The page shows which keys exist, what the agent is holding, and the hosts this
machine has met, with a two-press forget for a host whose key has changed.
Nothing here reads private key material. Fingerprints and comments come from the
.pub file, and "does this key need a passphrase" is answered by asking
ssh-keygen to derive the public half with an empty one -- it succeeds for an
unencrypted key and fails for an encrypted one, and either way the only thing it
can emit is public. The contract checks that against the payload that actually
reaches the page rather than against the source, because what the code intends
and what it ships are different claims.
Unloading a key from the agent is refused, with its reason. On this desktop
`ssh-add -d` prints "Identity removed" and the key is still offered a second
later: gnome-keyring's agent lists every key it finds in ~/.ssh, so a removed
one comes straight back off disk. That was measured rather than assumed -- a
plain ssh-agent removes durably, this one does not -- and a button reporting
success while changing nothing is worse than no button. The page says so and
names the thing that does work: move the file out of ~/.ssh.
SSH_AUTH_SOCK is not set in a normal shell here, so a naive check reports "no
agent" while one is plainly running. The helper falls back to the keyring
socket, and an agent started by hand still wins. That gap is the same one that
made reaching these servers awkward in the first place.
Generating a key is deliberately absent. A passphrase cannot reach ssh-keygen
without going somewhere it should not -- -N puts it in argv, which every process
on the machine can read -- and driving the prompt over a pty did not work.
Offering to generate an unencrypted key instead would be a downgrade dressed as
a feature, so the page does not offer to generate at all.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
gapsIn and gapsOut were the only two compositor settings the write sweep had
never verified: Hyprland answers for them in CSS shorthand, "5 5 5 5", and the
sweep had no way to compare that. The preference behind each is a single int
that Hyprland expands to four sides, so a uniform reading compares exactly. A
non-uniform one is not something the preference can express, and is skipped
rather than collapsed to a number it never wrote. 63 of 63 verified live now,
none skipped.
Wallpaper thumbnails are cached. The report that five of them sat at "Loading…"
was a screenshot taken 1.1 seconds after the page opened -- decoding one of
these at tile size takes between 1.2 and 2.6 seconds and about ten start at
once, which the code already said. Measuring it did turn up something real
though: without a cache, scrolling back up pays that decode again for every
tile. The tradeoff is a wallpaper replaced in place showing a stale thumbnail
until restart, which is worth it for a directory of files that are added rather
than edited.
A chord already in use is now named rather than taken: "Super+Q is already
Terminal". Two actions on one chord means whichever Hyprland reads last wins,
which is not a thing to find out later by pressing it. Rebinding a shortcut to
the chord it already holds is correctly not a conflict.
Also: Open Appearance lands on the Windows tab now that the page has tabs,
Storage points at reclaimable container space, and a dock row shows its desktop
id only when two pinned applications share a name -- it is developer text, and
repeating it under fifteen recognisable names made the list harder to scan.
Written down because it cost the shell: QML has no default parameter values, and
`function openSettings(page: string, section: string = "")` fails the entire
configuration rather than the one function -- so the bar and dock went with it,
and 43 contracts failed at once pointing at the same line. qmllint --bare passes
that, which is why the usual check before touching the running shell did not
catch it. openSettingsSection exists as a separate function for that reason.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Every container on this machine is created by rootless podman-compose and
labelled with the project it belongs to, so the grouping is read from the
labels rather than invented. State then decides prominence within that
grouping -- running containers get rows, stopped ones collapse to a line --
which is why neither axis had to be chosen over the other.
Acting on a stack uses plain podman over the labelled set, never
`podman-compose down`, which would remove containers this shell did not
create. The compose file is the source of truth for what exists and belongs
to the repository. Nothing here needs privilege.
The findings on top are the crossing the Firewall page reports, seen from the
side that can close it: the firewall knows only that something is listening,
while this page knows which container, which compose file, and which token is
missing from it. So `bind-local` prepends a loopback address and leaves the
line byte-for-byte -- variables, quoting and style intact -- then re-parses and
rolls back unless exactly those ports moved. It refuses anything ambiguous
rather than guessing. Rewriting the mapping to the port podman reports today
would have deleted the ${POSTGRES_PORT} indirection that makes it
configurable at all.
Unused volumes are read from podman's own dangling filter. The first version
used MountCount, which is a runtime lock counter and not a usage signal: it
reads zero for a volume a running container has mounted this second, so
"remove unused volumes" offered to delete the live Command Center database.
The cross-check against `podman system df` is what exposed it. The contract
reintroduces that bug deliberately and fails if the guard does not catch it,
because a guard nobody has seen fail proves nothing.
Every mutation in the contract runs against a stubbed podman. Nothing in the
suite starts, stops or removes a real container, image or volume.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Listing zones and services is what firewall-cmd already does. The
question it does not answer needs both halves at once: a port is
reachable only when something is LISTENING on a network address AND the
firewall permits it.
On this machine that crossing is the whole story. The rules look
unremarkable -- one zone, three services, a port range -- and what they
mean is that PostgreSQL and Redis, published by rootless containers on
every interface, are reachable by anyone on the network. Neither half
says that alone, which is exactly how a tidy rules list coexists with an
open database. Nothing was misconfigured: Fedora's default zone met
podman's default publish behaviour.
Ephemeral client sockets are excluded. A browser's outbound UDP port is
indistinguishable from a service in ss, and listing twenty of them
buried the two rows that mattered.
Closing the port range names what it would cut off, by service, before
doing it, and removing ssh says so when someone is connected over it.
Rich rules are shown and never edited: a syntax is not a setting, but
hiding it would misrepresent the configuration.
The contract needed a recorded firewall, and the reason is worth
keeping. The rule this page exists for cannot be tested against this
machine -- its zone permits everything above 1024, so "listening" and
"listening and permitted" give identical answers, and a blocked listener
needs a port below 1024, which needs root. With the crossing deleted,
the contract passed. It now runs against a fixture where two listeners
are blocked, and catches it.
Also here: polkit response files are written 0600 rather than at the
default mask, the agent sweeps requests left by an instance that did not
exit cleanly, and the write sweep waits for its harness to be ready
instead of reporting the startup race as settings that failed.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Live first, because unlike every other page here this one has a live
dimension: card temperature, power draw, whether Game Mode actually
engaged. It polls only while it is open, since a settings page nobody is
looking at has no business waking the CPU.
The part that makes it Panama's page rather than a gamemode config
editor is the hook. gamemode runs a script when a game asks for it and
another when the game exits, so the power profile switches to
performance and notifications go quiet for exactly the duration of a
game -- and afterwards both go back to what they WERE, not to a default.
A Do Not Disturb someone set by hand survives a game; a power profile
someone chose is restored rather than replaced. Verified against real
gamemode activation, not merely by calling the hook.
Two things the page reports rather than hides. Game Mode's headline
trick is switching the CPU governor to performance, and this machine
already runs performance, so it says so instead of implying it helps.
And Proton builds are listed but never chosen: Steam picks the runtime
per game, and a control here would claim an authority this page does not
have.
The hook first called a notifications function that did not exist, and
the one that did was a TOGGLE -- the wrong primitive entirely, since
toggling at game start would unsilence notifications that were already
silent. The shell gained an explicit setter and reader.
search-routing-contract kept its own hand-written list of every page,
which made adding one fail as "not a known page" -- a sixth place to
register a page and a sixth chance to forget. It now derives the mapping
from the shell, which already knows it.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Three sources that fail independently, so they are counted and applied
separately: a flatpak mirror being down says nothing about whether a
kernel security fix is waiting. Blending them into one number would hide
exactly the case that matters.
Checking costs about nine seconds, which is too long to spend every time
a page opens, so the page opens on the last result and says when it was
taken. A first visit with nothing cached goes and finds out rather than
showing a confident "up to date" it has no basis for.
Installing packages takes a snapshot first, named after what is about to
happen, so Snapshots shows "before 32 package updates" rather than a
timestamp. Best effort: a machine without snapper still updates, because
an update that refuses to run when a nicety fails would be worse than
one without a restore point.
Automatic updates cover applications only, through a Panama-owned user
timer running daily with a randomized delay. Packages still ask, and
dnf-automatic is reported as absent rather than offered, because
installing software is not a settings action.
Health gained a check, and that is where the bug was: it first returned
status "degraded", which is not in the doctor's vocabulary of ok,
warning, error and unconfigured. It was counted as nothing at all while
the summary still said healthy -- the same silent no-op this codebase
keeps relearning. A contract now asserts every status a check can return
is one the doctor counts, and the doctor's own contract knows about the
new check rather than failing on its arrival.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
The machine already had snapper running hourly on btrfs, so the tool was
never missing. What was missing is that snapper's only configuration
covered / -- and /home is a separate subvolume with no configuration at
all. Six hundred and forty-three snapshots existed and not one of them
contained a document. Anyone reaching for file history would have found
their system and none of their files.
/home now has a configuration on the same hourly timeline, with
deliberately conservative retention: Steam's 1.2 TB lives on that
subvolume and churns on every game update, so keeping five hourly and
seven daily bounds what those updates can pin.
Per volume, because on this machine "one is covered and the important
one is not" was the news, and a timeline opening on system snapshots
would have buried it. Inside a volume the timeline is the familiar view:
points in time, newest first, each openable as a folder tree to take a
file out of.
Restoring sets the current version aside as .before-restore-N rather
than overwriting it. A restore that destroys the thing you were about to
compare against is how someone loses the work they were trying to save.
Rollback is deliberately absent. snapper's rollback changes the btrfs
default subvolume, and this system's fstab pins subvol= explicitly,
which overrides it -- so a rollback would report success and change
nothing after a reboot. A recovery feature that silently does nothing is
worse than not having one, and making it work means editing fstab and
the bootloader, whose failure cannot be repaired from inside the
desktop.
Per-snapshot size is reported as not measured, because measuring it
needs btrfs quotas that cost performance on every write. Free space is
shown instead, which is the number that decides anything.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
The fourth panel this desktop handed to GNOME Settings, and the last one
worth owning.
Driverless only, deliberately. Adding a printer that describes its own
capabilities over IPP is supported; choosing a PPD or fetching a vendor
driver is not, and the page says so rather than pretending. That
restraint is the whole design: a wrong driver produces a printer that
accepts jobs, reports success, and prints nothing, which is the worst
failure this page could ship because it looks like it worked. A printer
old enough to need a PPD stays a job for the system printer tool.
Printers and the queue are separate cards because they answer separate
questions. Which printers exist is one; where a document went is the
other, and it is the one that actually brings someone here -- so the
queue is a single list across every printer.
Device URIs are validated by scheme before reaching CUPS, whose backends
run as root. file: and pipe: do not lead to a printer and are refused
here rather than further down.
This machine has no printer, so the page was built against a temporary
CUPS queue that was created, exercised through the service, and removed;
the service was confirmed to observe the removal rather than merely
perform it. Discovery and the driverless add path are verified by their
refusals rather than against hardware.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Two of the panels this desktop still handed to GNOME Settings.
Users manages the account through accountsservice -- the same daemon
GNOME's panel drives, so a name or picture set here is what the login
screen and lock screen read. Name, picture, account type, password,
automatic login, and adding or removing other accounts. Every change is
authorized by polkit through the agent this session already runs; a
dismissed prompt is a normal outcome and says so.
A new password is read from the helper's stdin, hashed by openssl
reading its own stdin, and handed over D-Bus from inside that process.
It is never an argument: argv is world-readable through /proc, so a
password passed that way is published to every process on the machine.
Removing an account takes two presses and says it destroys their files;
the last administrator cannot be removed or demoted, because a machine
nobody can administer is not a state to offer.
Sharing reports what is actually true, including "the software for this
is not installed" -- the honest answer for Samba here, and the case the
panel it replaces shows as a switch that does nothing. Password sign-in
is reported from sshd's configuration rather than assumed: claiming
"keys only" when the file is silent would state a security property that
cannot be backed up.
The Control Center now draws the account's real picture and name. A
generic glyph sat there while a real avatar was already set, which made
the desktop look like it did not know whose it was.
Also here: the KDE Connect contract no longer requires a phone to be
awake. kdeconnectd drops its device objects for a phone it has not seen
recently while the pairing survives in its config, so demanding one
failed whenever the phone was off.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Nothing showed what was using the drive, and removable media was handled
by a tray helper with no surface in Settings at all.
One scroll rather than tabs: space above, the device below. Every other
settings page is a scrolling card stack, and a tab would not be
deep-linkable from the launcher command or from search.
Three things the page has to get right, each now pinned by a contract,
because each is a way it could quietly lie. / and /home are one btrfs
filesystem sharing one pool of free space, and a page that copies df
shows double the free space that exists. zram is a block device and is
not storage; counting it as a drive overstates this machine by 8 GB.
Unmount and eject refuse anything not on a removable drive, because the
UI is what asks and a UI can be wrong.
The cheap read -- layout, usage, health -- runs when the page opens, at
around 90ms. Measuring what is filling the drive means walking every
file, so it happens on request and says so rather than showing an empty
list that reads as "nothing here".
Partitioning and formatting are deliberately absent. A settings pane is
the wrong place to put erasing a disk two clicks deep; the page opens
GNOME Disks for that.
Adding the page found a fourth hard-coded page list in ShellState. A
page missing from it does not error -- openSettings() falls back to
"home", so the launcher opens the wrong page and logs nothing. A
registry contract now holds the three lists together.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
GNOME Online Accounts is a daemon plus a D-Bus API, and the daemon
already runs in this session -- gvfs activates it, and all four accounts
on this machine work without gnome-shell involved anywhere. Only the
PANEL was GNOME's. The accounts themselves are ordinary D-Bus objects
that anything may read and modify.
So everything except the initial sign-in is now native: the account
list, per-service toggles for mail, calendar, contacts, files, photos,
music and chat, and removal. That is the whole Online Accounts panel
apart from one OAuth handshake.
Signing in is the exception, and only for OAuth providers. The daemon's
AddAccount takes credentials as an argument -- it stores them, it does
not obtain them -- and the code that runs Google's OAuth exchange lives
in libgoa-backend, which Fedora ships without a GIR binding, so it is
reachable from C only. Reimplementing it would mean our own Google
client credentials. That step is handed to GNOME's panel and the page
says so, because a hand-off the user does not expect reads as a bug.
Password-based providers (Nextcloud, IMAP, WebDAV) could be added
natively later; their credential keys are known now.
Accounts needing re-authentication are surfaced first, which turned up
something immediately: both Google accounts on this machine report
attention_needed, meaning their tokens have expired and they have
stopped syncing. GOA has known that all along and nothing outside its
own panel ever said so.
Every write re-reads the account list rather than assuming it landed.
GOA can refuse, and a toggle that springs back is the honest outcome.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
More GNOME Settings parity.
Region & Language is new. The locale is localectl's, and Panama stores no
copy of it -- there is exactly one system locale, so a preference here
would be a second source of truth that drifts the moment anything else
changes it. Codes are resolved against iso-codes into "Portuguese
(Brazil)" the way GNOME does, with the code kept visible because it is
what actually gets written and someone choosing between two Spanish
variants needs to see it. Changing it is privileged and only applies to
programs started afterwards, so the page says a sign-out is needed
rather than claiming the new language is in use.
The service is called SystemLocale, not Locale: QML has a built-in
Locale value type that silently shadows a singleton of that name, and
every binding then reads properties off the wrong thing. The page
rendered empty with nothing but "cannot read property of undefined" to
explain it.
About now answers what GNOME's About answers -- model, processor,
memory, disk, OS, kernel, windowing system -- where before it listed
only Panama's own component versions. Graphics is joined from
GraphicsDevices rather than read again, because two readouts of the same
hardware are two things that can disagree. Placeholder DMI strings
("To Be Filled By O.E.M.") are filtered out, and unreadable facts are
omitted rather than shown as "Unknown".
The Fedora hand-off card was one row listing five subjects that opened
the network panel regardless. Naming a panel and then not opening it
reads as a broken button rather than a deliberate hand-off. Each subject
now opens the panel that owns it, and openGnomePanel takes an optional
subpage so "Users" reaches System's users page the way GNOME's own
desktop entry does, instead of dropping the user on System's front page.
Printers and online accounts are not repeated here; they stay with the
network hardware on Network & Devices.
One bug this surfaced, caught by the hyprland write contract: the Lua
config key and the hyprctl option name genuinely differ for tap to
click. hl.config wants input.touchpad.tap_to_click; getoption answers to
input:touchpad:tap-to-click. Either spelling used for both fails -- a
hyphen is not a Lua identifier, and the underscored name is not a known
option -- which is what the schema's two separate fields are for.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Continuing towards GNOME Settings parity. GNOME's Privacy panel covers
screen lock, camera and microphone access, file history, trash, and
device security; Panama had no equivalent page at all, despite already
tracking camera and microphone use for the bar indicator.
Device security is a new read-only readout: Secure Boot, TPM, disk
encryption, SELinux mode, and the firewall. None of these is a
preference -- they are set in firmware, at install time, or by system
policy, and a switch offering to change them would either fail or do
something far-reaching from a control that looks like every other
control. What it answers is "is this machine set up the way I think it
is", which otherwise takes five commands and root. Facts that cannot be
determined report Unknown rather than guessing, because a security
readout that quietly says "fine" when it failed to look is worse than
no readout.
File history and trash retention are deliberately NOT offered as
switches. They are GNOME preferences enforced by gsd-housekeeping, which
does not run in a Hyprland session -- verified, it is not running here.
Toggling them would store a preference, change nothing, and give no sign
of it. They are delegated to GNOME Settings by name instead.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Working towards parity with GNOME Settings, which splits pointing
devices into their own panel. Panama had pointer speed and focus-follows
buried under a page called "Input & Shortcuts", and had nothing at all
for scroll direction, acceleration profile, scroll speed, left-handed
buttons, or any touchpad setting -- all of which could only be changed
by editing hypr/input.lua by hand, which is the thing this app exists to
stop.
Every new mapping was read back off the running compositor rather than
assumed, and two were not what they look like: touchpad drag lock is an
int with three states, not a switch, and scroll factors are floats even
at their default of exactly 1. Getting either wrong makes every write to
that setting look rejected. The shape contract now covers 35 mapped
options, up from 23.
The touchpad card renders only when a touchpad is attached, which is
what InputDevices is for. On a desktop it would be worse than useless:
every switch on it would appear to work, because the preference is
stored and Hyprland accepts an option for a device class it has no
member of, so the settings would silently affect nothing.
Keyboard layout was read-only text, justified by a note saying changes
needed a compositor reload. That is not true in 0.56.2 -- setting
input:kb_variant through hl.config re-keymaps attached keyboards
immediately, verified by watching active_keymap on a real keyboard
change to "English (US, intl., with dead keys)" and back. So layout,
variant, and options are now real controls, joined by a TextEntryRow
that commits on Enter or focus loss rather than per keystroke, since
half a layout name is a valid string meaning something else.
Rejected input is shown as rejected rather than sanitised: these strings
are serialised into an hl.config payload, where stripping an unexpected
character would turn a typo into a different working setting.
Verified each new pointer option applies and reverts against the live
compositor. Schema, search, commit/reset, and system contracts pass.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Wires the schema entries and routing the codex agent needs for default
applications, weather, vitals refresh, notification timing, and capture,
so its pages can be written against keys that already exist.
Capture directories and encoder arguments are enums rather than free
text. Both are handed to a recorder process, and a settings page has no
reason to expose an arbitrary string there.
ApplicationsPage.qml is a placeholder so the page id can be routed,
registered, and searchable before the real page lands. It is owned by
the other agent and expected to be replaced wholesale.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Continues the settings expansion toward replacing GNOME Settings for
everything Panama actually owns.
Wallpaper. A thumbnail grid rather than a path field: the value of this
setting is the picture, so typing a path to something you cannot see is
the worst version of it. Two things about hyprpaper 0.8 shaped this. Its
IPC is much smaller than older documentation suggests -- preload,
listloaded, unload, and reload all answer "invalid hyprpaper request",
so setting is a single call with no preload. And the "<empty>,<path>"
form that used to mean every output is silently ignored, so a wallpaper
set that way appears to succeed and never changes; outputs are walked
explicitly instead. hyprpaper.conf lives in the repo through the
~/.config/hypr symlink and so cannot hold machine state, which is why
the choice lives in the shared settings store and is re-applied at
startup.
Power & Lock. hypridle has no IPC for reconfiguration and its config is
hyprlang rather than the shared JSON, so scripts/panama-idle generates a
config from the settings store and restarts the daemon. The generated
file lives under XDG_STATE_HOME for the same symlink reason, with a
systemd drop-in pointing hypridle at it. Management is a real state and
the page says which one you are in rather than showing sliders that
quietly do nothing. Zero means never for all three timers, which a naive
template would render as "immediately".
Date & Time. Deliberately not stored in Panama's settings: the timezone
and network time belong to the machine and are shared with sessions that
never see this file. Storing a copy would create a second answer to a
question the system already answers. Reads and writes timedatectl
directly; a cancelled polkit prompt surfaces as an error rather than as
a value that appears to have been accepted.
Accessibility. Pointer size and text scale have to agree across three
consumers with no shared configuration -- the compositor, GTK, and the
shell -- so the store is the source of truth and the values are pushed
outward to gsettings and hyprctl setcursor.
Search now indexes the schema instead of the twelve page labels. "gaps",
"wallpaper", and "screenshot" previously found nothing on an app that
has all three, which is the clearest way a settings app feels smaller
than it is. Shortcuts are indexed by what they do. A contract asserts
every non-internal schema label is reachable, so a new setting cannot be
added in an undiscoverable state.
The GNOME delegation allow-list was widened to the panel names
gnome-control-center actually reports; the previous list contained
"users", which is not one of them and so opened nothing.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Panama had grown into three configuration surfaces that only agreed because
they had been typed to agree: looks.lua hardcoded values, DesktopPreferences
independently defaulted the same values, and SystemSettings replayed them at
startup. Nothing kept them in sync, and the Lua side read no shared state at
all. This lands the first three stages of docs/superpowers/plans/2026-08-17-panama-cohesion.md.
Fix silently failing Hyprland writes. On a Lua-configured Hyprland, hyprctl
keyword refuses the write, prints the refusal to stdout, and still exits 0, so
the HDR, VRR, and direct-scanout toggles persisted their value and reported
success while the compositor never changed. Writes now go through hyprctl eval,
which has the same hazard on syntax and runtime errors, so success is defined
as reading the value back and finding it equal. The existing contract passed
throughout the outage because it re-applied the values already in place; the
new one flips each value to something it does not hold.
Derive preferences from a schema. Every setting used to be restated four times
-- a property alias, a JSON adapter property, a change handler, and a line in
reset -- where omitting any one failed silently. PreferenceSchema.qml is now
the single source, and persistence, validation, reset, and the Hyprland mapping
all derive from it. Unknown keys on disk survive a write so a rollback does not
discard a newer build's settings, and a corrupt file falls back to shipped
defaults. The store moved to ~/.config/panama/settings.json, migrating from the
old state directory without deleting it.
Share that file with Hyprland. prefs.lua reads it at config time with every
shipped literal kept as the fallback, so the config still stands alone. The Lua
is the default, the JSON is the truth, and Settings is the editor. The
compositor-adjustable surface goes from 3 keys to 23.
Also fixes two test-hygiene bugs found by running the suite end to end for the
first time: settings-pages-contract could see the window settings-window-contract
leaves behind, and the new write contract was persisting its deliberately-wrong
values into the user's real store.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Gabriel Brown