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
The shell named a font, a pointer theme and a wallpaper that no package list
installed and no stage placed. It went unnoticed because this machine collected
all three under sunhat and never lost them; a fresh Fedora box would have come
up with tofu for every shell glyph, the default pointer, and no wallpaper --
while Wallpaper.qml called that missing file `shippedPath`.
The dependency contract reported PASS throughout, because it reads commands that
scripts invoke and none of these are one. The new assets contract covers what it
structurally cannot: fonts and pointer themes named in configs and gsettings, a
shipped wallpaper that must exist, commands launched from QML, and directories
nothing puts into service. Written against the broken tree first, where it found
ten faults.
Four of those were packages nobody had noticed were missing -- gnome-calendar,
podman, pipewire-utils and flatpak -- alongside gnome-control-center, which backs
fifteen rows of Panama's own Settings and is commented so it is not mistaken for
GNOME-session residue later.
Fonts turned out to need no install stage at all. Terra, which install-packages
already enables, packages every Nerd Font, so sunhat's wget-and-unzip is replaced
by five lines in a package list. The pointer theme had no such luck: it is
packaged nowhere, so it is vendored rather than downloaded from a URL that can
rot.
espanso stays undeclared. It is in no enabled repository, and building it from
source is the install method this repository is trying to stop using.
Claude-Session: https://claude.ai/code/session_01Q84axqUE5inJhf5Jz9CFy1
The initial package list was quoted into a single bogus dnf argument
and every dnf error was discarded, so a fresh install silently skipped
most of it. Two package lists were never wired into the pipeline at
all, and change-settings ran before install-packages, so the vicinae
theme step was permanently skipped. Fixed the ordering, the quoting,
and stopped swallowing errors.
Separately, hypridle could start with its WAYLAND_DISPLAY condition
unmet if it raced the env-publish call, silently never starting --
and it's the only listener for the logind Lock signal. Made the start
wait on the environment synchronously. panama-idle also wrote its
generated config to a fixed temp path with no locking, so concurrent
applies could interleave into a corrupt file; switched to mktemp plus
an atomic mv.
Claude-Session: https://claude.ai/code/session_01E6TJUAh41HaP25MVHWkhRZ
Two things stood between this repository and a machine that could
install it.
The installer aborted on its own first question. The hostname prompt
defaults to N, and the N branch ran `exit` -- so pressing Enter, the
obvious answer when you do not want to rename your machine, skipped the
entire installation and said nothing about it. Declining now just
declines. The installer is also safe to re-run, which is the upgrade
path too: it reports which stages failed instead of scrolling the
failure past twenty minutes ago, and restores the idle settings on every
exit path rather than only on success.
The package lists had drifted badly from what the configs and helpers
actually use. jq alone has thirty-one call sites across the helpers and
the contracts; kitty has a full shipped config and a dock pin; tmux and
btop have shipped themes the colour scheme switches; ddcutil, qrencode
and orca back features added today. None were declared. Neither were
fontconfig, pciutils, libselinux-utils, libnotify, wireplumber, fwupd or
python3-dnf, all of which shipped scripts invoke by name. A fresh
machine following this repository's own instructions would have got a
desktop whose features quietly were not there -- the helpers report "not
installed" rather than crashing, which is good behaviour and completely
silent.
So the lists are corrected and a contract now checks that every external
command Panama's scripts invoke is installed by Panama's packages.
Writing it was instructive about its own blind spots. The first version
reported `then`, `esac` and `done` as missing packages, burying the real
findings. The second passed while jq was undeclared, because the pattern
required three characters and jq is two -- a dependency checker with a
blind spot for short names is worse than none, since it reports PASS.
The third missed ddcutil, which is only ever invoked as `timeout 10
ddcutil` and so never appears statement-initial. It now also reads
`command -v X`, which is how these helpers probe for a tool and
therefore the clearest statement of a dependency there is. Verified it
catches jq, ddcutil and qrencode individually.
Also replaced a fixed 0.3s sleep in the write contract with a bounded
wait. It was failing about one run in three with "a rejected value did
not surface an error" when the error had simply not arrived yet, which
reads as a missing guard rather than a slow one.
Claude-Session: https://claude.ai/code/session_01BRvzt4H8XXLPVH5MyYdk9L
Gabriel Brown