Remote Machine Access

Manage remote machines defined in .remote-nodes/<host>.yaml: discover them, connect via SSH or Tailscale, run commands, transfer files, and add new nodes.

Rules

• MUST NOT store passwords, private keys, or secrets in configs. Authentication comes

only from SSH keys, ssh-agent, ~/.ssh/config, or Tailscale SSH.

• MUST run non-interactive commands with ssh -T (no PTY) — otherwise SIGHUP can kill

backgrounded processes when ~/.ssh/config sets RequestTTY=yes.

• MUST clean up any temporary tunnel you create (see Tunnels).
• MUST name each config file <host>.yaml, matching its host: field.
• Node configs MUST hold only connection info. MUST NOT add component-specific

metadata (model lists, Docker images, recipes, GPU specs, project paths) — a parent config merges into every sub-component's context, and such data duplicates what the component owns and goes stale.

Building the SSH Target

Every ssh/scp/rsync/tailscale ssh command uses the same target:

• If user: is set in the config, use user@hostname.
• If user: is omitted, use bare hostname and let ~/.ssh/config supply the user

(verify with ssh -G <host>).

Examples below use a bare host (gpu-box); substitute user@gpu-box when a user is set.

`hostname` resolution order (try until one connects):

1. DNS / mDNS (e.g. gpu-box.local)
2. ~/.ssh/config Host alias
3. Tailscale MagicDNS (if Tailscale is up)
4. tailscale_ip fallback
5. lan_ip fallback

Fall back to tailscale ssh when direct ssh fails or the node is only reachable on the tailnet.

Config Format

A config is a flat YAML file at .remote-nodes/<host>.yaml defining one machine.

Default to the minimal form; add fields only when ~/.ssh/config doesn't supply them:

# minimal
host: gpu-box
hostname: gpu-box

# with explicit fields
host: gpu-box
user: alice
hostname: gpu-box
port: 22
identity_file: ~/.ssh/id_ed25519
description: GPU workstation
lan_ip: 192.168.1.50
tailscale_ip: 100.64.0.1

Parser note: keep configs flat (no nested mappings) and put comments on their own lines, not inline after a value — some tooling reads these with a simple line parser, not a full YAML library.

Discovery & Resolution

To resolve a host the user names ("run this on gpu-box"):

1. Look in ./.remote-nodes/ first, then the parent dir's .remote-nodes/ (**one level

up only**). Merge both; the closer dir wins for a duplicate host.

1. Read the matching <host>.yaml, build the target, try ssh then tailscale ssh.

Sharing one config across deeply-nested components: because discovery scans only one level up, a component more than one level below a shared config can't see it. Symlink rather than copy, to keep a single source of truth:

# from the component's .remote-nodes/ dir; adjust ../ depth to reach the shared config
ln -s ../../../.remote-nodes/<host>.yaml <host>.yaml

Adding a Node

1. Pick the directory: the one the user names; else the current dir's .remote-nodes/;

else a named subproject's .remote-nodes/. Create it if missing.

1. Write <dir>/.remote-nodes/<host>.yaml (filename = host) using the minimal form.
2. Validate, then verify reachability:

python3 -c "import yaml; c=yaml.safe_load(open('.remote-nodes/gpu-box.yaml')); \
  assert c.get('host') and c.get('hostname'), 'missing host/hostname'; print('config OK')"
ssh -T gpu-box "hostname"        # or: tailscale ssh gpu-box "hostname"

Running Commands

ssh -T gpu-box "nvidia-smi"                  # non-interactive (always -T)
ssh -T gpu-box -i ~/.ssh/id_ed25519 "docker ps"
ssh gpu-box                                  # interactive session
tailscale ssh gpu-box "uptime"               # via Tailscale

# multi-line
ssh -T gpu-box <<'EOF'
  cd /path/to/project && git pull
EOF

Tunnels

Use -N -f for background tunnels. Always kill temporary tunnels when done.

ssh -L 8000:localhost:8000 gpu-box -N -f     # local: localhost:8000 → remote:8000
ssh -R 3000:localhost:3000 gpu-box -N -f     # remote: remote:3000 → local:3000
ssh -D 1080 gpu-box -N -f                     # dynamic SOCKS proxy

pkill -f 'ssh -[LRD]'                         # kill all tunnels (or kill <pid>)
lsof -i :8000                                 # verify gone

File Transfers

scp ./file gpu-box:/remote/path/             # push (scp gpu-box:/path ./ to pull)
scp -r ./dir gpu-box:/remote/dir/            # recursive
rsync -avz ./dir/ gpu-box:/remote/dir/       # sync (add --dry-run to preview)
rsync -avz --delete ./dir/ gpu-box:/remote/dir/   # mirror (deletes extra dest files)
rsync -avz --exclude='.git' ./ gpu-box:/remote/proj/
rsync -avz gpu-box:/remote/dir/ ./local/     # pull