llama-swap holds the request until a swapped model finishes loading; the
default first-event watchdog aborted mid-swap and triggered a fallback-model
switch that then cold-loaded too, repeating the stall. Env vars live in
.omp/agent/.env (decrypted from this blob); 0 would disable the watchdog.
Pins git.melonbread.xyz (Forgejo) to port 222 with the machine key
only — the forge SSH daemon does not listen on 22. All boxes pushing
to the forge need this stanza; known_hosts/authorized_keys stay
per-machine and unmanaged.
update-neovim.sh and the topgrade [commands] entry that calls it are
only meaningful where neovim is not a distro package. On arch-family
boxes the template renders empty, so chezmoi removes the script
instead of deploying a no-op, and topgrade no longer runs a step that
could fail on unmanaged paths.
paru's SudoLoop keeps sudo credentials alive for the whole run instead of
prompting per operation. Disable it — matches the local edit in
~/.config/paru/paru.conf.
- Merge active settings from shadow ~/.config/topgrade.toml (set_title,
assume_yes, disable, git repos) into chezmoi-managed file; topgrade
prefers the former path, so the managed config was never read.
- Replace invalid 'oh-my-zsh' step with 'shell' (no per-framework
variant) — config failed to deserialize, so no steps were disabled.
- Delete shadow config; topgrade now reads the chezmoi-managed file.
- omp: re-add config.yml as private_config.yml (matches live fallback
chains; minimax removed, slow -> deepseek/deepseek-v4-pro:high)
- omp: re-encrypt models.yml; llama-swap.miche static list synced to
live /v1/models (21 models), stale entries dropped
- pi: fix broken settings.json tracking — dir_dot_pi/ mapped to a
literal ~/dir_dot_pi, real ~/.pi/agent/settings.json was never
managed. Track it via dot_pi/agent/settings.json.tmpl
- Remove static model lists from llama-swap.byte and llama-swap.miche
(discovery: type: proxy fetches live models from /v1/models)
- Update byte model IDs in config.yml to current roster:
qwen3.6-35b-heretic-mtp -> qwen36-35b-genesis
gemma-4-12b-heretic -> gemma4-12b-hauhau
qwen3.6-35b-a3b-mtp -> qwen36-35b-genesis
Both run_once_20-install-user-packages-gentoo.sh.tmpl and
run_once_40-install-sway-gentoo.sh.tmpl were missing the os_family
template guard (the overlays script got it in the previous fix but
these two slipped through). On non-gentoo boxes — arch, debian —
chezmoi apply would run them and they'd try to 'emerge', failing
with 'sudo: emerge: command not found' and exit status 1, which
breaks topgrade.
Wrap each body in '$0{ if eq .os_family "gentoo" }' / '$0{ else }',
mirroring the chaotic-aur script's pattern and the overlays fix from
the previous commit. On non-gentoo boxes the rendered script reduces
to a single 'skipping' log line and exits 0; on gentoo boxes the full
body (including the existing .sway_setup gate on the sway script) is
preserved.
Verified by rendering both scripts with a gentoo-config override:
arch (live) -> 'skipping gentoo user-packages' / 'skipping sway'
gentoo -> full body (293 / 87 lines)
Both pass bash -n and actually exit 0 when run on this arch box.
Previously 'node' failures were silently skipped. With this removed,
any non-zero step exit code aborts the run, matching the user
preference of 'I dont want it to skip anything'.
The script's body was never wrapped in an os_family template conditional
despite the header claiming it was 'guarded by the chezmoi template
engine'. So topgrade/chezmoi apply would run it on every box and die
with '/var/db/repos/gentoo missing' on arch/debian.
Wrap the body in '{{ if eq .os_family "gentoo" }}' / '{{ else }}' (same
pattern the arch-only chaotic-aur script already uses). On non-gentoo
boxes the rendered script reduces to a single 'skipping' log line.
Also fix two stale comments in the header (filename was 05, not 10).
Verified by rendering with --config override: arch → no-op log line,
gentoo → full body intact, both pass bash -n.
chezmoi parses attribute prefixes left-to-right and 'encrypted_' must
precede 'private_'. The 'private_encrypted_' order silently breaks
decryption: chezmoi consumes 'private_', then treats 'encrypted_foo.age'
as a literal filename and copies the ciphertext verbatim instead of
decrypting it.
models.yml was named 'private_encrypted_models.yml.age' since commit
3c3fab7 and was never decrypting — a stale 'encrypted_models.yml.age'
blob was sitting in ~/.omp/agent/ and the plaintext models.yml was an
unmanaged leftover. .env and zai.key used the plain 'encrypted_'
prefix and were decrypting at umask 0644 (world-readable in isolation;
the 700 ~/.omp/agent/ dir was the only thing shielding them).
This commit:
* renames all three to encrypted_private_{zai.key,.env,models.yml}.age
so chezmoi decrypts AND lands them at 0600 natively
* rewrites run_onchange_35 as 'ensure-omp-secret-perms.sh' covering
all three, as belt-and-suspenders for any box where a secret still
sits at 0644 from a prior apply
* removes the stale encrypted_models.yml.age verbatim blob and its
orphan state entry
* corrects the README perms section to document the prefix-order
gotcha (was misleadingly claiming 'private_' alone gave 0600)
Verified end-to-end on this box: chezmoi managed lists all three as
decrypted targets, scoped apply writes them at 600, chmod script is
idempotent. Other boxes need a 'chezmoi apply' to pick up the rename
and the onchange chmod.
The actual cause of 'caffeine notification doesn't show' was the
-t 2000 timeout: the notification popped for 2 seconds and self-cleared
before the user could see it. Manual notify-send from a terminal (no
-t flag) used mako's default-timeout=5000 and was visible — so the user
saw my test notifications but not their own clicks.
Fix: drop the -t flag everywhere so mako config owns the duration.
Added: source restore-wayland-env.sh in toggle-dropdown.sh and
wifi-menu.sh — same env-loss pattern as caffeine.sh, would have
been next on the bug list.
Verified on tadbit: -t 2000 notification visible at +0.5s, gone at
+2.5s (user missed it). No -t flag: visible at +0.5s, gone at +5.5s
(visible long enough to read).
caffeine.sh click was working (icon toggled, swayidle restarted) but
the 'Caffeine ON/OFF' notification never appeared. Same root cause as
the prior swayidle bug: waybar's on-click context strips WAYLAND_DISPLAY
from the spawned script's env. notify-send writes to D-Bus (exit 0) but
mako can't display because it has no Wayland socket to render on. Click
appears to do nothing visually.
Fix: extract the /proc/*/environ probe into a shared helper
(restore-wayland-env.sh) and source it from both caffeine.sh AND
start-swayidle.sh. Now any future on-click script in this repo can
`. restore-wayland-env.sh` to inherit the right env.
Verified on tadbit:
- click 1 (caffeine ON): swayidle killed, flag set, 'Caffeine ON'
notification appeared in mako
- click 2 (caffeine OFF): swayidle restarted, flag cleared, 'Caffeine
OFF' notification appeared in mako
Both are visible in makoctl history now.
caffeine.sh click was silently breaking idle lock. Two bugs that
compounded:
1. caffeine.sh inlined its own swayidle command (with swaylock -f -i
, not lock-fancy.sh), drifted out of sync with sway/config
which used lock-fancy.sh. Toggle off would restart swayidle with a
different lock command than the rest of the session.
2. WAYLAND_DISPLAY isn't set in waybar's on-click context. When
caffeine.sh called 'swaymsg output * power off' (part of the
restart command) and later called swaylock / grim (via lock-fancy),
those failed silently. swayidle itself can run without
WAYLAND_DISPLAY but it can't actually monitor input/output, so it
exits immediately. Result: user clicks caffeine off, flag clears,
icon goes back to 'inactive', but auto-lock is silently dead.
The user thinks they turned caffeine off. They didn't. They're
just unprotected.
Fix:
- new start-swayidle.sh holds the canonical swayidle command
- sway/config: exec $HOME/.config/sway/start-swayidle.sh
- caffeine.sh: killall swayidle; start-swayidle.sh &
- start-swayidle.sh probes /proc/*/environ for a Wayland client's
env (mako always has it) and exports WAYLAND_DISPLAY /
DBUS_SESSION_BUS_ADDRESS / XDG_RUNTIME_DIR / DISPLAY if missing
before exec'ing swayidle
Single source of truth + env restoration = click works on every box
that has the same Wayland-capable process tree (mako, swaybar, etc.),
no matter who/what is calling the script.
Verified on tadbit: pre-fix toggle off killed swayidle permanently.
Post-fix: toggle off restarts swayidle successfully, env inherited
from mako's /proc/PID/environ.
