install_rust(): official rustup script as the user (stable, default profile,
--no-modify-path), rustup update on a re-run, rust-analyzer + rust-src added.
libssl-dev (apt) / openssl@3 (brew) for openssl-sys crates. bashrc-linux,
bashrc-osx and zshrc-osx source ~/.cargo/env when present. README, CHANGELOG,
CLAUDE.md, TODO plan.
harden_sshd ended with an unconditional 'systemctl reload ssh'. On WSL (and any
host where sshd was never started) the unit is inactive, systemctl exits with
'ssh.service is not active, cannot reload.' and set -e killed the whole install
right after the drop-in was validated.
- reload only when 'systemctl is-active --quiet ssh'; else say the drop-in is in
place and applies at the next sshd start
- try-reload-or-restart instead of reload: exits 0 if sshd stopped in between
- create /run/sshd before 'sshd -t': on Ubuntu 24.04 the privsep dir only exists
while ssh.service runs, so the validation failed on an inactive host and the
valid drop-in was removed as 'rejected'
ubuntu-desktop-minimal and gnome-remote-desktop were installed unconditionally
in the Linux block. On WSL they are useless (Windows owns the display) and the
desktop's NetworkManager/systemd-resolved rewrite /etc/resolv.conf, which lost
DNS mid-install (fail2ban/whois fetch failed on 'Temporary failure resolving').
- confirm() takes an optional default (y → [Y/n], Enter = yes, no tty = yes)
- is_wsl(): WSL_DISTRO_NAME or 'microsoft' in /proc/version
- offer_desktop(): desktop [Y/n], then RDP [Y/n]; WSL flips both to [y/N]
- README step 11 + table + packages, CHANGELOG Unreleased
docker.service + docker.socket disabled and /var/run/docker.sock removed before
dockerd-rootless-setuptool.sh install, as the docs recommend; skipped when no
rootful unit exists.
docker-ce-rootless-extras joins the Docker package list; offer_docker_rootless
installs uidmap/dbus-user-session/slirp4netns, runs dockerd-rootless-setuptool.sh
install (apt variant: /usr/bin, no PATH export) and enables linger. The rootful
daemon is left alone. bashrc-linux exports DOCKER_HOST only when the user socket
exists. README step 16 + package lists.
gitconfig template: [gitflow] autopush = @AUTOPUSH@ (exact true/false, the
hooks fail closed on anything else) and a fixed core.hooksPath =
~/.claude/githooks (created by `make link` in claude-config, git expands ~).
install.sh: resolve_autopush at the identity step. A true/false already in
~/.gitconfig wins silently, else DOTFILES_GITFLOW_AUTOPUSH (any other value
aborts before a file is touched), else a prompt that re-asks until the
answer is exactly true or false (Enter = true), else true. render_gitconfig
refuses to write when the mode is not a boolean or @AUTOPUSH@ survives.
Fix: deploy_gitconfig received the repo bashrc template, so every install
wrote `name = @USER@` / `email = @EMAIL@`. It now takes the resolved
identity directly. patsub_replacement is turned off so an `&` in a name is
not expanded on bash >= 5.2.
- choose_macos_shell answers zsh on Enter, on no terminal and on an
unknown value; MACOS_SHELL=bash still picks bash.
- remote-install.sh hands /dev/tty to install.sh when one can be opened:
piped into bash, stdin was the script, so install.sh saw no terminal
and silently skipped the identity, shell and offer questions.
- remote-install.sh cloned master by default; origin only has main
(the documented raw URL 404s on master, 200 on main). README updated.
bin/repo-sync ports the Alphalink dotfiles repo/repo-reset zsh functions to
bash + zsh on Linux and macOS, for several forges at once: GitLab, GitHub
(+GHES), Gitea/Forgejo, Bitbucket Cloud, via curl + jq (no per-forge CLI).
Tokens live in ~/.config/repos/forges.conf (0600, parsed line by line, never
tracked), written by `repo-sync add`. Cache ~/.cache/repos/list refreshed
at most once a day (mkdir lock: flock is not on macOS), same namespace/project
on two forges kept once, archived and mirrors skipped. Layout
~/repos/<namespace with / as @>/<project>.
rc files: `repo <project> [namespace]` clones on demand and cd's, with
completion (bash compgen, zsh _describe), background refresh at shell start.
install.sh: jq in the apt and brew lists, chmod repo-sync.
Verified: shellcheck + bash -n, zsh -n; stub-curl harness over the four forges
(pagination, dedup, auth headers, lock, stale cache, path/clone); live GitLab
refresh (139 projects, 3s).
Git never expands $VARS in its config, so gitconfig carries @USER@ and
@EMAIL@ placeholders that install.sh fills from the deployed bashrc's
USER/EMAIL exports. A differing ~/.gitconfig is kept as
~/.gitconfig.backup-<date>; an identical one is left alone. A repo's
.git/config still overrides it. excludesfile uses ~ (git expands it,
not $HOME).
The rc files now override $USER, so the installer takes the login name
from id -un for dscl and the code-server unit.
gitleaks, web stack (mariadb-server, imagemagick, unversioned php-* modules),
ubuntu-desktop-minimal before the RDP setup, and a lspci-gated
install_nvidia_driver() that runs ubuntu-drivers install. README + TODO updated.
Always applied in the Linux block, no prompt, idempotent:
- install_fail2ban: fail2ban + nftables, etc/fail2ban/jail.d/local.conf.
sshd jail reads the journal (backend systemd, works with or without
auth.log) and bans the offender on every port, so the SSH port is
irrelevant: the previous server's jail banned 22 while sshd listened
on 337. 5 failures / 10 min / 1 h. Loopback + RFC1918 never banned.
- install_unattended_upgrades: package + 20auto-upgrades (the file
dpkg-reconfigure writes, without the prompt).
- harden_sshd: sshd_config.d/20-hardening.conf (PermitRootLogin no,
MaxAuthTries 3, LoginGraceTime 20), sshd -t gated: a rejected file is
removed and the install continues with a warning. Auth methods, port
and user lists untouched.
Docs: README table + step 13 + packages, CLAUDE.md layout.
/tmp is a RAM-backed tmpfs on Ubuntu (50% of RAM). Agent runs fill it: half
the RAM goes, then every temp-file creation fails with ENOSPC and shells
break. Swap does not lift the cap, so the fix is /tmp on disk.
End-of-install offers (Linux, [y/N], skipped without a terminal, idempotent):
- offer_tmp_on_disk: mask tmp.mount + etc/tmpfiles.d/tmp.conf (wipe at
boot, 10-day purge, /var/tmp rule kept). Effective at next reboot.
- offer_ssh_memory_guard: the previous server's rules. ssh.service drop-in
(OOMScoreAdjust=-1000, MemoryMin=256M) + earlyoom with --avoid sshd and
--prefer node/java. MemoryMin covers sshd only; earlyoom is the real guard.
install_cloudpex deploys the NAS mount helper in the Linux block.
Docs: README steps 12-14 + table, CLAUDE.md layout + lint command.
VS Code Remote-SSH integrated terminals are non-login interactive shells:
they source ~/.bashrc but never ~/.profile, where the resume menu was wired
(login-scope, BDR-007). The auto-check therefore never fired in the user's
actual environment, even after install.sh.
Source dtach-router from bashrc-linux (every interactive shell) instead, and
turn install.sh's wire_dtach_profile() into unwire_dtach_profile(): it now
strips any stale ~/.profile block so a plain SSH login (which reads ~/.bashrc
via ~/.profile) does not prompt twice. README updated to match.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CN1KSmsuLG6TxSeN5m8xvM
Executing dtach-router broke its return-based interactive guard and
errored on /dev/tty in non-interactive login shells (bash -lc, cron,
scp). It is now sourced via a guarded, idempotent ~/.profile block
(case $- in *i*) ... . dtach-router) installed by wire_dtach_profile(),
which migrates the old execute-based block. Also adds cc (create) and
d (re-summon) aliases to bashrc-linux.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CN1KSmsuLG6TxSeN5m8xvM
Add etc/profile.d/disk-usage-warning.sh (POSIX sh) that warns in bold
red at login when / or /home cross 85% usage. Deployed system-wide via
install_disk_warning() (sudo install -D -m 0644), gated inside the
apt-get block since df --output=pcent and /etc/profile.d are GNU/Debian
conventions absent on macOS. Idempotent and re-runnable.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CN1KSmsuLG6TxSeN5m8xvM
Replace the Wayland-incompatible xrdp approach with gnome-remote-desktop system
"Remote Login": idempotent self-signed TLS cert, RDP enable, and an interactive
gate-credential prompt (ensure_rdp_credentials) — skipped when already set, never
hardcoding the secret, with a non-TTY fallback that prints the manual command.
Without gate credentials the RDP server rejects every connection (mstsc 0x904).
Also bundle prior uncommitted installer work: install code-server (skip if
present), swap wkhtmltopdf -> weasyprint for PDF rendering, and append a
login-time dtach auto-router to ~/.profile.
Verified: bash -n + shellcheck clean. Not run live; RDP connection untested.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JSFhCiEgndbRrMf7s8mmth
The bin/ scripts and pipx CLIs land in ~/.local/bin, which was not on
PATH in either deployed bashrc. Add an idempotent PATH guard to
bashrc-linux and bashrc-osx so dt, claude-provider, pymupdf, etc. are
found after login. Update the install.sh and README notes to match.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Rework install.sh:
- detect OS via uname (Darwin -> bashrc-osx, else bashrc-linux);
remove the {server|linux|arm|osx} argument and bashrc-server
- fix broken /tmp/config paths that made server+osx install no bashrc
- switch #!/bin/sh -> bash, add set -euo pipefail, quote expansions
- clone nerdtree into bundle (was a no-op cp of a dir), drop the
redundant molokai clone (already tracked in vim/colors)
- guard apt-get behind command -v so macOS skips it
- make re-runnable; drop the no-op trailing 'source ~/.bashrc'
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>