224 Commits

This Branch

This Branch

All Branches

Author	SHA1	Message	Date
jaap-jan	6d6edb02c1	Keep an open editor's pickers in step with the vault ... The host editor's four pickers were snapshots taken when it opened, and the comment on EditorAuthenticationChoices said why: a picker whose contents move under somebody halfway through a form is worse than a list a minute stale, and only one editor could be open at a time anyway, so the only way to add a key was to close this one. The second half of that stopped being true when AHostEditorIsInTheWay was split from AVaultEditorIsInTheWay. The host editor is the Hosts screen's business and the keychain's editors are the Vault screen's; neither refuses the other now, which was the right split — it stopped three quarters of a screen going inert over an editor the user was not looking at — but it left the assumption those snapshots rested on false and nothing to notice. So the ordinary way of using the feature was the broken one. Somebody starts editing a host, finds there is no key to bind it to, goes to KEYS, makes one, and comes back to a picker that does not have it — with the fix being to throw the form away and start again. The same for a password, a tag, a group, and for a whole vault made on the Teams screen because the host being typed belongs to the team rather than to the person typing it: the vault they had just made for it was the one place they could not file it. RefreshOpenEditors refills whichever editor is open, and it hangs off ReloadAsync rather than off the twenty-odd commands that write to the vault. That is the choice worth stating, because it is what makes a sync count as well as a save: a key pulled from another machine reaches the open editor by the same path a key typed here does, and a place that wrote to the vault without refreshing the editor would be a bug nobody would find for months. What the old comment was protecting against is real, so every picker is put back onto what it was already showing, by id, and not one typed field is touched. An editor that reset its own bindings because a background sync landed would be a worse bug than the stale list this fixes — it would rebind a host as a side effect of somebody else's work. The placeholder entries go back too, which is the case 3.4 measures: a group deleted on another machine mid-edit still cannot unfile the host when the form is saved. The group editor gets the same treatment for the same reasons; it shares the drawer, and its default binding is lent to every host under it. The snippet editor's vault picker was the same copy of the same list and went stale the same way. It watches TargetVaults rather than the reload, because that screen has always been a wrapper over the vault's collections and has no reload of its own to hang off — which is how it already follows Snippets. The move panels are deliberately left alone. A vault arriving from a sync while one is open still will not appear in it, but a move panel is opened by the act that fills it and its picker resets its selection to the first entry on every rebuild, so refreshing it would move a destination somebody had chosen. Same class of bug, different answer, and not this change. Five tests, and four of them were checked failing with the RefreshOpenEditors call commented out: a key reaching the open host editor and binding when chosen, an item arriving without moving a selection that was already made, a tag arriving as an unworn chip, a key reaching the group editor, and a vault reaching the host and snippet editors without moving either. Manual check 7.12 sits beside 7.11, which is this same bug on the files screen's picker, and says what the worse failure would look like: a picker that moves rather than one that does not notice.	2026-08-06 12:08:25 +02:00
jaap-jan	f1d6499bb5	Merge branch 'main' ... Two of main's changes land in files this branch rewrote, and both needed carrying across by hand rather than by the merge. The phone's nav staying up on Connections with nothing running is a fourth input to RefreshChrome, which this branch had already given two more — whether hosts are ticked and whether the host editor is filling the screen. They compose: the rail and the bottom bar now ask (pages || connectPage) && !editing, so a page-shaped terminal surface keeps its way off the screen and the editor still takes the whole display. The key question under the host's move panel is the harder one, because this branch deleted the panel it was added to. The connect card is gone and the phone's only route to a move is the action bar, so leaving the merge to take this side would have removed a capability main had just shipped — silently, since nothing would fail to build. It is asked in the action bar's own picker instead, in two shapes fewer than the desktop's: one host, because which key to carry is a fact about one machine and a selection of six has six answers, and a move rather than a copy, because taking the key out from under an original that is staying put would leave that original unable to connect. BindingOfTheMovingHost splits into MovableBindingOf so both heads answer it the same way from different panels. Main also fixed a real trap in the same commit — a host that only inherited its key from its group arrived in the destination naming nothing at all, because the group stays behind — and the batch move had the same bug for the same reason. It goes through Detached now, which is where that fix lives. The carried host is written as the carry left it rather than being detached again, which is the one thing worth measuring: the key takes a new id over there, so a run that rebuilt the payload from the row would send the machine across naming a tombstone. Both directions are pinned, along with the rule about which shapes the question is asked in at all.	2026-08-06 09:30:00 +02:00
jaap-jan	c882fa0cd3	Give the phone a selection instead of a card under the list ... A long press on a host raised a connect card over the bottom of the list: a password box, CONNECT, EDIT, MOVE and DELETE. It was the right idea in the wrong place. It covered rows, it had room for five things and never a sixth, and every one of them was about exactly one machine — so filing eleven imported hosts under a group was eleven trips through a form, and there was nowhere to put a sixth action if anybody wanted one. A long press now chooses the host it landed on, and the actions move into a bar across the top of the screen, in the vault header's place rather than beside it. That is where Android has put them since contextual action bars existed, and it is the one strip a list can never grow into — but the real reason for it is that while it is up the screen is unambiguously about the ticked hosts and nothing else, which is what lets the count in the middle of it mean something. Left to right: the cross that leaves the mode, the count, the pencil, and a ⋯ holding Connect, Connect via SFTP, Move to vault, Copy to vault, Change group, Duplicate and Remove. A tap still connects and still raises nothing. Once anything is ticked it ticks and unticks instead, which is what every Android list does and is not merely a convention worth following: a tap that connected while five machines sat ticked would open a terminal on top of a selection somebody was halfway through building. Unticking the last host leaves the mode, so there are two ways out of it and the cross is only one of them. Both gestures now read the row from the element under the finger rather than from the list's selection, and that is a correctness change rather than tidying. A tap on a group heading moves the selection and the view model bounces it straight back to whichever host was chosen before — which answered "a host, or nothing" for free while a tap only ever connected. It stops answering it the moment a tap can tick one: the heading would tick a machine the user was not pointing at, into a set they are about to delete. Three of the seven entries are about one machine and are drawn only for one. A terminal, a file-transfer session and a form each have no reading over six, so they are collapsed rather than refused. The other four read better for a count than without one — it is the reason the set exists — and each of them says afterwards how many hosts it wrote and how many it left alone. Skipping beats refusing the whole run: a selection of eleven with one read-only row would otherwise do nothing at all and then report about the wrong ten. Copy to vault and Duplicate are new, and the difference between them is what each can safely carry. A copy crosses a key boundary, so it drops the group and the tags exactly as a move does — both are items of the vault being left, and a host arriving with either would point at something the destination does not contain, resolvable on the machine that sent it and dangling for everybody else. A duplicate stays in the same keychain, so everything it points at is still there and it keeps both. Change group is the write dragging a card onto a group already makes on the desktop, run over a selection; it refuses one spanning two keychains rather than half-filing it, which is the refusal a drop across that boundary already makes one host at a time. Connect via SFTP is the one action that leaves the vault. Which machine is a decrypted item and so is this object's business; the screen it leads to and the transfers view model behind it are the shell's — so it is an event, on the same division SessionOpened already draws for a shell. The host is re-found in that screen's own copy of the list, because the picker binds to rows in that copy and handing it the vault's object would select nothing. What is left of the card is the password box, and only because it had nowhere else to go: a host that authenticates with a typed password cannot be reached by a tap alone. That tap now raises a sheet rather than the bar, and the difference is that a sheet is up only while a question is on screen — the bar was raised by a long press and stayed, so it was a password box sitting over the list whether or not anything was being asked. Dismissing it empties the box, which is not tidiness either: a secret left behind would satisfy the emptiness check that decides whether to raise the sheet at all, so the next tap would dial with somebody else's password. The pencil moving into that bar takes the host editor with it. It was a card in the list's own row, under the search box and the sync line — twenty controls sharing a screen with two rows of chrome about the list it had replaced. It is a page now, and PhoneShell stands all four of its rows down for it, which is what "opens with all the options" means at 360dp. That needed a second subscription in that control: two of its flags are questions about the vault rather than about the shell, and the shell does not forward the vault's notifications. The ticks are held as entity ids rather than as rows, and written back onto the rows after every reload. Every row object in the list is replaced on every filter keystroke and every synchronisation pass, so a set of rows would empty itself once a minute under somebody choosing what to do with eleven machines. Ids that no longer resolve are dropped, so a colleague's deletion arriving mid-selection leaves a count that matches what is on screen. One caller had to change with it. ConnectToRecent opened the pane about a host, which was the desktop's drawer and the phone's card; the phone's answer is now a tick, and nothing on that list means "selected" any more — so arriving with the host merely selected would be arriving at a screen with nothing to press. Both are raised together, and the one the head in front of the user does not draw is inert.	2026-08-06 09:15:37 +02:00
jaap-jan	69858f82d1	Merge branch 'claude/vault-key-sync-sharing-d098aa'	2026-08-06 07:39:32 +02:00
jaap-jan	509a7c34f5	Merge branch 'claude/snippets-vault-sharing-470476'	2026-08-06 07:39:25 +02:00
jaap-jan	185790fb14	Let a key move to another vault, and ask whether it goes with the host ... Keys sync and keys are shared: SshKey is in the sync registry on both sides, the material rides in the sealed payload, and every generation of the vault key is wrapped to a new member. What was missing was the way in. Hosts and groups could move between vaults and keychain items could not, so a key typed into a personal vault before the team existed stayed there for good — and moving a host into the team's vault left it authenticating with something nobody else in that vault can read. The code said so and could do nothing about it: "the answer is usually to put a copy of that key in the destination vault", which meant pasting the private half into a second item and deleting the first. A private key on a clipboard, and two items nobody can tell apart afterwards. MoveAsync already existed on the generic repository and is now exposed for keys and passwords as it is for hosts and groups. What had to be built around it is the re-aim. An item re-sealed under another vault's key lands with an id of that vault's making, so every host bound to the old one and every group lending it as a default is left naming a tombstone — and a host bound to something its vault no longer holds refuses to connect rather than falling back to a typed password. A move without the re-aim would look like a success and break every machine on that key. It runs over every vault this session can write to, because a binding resolves across all of them, and it counts what it could not rewrite: an item from a newer client, or one in a vault this account may only read. Those are said in the sentence afterwards rather than swallowed. The host's move asks the question rather than deciding it. A binding resolves across vaults, so the moved host goes on working for the person who moved it whichever way this is answered; it is the colleagues they have just joined who hold one vault's key and cannot connect with a host whose key stayed behind. Unticked, and it stays that way on purpose: moving a key into a team's vault hands it to everybody holding that key, and this design does not default anybody into a disclosure. Under the box is the count of everything else that authenticates with that key, because a key twenty machines use is a different decision from one nothing else touches, and neither number is visible from the panel otherwise. The question is answered against the vault in the picker, so choosing a different destination re-asks it and a key already in the destination offers nothing. One thing fixed on the way. A host that inherited its key from its group arrived in the destination naming nothing at all — the group belongs to the vault it left — so a machine that connected before the move refused after it, with no sentence anywhere saying why. The resolved binding is now written onto the host as it crosses, and the stranded-binding warning reads the resolved binding too, which is the case where somebody is least likely to know a key is involved. MOVE is on both heads, for keys and passwords only: a tag, a bucket and a pin are read from the active vault alone, so "another vault" is not a question any of them has. Four tests cover the move and its re-aim, the host's move with the key brought and without it, and the inherited binding.	2026-08-06 07:39:15 +02:00
jaap-jan	3d9ed03b09	Let a snippet be shared to a vault, the way a host already can ... A snippet was a first-class vault item everywhere except where it mattered: the crypto, the sync, the server table and every registry already treated it exactly as they treat a host, and the screen read it out of the active vault alone. So the one command a team most obviously wants to hold in common — the incantation somebody worked out once and everybody else retypes — was the only item kind that could not leave the machine that wrote it. The read is the half that had to come first, and it is why this is not simply a MoveAsync. ReloadSnippetsAsync now lists every readable vault rather than the active one, in the shape ReloadHostsAsync and ReloadKeysAsync already use: the vault new items go into first, then by vault name, then by label, with a badge on the row only where there is more than one vault to tell apart. Without that, a snippet moved into a team vault would have disappeared from the very screen that moved it, and one a colleague wrote there would never have arrived at all — sharing would have looked like losing. Three writes were pinned to the active vault and each one broke differently once the list spanned several. The delete tombstoned in the wrong vault, which tombstones nothing and leaves the snippet on screen. The save is the bad one: an update sent to the active vault creates a second snippet there and leaves the team original untouched, so the person editing sees their fix and nobody else ever does. That is a fork with no symptom, which is why the vault is now a parameter and the screen latches it when the editor opens — the chosen vault for a new snippet, the row own vault for an existing one — rather than reading it back off a selection that can move under a half-typed form. VaultViewModel has carried editingHostVaultId for the same reason since hosts crossed vaults. Two controls rather than one, and that is the same line the host pane draws. The editor asks which vault a new snippet is filed into; MOVE re-seals an existing one under another key and tombstones the first. Putting the second inside the first would let somebody correcting a typo hand a command to a team by leaving a picker where they found it, so the picker is not drawn for an existing snippet at all. Both live on SnippetsViewModel rather than VaultViewModel because this screen owns its editor, unlike the host drawer; the writing they ask for is still the vault. A snippet crosses whole, which is the one way this is simpler than the host it copies. A host leaves its group and its tags behind because both are items of the vault it came from and would dangle for everybody in the destination. A snippet is a label, a command and a note, and none of them points at anything — so there is nothing to strip, nothing to report as left behind, and what the copy says instead is the thing that is actually at stake: who can read the command afterwards. For a command that may carry a hostname or a path, that is the whole decision. Two judgement calls worth finding later. A hidden vault now hides its snippets, filtered in the screen projection rather than in VaultViewModel.Snippets, which is the rule keys and passwords already follow: the list stays whole so nothing that resolves against it breaks, and the projection is what a preference about reading gets to change. And the nav rail count is left spanning vaults unfiltered, because Vault.Hosts.Count beside it is unfiltered too — filtering one of the four would make the rail disagree with itself. Four flow tests in VaultSharingTests, beside the host ones they mirror: the move re-seals with a new id and carries the runs-on-insert flag across, the move with nowhere to go refuses rather than opening an empty picker, the editor files into the vault chosen on it, and the edit of a shared snippet goes back to its own vault instead of forking. That last one is the regression the latch exists for and the only one whose absence has no visible symptom. Plus a layout test with the move panel open, since that paragraph wraps in a 300-pixel column and the desktop pane it lands in is measured. The whole suite passes: 1660 tests, none failing.	2026-08-06 07:39:03 +02:00
jaap-jan	cddfeb1f55	Keep the phone's nav under Connections when nothing is running ... The chrome stands down for a shell, and it was standing down for the whole terminal surface. Those parted company when that surface gained a connect page: with no tabs open it draws a box, a CONNECT button and the machines connected to before, which is a page in everything but which enum it is in. A third of the display is worth giving to a shell and is not worth giving to that. Worse, it is the one screen somebody arrives at by closing their last tab — so the state the collapsed bar was most likely to be seen in was the state where it left the system back gesture as the only route to Hosts or Settings. So RefreshChrome reads one more question. IsTerminalSurface with no tabs joins the pages in both flags, which keeps the rail and the bar in step: above 600dp the rail is the bar, and fixing only the narrow layout would leave an unfolded device on the same screen with the same nothing. The vault header is deliberately not part of it. The surface draws its own bar with back and the +, and a header above that is the second row of chrome this head exists to avoid. The Connections entry lights for the first time, on IsTerminalSurface. It was left unbound on the argument that the bar was never drawn while that surface was up, so a lit state was unreachable — that argument is now false, and the flag is unambiguous on a control that is only drawn in two situations: false on every page, true on the connect page, and never read while a shell is showing. A bar sitting under a screen it does not point at is the entry looking broken instead. Nothing here is testable on this head — the phone's rectangles have no coverage, for the reasons Phase 8 of manual-checks records — so 11.7 gains the check that the bar is there with Connections lit, and 11.1 keeps the one that it is gone with a shell up, which is the half that pays for the arrangement.	2026-08-06 07:38:31 +02:00
jaap-jan	174ef7c420	Merge branch 'claude/android-release'	2026-08-05 22:49:51 +02:00
jaap-jan	af0e29a98b	Give the desktop a nightly channel, the way the phone has one ... ADR 0014 gave the phone a nightly and ADR 0013 rule 3 gave the desktop none, so the two heads had different answers to the same question — how does somebody try what is on main? — for no reason except the order the work happened in. This is the desktop's answer: CI publishes a build from main on every push, and it installs beside the release one rather than over it. The phone gets its separation from the platform. Android refuses an update signed by a different key, so its two channels cannot replace one another whatever anybody does. Nothing refuses anything here: Velopack applies what its feed serves and verifies no signature. So all of it is construction, and there are four separations because each closes a different door. A pack id each, so the two install in different directories and neither feed's package can be applied to the other's install. A Velopack channel each — win and win-nightly — so neither build ever reads the other's release index; the name reaches the wire as releases.win-nightly.json, which is why the constant in VelopackUpdateChannel and the argument in ci.yml have to agree or the channel answers nothing forever with no error. A prerelease flag, so the release channel cannot see the nightly even by accident. And a profile directory each, which is the one that is easy to skip and would hurt most: the cache schema is migrated on every launch, before unlock, so a shared profile means a nightly quietly upgrading a database the release build then opens. Both are installed at once by design, so that is an ordinary Tuesday rather than a corner case. The prerelease flag turns out to be load-bearing across heads as well. The phone's release channel reads releases/latest, which skips prereleases — so a desktop nightly published as a stable release would become the newest release in this repository and every phone on the release channel would start failing its check against a release carrying no Android manifest. Which build this is arrives as assembly metadata, the same mechanism and the same reasoning as the Android head: the updater needs the string rather than a branch, and a value baked into the assembly is one a crash report can be asked for. Three things read it — the feed, the prerelease flag, and the profile — and one more shows it: the titlebar says DodoSSH Nightly. Everything else that distinguishes the two is somewhere nobody is looking while typing a passphrase into one of them. The version needed a floor and it is applied to the whole build rather than to the packaging. MinVer answers 0.0.0-alpha.0.N until the first v* tag and vpk refuses anything below 0.0.1, so the job lifts the patch digit and keeps the height — through MinVerVersionOverride, so the assemblies carry the same number the installer does. Packing a version the assembly disagreed with would put one string on the preferences screen and another in the feed, which is the screen somebody reads when asked which nightly they are on. Two things found by running it rather than reading it. -t:MinVer needs a restore first, because the target arrives with the package and MSB4057 on a clean checkout reads like a typo in the workflow rather than a missing restore; the release script had the same gap and now restores before it reads. And vpk rejects an empty --packVersion loudly, which is how a broken version handoff announces itself rather than shipping a package called 1.0.0. Rule 3 is untouched. The release channel still has no job, no token and no runner, and the two channels cannot see each other. What a nightly costs is written where somebody reads it before installing one: whoever can write a release here can put a build on every nightly machine, which is fine for a build being tried and is not fine for a build holding somebody's infrastructure credentials. Verified by running the job's own steps against a clone in a Linux container: DodoSSH.Desktop.Nightly-win-nightly-Setup.exe, and an index naming pack id DodoSSH.Desktop.Nightly at 0.0.1-alpha.0.144. The upload itself is the one step not exercised — it needs a real forge and a write token, and check 16.10 is what walks the half no runner can.	2026-08-05 22:35:51 +02:00
jaap-jan	8591035170	Merge branch 'claude/pipeline-curl-not-found-97e70f'	2026-08-05 22:31:42 +02:00
jaap-jan	ca7fee2358	Start the confirmation Android hands back, so an update can install ... Pressing INSTALL closed the application, installed nothing and said nothing. That is two independent faults in one method, either of which breaks it on its own, and they hid each other: the first kills the process before the second can be observed, and the second is silent by construction. The pending intent handed to commit was implicit — an action string with no component behind it. A mutable pending intent may not wrap one of those from API 34, and this head targets 36, so every current phone threw IllegalArgumentException before commit was reached. Nothing caught it, so it left the command handler, passed the dispatcher and took the process with it. That is the closing. Below 34, where it did not throw, it still installed nothing. An application holding REQUEST_INSTALL_PACKAGES rather than the privileged INSTALL_PACKAGES gets no verdict back from a commit: what the platform answers first is STATUS_PENDING_USER_ACTION, carrying the activity that draws the dialogue in EXTRA_INTENT for the application to start. Android does not draw it on its own. The comment here asserted the opposite — that a pending intent is required whether or not anything listens, and that nothing needed to — so no receiver was ever written, and the session was written, committed and left staged forever. So there is a receiver now, not exported because the only sender is this application's own commit, and the intent naming it is explicit, which is the same change that stops the throw. Sessions are abandoned when anything fails, since one created and neither committed nor abandoned stays staged against a per-application cap — a repeating fault would have started failing at CreateSession instead, which is the same bug wearing a completely unrelated face. The reporting is the part worth keeping even after the cause is gone. Where applying ends the process an exception has nowhere to go; where it does not, which is this head's whole shape, it goes out through the dispatcher. RestartNowAsync now answers the way CheckNowAsync already did, and the regression test asserts the absence of a throw rather than the presence of one. ADR 0014 rule 6 gets the correction in place: "asks Android to ask" is one step longer than it reads. Check 17.5 needed no rewording — it asks for the installer appearing by name, which is exactly the thing that never happened — so what it gets instead is the two symptoms named, because both present as a dead button. It is the only thing in the project that can catch either, and it plainly was never run against a real pair of builds. Note for whoever takes the next nightly: a broken updater cannot install its own fix. The phone is running the code this commit replaces, so the first build carrying it has to be sideloaded by hand; the ones after that install normally. Compile-verified and manifest-verified — the receiver reaches the generated manifest — and 321 tests pass. Not run on a device, which is what 17.5 is for.	2026-08-05 22:31:33 +02:00
jaap-jan	b4619db8d2	Merge branch 'claude/android-release'	2026-08-05 21:57:05 +02:00
jaap-jan	3d3d0bc95f	Package the Windows client in CI, on the runner that could not ... The build job published a win-x64 tree and stopped there, so the half of a release that fails in ways a compile cannot see was proved by nobody until a person was midway through cutting one. It now packs as well: vpk opens the published binaries and verifies the main executable really calls VelopackApp.Build().Run(), which is the check worth having — a refactor that drops that call compiles, tests green, and produces an application that silently never updates itself. The file said this was impossible on Linux, and also said it was fine, in comments forty lines apart. The claim that vpk needs Windows tooling to stamp the Setup.exe stub is the one that was wrong: vpk cross-compiles when told to, and the telling is a bracketed directive before the verb rather than a flag. Plain `vpk pack --runtime win-x64` on a Linux host refuses outright and says so in the message that names the fix. `[win]` must be quoted, or the shell reads it as a glob matching any one of w, i and n. Only signing needs Windows, and nothing here is signed yet. Fixing that does not move ADR 0013 rule 3 an inch, which is why the two reasons were recorded separately in the first place. What may not live on a runner is the token, not the build: Velopack clients apply what their feed serves without verifying a signature, so whoever can write a release can ship an update every install runs. The packages go to RUNNER_TEMP and die with the job. They are not offered as workflow artefacts either — an installer nobody has run should not sit somewhere that invites passing it on. Written out as shell rather than by calling scripts/release-windows.ps1. That script is a person's procedure and holds things a runner must not have and must not skip: it refuses a dirty tree, insists HEAD is tagged, downloads the previous release for deltas, and asks for the forge token. Calling it would mean either weakening it with CI switches or having CI satisfy conditions that only make sense at a desk. The constants the two now share — pack id, title, authors, channel, icon — are a contract with VelopackUpdateChannel and with every installed client, and both sides say so. Two things fell out of running the steps rather than reading them, and both were in code nothing had ever executed: dotnet msbuild -getProperty:Version answers 1.0.0. Without a target named it evaluates the project and runs nothing, and MinVer computes inside a target — so the read comes back as the SDK default on a full checkout with every tag present. That line is the tag check in this file, which is `if:` a tag ref, and there are no tags yet: the first release ever cut would have been refused by its own guard, which would then have blamed fetch-depth. release-windows.ps1 had the same line and would have demanded HEAD be tagged v1.0.0. Both now pass -t:MinVer. And MinVer answers 0.0.0-alpha.0.N until that first tag exists, which vpk rejects outright as below 0.0.1 — so packing the true version could not have worked on any build made today. The patch digit is lifted for the throwaway package only. The release script gets no such floor and must not: its version is the one users compare against, and there the refusal is the right outcome. Verified by extracting both steps from this file and running them against a real clone in a dotnet SDK container: Setup.exe, the portable zip, the .nupkg and releases.win.json, from a machine that is not Windows.	2026-08-05 21:56:45 +02:00
jaap-jan	205f946450	Give the nightly's publisher a curl to publish with ... The android job builds its APK in a glibc container and then does the publishing on the host, and the host is Alpine — busybox wget, and no curl anywhere on it. So the step that talks to the forge six times died on the first of those calls not written to tolerate a failure, with `curl: command not found` and exit 127, after the whole build had already been paid for. It took until the fourth call to say so, and that is the part worth reading. The lookup for an existing release and the tag delete before it both end in `|| true` with stderr discarded, which is exactly right for "there is no nightly yet" and indistinguishable from "there is no curl on this machine". Installing it is what makes those two different again. Dropping the `|| true` instead would fail the step on the first run of a channel that has never published, which is the one state that shape is there to handle. Not busybox wget in curl's place. The asset upload is a multipart -F, which busybox wget cannot send — so that reading ends in a release created with no APK on it, which is the failure the verification at the end of the step exists to catch, arrived at on purpose this time. Conditioned exactly like the step it serves, main only, because nothing else in this job wants curl and a pull request should not pay for a package it will not use. That is a coupling rather than a tidiness, and it is in the comment: the two conditions have to move together, and loosening the publish on its own puts the job back at exit 127 several minutes in. Unproven until the next main build, since the install path only runs there. What is checked is that the workflow still parses, that the script is valid under sh as well as bash like the ensure steps beside it, and that the two conditions read identically once parsed.	2026-08-05 21:39:58 +02:00
jaap-jan	5cb361ea13	Lay the phone out like the desktop when the surface is not a phone ... Three destinations in a bar and everything else behind SETTINGS is the right shape at 360dp, where a fourth entry costs the width of the three that are there. On a tablet, an unfolded foldable or a landscape phone it is the wrong one: there is room for every destination at once, and the hub becomes an extra tap between somebody and a screen they can already see space for. So at 600dp — Android's own boundary between a compact window and a medium one, in the density-independent units Avalonia lays out in — the bar stands down and PhoneRail takes the left edge with all nine on it. It is the desktop's NavRail arrangement rather than its file: the two heads cannot share a view, and this one draws the phone's destination set with the phone's palette and touch targets. The flags are computed in code rather than assembled in the markup because none of them is a single question any more, and Avalonia's bindings have no "and" — and the header's condition is an "or", which not even a wrapper can express. That header is the one worth reading twice: narrow it stands down behind SETTINGS, so the hub's screens can draw their own; wide there is no hub to be behind, so it stays up everywhere. Losing it on the keychain would be losing the only LOCK button on the surface. Removing the hub means removing the routes into it, and there were four kinds. The rail has no SETTINGS entry, because that screen is a menu of the rail. The five back arrows in the screens under it are hidden, since an arrow to a screen the layout removed is the one control on a header that leads nowhere. The system back gesture goes to Hosts instead. And unfolding while sitting on the hub moves to Hosts, rather than leaving somebody on a list of things now visible beside it. One bug fixed on the way: OnBodyResized returned early unless the keyboard was open, so a foldable would have opened to a phone layout until somebody typed something. The chrome is refreshed first and unconditionally; the early return belongs to the older job below it. What this does not do is use the width *inside* a screen — the host list is one column at any size. Two columns needs the row model to change, because that list is headings and hosts in one sequence and a heading has to span, and that model is shared with the desktop. Check 8.1 walks the rail; nothing here is covered by a test, for the reason 8.0 exists.	2026-08-05 20:50:18 +02:00
jaap-jan	0c61ea3a97	Let a vault be shared from the phone, not only read there ... This screen's own comment argued ADD out: an address typed into a box, a directory lookup, a role picker and a paragraph saying what adding somebody did not do, for an act a colleague at a desktop is already performing. That was a cost argument and it was wrong about who is holding what. The person who needs to let somebody into a vault is often the one away from their desk, and answering them with "go and find a desktop" is the thing this head exists to stop doing. Making a vault was already here on exactly that reasoning. Nothing shared changed — AddMemberCommand, the role and the chips are the same members the desktop binds — so what this is, is markup and the argument it reverses. Four rows under the members list: the box, three role chips rather than a picker because the answer is one of three short words, ADD, and the paragraph. Gated on being able to administer the vault, so a plain member sees nothing rather than a button whose only outcome is a 403. The paragraph is not the optional part. Adding somebody changes what the server will serve and nothing else; the key is still wrapped by a machine that holds one — which on an unlocked phone is this one, in the same press. A screen that offered the first and stayed quiet about the second would imply the server can hand out access, which is the single claim this product is built to refuse. What the phone still does not draw is anything that takes access away. REMOVE and WITHDRAW KEY act on the first press, and an irreversible revocation under a thumb with its explanation in a tooltip no touch screen can show is the wrong trade — which is the line this file already drew and this does not move. Check 12.4 walks it, including the locked-keychain case: the membership is made and the line says the key could not be wrapped, which is a state somebody can act on rather than silence.	2026-08-05 19:10:25 +02:00
jaap-jan	dc1ebf6afa	Let the recovery code be copied, and give the phone a clipboard to copy to ... Both screens had made the code selectable and both said why: a person who cannot get it out of the box photographs the screen, and a screenshot is a worse home for it than a clipboard. This finishes that argument. Selecting 64 characters of letter-spaced monospace with a thumb is the version of "possible" people give up on halfway — and on the phone the screen blocks screenshots, so the honest remaining options were retyping it or losing it. It is the one secret this application deliberately offers to a clipboard, and the contrast with the keychain's copy is the whole argument rather than an inconsistency. There, copying the private half is refused outright, because installing a key means pasting the public one and the private one has no business leaving the vault. Here there is no better route: the code exists for one screen, is stored nowhere, and has to reach a password manager. The clipboard is the intended destination rather than a way round the design. The sentence afterwards matters as much as the copy, and is asserted: a clipboard is a staging post, this screen is the only place the code exists, and the next thing copied replaces it. Somebody who copies and does nothing has not saved it. The phone had no clipboard delegate at all — the desktop passed one and this head passed null — so COPY PUBLIC KEY on the keychain answered "this machine has no clipboard" on a device that plainly has one. Nothing about that was platform shaped: Android has a clipboard and Avalonia surfaces it through the same TopLevel. Wiring it fixes that copy too. The test fixture built its shell without a clipboard, which modelled the bug rather than the product, so it has one now and the public-key test asserts what lands there instead of the refusal. The refusal keeps its own test, on a shell built without one, because the view model reads the delegate's absence rather than an empty result — and because a button that silently does nothing on this screen is worse than one that refuses.	2026-08-05 18:14:34 +02:00
jaap-jan	253c72d2b7	Name the organisation the repository actually lives in ... It moved to DodoTech-Public, and every address in the product still said DodoTech. That looked like it worked, which is the part worth writing down: Gitea leaves a 301 at the old path and HttpClient follows a redirect on a GET, so both update channels would have kept polling through it. What a 301 does not survive is a POST. `vpk upload gitea` publishes the desktop release by POSTing to that URL, so the stale address would have failed at the one step the whole feature depends on — and a redirect is a thing an operator can delete, which turns "works today" into the same silent outage this session has already spent two commits on. So both channel constants, both release scripts, the workflow's REPO, the image's source label and the curl in phase 16 all name the live path. The local remote too, which had been printing a redirect warning on every push. Measured after the move: the org, the repo and the nightly release all answer 200 anonymously, and that release now carries both assets — the manifest and a 54 MB APK. The upload going through also answers the open question about the reverse proxy's body-size limit, which nothing local could test.	2026-08-05 12:51:19 +02:00
jaap-jan	23f1db9dc8	Stop the terminal's accessory keys taking the keyboard off it ... Ctrl, Esc, Tab, the arrows and the two text-size keys were ordinary Avalonia buttons sitting over a NativeWebView. An ordinary button takes focus on tap, which takes it off the WebView — and the package's own OnLostFocus then calls the adapter's ResignFocus(). So pressing Tab handed the terminal one byte and took the keyboard away from it, and everything typed afterwards went nowhere. What makes it worth more than a one-line fix is the symptom. The row goes on working, because its keys are pressed rather than typed into, so what you see is a terminal that answers the buttons and ignores the keyboard — which reads as the session having died rather than as anything to do with focus. Focusable = false is what a toolbar button is: these keys are an extension of the keyboard, not a place it should go. The focused element then never changes, so nothing resigns and nothing has to be handed back — which matters, because the hand-back is the direction platform-flags already records as the hard one. The flags file gains the phone's half of that entry, and check 11.10 is the measurement: this needs a paired hardware keyboard and there is no test on this head that could stand in for one.	2026-08-05 12:45:17 +02:00
jaap-jan	1bcf422bbe	Build the phone once per run, and stop rebuilding the toolchain image ... The android job compiled everything twice. The plain `dotnet build` before the packaging step looked like a cheap check ahead of an expensive one and was neither: SignAndroidPackage depends on Build, so the packaging line compiles everything anyway — and the build above it ran with no -p:DodoChannel, which means it ran as the *release* channel. Different application id, different version, different assembly metadata; MSBuild treats a different set of global properties as a different project instance, so not one output was reused. It was a full second compile of the reference closure, producing an APK for the one channel this job must never build, thrown away unread. Measured in the toolchain image with a warm package volume, same commit: two builds 2m52 + 2m26 5m21 total one build 3m01 3m04 total Byte for byte the same artefact out of both — versionCode 203, versionName 0.0.0-alpha.0.136, dev.dodotech.dodossh.nightly. The toolchain image is now built once per Dockerfile rather than once per run. The tag is the Dockerfile's own digest and docker applies a tag only on success, so an existing tag is by construction the right image and `docker image inspect` is a sound check rather than a guess. What that trades away is the JDK from apt drifting; everything that decides what is in the image is pinned in the Dockerfile, so anything that matters changes the digest. It is a build tool, not something shipped — the image job takes the opposite trade with --pull, because what it builds is what users run. And the build job now says whether its package cache did anything. setup-dotnet's cache: true is actions/cache underneath, which needs a cache server act_runner ships and can have turned off — and when it is off it does nothing and says nothing about it. A cold restore and a perfect cache look identical from outside: both are green, and the difference is minutes. One `find` before anything writes to the folder turns that from a belief into a line in the log. It does not fail the build, because a runner without a cache server is slow rather than wrong. What is deliberately not cached: the apt installs in each job's preamble, which need the runner's image fixed rather than a workflow change and already say so; the Testcontainers pulls and the API image's layers, which the daemon already caches on a persistent runner; and the android obj/bin, which would not help — the source arrives by `docker cp` with fresh timestamps, so MSBuild rebuilds it whatever is in there.	2026-08-05 12:40:28 +02:00
jaap-jan	33d4c3ff48	Check the organisation as well, because in Gitea the organisation wins ... Setting the repository public changed nothing: /api/v1/orgs/DodoTech answers 404, and a Gitea org's own visibility gates everything under it — a public repository inside a Limited or Private org is invisible to anyone not signed in. So 16.0 now checks both, and says which answer means which. It also names the signal that tells this apart from an instance requiring sign-in for everything: /explore/repos answering 200, which this one does, so what is hidden is hidden on purpose rather than by policy.	2026-08-05 12:06:09 +02:00
jaap-jan	e0655dbb31	Look the host list up by name, rather than off a field that is never assigned ... Nightly 0.0.0-alpha.0.133 died before its first frame. The long press I added attached itself in HostsScreen's constructor through the field the Avalonia name generator declares for `x:Name` — and that field is assigned by the generated InitializeComponent, which no view in this repository calls. Every one of them loads its XAML directly. So the field compiles, resolves in the editor, and is null at run time; PhoneShell builds this control on the way up, so the NullReferenceException took the launch rather than the hosts screen. PhoneShell and TerminalScreen both use FindControl, and PhoneShell carries a <remarks> saying exactly this and naming exactly this consequence. I read neither and wrote the field. So the rule is in docs/platform-flags.md now as well. A comment on the control that already got it right is not where somebody writing a new one is looking, which is the whole of why two correct examples and one warning were not enough. And phase 8 opens with "it launches at all". Nothing on this head is covered by a test — no test project, no headless surface — so a view that throws while being built takes the launch with it and no gate anywhere says so. Thirty seconds, and it would have caught this one before it was published.	2026-08-05 12:00:20 +02:00
jaap-jan	9a7e3bbd5c	Let a failed update check say so, instead of reporting good news ... The phone reported every build as current because the release repository is private. Gitea answers 404 rather than 403 for a repo you cannot see, the client reads that address anonymously, and AndroidUpdateChannel caught the failure and returned null — which IUpdateChannel documented as meaning "this build is the latest". The check had never once succeeded on any phone and nothing anywhere said so. Two faults, and the second is why the first lasted. The seam said null was the honest answer for an unreachable channel, on the reasoning that the caller does the same thing either way. That is true of the six-hourly pass and false of CHECK NOW. UpdateViewModel already draws the line correctly — silent on the timer, the exception's message on the button — and it could only ever draw the first half, because nothing was ever thrown at it. The desktop's channel does not catch, so the interface described neither implementation. So CheckAsync throws now, and null means one thing. A release that is reachable but missing its manifest or the APK it names throws too: "you are up to date" about a half-published feed is the same lie in a smaller costume, and the self-healing that argument protected is untouched, since the timer still swallows everything. The precondition is written down where somebody would look, rather than left as a sentence about where a token could live. ADR 0013 §4 already said a private release repository was incompatible with this design; nobody checked which side of it this repository was on. It is one curl, and manual-checks phase 16 now opens with it — pointedly not against /api/v1/version, which answers 200 from a forge that is up whatever is readable on it, and which is what made this look like nothing was wrong. Phone check 17.4 was the one that passed all along. It now presses CHECK NOW with the network off as well as on, because two different answers are the whole of what makes the first one worth reading.	2026-08-05 11:17:44 +02:00
jaap-jan	ca07d63585	Offer to bring the keys an ssh_config points at ... An import that recorded a key path and left every host asking for a password was an import whose result did not connect. The answer to that was a manual paste per key, which is the sort of thing people do once and then stop importing. So there is a tick, and it starts off. With it off nothing changes: an IdentityFile becomes a note and the host asks for a password. With it on, IMPORT reads each host's first IdentityFile out of ~/.ssh, stores it in the vault encrypted like any other key, and binds the host to it. Three things about how it is drawn are load-bearing rather than tidy. It is a default nobody arrives at by accident. The sentence beside it names the directory rather than saying "your keys", because that is what somebody is agreeing to. And nothing is read during SCAN — tick it, read what it says, untick it, and no private key has been opened. This is the only place the application opens key material out of a directory the user did not point at file by file, and the whole of what makes that acceptable is that it took a deliberate press. One vault key per file, however many entries named it: an ssh_config pointing twelve hosts at one id_ed25519 is the ordinary shape, and twelve copies would be twelve things to rotate and eleven to forget. A file whose material is already in the keychain is bound to rather than stored again, which is what makes running the import twice harmless. What cannot be read off a disk is a passphrase, so a protected key arrives without one — and the report under the button names those files rather than leaving a host to fail at connect time with a message about a malformed key. Telling them apart means decoding for OpenSSH's own container, whose cipher name is the first field inside the base64 rather than anything in the armour, and that is the format ssh-keygen has written by default for years. The 88 base64 characters it decodes need 66 bytes, not 64: with the smaller span every protected key came back unprotected, which the tests now pin. A path that is not on this machine leaves its host imported and unbound, exactly as it would have been with the tick off, and is named in the same report. A config carried from another machine is the ordinary case, not an error.	2026-08-05 08:58:28 +02:00
jaap-jan	746711da9d	Let a tap on the phone's host list mean connect ... Choosing a machine raised the connect bar over the bottom of the list: a password box, CONNECT, EDIT, MOVE and DELETE. Five controls in the way of the one thing a tap on a machine's name obviously means. So the gestures split. A tap connects. A long press raises the bar, with all five. The pencil in the phone's header — its only persistent chrome — edits whichever host is chosen, which is the one of the five common enough to be worth a control that is always in the same place. The flag doing it is the desktop's own IsHostPaneOpen rather than a second one. That head made exactly this move when a selection stopped opening its drawer, and the question both are asking is "has somebody asked about this host" — answering it twice is how two heads come to disagree about what a selection means. One tap cannot finish: a host that authenticates with a typed password has nowhere on a list to be given one. That tap raises the bar with the box in it and says so, and a second tap with the box filled in connects. The branch is in the view model rather than in the head, because "can this machine be reached without asking for anything" is the same question the bar's own password box answers, and a copy of it in a view would be a second reading of a binding chain that has one. Two mechanics worth knowing. Avalonia raises Tapped on release whatever the press lasted, so a long press would open the bar and then connect — one touch firing both gestures — which is why HostsScreen tracks the hold and swallows the tap it precedes. And Holding only fires once IsHoldingEnabled is set, so that and the handler are attached together rather than one in markup and one in code. ConnectToRecent now opens the pane rather than selecting the row. On the phone it has to: a selection alone raises nothing now, so going back to a recent machine would land on a screen with nothing to press.	2026-08-05 08:43:33 +02:00
jaap-jan	69bc9e270b	Let a team be joined only by somebody who is already here ... An invitation decided access from an assertion about an address. Everything else in this model decides it from something a person did — an admin naming an account, a key holder wrapping a vault key to a key they verified — and this was the one place a token's email claim was the thing that let somebody in. It was guarded as tightly as that can be guarded: the claim was refused outright on an unverified or absent `email_verified`, with no setting to relax it. But the guard and the risk were the same shape. The whole defence was one boolean sent by a system the deployment does not control. So `POST /teams/{id}/members` is the only way in, and an address with no account is refused with `no-such-account` — which is now the end of the road rather than the signal to invite. Both clients say the remedy: that person signs in here once, which is what creates the account, and then they can be added. The desktop leaves the address in the box, because a message telling you to come back later is one you act on later. Gone with it: the `team_invitation` table, the claim hook in the sign-in path, and `Oidc:EmailVerifiedClaim`, which that hook was the only reader of. Nothing in the server now reads the email claim to decide anything. Pending invitations are dropped rather than converted. Converting one would mean creating a membership because an address matched, which is the property being removed — and an invitation to an address that did have an account here had already been claimed by the hourly sweep, so what is left is offers to people who never arrived. Two tests carry the property rather than the feature: the endpoint inventory asserts the three routes are absent, and the API suite adds an address that has no account, watches the refusal, then signs that address in and checks it joined nothing. Without the second half, a server that merely renamed the deferred path would pass.	2026-08-05 08:28:57 +02:00
jaap-jan	7dc3b8950d	Read the release id out of the front of the response, not the back ... The nightly release was created and the step said it had not been: Gitea accepted the release call and returned no id: {"id":1,"tag_name":"nightly",…,"author":{"id":-2,…},"assets":[]} `sed -n 's/.*"id":\([0-9]*\).*/\1/p'` — the leading .* is greedy, so it walked past the release's own id to the last "id": in the document, which belongs to the embedded author object and is -2. [0-9]* then matched no digits at all and the answer was the empty string. The release is real; only the reading of it was wrong, which is why the page exists and carries nothing but source tarballs. It broke both readings and the other one silently. The existing-release lookup got the same empty id, so the delete never fired, so the next run would have failed to create a release for a tag that already had one — a rolling channel that works exactly once. Both now go through one function, and grep matching left to right with [0-9]+ cannot reach the author's -2 at all. And the step now asks for the release back and checks its own uploads are on it. The failure this channel is exposed to is a release that exists and carries nothing: a phone reads that as a feed it can never update from and a person reads it as a page offering source tarballs. The run before this one left exactly that behind, and it would have reported success for every upload it never made. Checked against the response body the failing run printed: the old expression answers empty, the new one answers 1, and the verification rejects that release and accepts one with an asset on it. The bad release is not cleaned up by hand — the next run on main deletes and recreates it, which is what the rolling tag does anyway.	2026-08-05 07:44:48 +02:00
jaap-jan	a18ca56fde	Copy the checkout into the container, because a socket is not a shared filesystem ... The container started and could not see the repository: MSBUILD : error MSB1009: Project file does not exist. Switch: src/DodoSSH.Client.Android/DodoSSH.Client.Android.csproj `-v "$PWD:/build"` cannot work here. This runner is itself a container holding the host's Docker socket, so the workspace path it reports — /root/.cache/act/<hash>/… — exists in the runner and not on the daemon's host, which is where Docker resolves a bind source. It finds nothing, creates an empty directory, and mounts that. Nothing about the failure says so. Nothing earlier in this workflow would have caught it either, and that is the part worth keeping: the image job's `docker build` sends its context over the API and Testcontainers mounts nothing, so neither of the two places this repository already used Docker proves a bind mount would work. I read a working daemon as a shared filesystem, and they are not the same claim. `docker cp` goes over the same API and so does not care where the daemon lives. In with the whole checkout, .git included, since MinVer and the versionCode both read it — the repository is well under a megabyte packed. Out with the staged package. The NuGet cache becomes a named volume for the same reason: it lives on the daemon and needs no path either side has to agree on. The two container steps collapse into one, since with a copy in and a copy out there is nothing to be gained by paying for both twice, and scripts/ci-android.sh is now what runs inside — a file that can be read and executed on its own rather than a heredoc inside a workflow. Exercised locally against a real clone, every step as the job runs it: create, cp in, start --attach, cp out, parse the manifest on the outside. package: name='dev.dodotech.dodossh.nightly' versionCode='196' versionName='0.0.0-alpha.0.129' The second run took 2m25s against the first run's 8m, which is the named volume doing its job.	2026-08-05 07:29:53 +02:00
jaap-jan	30a3edb1d4	Build the phone in a container, because this runner cannot build it at all ... The runner is Alpine, and .NET for Android does not work on musl. Not "needs setting up" — the SDK's own MSBuild tasks pull glibc shared objects out of the workload pack into the build process, and a musl-linked dotnet will not load one: error XARLP7000: Error relocating .../libZipSharpNative-3-3.so: __snprintf_chk: symbol not found That is a glibc fortify symbol musl does not implement, reached through a DllImport rather than an exec, so gcompat is no help: it gets a glibc *executable* started, which is a different problem. There is no musl variant of the pack. Everything this job did on the host to make Android work was therefore treatment of symptoms, mine included. The missing aapt2 was present. The "unsupported version" was of a binary that had never run. Both were this one sentence in a different accent, and the loader was the accent, not the sentence. So the toolchain moves into build/android-build.Dockerfile — Microsoft's own sdk:10.0-noble plus a JDK, the Android SDK and the workload — and the job keeps on the host only what the host is good at: checkout, git, publishing. The daemon needed no arranging, since the image job already builds with it and every Testcontainers suite reaches it over the socket. The image is tagged by the digest of the Dockerfile that made it, so on a persistent runner every run after the first is a cache hit, and a change to the toolchain is the only thing that buys a new one. Built rather than pulled: a community image with the Android SDK already in it would put a stranger in the path of a package this project signs and publishes. Eleven lines of apt and sdkmanager is the cheaper trade. Verified end to end in that image against a real clone rather than reasoned about, which after three rounds of reasoning seemed the least I could do. Restore under locked mode, Release build, then SignAndroidPackage: package: name='dev.dodotech.dodossh.nightly' versionCode='195' versionName='0.0.0-alpha.0.128' Signer #1 certificate SHA-256 digest: a9f067877724ddb0fdc04b637fbd5bfb97df753976616f100b48b522e132ba22 which is the keystore in build/. The versionName carries MinVer's height, so the csproj's target fires in the container too, and the manifest the feed publishes parses back on the host. Staging moves from RUNNER_TEMP to artifacts/, which is forced rather than preferred: the package is made inside a container and read outside one, so it has to land under the bind-mounted checkout.	2026-08-04 23:28:47 +02:00
jaap-jan	65256fa337	Take the phone's aapt2 from the SDK the job installs rather than the workload's ... The android job got past restore and died in the .NET Android SDK's own tooling resolution: warning : An error occurred trying to start process '.../packs/Microsoft.Android.Sdk.Linux/36.1.69/tools/Linux/aapt2' … No such file or directory error XA0111: Unsupported version of AAPT2 found at path '.../tools/Linux' The error names the wrong problem. Nothing was found, so nothing had a version, and XA0111 points at an Aapt2ToolPath in the project file that has never been set. The warning above it is the real message and it is only a warning. The pack was incomplete, and on this runner it would have stayed that way: act's host executor keeps /usr/share/dotnet between runs, and `dotnet workload install` reads the installed-workload records and does nothing when android is listed, whatever is on disk. So two changes, each of which stands alone. The build and the packaging step are now given -p:Aapt2ToolPath pointing at build-tools, which this job installs itself and can therefore vouch for. That is already the aapt2 the packaging step shells out to for `dump badging`, so this makes one tool of what were two, and the manifest the feed publishes is now read by the binary that wrote it. Checked rather than assumed: build-tools 36.0.0 answers aapt2 2.20 and .NET for Android 36.1.43 builds and packages this head against it with no complaint. The version is named once, in the step's env, because three things now depend on it agreeing with itself. And the workload step probes for the pack file that went missing and repairs the workload when it is absent. The probe is a witness rather than the point — the build no longer touches that binary — but a 5 MB file near the end of a 130 MB package is what a truncated extraction loses first, and r8.jar and manifestmerger.jar are what it loses next.	2026-08-04 22:32:52 +02:00
jaap-jan	7e0a1b2af8	Bring the phone's lock file back to the graph it actually restores ... The android job's restore has never been reachable — the runner had no JDK and no SDK, so it failed before it got there — and DodoSSH.Client.Android is deliberately outside DodoSSH.slnx, so the solution restore that keeps the other fourteen lock files honest has never seen this one either. It went stale for a whole release and nothing could say so. Two things had drifted by the time the repaired job reached the step: MinVer, added to Directory.Build.props for the desktop updater, and the ABI set, which grew when the -r android-arm64 pin came off the packaging step so that the nightly is installable on more than an arm64 handset. --force-evaluate on this project alone, and the fourteen it references come back byte-identical: an android-* RID is not a graph any of them has a package for. Checked rather than assumed, because this is the same mechanism that once put win-x64 into the server's lock files and broke its image build. docs/platform-flags.md records the lasting half, which is not this fix: any change to a shared props file touches a lock file this repository cannot verify from a machine without the Android workload.	2026-08-04 21:58:02 +02:00
jaap-jan	b4a6c19ac1	Let the phone replace itself, and give CI a channel it may sign ... The Android head had no updater and no release path, and the two are one problem: Android refuses an update signed by a different key, and CI generates a fresh debug key in every container. An APK released from a workflow could be installed once and never updated again — each new one an uninstall, which on this product means losing the cache, the outbox and the device key. So there are two channels, and they are two applications because the platform gives no third option. dev.dodotech.dodossh is cut from a v* tag by a person running scripts/release-android.ps1 with the key ADR 0011 rule 1 keeps off runners. dev.dodotech.dodossh.nightly is cut from main by CI and signed with a keystore committed here in the open — a key everybody has cannot be stolen and grants nothing by being held, which is why putting it in CI does not touch the rule. Neither can update the other, by construction. See ADR 0014. The android job assumed an image with a JDK and an Android SDK on it, which is what a GitHub runner is and what this project's is not. It now installs a JDK, fetches Google's command-line tools, accepts the licences and installs API 36 — each a no-op where it is already satisfied, and each cached by the persistent runner's own disk rather than by an action that would move a quarter of a gigabyte to rebuild a directory that never left. The client reads a small JSON manifest beside the APK, the counterpart of releases.win.json, and compares Android's versionCode rather than a version name: that integer is what the platform itself uses to accept or refuse an install, so comparing anything else would offer updates the phone then rejects. It fetches, and then asks Android to ask — the system draws its own confirmation, and from API 26 will not draw even that until unknown sources is on for this application. IUpdateChannel gained ApplyingEndsTheProcess. On Windows applying replaces the files and restarts, so the shell disposes the vault first and that is what zeroes the keys. On the phone the install is a request and the answer may be no, so disposing first would answer "not now" with a locked keychain and every shell closed — a punishment for declining an update. Two measured bugs found on the way, both older than this work and both invisible to a -getProperty check. ApplicationDisplayVersion is read by the Android targets in a top-level PropertyGroup, so the target setting it from MinVer ran after the only thing that reads it: every APK ever built here said versionName 1.0.0. And nothing found so far varies the launcher name per channel — four mechanisms tried, all of them recorded in platform-flags, none of them reaching the label the launcher shows. The two channels share an icon name for now and are told apart by package name, version, and what the preferences screen says.	2026-08-04 21:46:01 +02:00
jaap-jan	f90c331334	Merge branch 'claude/android-catch-up'	2026-08-04 19:58:16 +02:00
jaap-jan	50fa6fba38	Let the phone delete a host, and move or remove a group ... The desktop gained three things the phone did not follow: moving a group to another vault, the second question asking whether a group's deletion takes its hosts with it, and — since long before either — deleting a host at all. What that left was a head whose v3 + can fill a keychain and whose editors can correct one, with no way to empty either. The commands could not simply be bound. DeleteGroup and MoveGroup aim at GroupTarget, which is the selected card or the open group, and the phone has neither: its list draws headings, and a heading's selection deliberately bounces back to the host. Called bare on that head they would have returned having done nothing — a DELETE that appears to have been pressed and has not. Both now take the row and fall back to GroupTarget for the desktop's menu, and ConfirmMoveGroupAsync resolves from the panel's own movingGroupId rather than from the selection, which is also the honester answer on the desktop: what moves is the shelf the panel was opened on. The heading's pencil became a menu. Three icons after a chevron, a name, a vault badge and a count is what would be left of the name at 360dp, so the ⋯ raises the add sheet's shape carrying Edit, Move to another vault, a rule, and Delete — the desktop's card menu, in the one idiom this screen already has. It does not carry Open: the desktop's grid holds one level of the group tree and this list holds all of it flattened, so there is nowhere to open a group into. DELETE under a host sits on a row of its own beneath EDIT and MOVE rather than beside them. A phone has no hover and no tooltip, so where a thumb lands is the only thing separating a destructive control from an ordinary one. Both questions take the controls that asked them — ShowsConnectControls, which is the phone's half of the rule ShowsHostPaneActions already carries for the desktop's drawer — so DELETE cannot be pressed a second time underneath its own confirmation. Preferences gained the running version, and the sentence saying this head does not replace itself and that no DodoSSH server will ever offer one. It reads Updates.CurrentVersion off the same view model the desktop's UPDATES section does, over the null channel that reports itself unsupported. Nothing was needed for the realtime push: it is composed in ServerConnection, which both heads use. Seven tests, all phone-shaped — a group acted on with nothing selected, the menu waved away leaving nothing armed, the ungrouped heading raising none, and the bar's three states. The rectangles remain unmeasurable for the reason phase 8 gives; the checks for them are 8.10 to 8.13 and 13.6.	2026-08-04 19:58:10 +02:00
jaap-jan	e923b12b7f	Merge branch 'claude/desktop-auto-updater-0264a3'	2026-08-04 17:56:50 +02:00
jaap-jan	0b6059d4a8	Stop a late progress report from putting the download bar back ... `Progress<T>` does not invoke its callback inline — it posts to the captured synchronisation context — so a report can be delivered after the download it belongs to has already returned. The unguarded handler then wrote a stale smaller number over the 100 set on completion, and nothing reports again, so it stayed there: a finished download showing 50% next to a banner offering the restart. Caught by running the suite a second time on another checkout, which is the only reason it was caught at all. The assertion that failed is the one that says a found update is fetched without being asked about, and it had passed on every previous run — the ordering it depends on is real and simply usually goes the other way. Guarded rather than made synchronous. The posting is wanted: in the application these callbacks arrive on whichever thread Velopack downloads on, and setting an observable property off the UI thread is a binding exception rather than a stale number. Monotonic is what the bar should have been anyway, since progress that goes backwards is not progress. Ran five times over to confirm it is settled rather than merely rearranged.	2026-08-04 17:56:39 +02:00
jaap-jan	f37d7ee12e	Merge branch 'claude/desktop-auto-updater-0264a3'	2026-08-04 17:53:29 +02:00
jaap-jan	3ead865f01	Merge branch 'main' into the desktop updater, and give way on two numbers ... Main landed a realtime push feature while this branch was building the updater, and the two collided in three places. Every one of them resolves the same way: main got there first, so this branch moves. **Two ADRs were both numbered 0012.** Main's is realtime push; this one is now [ADR 0013](docs/adr/0013-desktop-distribution-and-updates.md). Git did not call this a conflict — the filenames differ — so it would have merged quietly and left the directory with two 0012s and every cross-reference ambiguous. Renumbered here along with the nine places that point at it. **Two manual-check phases were both numbered 15**, and that one git did catch. Main's "Changes that arrive without a timer" keeps 15; installing and updating the desktop client becomes Phase 16, with its checks and every reference to them renumbered. The file's own rule is that a number is for life, which is exactly why the one that had not been pushed is the one that gives way. **The merge rewrote several files with CRLF**, and `.editorconfig` asks for LF on everything except `*.ps1`. That is not cosmetic here: IDE0055 is an error and `EnforceCodeStyleInBuild` is on, so it failed the build on three lines of App.axaml.cs whose only change in this branch was an ADR number in a comment. Forty-six files normalised back to LF; the release script keeps CRLF, which is what `.gitattributes` and `.editorconfig` both already say for a PowerShell file. Nothing else conflicted. The updater does not touch the sync loop or the event stream, and the one file both sides edited heavily — MainWindowViewModel — merged without a hunk in common. Verified after merging: the solution restores locked and builds clean, and 304 shell, 100 layout, 54 session, 28 client-api and 25 contracts tests pass. The first two counts are higher than before the merge because main's own tests came with it and pass alongside these.	2026-08-04 17:52:57 +02:00
jaap-jan	5a8731ca8c	Merge branch 'claude/group-vault-migration-deletion-4af5f4'	2026-08-04 17:11:33 +02:00
jaap-jan	780f4bf892	Merge branch 'main' into the group's move and its deletion question ... Main took the group's EDIT and DELETE off the GROUPS heading while this branch was adding a MOVE beside them, so the conflict was about the same six pixels from both directions. Main's answer wins outright, and it is the better one for the reason its own message gives: a button beside a heading has no card under a pointer to mean, and had to work its subject out from the selection or from the trail. Moving a group had that problem worst of all — the thing it takes with it is everything on the shelf, and "which shelf" is not a question a button there could answer plainly. So the MOVE button is gone and the menu entry it was drawn beside is the whole of it. That entry was already in this branch, above the separator DELETE sits below, and it needed no change: the card menu selects whatever was right-clicked before it runs anything, which is exactly the aiming a group move wants. Three things went with the button. ShowsGroupActions, which main deleted because hiding buttons was all it did, and which this branch had extended to hide them for the move panel as well. CanMoveGroupTarget, which existed to answer whether that button was worth drawing — CanMoveSelectedHost stays, because the phone really does leave the host's MOVE out rather than offer a refusal, and a menu whose entries came and went would be a menu whose items move. And the two test assertions that read them, which were describing the button rather than the behaviour; what they were guarding is that the two panels never share the moment, and IsConfirmingGroupDeletion says that directly. The move panel and the deletion question both keep their place under the heading, which is where the buttons were and is now simply where that section puts things. They still exclude each other, by disarming rather than by a visibility flag: MoveGroup clears a pending deletion and DeleteGroup folds the move panel away. Manual checks 3.3 was rewritten by main for the menu and by this branch for the tick, and now says both; 3.3a is new and walks a two-level shelf across a vault boundary, which is the half of this feature no headless test can watch land.	2026-08-04 17:11:06 +02:00
jaap-jan	6728a0a597	Let the desktop client replace itself, and give the repository one version ... Packaging for Windows, and the updater that only exists once something is packaged. Velopack, win-x64, fed from the project's own forge — never from the deployment a client signs in to, which is ADR 0011 rule 2 carried over unchanged and is why the feed address is a constant in the code rather than a setting. See docs/adr/0012-desktop-distribution-and-updates.md. **Nothing is ever installed while somebody is using it.** A newer build is found on a six-hourly pass, downloaded in the background, and then waits — for a restart the user presses, or for the next launch they were going to do anyway. That is a policy rather than caution: this application argues at length that locking keeps shells running, because a lock that destroyed work would stop being used, and a restart does not keep them. Having taught that, it owes the user the choice at the one moment it stops being true, and the sentence saying so counts the shells it would close. **The version is now derived from the v* tag**, by MinVer, for everything. There was no version before this — no property anywhere, so every assembly reported the SDK's 1.0.0 and the API served that string as its serverVersion to every client that asked. The tag was already the version of record for the container image; this makes it the version of record full stop. MinVer's failure mode is answering plausibly rather than failing, and here a wrong version is a client that never updates, so it is guarded twice: fetch-depth 0 on every checkout, and a step that fails a tag build when the tag and the computed version disagree. **The pack id is DodoSSH.Desktop and not DodoSSH**, which is the one decision here that would have destroyed data. Velopack installs to %LOCALAPPDATA%\<packId> and removes that whole directory on uninstall, and %LOCALAPPDATA%\DodoSSH is where ClientPaths keeps the encrypted cache, the outbox of changes not yet pushed, and the device key. The obvious id would have had the uninstaller silently delete work the server has never seen — the thing the application refuses to do without a counted confirmation. Velopack's own advice to move user data to roaming %APPDATA% is declined for the reason ClientPaths already gives. **Releases are cut by a person, and CI gains no job that could.** The tempting argument is that a forge write token is not a signing key. It does not survive contact with what the token does: Velopack clients trust their feed and do not verify a package signature when they apply one, so whoever can write a release can ship an update every install runs. That is the capability ADR 0011 rule 1 puts on a machine which is not a runner, reached through a different door. The mechanical objection — vpk needs Windows and the runners are Linux — is the smaller of the two and is recorded beside it, because somebody will fix one and believe they are done. Unsigned for now, deliberately and with the cost stated where a user reads it: SmartScreen warns once per person, on Setup.exe, because Mark-of-the-Web is applied by the browser that downloaded it. In-app updates are fetched by the application and applied from a local file, and never trip it. The banner is a fourth row of the window rather than an overlay. Anything drawn in the terminal's rectangle is sliced by the native child window that composites above it — the defect this window has shipped once — and a sibling row is the arrangement TitleBar and StatusBar already prove works. ---- Three defects surfaced on the way, none of them in the feature being built. **A settings key absent from the file came back as the CLR default, not the declared one.** The JSON source generator builds a record through a synthesised parameterised constructor and assigns every property from its argument array, so a property initializer runs and is then overwritten by a default for anything the file did not contain. A settings.json of {} read back a font size of 0, clamped up to the 8px floor rather than the 13px the renderer draws at. It could not bite while there was one setting, because that setting was written on every save and so was never absent; adding a second would have turned automatic update checks off for every existing profile, silently, the opposite of the documented default. Reflection-based deserialisation of the same JSON answers correctly, which is why every way of checking it by hand agrees except the one that ships. The defaults now live on the constructor parameters, which is the only place the generator reads them from. **Declaring a RuntimeIdentifier on the desktop head broke the server's image build.** It is the obvious way to let a self-contained publish restore under locked mode, and it writes a net10.0/win-x64 target into the lock file of every project the head references transitively — including DodoSSH.Contracts and DodoSSH.Crypto, which the API builds too. The Dockerfile restores those with no RID and fails NU1004. Found by running docker build rather than by reading. The RID stays out of the committed state; the two commands that need one ask for it unlocked, and the release script puts the lock files back. **A Docker ARG named VERSION silently sets MSBuild's Version.** An ARG is an environment variable for the rest of the stage, MSBuild reads environment variables as properties, and property names are case-insensitive. With the workflow passing main-<short sha> on a main build the publish died with NETSDK1018 pointing at DodoSSH.Contracts, a project nobody had touched. The build stage's argument is ASSEMBLY_VERSION now, empty except on a tag build. All three are in docs/platform-flags.md, which is where the next person will look. ---- Verified: the whole solution builds and restores locked; 289 shell, 93 layout and 54 session tests pass, including the regression test for the settings defect and a measurement of the banner at the window's minimum width. vpk pack runs end to end and reports "Verified VelopackApp.Run()" against Program.Main. The API image builds correctly both as a main build and as a tag build, carrying 1.0.0 and 0.1.0 respectively. Not verified, and it needs a published release to be: installing, updating and uninstalling on a real machine. That is Phase 15 of docs/manual-checks.md, and the pack id and the WebView2 profile fix are reasoned and commented but only proved by walking it. Two things to watch at the first upload — the reverse proxy's body-size limit for a 64 MB asset, and whether vpk upload gitea is happy with Gitea 1.27.1.	2026-08-04 17:04:41 +02:00
jaap-jan	a86731ee08	Move the shelf as well as what is on it, and ask what a deletion takes ... Two things about a group, and they turn out to be the same argument twice. **A group can be moved to another vault, and it takes everything on it.** The host's move shipped last week and stopped one level too low: moving twenty machines into a shared vault meant twenty trips through a menu, and each one arrived stripped of the group it had been filed under, so the shelf had to be rebuilt by hand on the other side. Moving the shelf is what people were attempting. MOVE sits between EDIT and DELETE over the group cards, and on the card's own menu above the separator DELETE is below — the same place, and the same reasoning, as the host pane's ⋯ entry. **The whole subtree goes, and taking less was never coherent.** A group's children are items of the vault it is leaving, so a parent moved alone leaves them naming a tombstone and they surface as roots in the vault the user has just emptied: half a shelf here and half there, from one gesture that said "move this". The hosts are the same argument and are the half the request was about. The groups go first, top down, and the hosts last. Each item is re-sealed under the destination's key and takes a new id — VaultItemRepository.MoveAsync, which HostGroupRepository now exposes — so nothing pointing at a group can be written until that group has landed and its new id is known, and a child's parent must already be over there. What an interruption leaves is therefore hosts still in the vault they started in, under UNGROUPED: visible, and re-movable. The reverse order would leave hosts in the destination filed under nothing. The parent stays behind and the tags are dropped, which is the host move's rule one level up: both are items of the vault being left, so a reference carried across would resolve on the machine that moved it and dangle for everybody else in the destination. The moved group arrives at the top level, and the panel says so before the press rather than the status line saying it after. Keys and passwords are kept — those genuinely resolve across vaults, and clearing them would take a working host and make one that cannot connect — and any now outside the destination is named, because that is precisely what the other members of it will not be able to resolve. Refused as a whole where anything under the group was written by a newer client, rather than skipped item by item: a move that left behind what it could not re-encode would file some of the shelf in one vault and the rest in the other, which is the state this exists to prevent. Refused with a host editor open, as the drop gesture is, because it rewrites hosts. And the walk carries a visited set, for the reason every walk over this tree does: a group that is its own parent — which two offline clients can build and no editor was ever shown — would otherwise be appended to the move list for as long as there was memory. **Deleting a group now asks what should become of the hosts under it, and that reverses a decision this repository had written down.** The deletion did not touch them: the reference was left dangling, the list resolved it to nothing, and the machines turned up under UNGROUPED. That was right for one of the two things people delete a group for and wrong for the other — a heading being tidied away should leave its machines alone, and a project that has been decommissioned is a shelf and everything on it — and nothing in the code can tell which of the two it is looking at. So it is asked. A tick rather than a pair of options, because the two answers are not equally weighted: keeping the hosts is recoverable and deleting them is not, so the safe answer is the one that needs no decision. It is off on every question, including the one that disarms it, or a tick left standing would destroy the next group's machines on the strength of a decision about the last one's. Once the deletion knows which hosts it means, leaving them naming something that has gone is a state kept for no reason, so the unticked answer writes too: N hosts with the reference cleared, where the ticked one writes N tombstones. That is the N writes HostGroupRepository refuses to hide behind a DeleteAsync overload, made where somebody asked for them and where the count is on screen first. The nested groups take the deleted group's place in the tree rather than being orphaned to the top level. A read-only host is skipped, counted and named, because unfiling it would re-encode a payload this build cannot represent — and the cost of skipping is a dangling id, which every reader here already survives. **Both are the desktop's alone**, and that is not an omission. The phone draws groups as headings in the host list and has never had a way to delete or move one; the two panels take the row of buttons over the group cards, and there is no such row on a 360dp screen to take. One test had to change its premise rather than its assertion. EditingAHostWhoseGroupIsGone built its dangling reference by deleting the group, which now unfiles instead — so it imports a host naming an id nothing resolves, which is what a group deleted on another machine actually looks like and is the only way that state still arises. The picker's placeholder is still needed and still covered. Four places said an item could not be moved between vaults. Two were about a group and were true when written; the other two were left stale by the host's move. All four now say what is true, including the design gaps document, where the chevron beside the vault name stays undrawn for the reason it already had.	2026-08-04 17:04:06 +02:00
jaap-jan	6fad84c484	Merge branch 'claude/remove-group-edit-delete-buttons-3d38ec'	2026-08-04 16:40:56 +02:00
jaap-jan	d02d103569	Take the group's EDIT and DELETE off the heading row ... A pair of buttons sat at the end of the GROUPS heading, and the card's own right-click menu arrived later offering the same two things. Two controls for one job, and the buttons were the harder of the two to read: a button beside a heading has no card under a pointer to mean, so it had to work its subject out — the selected card, or failing that the group the trail ends with, which once a group is open is not a card on screen at all. The menu never has that problem, because opening it is what aims it. The menu is the whole of Edit and Delete on the desktop now. ShowsGroupActions went with the buttons, since hiding them was all it did. GroupTarget stays: the menu's two entries read it after the code-behind has selected whatever was right-clicked, and its fallback to the open group is what makes + NEW HOST open on the group somebody is standing in rather than on none. One case changes shape. Opening a group with nothing inside it folds the card grid away, so from in there nothing can be right-clicked — renaming that group means pressing the trail back one level, to where it has a card of its own. The buttons used to cover it through the fallback. The trail was already the way out of an empty group, and 3.2a says so now. The test that pressed EDIT through its binding is replaced by one holding that no button on the screen commands either of them, which is the failure worth catching: a button coming back is not a compile error, and it would draw itself in place, aimed at the group the trail ends with. What that test covered — a command that has to accept an empty parameter, and act on the card the pointer was on — the two menu tests beside it already do.	2026-08-04 16:39:14 +02:00
jaap-jan	589300253d	Merge branch 'claude/vault-realtime-push-d64c61'	2026-08-04 16:38:42 +02:00
jaap-jan	4b706bc3c3	Say when a vault has moved, so nobody waits out the minute ... The delta pull was cheap enough to run on a timer and the client did, once a minute. That is fine for a machine and wrong for two people: an edit a colleague makes is up to a minute stale, which is long enough for both of them to make it and produce a conflict neither needed to have. Shortening the interval is the obvious answer and the wrong one — it costs a request per client per interval whether or not anything happened, and it converges on a busier server that is still late. So the server now says so. A client holds a WebSocket open at GET /api/v1/events, subprotocol dodossh.events.v1, and gets a line down it when something it can read has changed. ADR 0012 has the reasoning; three parts of it are worth repeating here, because they are what everything else rests on. **What crosses the socket is a notice, never data.** A frame names a vault and how far its change log has got. No item, no ciphertext, not even which item it was. The client's answer is the delta pull it would have run anyway, so there is still exactly one code path that applies a change to a keychain, and it is not this one. Pushing the items themselves would save a round trip and fork that path in two, with the cursor, the merge and the tombstone rules duplicated across both — ADR 0003 put every mutation through one write path for that reason, and this keeps every read on one for the same one. It also makes a dropped notice harmless, which is what lets the fan-out below be as simple as it is. **Polling stays, and is what guarantees a pass.** The minute timer is unchanged. A network that eats WebSockets, a server with Events:Enabled off, an older server, a proxy that will not upgrade, a notice dropped under backpressure — every one of those leaves a client behaving exactly as it did before this commit. Nothing is reachable only over the socket and nothing is meant to become so; VaultViewModel's AutoSyncInterval remark now says that where somebody changing it will read it. **The bearer token authorises the upgrade, unlike the relay's ticket.** Not an inconsistency with ADR 0004: the relay's socket is a byte pipe whose whole authorization decision — which host, which IPs, which port — is made before it opens and never revisited, and it is the extraction seam for a process that must hold no ACL code. This one is a view of the caller's own vault list and has to keep answering "what may this account read" for as long as it is held. A ticket would carry that answer in a token and be wrong the moment the account's access changed. The two bounds that arrangement needs are met rather than waved at: the socket is closed at the token's exp with close code 4401 and the client comes straight back with a fresh one, and the vault set is re-resolved every few minutes as well as on the changes known to affect it. Both bound *metadata*, because a notice contains nothing else and reading a vault still needs a key this server has never held. **The fan-out.** VaultEventHub is a singleton holding the sockets this node accepted; publishing walks them and asks each whether it cares, rather than keeping a vault-to-subscriber index that every re-subscription would have to move entries between under a lock publishing also takes. At a few hundred sockets per node and an event rate bounded by how often people edit keychains, the walk is not measurable and its races are obvious. Per-connection queues are bounded and drop the *oldest*: a notice means "pull vault X, which is at least at sequence N", so the newest subsumes what it displaces and the client's answer is identical either way — which is what lets the publish path be void, never block, and never fail. Announced from the endpoint rather than from SyncService, and that placement is the point: by then the push has committed and released the per-vault advisory lock. From inside it would name a sequence no reader can see yet and would hold the lock that serialises writers across a socket write. Only the highest *applied* sequence, so a batch of pure conflicts announces nothing, and a duplicate — already announced when it first landed — announces nothing either. Grants and membership publish too, and those take the *recipient* rather than the actor. This is what AdmitNewVaultsAsync has been apologising for since sharing shipped — "the recipient is handed nothing, there is no push channel" — and the README with it. A vault shared with somebody now turns up as it is shared. The comment and the README paragraph both say what is true now, and both keep saying that the pass is what *discovers* the vault, because a client with no socket has to arrive at the same place. **On the client**, VaultEventStream is really a reconnection policy wrapped round a ClientWebSocket: a dropped socket is the ordinary case here — laptops sleep, proxies time out, tokens expire, servers are redeployed — so nothing in it treats a failure as exceptional, and every path ends in "wait, then dial again". A connection that lived long enough to say hello resets the backoff, so a laptop that woke, worked, and lost its network an hour later does not inherit a minute-long wait it has already proved it need not take. A 4401 close skips the backoff entirely and asks the token provider again, which is the whole reason that close code is distinct. A server that does not advertise the events feature gets IdleVaultEventStream, which never delivers — so IVaultServer.Events is never null and every caller stays on one shape, because the correct behaviour without a socket is the behaviour with a silent one. The shell's background loop now selects between the timer and a notice, and both waits are held across iterations. That is load-bearing rather than tidy: PeriodicTimer permits one outstanding WaitForNextTickAsync and throws on a second, and an abandoned channel read stays registered and consumes the next notice written. Either defect leaves the first notice working and every one after it silently lost, which is why NoticesKeepWakingTheLoop_NotJustTheFirst pushes three and not one. Notices are coalesced over a quarter of a second, so one person's save — a host and its log entry are two items — and a colleague clearing a folder each cost one pass rather than a dozen. **The kind is a string, not an enum**, and that is a compatibility decision. UseStringEnumConverter throws on a value it does not know, so a newer server sending a kind an older client had never heard of would not add an unreadable frame — it would break that client's socket outright. A string is ignored instead. ProblemCodes is the same shape for the same reason. **Tested on both sides, through the real pipeline.** The endpoint suite opens a genuine socket against TestServer and proves a push produces a notice, that another account's push does not reach it, that a ping is answered, and that a frame this server cannot parse does not end the connection. Two of those assert on *ordering* rather than on absence within a timeout — the stranger's write goes first, so a socket that leaked would have announced it before the one the test waits for — because "nothing arrived in two seconds" is a test that passes on a slow machine for the wrong reason. And ANoticeCarriesNoCiphertext asserts on the bytes that crossed the wire rather than on the record's fields, since the latter would only prove that this type has no payload member, which is a tautology; the former is what catches a field added later without anybody thinking about disclosure. The client suite drives VaultEventStream through an injected connector, because the one thing a test cannot do to a real network is make it fail on cue — and failure is the entire subject. The shell suite proves a notice produces a pull inside ten seconds against a sixty-second timer, so the timer cannot be what caused it. **Two limits, stated rather than left to be discovered.** Fan-out is in-process, so a deployment running more than one API replica only pushes for writes its own replica handled and the rest arrive on the timer. IVaultEventPublisher is the seam a PostgreSQL LISTEN/NOTIFY backplane implements and it is deliberately not implemented: an untested backplane is worse than a documented gap, and multiple replicas degrade to the behaviour before this commit rather than breaking. And a client is notified of its own writes; it pushed, so it already pulled, and the extra pass finds nothing. Suppressing that echo correctly needs a per-device identity on the socket, and the same user's other machines must still be told. Manual checks phase 15 covers what no test here can reach, which is the network in between: a proxy that will not upgrade, one that drops an idle socket without telling either end, a laptop lid, a token expiring. Every one of those is invisible inside a test host, and every check there passes only if the change arrives quickly *and* still arrives with the socket taken away. ADR 0012 also fixes one thing about the shared terminal session this is the transport for, so it need not be renegotiated later: session data will be binary frames on this same socket, because base64 in a JSON envelope is the wrong shape for the one payload here that is continuous rather than occasional. Two questions it explicitly does not answer by implication — whether those bytes go through the API at all, and what end-to-end encryption means when the second party watches a stream rather than holding a key — are ADR 0001 questions and get their own decision. 1512 tests pass. DodoSSH.SystemTests was not run — it needs the whole compose stack — so the end-to-end path is unverified for this change beyond what the manual checks describe.	2026-08-04 16:37:41 +02:00
jaap-jan	742f65c204	Merge branch 'claude/card-selection-state-sharing-f9348c'	2026-08-04 16:28:34 +02:00
jaap-jan	0258ec3029	Merge branch 'claude/groups-vault-sharing-e4b154' ... # Conflicts: # src/DodoSSH.Client.Shell/ViewModels/VaultViewModel.cs	2026-08-04 16:28:29 +02:00
jaap-jan	3f419cb19b	Let one card be selected at a time ... The hosts screen draws two grids, one above the other, and each is a ListBox with a selection of its own. Nothing joined them, so a group card and a host card could be lit at the same moment — two chosen things, under two pairs of buttons of which only one would act on whichever the eye had settled on. Both grids mark a selection the same way, so there was nothing on screen to say which of the two the next press belonged to. They share one mark now. Selecting a host clears the group and selecting a group clears the host, and that second one takes the detail pane down with it: a pane about one machine cannot go on standing beside a marked group, because nothing on it would be about what is selected. Losing a selection deliberately clears nothing. A null arrives whenever either list is rebuilt — every keystroke in the filter box and every background sync — and treating that as somebody deselecting would take the mark off a group card because a search emptied the grid beneath it. The reload needed the same guard for the same reason. It falls back to the first host when nothing is selected, which is what puts a target under CONNECT on a fresh unlock; with one mark between the two grids that fallback would have unselected a group nobody had touched, once a minute. It is skipped while a group holds the selection, and it moved into a method of its own because the comment saying why pushed ReloadHostsAsync past sixty lines. One consequence needed handling rather than accepting. The phone's only route into the group editor is a button on a heading in the host list, and it worked by selecting the group first — which under this rule takes the highlight off the machine somebody was about to connect to, on a screen that draws no group cards to say where it has gone. EditGroup takes the group as an argument now: the heading passes its own row, and the desktop's button beside the cards passes nothing and still means "the card that is selected". What the tests hold is the half that lives in the controls. Clearing the property has to reach the list that is drawing the card, and a selection nulled in the view model while the card stays highlighted is the exact failure this is about — so the rule is driven on the real screen, in both directions, against the ListBoxes' own SelectedItem. The desktop's EDIT is pressed through its binding for the same kind of reason: a command refusing the button's empty parameter would be a button that never fires, and nothing about the markup would say so.	2026-08-04 16:27:20 +02:00