t3-fleet-onboard
Onboard, re-check, or repair one fleet box, $H, over SSH from another box.
Every step is safe to re-run; on a healthy box the whole skill is a health
check. SKILL_DIR is this skill's directory. The roster and per-box facts
(paths, quirks, which box is dormant) are in fleet.md in
T3_code_improvements; read it first when you have it.
Run every remote command as ssh -o BatchMode=yes blaise@$H '...'. Find out
the box shape first with uname -s, because the steps differ:
Linux (archlinux, war, bigarch, littlearch) |
Mac (blaises-mini, blaises-macbook-air) |
|
|---|---|---|
| T3 | CLI service t3code.service (systemd user unit) |
desktop app T3 Code (Beta).app, updates itself |
t3 CLI |
npx t3@<version> |
none on PATH; use the app's UI |
| Serve mount | T3CODE_TAILSCALE_SERVE=true drop-in |
tailscale serve by hand |
| Skills sync | systemd user timer | LaunchAgent |
Don't install a service on a Mac or reach for systemctl there.
Preflight
- Ownership. Only the six boxes above.
nuc,teddy-computer,lovelandnuc, andm1airlinuxare shared or out of scope: never install on them, even if SSH reaches them. - Reachability.
$Hshould appear intailscale status. The real test isssh -o BatchMode=yes blaise@$H true. If SSH fails, stop and tell the user. Don't fall back to a password or another username. - Names.
warresolves some fleet names to stale LAN addresses. If SSH works but HTTPS hangs, comparegetent hosts $H.tail2b35ba.ts.netwithtailscale statusand dial the100.xaddress.
Mac that refuses SSH: macOS Remote Login is off. Only the user can fix
that at the keyboard: System Settings > General > Sharing > Remote Login,
then append war's and archlinux's ~/.ssh/id_ed25519.pub to the Mac's
~/.ssh/authorized_keys. Tailscale SSH doesn't replace it: on a Mac,
tailscale set --ssh succeeds but nothing listens.
Linux Tailscale SSH, as a backstop if keys break: tailscale set --ssh
on the box, then prove it from another box with
tailscale ssh blaise@$H hostname. It needs OperatorUser set to blaise
and an ssh rule in the tailnet ACL. If it fails for either reason, hand it
to the user.
Install
1. Node 24, user-local (needed on both shapes; Macs ship no node):
ssh -o BatchMode=yes blaise@$H '
set -e
case "$(uname -s)-$(uname -m)" in
Linux-x86_64) P=linux-x64 ;; Linux-aarch64) P=linux-arm64 ;;
Darwin-arm64) P=darwin-arm64 ;; Darwin-x86_64) P=darwin-x64 ;;
*) echo "unknown platform" >&2; exit 1 ;;
esac
test -x ~/.local/share/t3-node/bin/node || {
mkdir -p ~/.local/share/t3-node
curl -fsSL "https://nodejs.org/dist/v24.19.0/node-v24.19.0-$P.tar.xz" \
| tar -xJ --strip-components=1 -C ~/.local/share/t3-node; }
~/.local/share/t3-node/bin/node --version'
2. T3. On Linux:
ssh -o BatchMode=yes blaise@$H '
export PATH="$HOME/.local/share/t3-node/bin:$PATH"
env -u T3_SERVICE_LAUNCHER_CONTEXT npx -y t3@latest service install'
The CLI you run must match the version the box's service runs, or it fails
with The service launcher started a different t3 version. For a box held
at a version, use the version in ~/.t3/runtime/versions/. Back up
~/.t3/userdata/statev2.sqlite before an upgrade. On a Mac, only check that
the app is up; if it isn't, ask the user to launch it.
3. Serve the web UI at https://$H.tail2b35ba.ts.net/. On Linux, write
the drop-in (systemctl --user edit hangs under BatchMode):
ssh -o BatchMode=yes blaise@$H '
mkdir -p ~/.config/systemd/user/t3code.service.d
printf "[Service]\nEnvironment=T3CODE_TAILSCALE_SERVE=true\n" \
> ~/.config/systemd/user/t3code.service.d/override.conf
systemctl --user daemon-reload && systemctl --user restart t3code.service
loginctl enable-linger blaise'
On a Mac, run tailscale serve --bg --set-path=/ http://127.0.0.1:3773. Also
pin the name with tailscale set --hostname=$H if tailscale debug prefs
shows an empty Hostname. An unpinned Mac renames itself after a
tailscaled restart, and serve keeps the old name.
4. Providers. An installed CLI isn't enough; it must be signed in. The
pong check under Verify is the proof. Fix it with ssh -t blaise@$H claude
and /login, unless fleet.md says the box uses a token or the proxy.
Skills
Use the HTTPS URL, because fleet boxes have no GitHub SSH key:
ssh -o BatchMode=yes blaise@$H '
set -e; export PATH="$HOME/.local/bin:/opt/homebrew/bin:$PATH"
claude plugin marketplace add https://github.com/BlaiseBaptist/blaise-skills.git
claude plugin install t3-fleet@blaise-skills --scope user'
Every box gets t3-fleet. Only desktop boxes get desktop@blaise-skills.
The hourly sync then updates Claude and Codex and installs Codex's copy
itself. Install the sync script and its timer:
S="$SKILL_DIR/scripts"
ssh -o BatchMode=yes blaise@$H 'mkdir -p ~/.local/bin && cat > ~/.local/bin/claude-skills-sync && chmod +x ~/.local/bin/claude-skills-sync' < "$S/claude-skills-sync.sh"
# Linux
cat "$S/claude-skills-sync.service" | ssh -o BatchMode=yes blaise@$H 'mkdir -p ~/.config/systemd/user && cat > ~/.config/systemd/user/claude-skills-sync.service'
cat "$S/claude-skills-sync.timer" | ssh -o BatchMode=yes blaise@$H 'cat > ~/.config/systemd/user/claude-skills-sync.timer'
ssh -o BatchMode=yes blaise@$H 'systemctl --user daemon-reload && systemctl --user enable --now claude-skills-sync.timer'
# Mac
sed "s|HOME_DIR|/Users/blaise|g" "$S/com.blaise.claude-skills-sync.plist" | ssh -o BatchMode=yes blaise@$H '
P=~/Library/LaunchAgents/com.blaise.claude-skills-sync.plist; mkdir -p ~/Library/LaunchAgents; cat > $P
launchctl bootout gui/$(id -u)/com.blaise.claude-skills-sync 2>/dev/null || true
launchctl bootstrap gui/$(id -u) $P'
Judge a run by its log (journalctl --user -u claude-skills-sync or
~/Library/Logs/claude-skills-sync.log), not the exit code. Updates apply
only to sessions started afterwards.
Verify
All three must pass:
- T3 up. On Linux,
systemctl --user is-active t3code.servicereturnsactive. On a Mac,curl -fsS -o /dev/null -w '%{http_code}' http://127.0.0.1:3773/returns200. - Served. From another box,
curl -fsS -o /dev/null -w '%{http_code}' https://$H.tail2b35ba.ts.net/returns200. Check fromarchlinuxrather thanwar(stale DNS). - Signed in.
claude -p "Reply with exactly: pong"printspong. Read the output, because it exits 0 even when auth has expired. On Linux, wrap it intimeout 90. Macs have notimeout, so bound it with-o ConnectTimeout=10from your side, and exportPATH=$HOME/.local/bin:/opt/homebrew/bin:$PATHfirst.
Then make sure the box has a T3 project for its workspace. Add it in the web UI if it's missing.
Pairing and Connect
To pair a browser or phone over the tailnet on Linux, run
env -u T3_SERVICE_LAUNCHER_CONTEXT npx -y t3@<version> pair --tailscale.
It prints a /pair#token= URL that expires in 5 minutes, so mint it right
before use. On a Mac, pair from the app. T3 Connect (t3 connect, the relay)
is a different mechanism. The fleet uses it, so don't link, unlink, or relink
it during onboarding.
Traps
T3_SERVICE_LAUNCHER_CONTEXT. A session hosted by T3 inherits it, and everyt3CLI call then fails withipc-unavailable. Always runenv -u T3_SERVICE_LAUNCHER_CONTEXT.- Mac
PATHover SSH. It isn't a login shell, soclaude,tailscale, andbrewlook missing. Export~/.local/bin:/opt/homebrew/binin every remote command. - Versions. There's no fleet-wide pin, and boxes may differ, but a 0.0.x
(V1) box can't talk to a 0.1.x (V2) one. Pinning one box means installing
an exact
t3@<version>. Nothing re-pins it later.