, but the test runs on a pool thread, and that
same thread is free to pick up the queued probe as soon as the test awaits.
The probe now waits on a release the test only gives once construction has returned. A probe run on
the constructing thread could never be released, and the test fails after ten seconds instead of
passing by luck. Twenty runs pass; an inline probe fails it.
The desktop nightly has failed on every push to main since 2 October, at 'publish the nightly
release'. vpk 1.2.0 creates the draft release and then reads release.Assets from the forge's
answer, which the upgraded server (28.0.0) no longer fills in for a new release, so it threw a
NullReferenceException at GiteaRepository.cs:136. The step deletes the previous nightly-desktop
release before it uploads, so desktop clients have had no nightly to update from since.
1.2.161 reads the assets null-safely. The library moves with the tool, since Velopack expects the
two to match. Every vpk option the workflow and the release scripts pass is present in 1.2.161;
the only renamed ones are two MSI options nothing here uses.
The key bar above the software keyboard is now the user's own: which keys, in what order, up to sixteen
in two rows that never scroll. Ctrl and Alt latch across the bar and the phone's keyboard, and Ctrl or Alt
with a cursor key sends the modified CSI form. Seven predefined groups (essentials, cursor, control,
symbols, tmux, vim, function) and keys of one's own, written in \e / ^C notation, are arranged on a
customise sheet that saves as it goes.
A grid button swaps the software keyboard for a quick-access keyboard of the same height, with KEYS,
SNIPPETS, HISTORY and APPEARANCE tabs; it reopens to the tab last used. Snippets go through the shell's
existing InsertSnippet. History keeps a typed command only once the remote has echoed it, so nothing
typed at a password prompt is ever recorded; it is memory-only and LOCK clears it. Each session pump now
keeps its last kilobyte of output and the alternate-screen state for that test (OutputTail).
Long-press and drag on the terminal sends arrow keys, as a joystick with three speeds by distance;
scrolling is unchanged and there is no mode switch. It takes the touch before Chromium's own long press,
so a held finger no longer pastes; paste is on the bar by default instead. A switch in APPEARANCE turns
the gesture off. The text-size buttons moved from the bar to APPEARANCE.
On wide screens the same tabs open as a side panel beside the terminal, toggled from the top bar, with
the bar's own keys repeated at the top of KEYS while the bar stands down.
The model, gesture arithmetic, echo reading and history are in DodoSSH.Client.Shell and covered by
TerminalKeyboardTests and OutputTailTests. Builds clean; not yet exercised on a device, because the
emulator's install is locked behind a passphrase that was not recorded.
REPORT.md is the full-app sweep and what became of each finding. ANDROID-DESIGN.md is a walk through
every phone screen on the emulator, from a cleared install to SFTP, with findings ranked by severity
and a screenshot for each. The home-folder listing in the sweep's transfers screenshot is blurred.
The server now forgets a deletion past its retention and refuses cursors
from before it. The client then reads everything again, but everything
again says nothing about what went. So a machine away for longer than the
retention kept those items for good, and its next edit of one put it back
on the server for everyone.
A read from the beginning now marks each item it reads, keeping the mark in
the sync state because the read can span passes. When it finishes, it
removes the items of the synced types it never marked. Two kinds are kept:
items newer than the read's bound, which this machine pushed during it, and
items with a change still waiting, which are the user's work. For those,
the push finds the server without the item and sends it as a create, as
before, but now records a RecreatedAfterCollection conflict and counts it,
so the sync message says so instead of undoing someone's deletion silently.
The bound is the highest sequence the cache held when the read began, or
anything higher it has read since. Taking only what the read returned would
have left the last item of a vault in place for ever.
A cursor behind the horizon was refused on position alone. But the read a
refused client starts from the beginning is itself behind the horizon on
every page until it passes the collected deletion, since older untouched
items come first. So its second page was refused, it started over, and it
never finished, on any vault larger than one page.
The cursor now records the horizon read before the page that produced it,
and only a cursor issued under an older horizon is refused. A cursor issued
under no horizon keeps the v1 format, so a server from before this still
reads it.
laurensandClaude Opus 5.5
job
jaap-jan