OpenAI ships an official Linux RPM now, so the community wrapper goes away:
`panama app chatgpt-desktop` built codex-desktop from the upstream macOS disk
image and ran a local rebuild daemon to keep it current, and the official
package comes from a repository that upgrades with everything else. The app
file, the help example and the dock's pinned id all move over, and a migration
replaces the build on machines that already have it -- official package on
before the community one comes off, so a failure part-way still leaves an app.
The install itself does not follow upstream's instructions. Those are "download
this RPM and install it", and the RPM's own root scriptlet is what writes the
repository file and drops the signing key into /etc/pki/rpm-gpg -- so root runs
an unverified download and then learns from it what to trust. That is the shape
the repository audit forbids: no network response is executed as root without a
verified digest or signature first.
OpenAI publishes no key and no fingerprint anywhere an install could fetch and
check them, so the key is pinned here instead. setup/keys/ carries it and says
where it came from, including the honest part -- this is trust established on
first use and then held, not trust verified against the publisher. setup/lib/
chatgpt-package verifies that copy's fingerprint, installs it, and writes the
repository with gpgcheck and repo_gpgcheck on before anything is installed, so
dnf checks the metadata signature and the package signature itself. It is byte
for byte the repository the scriptlet would have written, so nothing churns
afterwards, and every later upgrade goes through the same key. Both callers use
it; a verification failure skips ChatGPT rather than installing it anyway.
The contract proves the pinned key is the key the library names, that a
missing, unreadable or mismatched key writes nothing at all, that what is
written actually turns the checks on, and that neither caller hands root a
downloaded RPM.
Claude-Session: https://claude.ai/code/session_017zzbtfnMLoYrB8WesqANFY
A machine's role is now the interview's first question and the one answer
Panama records. Servers get the same shell minus the screen: core packages,
nvm, Bun, Claude Code and Codex (desktops get Codex too), linger, rootless
ports from 80, firewalld, the nginx-bridge network, and a nightly image
updater that replaced watchtower for cause.
server/containers/ carries junior's 23 compose services -- secrets moved to
per-machine .env files that never enter this public repo, every transformed
compose proven to render byte-identical to what is live. 'panama server'
enables, disables and relinks them; nothing here restarts a running service.
'boot --server' walks a fresh VPS from its root login to a normal install.
Five new contracts pin the secrets rule, the catalog's shape, panama-server's
behavior, the role plumbing, and the dotfile classification.
Claude-Session: https://claude.ai/code/session_01NU5JGiN3JfzqrLQB6wmJ1E
Desktop & Dock becomes Shell — Bar, Dock, Control Center, Tiling,
Workspaces — the home for everything Quickshell draws. The settings-
management cluster moves to System as Sync & Backup, Appearance's
Shell tab dissolves, and 24-hour time finally lives on Date & Time,
which always owned it.
The bar gets what it never had: a way to survive the wallpaper. A
second neutral text family (follow theme, or forced light or dark),
a one-layer shadow under every glyph, and a gradient scrim for
wallpapers nothing else survives — all off by default, pixel-identical
until asked. Widgets earn toggles (weather, media, clipboard, calendar
countdown), the vitals cluster stops leaving a dead pill behind, and
Control Center's sections learn to step aside.
The dock graduates from MVP: a context menu with window rows, pin,
unpin, quit and new-window; scroll an icon to cycle its windows; drag
to reorder on the dock itself; hover previews with one-shot captures;
and "Add App to Dock" in the launcher. Three real bugs died en route —
menus that slid away with the autohide, a readonly-property crash on
every menu open, and a drag that drifted half a slot per icon on side
docks. The pinned-apps editor in Settings becomes a drag strip.
166 contracts; the full suite is green except two live display and
switcher tests that cannot run behind a locked session — re-verified
on unlock.
Claude-Session: https://claude.ai/code/session_01Ms2FbjQy31TVf3CEvQhGM8
Appearance now opens on Themes: light and dark side by side, each
remembering its own choice, over galleries of ten shipped themes —
Tokyo Moon and Day joined by Moon Rose, Catppuccin, Nord, Gruvbox and
Everforest in both modes. A theme is a complete palette: the catalog
lives in themes.json, Theme.qml reads every color token from the
active record, and one render pipeline carries it to kitty, tmux,
btop, GTK, Vicinae, Firefox's chrome, and the lock screen. The Theme
editor builds new ones from four wells — wheel, hex, or eyedropper —
with derived surfaces, a saturation slider, debounced fine-tune, and
effects that save with the theme. Custom edits finally keep GNOME's
accent, kitty's border, and hyprlock in sync.
Wallpapers can be video: mpvpaper per output, hardware-decoded, muted
and looped, supervised and respawned. Panama owns the pausing — games,
battery, and a bar pill for right now — because the compositor
rebuilds full-screen blur for every frame a video wallpaper draws.
The lock screen gets a still frame.
Titlebars stop lying. GNOME apps get close-only on your chosen side,
the maximize and double-click settings are gone, the Settings window
obeys the same rules, and its titlebar can be turned off entirely.
Typography becomes five labeled dropdowns instead of a wall of
samples.
Contracts updated and written throughout (165 now); per the redesign
workflow none were executed — the full sweep runs once at the end.
Claude-Session: https://claude.ai/code/session_01Ms2FbjQy31TVf3CEvQhGM8
Home is now Overview | My Home | Phone. Overview leads with quick-action
tiles (focus, Do Not Disturb, health, snapshots, storage), keeps the
findings card — updates fold in, the reclaim-space prompt is gone on
purpose — and adds glance cards, the next calendar event, and weather.
My Home groups every light by Home Assistant area: the helper gained an
`areas` command (one REST template render, no websocket), and the rooms
degrade to a flat list on setups without areas. The favorites editor and
connection card moved intact. Phone gains a vitals strip — battery and
cell signal read from KDE Connect's plugin D-Bus objects, where absence
is data, not an error — beside ring, clipboard, send-a-file, and the
BlueBubbles handoff.
The retired home-phone id resolves to my-home forever via a new alias
map in SettingsRoutes (with a hasOwnProperty guard so prototype names
cannot leak into settingsPage). Storage no longer claims 0 B free — the
old page read a field the disks helper never emitted.
Contracts updated alongside; per the new workflow, the full suite runs
once at the end of the redesign (see the test backlog note).
Claude-Session: https://claude.ai/code/session_01Ms2FbjQy31TVf3CEvQhGM8
A sweep for controls wired to targets the backend does not handle -- the shape
of the dictation bug -- found one more: "Start Orca" on the Accessibility page
called openApplication("orca") with no "orca" in the command map, so it set an
error and launched nothing. orca ships in hyprland-packages; it now runs.
That is the whole count. Every service-method call across 84 settings pages
resolves, every helper subcommand a service invokes is implemented, every
openGnomePanel handoff is allow-listed. Two dead buttons existed in the entire
settings app -- the dictation Download and this -- and both are fixed.
settings-buttons-contract pins the class: every openApplication id and
openGnomePanel panel a QML button passes must be present in SystemSettings'
dispatch maps, both of which return false silently on an unknown name.
Verified it catches the orca button when the fix is reverted.
And a dependency-contract exception the ffmpeg fix needed: ffmpeg is provided
by a swap (ffmpeg-free -> ffmpeg), never a list entry, because listing it is
the conflict that fix removed -- so panama-transcode's use of it is satisfied
without a package name to point at.
Claude-Session: https://claude.ai/code/session_01Epx9ZC1gwm81K3jm9x9CKh
Logging out of plain Hyprland left the desktop running. autostart.lua starts
about ten user units, none of them children of the compositor, and the shutdown
handler that was meant to stop them stopped hyprland-session.target -- a unit
Fedora does not ship. systemctl says 'Unit not found', the handler reports
nothing, and hyprpaper, hypridle, vicinae and the polkit agent keep running
after logout. Each of them Requires=graphical-session.target, so that target
stayed active with no session behind it.
The symptom was not a broken Hyprland. Both uwsm and gnome-session refuse to
start into a session that is already running, so 'Hyprland (uwsm-managed)' and
GNOME both bounced back to the login screen, and plain Hyprland -- which makes
no such check -- was the only one that worked. The session that looked healthy
was the one leaving the mess, and it was the one Panama tells people not to use.
Now stops graphical-session.target, which every one of those units is PartOf,
so a unit added to the start handler later cannot be forgotten here.
panama-crash-watch had the same leak and no PartOf at all: a journal follower
per login, and a 'once per program per session' dedup that had quietly become
once per boot.
The audit's first tier, in one change: every case found where the interface
asserted something the system did not do.
Twenty-one compositor-owned preferences -- the whole Mouse & Touchpad page,
plus window layout, snapping, dim-inactive and the magnifier -- had the live
half (hyprctl eval) and not the config-time half, so they quietly reverted on
every hyprctl reload. All 70 hypr-backed keys now have a prefs.get() in the
Lua, and hypr-prefs-contract pins both the presence and that the Lua fallback
equals the schema default, which is how the touchpad page misreported natural
scrolling on first boot.
The idle generator fell back from unwritten battery keys to the AC values
while the Power page displayed the schema defaults: a fresh laptop showed
"suspend at 20 minutes" and generated no suspend listener, then discharged to
zero in a bag. Unwritten keys now use the defaults the page shows
(idle-defaults-contract pins generator to schema; idle-config-contract
re-pinned to the new rule with the tradeoff recorded), and change-settings
enables managed idle on any machine with a battery -- without starting
hypridle in whatever session the installer runs under.
The per-app lock-screen notification switches wrote fields nothing read:
hyprlock cannot render notifications. Removed, with the rule model shrunk to
{enabled}, stale stored fields dropped at normalization, and the contract now
forbidding the page from growing lock-screen switches it cannot honor.
The battery warning thresholds were searchable, documented as "Found on
Power & Lock", and rendered nowhere -- and crossing the low threshold changed
only a glyph's color. Both sliders now exist where search was already sending
people, and low battery publishes a real notification at important priority.
Three handoffs opened GNOME panels that are inert in a Hyprland session. The
keyboard handoff is gone (that panel writes gsettings nothing here reads, and
the working controls sat on the same page); Connectivity gains a Wi-Fi row
that opens GNOME's actual Wi-Fi panel -- hidden SSIDs and 802.1X finally have
a road -- beside the network row that legitimately drives NetworkManager; the
universal-access handoff is gone, its few working toggles being controls this
app already owns. And the accessibility page now gives the true reason sticky
keys are missing: each Wayland compositor implements its own and Hyprland
does not yet -- not "an X11 feature with no Wayland equivalent," which sent
people to the wrong conclusion about the platform.
Claude-Session: https://claude.ai/code/session_01Epx9ZC1gwm81K3jm9x9CKh
Three surfaces the first laptop install showed were missing.
The bar's battery icon gets an optional exact number beside it -- GNOME's
"Show Battery Percentage", off by default for GNOME's reason, one color with
the icon so it reads as one indicator.
The Power page says what closing the lid does. The policy already existed
(LidPolicy holds a suspend inhibitor while an external display is connected)
but was surfaced nowhere, so the machine's most physical behavior was
undiscoverable -- and the deliberate absence of an override deserves stating
rather than leaving someone to hunt for a switch that does not exist.
And the Users page grows a Fingerprint card, because fingerprint login is two
systems that fail silently when they disagree: fprintd holds the enrolled
prints, authselect decides whether PAM ever asks the reader. This machine
arrived with a finger enrolled from its GNOME days and with-fingerprint off,
which reads as "the reader is broken". The card shows both facts, flips the
authselect feature through polkit with a stated reason, and hands enrollment
to GNOME's Users panel, which owns the only good capture dialog -- a named
exception in the handoff contract. Everything through scripts/panama-fingerprint,
pinned by a stub-driven contract.
Claude-Session: https://claude.ai/code/session_01Epx9ZC1gwm81K3jm9x9CKh
Two bugs, one story: ./install on a machine that had been half set up before
produced no Hyprland at all, and said so in one line among twenty minutes of
scrollback.
Terra bootstraps itself with --repofrompath, which defines a throwaway repo id
just long enough to install terra-release. Run it again on a machine that
already has terra-release and dnf5 refuses the whole transaction -- the
throwaway id collides with the real one. That step sits above everything, so
set -e ended the stage before a single package was considered. It is skipped
now when terra-release is already installed.
The rest is the reason one failed repo cost the desktop. Hyprland was installed
near the bottom of the stage, below a codec swap, two group updates and a
GStreamer glob, any one of which can fail for reasons outside this repository.
It now installs directly after the packages it needs and before anything
optional, and everything fragile below it runs through a soft helper that logs
and continues rather than ending the run. What was stepped over is listed at
the end, because tolerating a failure only beats aborting on it if somebody is
told.
A missing Hyprland is still fatal, and now says so in words.
Also removes the leftover disabled solopasha/hyprland COPR, which would mix
with lionheartp's the moment anyone enabled it while debugging.
Fixes the usage widget reading 1500%: the endpoint reports percentages, not
0..1 fractions. Clamped as well, and the widget answers a click now -- it set
interactive:false, which disables the mouse area its own handler needed.
Agent instructions, skills, SSH host aliases and expansion triggers are worth
having identical on every machine one person owns, and belong in none of the
shared configuration. They live in user/ now, with a manifest saying where each
piece goes and a link-user stage that puts it there.
That stage does nothing unless the machine said yes. Somebody who clones Panama
to try the desktop keeps their own ~/.claude/CLAUDE.md exactly where it was;
the question names the destinations and defaults to no. Anything displaced goes
to config/old rather than being deleted.
~/.claude/CLAUDE.md and ~/.codex/AGENTS.md were byte-identical copies of one
file, which is the drift this exists to prevent.
Also adds the vitals toggles for the battery and Claude usage readouts, which
had preferences and no way to reach them.
The last of Section F, and the only thing in Panama that reads an
authentication token, so most of the design is about that rather than
about the number.
It never refreshes the token and never writes to the credentials file.
That token expires roughly hourly and Claude Code refreshes it on
demand; if this refreshed it too, two processes would be rotating one
credential, and a rotation invalidates the other holder's copy. The
failure mode is being silently signed out of Claude Code by a status
widget, which no bar indicator is worth. So it reads the token, uses it
while valid, and reports "waiting for Claude Code to refresh" when not
-- which covers the case that matters, because while you are using
Claude Code the token is fresh, and while you are not there is nothing
to watch.
The token never reaches argv either: curl takes the Authorization
header on stdin through --config, because a header passed as an
argument sits in /proc/<pid>/cmdline for the length of the request.
Same rule the password and MOK paths already follow. And it never
reaches the output: the record carries percentages and timestamps and
nothing else. Both are pinned, and both were checked by sabotaging the
collector to pass -H and watching the contract name it.
Off by default. It is a coding-tool readout, not something a
general-purpose desktop shows without being asked, and it hides unless
the collector has real numbers rather than displaying "unknown".
Two of Section F.
Hooks are the pressure valve. "Can Panama also do X when the theme
changes" is now a five-line file in ~/.config/panama/hooks rather than
a fork, a feature request, or a patch somebody rebases forever. Each
name takes a single file and a .d directory so several things can react
without fighting over one, and a broken hook is reported and stepped
over: somebody's script must never cost a theme change, an upgrade or a
login. Wired at theme-set, post-upgrade and post-migrate. This is the
thirty-line version of the plugin host the upstream ledger defers, and
it has no API to keep stable beyond "we will run your script and tell
you what happened".
Testing it caught a real bug the reading would not have: run_one
captured the script path but never shifted it off, so every hook got
its own filename as $1 and the real arguments arrived one place late. A
hook reading $1 as the colour scheme got a path.
The crash watcher notices when a program dumps core and says so. Under
GNOME, ABRT does this; here nothing did, and applications died silently,
which is most of how "Linux is flaky" gets earned.
Once per program per session is the entire design, not a nicety. This
machine's portal backend crashes between eleven and sixty times a day,
and a notification per crash would be one every few minutes for
something nobody can act on. The first is news; the fortieth is why
people turn notifications off. The health page keeps the running count.
It waits for the notification server before reporting, because the
crash most worth hearing about is the one that took the shell with it,
and it names the executable rather than the kernel's comm field, which
truncates at fifteen characters. Verified against real segfaults.
Section G, the closeout.
UPSTREAM-INSPIRATION gains the second and largest borrowing round: what
was taken from omarchy, end-4, caelestia, ML4W and HyDE, what was
rebuilt rather than transplanted, and -- the part that matters more --
what was declined and why. Wallpaper-derived colour is declined on
identity, because dynamic colour is how rices end up looking like each
other. Audio visualisers are declined because a continuously repainting
equaliser is the GPU-pegging animation this project avoids. Patching
application internals to theme them is declined because the maintenance
is unbounded and lands on whoever is holding the desktop that day.
DESKTOP-PARITY gains a Portable machines section, which it has never
had, because until yesterday Panama was a desktop configuration in the
code as well as on the machine. Nine rows, one of them honest about not
having run on hardware with a lid. The remaining-gaps list gains the
four things now known to be missing rather than the one.
The README stops describing a desktop from before yesterday: migrations
in the install flow, the welcome and cheatsheet and manual, and laptops
working with every surface hiding itself where the hardware is absent.
The plan records its own outcome per section, including the two places
the work departed from it: D2's template renderer deferred once D1
removed the duplication it existed to fix, and E1 inverted after using
it for an afternoon.
Making the plain application keys focus an existing window was the
wrong call. It reads well in a demo and it is what macOS does, but it
made "give me another terminal" the awkward case -- and on a tiling
desktop a second terminal beside the first is the normal way to work,
not an edge case. Reaching for the launcher to open a second file
manager is not an improvement on anything.
So the plain keys do what they always did, and SUPER+ALT is the new
capability rather than a tax on the old one: go to the terminal,
editor, browser, files, calculator or mail you already have, wherever
it is, and start one only if there is none.
ALT rather than SHIFT because SUPER+SHIFT is already the
window-manipulation space -- Files, Neovim and Settings would have
collided with Focus session, Taller and Shorter, and breaking two keys
out of the eight-key resize set to make room is the worse trade.
Also fixes a real trap found while using it. The Alt-Tab overlay
commits on SUPER release, which is a compositor bind running an IPC
call; if that call ever fails to land, the overlay stayed up with no
keyboard focus, no Escape handler and nothing clickable, so the only
way out was an IPC call typed into a terminal it was covering. Clicking
outside now dismisses it, clicking a row switches to that window --
which is the obvious thing to try and did nothing -- and an abandoned
switch closes itself after ten seconds. Keyboard focus still stays with
the compositor, because taking it mid-switch is what would break
stepping.
Right-click a file to send it to your phone, or a video to make it
smaller. macOS has both behind the Share sheet and Quick Actions;
Windows has "Send to"; a stock Linux file manager has neither, and the
usual answer for the second one is a web uploader or an ffmpeg
incantation looked up again every time.
Neither adds machinery. Sending reuses panama-kdeconnect, the same
helper the Home & Phone page and quick settings already drive, so
there is one way to talk to a phone rather than two. The entry appears
only when a phone is actually reachable: an item that is present and
fails is worse than one that is absent, because the absence explains
itself.
The transcoder's two rules are both about not losing work. It never
writes to its input, and it never writes over an earlier output -- a
second run produces -2 rather than eating the first result. Verified
against a real encode: 1920x1080 became 854x480, with an even width
because H.264 rejects an odd one at the very end of a long encode,
which is the worst possible moment to find out.
Menus decide by mime type rather than extension, act on one file at a
time, and refuse anything that is not a local path. nautilus-python
turned out to be declared already; it now says it carries Panama's own
extensions too.
Two of Section E's small wins.
panama-webapp gives a site its own icon, its own window with no browser
chrome, and its own launcher entry, which is most of what "installed"
means in practice and what both macOS and Windows now ship. It scrapes
the site's apple-touch-icon, falls back twice, and never fails an
install over a favicon: a web app with a generic icon still works.
Names are slugged, so "../../../../tmp/pwn" lands inside the
applications directory as tmp-pwn rather than anywhere else, and
remove refuses anything without the marker it writes -- sharing a name
with a real application must not delete that application. A browser
that cannot do app mode is told so rather than handed something that
opens an ordinary window and pretends.
The charger now announces itself through StatusEvents, which was
already the right layer and only wanted a producer. Ambient priority,
so Do Not Disturb quiets it, because a charger is exactly what DND is
for. A critically low battery is published at a priority DND does not
silence, because the one message you must not miss is the one saying
the machine is about to stop. Both join the existing silent-startup
window rather than announcing the state they found.
The keyboard-layout toast the plan also listed is deliberately not
built. Hyprland exposes the active keymap but not a change event
Quickshell already consumes, so it would need either polling or new
event plumbing, and this machine has one layout and could not test it.
The application keys ran their command unconditionally, so pressing the
browser key twice gave you two browsers. That is the single most common
"Linux feels wrong" moment and it is a twenty-line fix: match the
window class, focus it if it is there, launch if it is not.
Two details the obvious version gets wrong. Patterns are anchored,
because an unanchored "mail" matches gmail-notifier and the mail key
would raise somebody's notifier. And class alone is not enough here:
the terminal and the editor are both kitty, so the editor bind narrows
by title or it raises whatever terminal happened to be open. The
contract's fixture contains both traps.
Found and fixed a bug this shipped with earlier today. Hyprland 0.56
dispatches through Lua, so `hyprctl dispatch focuswindow address:0x...`
is parsed as Lua source and fails with a syntax error -- which is what
the launcher's Switch Window command had been doing since it landed.
Its own contract stubbed hyprctl and recorded the arguments, so the
call looked correct and never ran. Both now use the selector form
hl.focus actually accepts, and the contract asserts the shape rather
than only that something was dispatched.
They were written down five times: ThemeProfileModel.js for QML,
looks.lua for the compositor, and again in panama-theme-apps and
panama-lock. The GNOME accent-name mapping was a sixth list. Adding a
ninth accent meant editing all of them, and the file most likely to be
missed was the lock screen, which fails silently -- the machine locks
in last season's colour and nothing says why. panama-theme-apps
admitted it in a comment: "there is no shared source between QML and a
shell script".
config/palette.json is that source now. looks.lua reads it through a
new prefs.readJson, which uses the same never-raise parser the settings
store uses, so an unreadable palette costs the accent colours and never
the compositor config. The two shell generators read it through
scripts/panama-palette, which also carries the hex-to-rgb conversion
hyprlock needs and the GNOME member lookup.
QML keeps its table, because a .js module imported into QML cannot read
a file. That is still a copy, so the palette contract compares the two
value by value -- every accent, every field -- and fails on any
disagreement. Verified by planting a wrong hex and watching it name the
exact field.
The adwaita contract used to check the shell's own copy of the GNOME
mapping. It now checks that the shell resolves through the palette, and
fails if that copy ever grows back.
docs/ is engineering artifacts -- design specs, plans, an upstream
ledger -- and none of it is written for the person who has to live
here. Five chapters that are: coming from another desktop, the
keyboard, windows and workspaces, when something breaks, and making it
yours.
Rendered inside Settings rather than opened in a browser, so a chapter
that says "the Displays page reverts after fifteen seconds" is one
click from the Displays page. Qt's markdown renderer does the work;
one chapter per Text, because Text has an implicit texture size limit
and a document long enough to hit it goes blank rather than
complaining.
The chapters live beside the shell in manual/ rather than at the
repository root, which departs from the plan. The reason is the path:
the shell finds them through Quickshell.shellDir, which is correct
wherever the repository is, whereas walking upward out of the shell
directory is only correct by accident. The contract fails that pattern
if it comes back.
The contract also pins the set both ways -- every chapter listed
exists, every chapter that exists is listed -- because a renamed file
shows an error card where a chapter should be, which reads as a broken
manual rather than as a moved file.
Thirty settings pages is the opposite of the usual problem: a person
arriving from GNOME, macOS or Windows cannot tell which few things
matter. This is those few, once, on the first start.
Not a tour. Nobody reads a tour, and a multi-step wizard on a desktop
somebody just installed is one more thing between them and using it.
One card, five keys, and a way out.
The chords come from the live keymap rather than being written here, so
a machine whose owner has already rebound something teaches what they
actually have. A welcome screen is the one surface read by somebody
with no way to tell it is wrong, which is exactly why it must not be.
Two deliberate departures from how every other surface behaves. It does
not close on a click outside, because a stray click in the first thirty
seconds would throw away the only explanation on offer. And dismissing
by any route marks it seen, Escape included, because a desktop that
reintroduces itself every login has failed to take no for an answer.
It stays reachable from the launcher afterwards, since the moment
somebody wants it again is exactly when a one-shot has thrown it away.
Also teaches the keymap to spell punctuation: slash, period, comma and
the rest were rendering as their raw keysym names, so the welcome
screen offered "Super + slash" and the cheatsheet agreed with it.
The Shortcuts settings page answers "how do I change this", which is
worth opening a window for. This answers the other question, the one
you have with your hands already on the keyboard, so it is an overlay
on SUPER + / and the same key closes it.
It reads Keybinds.grouped() rather than a written-down list, so a
shortcut rebound in Settings shows its new chord here with nothing kept
in sync. A cheatsheet that lies is worse than none: it gets consulted
exactly when somebody does not already know.
Three columns, balanced by how many shortcuts each category holds. The
first attempt used a Flow, which wraps into as many columns as it likes
and made 120 binds across six uneven categories unreadable; it also
sized the card from a child that filled it, which is a circular binding
and produced a card taller than the display with its contents running
off the bottom. Both were found by looking at it rather than by a test,
which is the argument for looking at it.
Fixes a real bug on the way past: luaChord and formatChord appended the
key unconditionally, so the window switcher's modifier-only release
bind became "SUPER + " with a dangling separator. That matched neither
the chord keybinds.lua binds nor the one an override is keyed by, so
that bind could never be rebound and had no category -- it was sitting
in a seventh group of its own, which is how it was noticed.
The Shortcuts page grouped shortcuts by matching substrings in their
descriptions, which put "Close window" and "Close the notification
list" in the same group and left anything phrased unusually in
whichever bucket matched first. The cheatsheet that comes next would
have inherited the same guesswork.
keybinds.lua says it outright now. Its sections already were the
categories, so a section sets one and the binds below inherit it: one
line per section instead of one per bind, and a new bind lands in the
category of the section somebody wrote it in without having to
remember anything.
Hyprland reports a Lua bind's dispatcher as __lua with a bytecode
offset, so nothing can be attached to a bind that survives into
`hyprctl binds`. The config writes a manifest at load instead, keyed
by the chord actually bound so the shell can join on what it sees.
Writing never raises: a read-only state directory costs the grouping,
never the keymap, and the shell keeps the old derivation as its
fallback so a machine that has not reloaded its compositor still works.
The one failure mode is a section that forgets to set a category and
silently inherits the one above. That is not hypothetical -- it
happened while writing this, because the dictation section sits in the
middle of the media binds and its category leaked onto the volume,
media and brightness keys below it. The contract walks the file for
sections with binds and no category, and spot-checks the boundaries
where inheritance is doing the work.
logind handles the lid correctly except for the one case it cannot
see: an external display means a closed lid is a docked machine, not
one being put away. Its own docked test looks for an ACPI docking
station that modern hardware does not have.
Panama does not take the lid over to fix that. It holds a logind
handle-lid-switch inhibitor while an external display is connected and
releases it when the last one goes, which needs no lid watcher, no
polling, and no drop-in. The direction it fails in is the point: if the
guard dies, logind's default comes back and a docked laptop suspends,
which is annoying. A drop-in setting HandleLidSwitch=ignore plus a
watcher of our own fails the other way, leaving a lid that does nothing
at all on a machine being carried out of a building.
Locking on the way down needed no work: hypridle's before_sleep_cmd
already runs loginctl lock-session, so a lid-close suspend is a locked
suspend. The contract fails anything that duplicates it.
Not yet verified against a real lid, which is stated in the helper's
header rather than implied by silence. The decision logic, the
inhibitor's shape, and every machine that should hold none of it are
covered.
Panama had no idea whether it was running on a laptop. No upower, no
battery, no lid, no AC: hypridle.conf says "This is a desktop" in its
own header, and that was true of the code as well as the machine.
panama-hw answers hardware questions one at a time, exits 0 or 1, and
prints nothing, so scripts, services and contracts all ask the same
way. The definition the rest of the laptop work hangs on is one line:
clamshell is lid-closed AND an external monitor. A machine with no
mains supply at all reports as being on wall power, because a desktop
cannot run out of it.
The battery service follows Vitals: sysfs through FileView, an
availability flag, and no subprocess on the timer. Globbing is the one
thing QML cannot do -- a battery is BAT0 or BAT1 or CMB0, mains is AC
or ADP1 or ACAD -- so panama-battery resolves the names once and the
shell reads the files directly after. Nothing falls back to a
plausible zero: a desktop shows no indicator, no card, and no charge
limit control where the firmware has no ceiling.
Also repairs two contracts that were already failing and had not been
noticed, because only the full suite runs them. The dependency
scanner treated line-initial variable assignments, case labels,
comments and heredoc bodies as commands, and `count`, `host`, `cancel`
and `import` are all real binaries on Fedora, so `command -v` could
not filter them out. It now drops comments and heredoc bodies and
requires a command to be followed by whitespace. Verified it still
catches a genuinely undeclared dependency rather than passing quietly.
The launcher command contract had not been told about the fourteen
commands added earlier today.
./install only ever adds. It copies over /, links dotfiles, installs
packages -- and has no way to say "remove that file", "disable that
unit", "that symlink points nowhere now". So a machine set up months
ago keeps whatever this repository has since decided was wrong, and
the only thing that ever fixes it is somebody reading a commit
message. With a curl installer in the README, that stopped being
hypothetical.
A migration is one script that performs one repair, exactly once, on
the machines that need it. Named by the commit timestamp that authored
it, so glob order is chronological without a sequence number two
branches could both pick. Marked in ~/.local/state on success and only
on success, so a repair that failed stays pending rather than being
recorded as done and hidden forever. Ordered, and stopped at the first
failure, because a later repair may assume an earlier one landed. A
fresh install marks everything without running it, the way
Migrations.qml stamps a pre-versioning settings file at its baseline.
The first real one removes the dangling ~/.config/forge symlink left
behind when the GNOME session was cut: link-dotfiles could link it but
never unlink it. Verified both ways -- a no-op on a machine that never
had it, an actual repair on one that did.
Root work goes through panama-sudo --reason so the password prompt
names the repair, and the contract fails any migration reaching for
bare sudo.
A second IpcHandler with an already-used target does not error -- it
silently shadows the first, and for the polkit target that means the
agent's authentication requests stop reaching the prompt: every
password dialog on the desktop, gone without a message. That duplicate
nearly shipped once, because the handlers live scattered through a
long shell.qml. The contract pairs every IpcHandler with its quoted
target across the shell's QML (harnesses excluded -- each is its own
root), fails on any name declared twice, and refuses to pass on an
empty scan so a declaration-format change cannot quietly blind it.
The OS-parity batch from the vicinae plan, tasks 1 through 7. The audit
came back better than the plan guessed: the calculator already links
libqalculate, the built-in file index answers in under 100ms across all
of home, quicklinks and snippets ship as built-in stores -- so zero new
packages, and `vicinae dmenu` replaces the planned compiled extension
outright. What was missing gets built: a power menu (lock, suspend, log
out through uwsm, restart, power off), reminders as transient systemd
timers with a pick-to-cancel list, a color picker over hyprpicker, and
dmenu pick-lists for window switching, force quit, kill process, SSH
hosts, and recent files -- all through one panama-pick helper. The
launcher commands contract exercises the reminder parsing and every
pick-list against stubs, including killing its own sacrificial sleep.
panama-sudo is pkexec with a stated reason: the reason travels to the
shell over the existing polkit IPC target, and the prompt renders it
labeled "Stated reason (unverified)" beside polkitd's real action
message -- beside, never instead of, because any process can claim any
reason and the action text is the trust anchor. Reasons are single-shot
and expire in ten seconds, so a stale one cannot dress up an unrelated
prompt; without a reason, a running shell, or qs the wrapper is exactly
pkexec. Built for agents, so the person typing their password learns
what for. Verified live end to end -- reason shown, consumed once,
expired when stale, cleared on dismissal -- and pinned by the polkit
reason contract.
`boot` is the script the README now leads with: curl it, and it installs
git if the machine lacks it, clones the repository to PANAMA_PATH, and
hands off to ./install -- reattaching the terminal first, because a
piped stdin would strand the interview. Deliberately dumb: a curled copy
leaves the repository the moment it runs, so nothing that can drift
lives in it. Re-running is the recovery path: an existing clone is
fast-forwarded, never re-cloned, and a refused fast-forward installs
from what is there rather than stopping mid-repair. All of it pinned by
the boot contract, against stub git and a throwaway clone.
Arrange the desktop, run "Save Layout as Project" from the launcher, name it.
"Open Project" lays it out again on free workspaces, so it never lands on top of
what you are already doing. Saved layouts are listed on the Desktop settings
page, which is where they are removed.
Recorded rather than written by hand, and a terminal's directory is most of why
it is worth having: without it a project opens three terminals in your home
folder and you change directory three times. This machine had two terminals in
the same project when it was written, and reopening there is the difference
between a layout and a working desktop.
Four things had to be found by running it, none of which reading would have
shown.
A terminal's directory is not the terminal's working directory -- that is where
it was launched from. The shell inside it is what has been cd'd. Reading the
wrong one looked correct for exactly as long as the terminals under test had
been started from the right place, which they had.
gtk-launch cannot place a window. It activates over D-Bus, so the process
Hyprland started exits and a [workspace N silent] rule has nothing left to apply
to; Nautilus landed on whatever workspace was in front of you. The Exec line
from the desktop entry is run directly instead. But DBusActivatable
applications do the same thing regardless, so the window is found afterwards and
moved by address -- which also means a window that never appeared is reported
rather than assumed.
/proc/PID/task/PID/children is a file of pids, not a directory. Listing it as
one always raised, so the fast path was never once taken and everything went
through pgrep.
And kitty's --directory needs an equals sign or the short -d; the
space-separated long form is accepted and silently ignored.
The contract exercises a save and open round trip against a stubbed compositor,
and reads a terminal's directory out of a real process tree rather than grepping
the source for a shell name -- an earlier version passed against a helper that
had been changed back, because the constant was still there.
Claude-Session: https://claude.ai/code/session_01Q84axqUE5inJhf5Jz9CFy1
Super+D holds the microphone open, releasing it transcribes on the GPU and
types the result wherever the cursor is. Roughly 150ms for a normal utterance
once the model is resident, measured rather than hoped for.
Getting there meant discarding two approaches. Fedora 44 cannot install any
GPU-capable Whisper for Python -- openai-whisper needs a numba that needs an
llvmlite that does not exist for 3.14, and faster-whisper needs a ctranslate2
nobody packaged. The whisper-cpp package IS built with HIP but ships libraries
with no binary and no bindings, and hand-writing ctypes for a large by-value
struct is a segfault waiting for a version bump. So a container, as suggested.
Vulkan rather than ROCm, and upstream's image rather than one built here. ROCm
is seven gigabytes and serves AMD alone; Vulkan compute runs on the AMD, Intel
and NVIDIA machines this config is used on, in a twentieth of the space. The
Vulkan tag already contains whisper-server, so there is no Containerfile to keep
working -- an earlier draft of this commit had one, and it was strictly worse.
Two bugs found by using it rather than by reading it. Whisper describes silence
as the literal text "[BLANK_AUDIO]", and the first working version pasted that
string into the clipboard; a transcription that is nothing but such markers is
now discarded. And the server answers with a line per segment, which typed into
a window is an Enter press -- sending the half-written message, submitting the
form. Whitespace is collapsed to one line.
Neither the image nor the model is installed by ./install. Together they are
over two gigabytes that want the network, and Settings offers both as one
action instead. Nothing starts at login either: whisper-server holds the model
from the moment it starts, so the first press of the key is what brings it up.
The contract pins both text bugs, that the server stays on loopback, and that it
does not start at login. Reverting the [BLANK_AUDIO] guard did not fail it at
first -- the check was still correct, it had simply stopped being called -- so
it now checks the call site too.
Claude-Session: https://claude.ai/code/session_01Q84axqUE5inJhf5Jz9CFy1
GNOME's Multitasking panel asked one workspace question worth reproducing, and
it is not which workspace goes on which screen. It is whether the second screen
participates at all: workspaces on the primary display only, or each screen with
its own. Ten rows of per-workspace assignment would be more powerful and worse.
Off is Hyprland's own behaviour and emits nothing. On pins workspaces 1 to 10 --
however many ALT+1..ALT+0 actually reach, read from keybinds.lua rather than
written down twice -- to whichever output is recorded as primary. With no
primary recorded, nothing is pinned: guessing one would move every workspace
onto whichever output happened to sort first, and this machine is in exactly
that state.
Applying is a reload, which is the part that shaped the design. Hyprland reads
workspace rules at config time and will not remove one afterwards -- a rule
written with an empty monitor keeps its old binding, which was checked rather
than assumed. Only a reload clears them, so the config is the only honest source
and the page cannot pretend a change has landed before one happens. Hence a
service that reads `hyprctl workspacerules` back rather than inferring success
from having written the preference, and a Reload row that exists only while the
two disagree.
Verified end to end against the live compositor and put back: off emits nothing,
on emits ten rules naming the primary, and turning it off clears them. The
settings file came back byte-identical.
Claude-Session: https://claude.ai/code/session_01Q84axqUE5inJhf5Jz9CFy1
Gabriel Brown