eageric/cursor-cloud-eag

EAG Proxmox over the Cloudflare SSH tunnel

The Proxmox VE 9.2 server hosts disposable test VMs. From this environment the host is eric-pve-ssh.eag.tools. Agents clone a template, run the test, and then ask before deleting anything.

The templates, clone steps, and guest logins match the LAN environment. What changes is the path: the cloud agent reaches the host through Cloudflare Access, and reaches every guest by jumping through that host.

Access

The environment start script writes the key pair from PVE_SSH_PRIVATE_B64 and PVE_SSH_PUBLIC_KEY:

File	Mode	Use
~/.ssh/pve_ssh	600	Private key. Proxmox host, Windows Administrator, Ubuntu erolleman.
~/.ssh/pve_ssh.pub	644	Matching public key.

/etc/ssh/ssh_config.d/eric-pve-ssh.conf already sets User root, IdentityFile ~/.ssh/pve_ssh, IdentitiesOnly yes, and ProxyCommand /usr/local/bin/cloudflared access ssh --hostname %h for eric-pve-ssh.eag.tools. Use that host alias. Leave the system file in place.

KEY="$HOME/.ssh/pve_ssh"
test -r "$KEY" || { echo "missing $KEY; the start script creates it from PVE_SSH_PRIVATE_B64" >&2; exit 1; }
PVE="ssh -o StrictHostKeyChecking=accept-new eric-pve-ssh.eag.tools"
$PVE 'qm list'

Guest SSH, for both 10.115.64.x (vmbr0) and 10.20.0.x (nat0):

JUMP='ssh -o StrictHostKeyChecking=accept-new -W %h:%p eric-pve-ssh.eag.tools'
VMSSH="ssh -i $KEY -o IdentitiesOnly=yes -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null -o ConnectTimeout=5 -o ProxyCommand=\"$JUMP\" USER@$IP"

Set USER to Administrator or erolleman and IP to the guest address before expanding VMSSH.

• Windows VM login: Administrator / Change_me, or ~/.ssh/pve_ssh.
• Ubuntu VM login: erolleman with ~/.ssh/pve_ssh. sudo is passwordless. docker needs sudo.
• The Proxmox web UI listens on the LAN at https://10.115.64.44:8006. This tunnel is SSH only. Administer the host with qm and pvesh over $PVE.
• Keep key material in the files above. Read neither ~/.ssh/pve_ssh nor PVE_SSH_PRIVATE_B64 into a command, a log, or a reply. Generate no replacement key.

PowerShell remoting over SSH uses the same jump. Add this to ~/.ssh/config (the guest Host block only; the system file already covers the tunnel host):

Host 10.115.64.* 10.20.0.*
  IdentityFile ~/.ssh/pve_ssh
  IdentitiesOnly yes
  StrictHostKeyChecking no
  UserKnownHostsFile /dev/null
  ProxyCommand ssh -o StrictHostKeyChecking=accept-new -W %h:%p eric-pve-ssh.eag.tools

Enter-PSSession -HostName <ip> -UserName Administrator -KeyFilePath "$HOME/.ssh/pve_ssh"

Hard rules

• Never start, modify, clone over, or delete VM 100 (ubuntu-2404) or VM 101 (ws2025-template). Both are templates.
• Do not change host configuration (storage, network, SDN, packages) unless the user explicitly asks.
• Do not re-run sysprep, windows-template/Copy-SysprepFiles.ps1, or ubuntu-template/cloud-init-prepare.sh on a clone. Those scripts prepare the templates and stay on the workstation. They are not in this environment.
• Tag every VM you create with ai-test.
• When testing is finished, ask the user whether to destroy the VMs you created. Never destroy a VM without the user's answer.
• Open no direct SSH session to 10.115.64.44, 10.115.64.x, or 10.20.0.x. Those networks are not routed to the cloud agent. Host commands go through eric-pve-ssh.eag.tools. Guest commands use the ProxyCommand jump above.

Host facts

The host is a Hyper-V virtual machine. Nested VT-x is exposed. On the LAN its address is 10.115.64.44.

Item	Value
Hardware	Intel Xeon Gold 6234 (8 cores / 16 threads), 94 GB RAM. Budget VMs accordingly.
vmbr0	LAN bridge, DHCP. Host address 10.115.64.44/24, gateway and DNS 10.115.64.1, search domain eag.lan. Both templates' NICs are on this bridge. Guest addresses are 10.115.64.x. From this environment, SSH to them jumps through eric-pve-ssh.eag.tools.
nat0	SDN VNet (simple zone nat), 10.20.0.0/24. Gateway 10.20.0.1. DHCP 10.20.0.100–10.20.0.200 (12 h) from dnsmasq on the host, DNS 10.20.0.1. SNAT out vmbr0 as 10.115.64.44. The Proxmox host can reach this subnet. From this environment, SSH to 10.20.0.x uses the same jump.
local-data	LVM-thin on VG vmdata, ~671 GB. VM disks, EFI, TPM, and cloud-init drives.
local-lvm	LVM-thin on the small PVE system VG, ~23 GB, empty. Leave it unused.
local	Directory on the root filesystem. Backups, container templates, imports.
isos	Directory /mnt/isos. Install ISOs only.

Which template

Need	Clone
Windows, RDP, or Windows PowerShell	101 ws2025-template
Linux, Docker, pwsh, or Nushell	100 ubuntu-2404

If the user does not say which OS, ask before cloning.

Windows clones enable RDP inside the guest. The Cloudflare tunnel carries SSH. Use SSH or PowerShell remoting to reach the guest.

Template: ws2025-template (VMID 101)

Windows Server 2025 Datacenter Evaluation. 4 vCPU (host), 4 GB RAM, UEFI + Secure Boot + TPM 2.0, 64 GB VirtIO SCSI disk on local-data, VirtIO NIC on vmbr0 (4 queues), VirtIO drivers and QEMU guest agent installed.

On first boot, every clone automatically:

• Gets a random computer name, Pacific time zone, en-US locale, and Administrator password Change_me.
• Auto-logs on once and runs Initialize-WindowsServer.ps1, which installs the latest PowerShell 7 from GitHub, enables OpenSSH server with a powershell subsystem (PS remoting over SSH), adds the SSH key to administrators_authorized_keys, enables RDP, sets Public networks to Private, and removes the Ctrl+Alt+Del requirement.
• Removes the auto-logon password and the answer file, writes C:\ProgramData\TemplateSetup\setup-complete, then restarts.

First boot takes several minutes and needs internet (PowerShell 7 download). Log: C:\ProgramData\TemplateSetup\Initialize-WindowsServer.log.

Template: ubuntu-2404 (VMID 100)

Ubuntu 24.04.5 LTS. 2 vCPU (x86-64-v2-AES), 4 GB RAM, UEFI (Secure Boot keys not enrolled), no TPM, 32 GB VirtIO SCSI disk on local-data, VirtIO NIC on vmbr0. QEMU guest agent, OpenSSH, PowerShell 7 (pwsh), Nushell (nu on /usr/local/bin), and Docker are installed. Podman is not installed. Time zone is UTC.

Cloud-init has been cleaned and limited to the NoCloud and ConfigDrive datasources. The disk has no netplan, so a clone has no network until Proxmox cloud-init runs. The account erolleman is already in sudo and already has the project public key. SSH is socket-activated. The guest agent starts from udev when the virtio port appears.

Proxmox's generated user-data also creates a locked ubuntu account with no SSH key. Log in as erolleman.

Workflow: create a test VM

- [ ] 1. Confirm the source VM is a template
- [ ] 2. Clone, configure, tag, start
- [ ] 3. Get the IP from the guest agent
- [ ] 4. Wait until first boot has finished
- [ ] 5. Run the test
- [ ] 6. Ask the user whether to destroy the VM

1. Confirm the template. If this prints nothing, stop and tell the user. Do not clone a VM that is not a template.

$PVE 'qm config 101 | grep ^template'   # Windows
$PVE 'qm config 100 | grep ^template'   # Ubuntu

2. Clone, configure, tag, start. Linked clones are the default. Add --full --storage local-data only when the VM must outlive the template.

Clones inherit 4 vCPU / 4 GB (Windows) or 2 vCPU / 4 GB (Ubuntu). When the test needs more, qm set the clone before starting it. Leave the templates unchanged.

Windows:

ID=$($PVE 'pvesh get /cluster/nextid')
$PVE "qm clone 101 $ID --name test-<purpose> && qm set $ID --tags ai-test && qm start $ID"

Ubuntu: attach the cloud-init drive before the first start. ide2 on the template is an empty CD; replace it. --ciupgrade 0 skips the default full package upgrade. Without --ipconfig0 ip=dhcp the clone never gets an address.

ID=$($PVE 'pvesh get /cluster/nextid')
$PVE "qm clone 100 $ID --name test-<purpose> && qm set $ID --tags ai-test --ide2 local-data:cloudinit --citype nocloud --ipconfig0 ip=dhcp --ciupgrade 0 --boot order=scsi0 && qm start $ID"

Isolated network, either OS: before start, put the NIC on nat0. Keep this only when the user wants the VM off the LAN. The guest address will be 10.20.0.x. SSH from this environment still uses the jump in Access; the Proxmox host is the one that can reach nat0.

$PVE "qm set $ID --net0 virtio,bridge=nat0"
# Ubuntu only. Windows DHCPs on its own. Point Ubuntu DNS at dnsmasq on the NAT gateway.
$PVE "qm set $ID --ipconfig0 ip=dhcp --nameserver 10.20.0.1"

3. Get the IP. The guest agent answers once the OS is up. Retry every ~15 s. Use the 10.115.64.x address for a vmbr0 NIC, or the 10.20.0.x address for nat0.

$PVE "qm guest cmd $ID network-get-interfaces" \
  | jq -r '.[]."ip-addresses"[]? | select(."ip-address-type"=="ipv4") | ."ip-address"' \
  | grep -vE '^(127\.|169\.254\.)'

4. Wait for first boot. Clones get new host keys and addresses get reused, so guest SSH skips host key checking. The jump host key is stored with accept-new.

Windows: SSH comes up during setup, and the script restarts the VM when it finishes. The VM is ready once setup-complete exists and Windows has booted again since it was written.

VMSSH="ssh -i $KEY -o IdentitiesOnly=yes -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null -o ConnectTimeout=5 -o ProxyCommand=\"$JUMP\" Administrator@$IP"
# -EncodedCommand (base64 UTF-16LE) avoids nested bash/cmd/PowerShell quoting.
PS_CHECK=$(cat <<'EOF'
$m = Get-Item C:\ProgramData\TemplateSetup\setup-complete -ErrorAction SilentlyContinue
if ($m -and (Get-CimInstance Win32_OperatingSystem).LastBootUpTime -gt $m.LastWriteTime) { 'READY ' + (Get-Content $m.FullName) } else { 'WAITING' }
EOF
)
CHECK="powershell -NoProfile -EncodedCommand $(printf %s "$PS_CHECK" | iconv -t UTF-16LE | base64 -w0)"
until STATUS=$($VMSSH "$CHECK" 2>/dev/null | tr -d '\r'); [[ $STATUS == READY* ]]; do sleep 15; done
echo "$STATUS"

READY succeeded means setup finished and the VM has restarted. If it prints READY failed, read the log and report it:

$VMSSH 'powershell -NoProfile -Command "Get-Content C:\ProgramData\TemplateSetup\Initialize-WindowsServer.log -Tail 30"'

Ubuntu: wait until cloud-init reports status: done. status: error means stop and read /var/log/cloud-init-output.log.

VMSSH="ssh -i $KEY -o IdentitiesOnly=yes -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null -o ConnectTimeout=5 -o ProxyCommand=\"$JUMP\" erolleman@$IP"
until STATUS=$($VMSSH 'cloud-init status 2>/dev/null' | tr -d '\r'); [[ $STATUS == status:\ done || $STATUS == status:\ error ]]; do sleep 10; done
echo "$STATUS"

5. Run the test.

Windows: the default SSH shell is cmd.exe.

$VMSSH 'pwsh -NoProfile -Command "Get-ComputerInfo | Select-Object OsName, CsName"'

PowerShell remoting over SSH (from pwsh on the cloud agent) uses the ~/.ssh/config guest block in Access:

Enter-PSSession -HostName <ip> -UserName Administrator -KeyFilePath "$HOME/.ssh/pve_ssh"

Ubuntu:

$VMSSH 'pwsh -NoLogo -Command "$PSVersionTable.PSVersion.ToString()"'
$VMSSH 'sudo docker info --format "{{.ServerVersion}}"'

6. Ask before cleaning up. Report the test results, list the VMs you created (VMID, name, IP), and ask the user whether to destroy them. Use the AskQuestion tool when available, with options "Destroy", "Keep running", and "Shut down but keep" (plus per-VM choices if you created several).

If the user chooses to destroy:

$PVE "qm stop $ID && qm destroy $ID --purge"

If the user chooses to keep but shut down:

$PVE "qm shutdown $ID --timeout 120 || qm stop $ID"

When VMs are kept, tell the user their VMIDs so they can be found later (they keep the ai-test tag).

Snapshots

Both templates' disks are on LVM-thin, so clones support snapshots:

$PVE "qm snapshot $ID before-test"
$PVE "qm rollback $ID before-test"
$PVE "qm delsnapshot $ID before-test"

Cleanup and inventory

$PVE 'pvesh get /cluster/resources --type vm --output-format json' \
  | jq -r '.[] | select((.tags // "") | split(";") | index("ai-test")) | "\(.vmid)\t\(.name)\t\(.status)"'

Only offer to destroy ai-test VMs you created in this session, unless the user asks for a broader cleanup. Always get the user's confirmation first.

Gotchas

• Windows is an evaluation edition with a 180-day license. Use it for short-lived VMs.
• An Ubuntu clone that boots once without the cloud-init drive comes up with no network, and cloud-init will not retry cleanly. Destroy it and clone again.
• Do not run sysprep on Windows clones, and do not convert clones of either template into templates.
• The host address is a DHCP lease. nat0 SNAT is pinned to 10.115.64.44. If that address changes, say so; do not edit SDN unless the user asks.
• cloudflared is invoked as /usr/local/bin/cloudflared by the system SSH config. A missing key file means the start script did not run; leave the secrets untouched and report that.
• Guest host keys change every clone. UserKnownHostsFile=/dev/null applies to the guest session. The tunnel host key stays in ~/.ssh/known_hosts.