ssh-server-ops

name	ssh-server-ops
description	Use when Codex needs to bootstrap, diagnose, transfer files, or operate SSH access from Windows, macOS, or Linux to Linux, POSIX, gateway, bastion, or Slurm/HPC targets using the bundled SSH Server Ops toolkit.

SSH Server Ops

This skill is the agent wrapper for the bundled SSH toolkit in scripts/. The toolkit is the product surface; the skill defines the safe workflow and which entrypoint to use.

Entrypoints

Choose the local OS entrypoint:

powershell -ExecutionPolicy Bypass -File .\scripts\sshops.ps1 <subcommand> ...

python3 ./scripts/sshops.py <subcommand> ...

Use the PowerShell entrypoint on Windows when you need Windows OpenSSH ACL repair. Use the Python entrypoint on macOS/Linux, or on Windows when cross-platform parity is more important than ACL repair.

Current subcommands:

• doctor: read-only SSH diagnosis and environment inspection
• configure: create or repair a host block in ~/.ssh/config
• bootstrap-key: install a public key over one password-backed session, then verify non-interactive SSH
• transfer: password-backed Paramiko upload/download fallback
• run: standardize remote command execution through ssh

When to Use

Use this skill when:

• a local Windows, macOS, or Linux machine needs to reach a Linux, POSIX, gateway, bastion, or HPC target over SSH
• SSH partially works and Codex needs to determine which layer is broken
• Codex needs to repair ~/.ssh/config, key auth, identity selection, or local OpenSSH assumptions
• file transfer is failing and Codex needs a safe fallback path
• the task lives on a login node or behind a bastion, managed gateway, or Slurm scheduler

Do not use this skill for Windows remoting, RDP, or cases where the user explicitly wants to avoid SSH automation.

Security Rules

• Never ask the user to paste passwords, tokens, private keys, or MFA codes into chat.
• If password bootstrap is required, the password must stay in the local terminal session only, usually in SSH_SERVER_PASSWORD.
• Never store passwords in ~/.ssh/config, SKILL.md, prompt templates, repo files, or logs.
• Treat destructive remote actions the same way as destructive local actions: explicit user intent first.
• doctor is the preferred first step because it is read-only.

Workflow

1. Gather facts

Collect:

• alias, host or IP, port, username
• local OS: Windows, macOS, or Linux
• whether the target is a normal Linux host, jump host, managed gateway, or HPC login node
• whether key auth, interactive password auth, or both are available today
• whether the task is diagnose, configure, bootstrap, transfer, run, sync, or HPC job control

2. Run read-only diagnosis first

Prefer doctor before changing anything.

Windows:

powershell -ExecutionPolicy Bypass -File .\scripts\sshops.ps1 doctor -Alias <alias>

macOS/Linux:

python3 ./scripts/sshops.py doctor --alias <alias>

If the alias is not configured yet, use direct mode.

Windows:

powershell -ExecutionPolicy Bypass -File .\scripts\sshops.ps1 doctor `
  -HostName <host> `
  -Port <port> `
  -User <user>

macOS/Linux:

python3 ./scripts/sshops.py doctor \
  --host-name <host> \
  --port <port> \
  --user <user>

Read the structured output, especially:

• local_tools
• network
• auth
• transfer
• remote_shell
• likely_root_cause
• recommended_next_step

Use the diagnosis to decide the next safe action instead of guessing.

3. Configure the host entry only when needed

If the host is not saved or the alias is wrong, write a focused ~/.ssh/config block.

Windows:

$keyPath = Join-Path $HOME ".ssh\id_ed25519_<alias>"
powershell -ExecutionPolicy Bypass -File .\scripts\sshops.ps1 configure `
  -Alias <alias> `
  -HostName <host> `
  -Port <port> `
  -User <user> `
  -IdentityFile $keyPath `
  -PreferredAuthentications publickey

macOS/Linux:

python3 ./scripts/sshops.py configure \
  --alias <alias> \
  --host-name <host> \
  --port <port> \
  --user <user> \
  --identity-file ~/.ssh/id_ed25519_<alias> \
  --preferred-authentications publickey

Report what changed because this modifies local SSH state.

4. Bootstrap key auth only when password auth exists locally

If non-interactive SSH is still broken but the user has a valid password locally, bootstrap once.

Windows:

$env:SSH_SERVER_PASSWORD = "<set-locally-in-terminal>"
powershell -ExecutionPolicy Bypass -File .\scripts\sshops.ps1 bootstrap-key `
  -HostName <host> `
  -Port <port> `
  -User <user> `
  -PublicKey <path-to-public-key> `
  -PrivateKey <path-to-private-key>
Remove-Item Env:SSH_SERVER_PASSWORD

macOS/Linux:

export SSH_SERVER_PASSWORD="<set-locally-in-terminal>"
python3 ./scripts/sshops.py bootstrap-key \
  --host-name <host> \
  --port <port> \
  --user <user> \
  --public-key <path-to-public-key> \
  --private-key <path-to-private-key>
unset SSH_SERVER_PASSWORD

Do not ask the user to put the password in chat. If the password is not already available in the local terminal, tell the user what variable to set and wait.

5. Use the toolkit for execution and transfer

Remote command execution:

python3 ./scripts/sshops.py run \
  --alias <alias> \
  --remote-dir ~/repo \
  --command "git status --short" \
  --bash \
  --batch-mode

Password-backed file transfer fallback:

export SSH_SERVER_PASSWORD="<set-locally-in-terminal>"
python3 ./scripts/sshops.py transfer \
  --host-name <host> \
  --port <port> \
  --user <user> \
  --direction download \
  --remote-path <remote-path> \
  --local-path <local-path>
unset SSH_SERVER_PASSWORD

Prefer direct OpenSSH operations when they are already healthy. Use Paramiko fallback when diagnosis shows OpenSSH auth or transfer is the weak link.

6. Treat HPC as a first-class workflow

On HPC or Slurm-backed systems:

• distinguish login nodes from compute nodes
• inspect scheduler state before launching work
• use squeue, sinfo, sbatch, srun, and log collection deliberately
• avoid treating a compute node shell like a durable host unless the scheduler says it is

See references/hpc-slurm-playbook.md.

Risk Levels

• doctor: read-only
• configure, bootstrap-key: low-risk local or account configuration changes
• run, transfer: medium risk, because they can affect remote state depending on command or target path
• delete, reset, overwrite, restart, package install, or service-impacting actions: high risk and require explicit user approval

Output Expectations

Report:

• the target and how it was addressed, alias or direct
• the diagnosis result, including likely_root_cause and recommended_next_step
• which toolkit subcommands were used
• any local or remote state changes
• exact verification commands and observed results
• residual risks, especially on shared or HPC systems

References

• README.md: public project framing and direct toolkit usage
• SECURITY.md: secrets handling and publishing checklist
• references/windows-linux-ssh-playbook.md: common Windows-to-Linux run, sync, and diagnose patterns
• references/hpc-slurm-playbook.md: login-node and Slurm workflows
• references/prompt-templates.md: safe prompt templates that avoid putting secrets in chat