Claude Skill

network-device-cli

Use when SSHing into a router, switch, or firewall CLI.

LLM Mart · 0 points · 2 views 0 listing impressions 0 install-command copies
Download dandi41-ntnroute-skills_network-device-cli-aeff5d4.zip · 10 KB

Install

skills CLI npx skills add https://github.com/Dandi41/ntnroute/tree/main/skills/network-device-cli
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install dandi41-ntnroute@llmmart
Git git clone https://github.com/Dandi41/ntnroute.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole dandi41/ntnroute collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Network Device CLI (Cisco IOS and other appliances)

When to Use

  • The user gives a host + credentials and asks for a command (ping, show) to be run on the device — router, switch, firewall, appliance.
  • An SSH attempt to network gear fails with Permission denied or no acceptable kex algorithm.
  • You need device-side reachability (can the router reach 8.8.8.8?) rather than host-side.
  • A request wants device config changed (interface, VLAN, route, ACL), not just read — see "Before a config change" below.
  • You are rebuilding a device-inventory / automation workflow that must survive session resets — the inventory (hostname, ip, port, credentials, config_list, checkpoint) belongs on disk, never in the conversation.

Not for a Linux server over SSH — plain ssh user@host works there.

Network gear is not a Linux box. It runs an exec-mode CLI with paging, restricted exec commands, and — on anything a few years old — SSH algorithms that modern OpenSSH disables by default. Follow this order; each step fails in a way that misdiagnoses as "wrong password" if you skip it.

1. Confirm reachability, then read the banner

ping -c1 -W2 <host>
ssh -v -o ConnectTimeout=10 -o StrictHostKeyChecking=no \
    -o UserKnownHostsFile=/dev/null <user>@<host> exit 2>&1 \
  | grep -iE "remote software|kex: algorithm|Authentications that can continue"

remote software version Cisco-1.25 plus kex: algorithm: (no match) means legacy crypto, NOT bad credentials. Solve the crypto before you touch the password — otherwise every auth attempt reports Permission denied and you burn time guessing passwords that were right all along.

2. Force legacy algorithms per connection

Modern OpenSSH still ships these algorithms; it only stops offering them. So the fix is flags on the invocation, not an install and not a config edit:

-o KexAlgorithms=+diffie-hellman-group14-sha1,diffie-hellman-group1-sha1,diffie-hellman-group-exchange-sha1 \
-o HostKeyAlgorithms=+ssh-rsa -o PubkeyAcceptedAlgorithms=+ssh-rsa \
-o Ciphers=+aes128-cbc,3des-cbc -o MACs=+hmac-sha1 \
-o PreferredAuthentications=password -o NumberOfPasswordPrompts=1

Do not edit /etc/ssh/ssh_config.d/*.conf for this: it needs root, and it weakens every future SSH connection instead of one. Keep the legacy flags inside the invocation (or a Host block in ~/.ssh/config scoped to that IP).

3. Password auth without sshpass/expect

sshpass is usually absent and installing it needs root. OpenSSH's own askpass hook works fine and keeps the password out of argv and out of logs:

cat > /tmp/askpass.sh <<'EOF'
#!/bin/bash
printf '%s\n' "$SSH_PROBE_PW"
EOF
chmod +x /tmp/askpass.sh

SSH_PROBE_PW="$PW" SSH_ASKPASS=/tmp/askpass.sh SSH_ASKPASS_REQUIRE=force DISPLAY=:0 \
  ssh -tt <legacy flags> <user>@<host> '<command>'

SSH_ASKPASS_REQUIRE=force is what makes it work without setsid. Do not put the password inline in the script or on the ssh command line.

4. Issue commands the way the device accepts them

  • One command per connection on IOS. A single string with IOS arguments (ping 8.8.8.8 repeat 5 timeout 2) is rejected as Line has invalid autocommand "…" — exec mode takes the bare form only (ping 8.8.8.8, default 5 packets). For arguments, drive an interactive session instead of exec mode.
  • Hold the channel open. With < /dev/null the session closes the moment auth completes and you get a truncated reply (a lone ! from a 5-packet ping). Pipe a sleep into it instead:
    { sleep 25; } | ssh -tt <flags> <user>@<host> 'ping 8.8.8.8' 2>&1 | tr -d '\r'
    
  • -tt allocates the PTY the CLI expects; without it IOS output arrives mangled or not at all. tr -d '\r' afterwards makes the output readable.
  • Paging: terminal length 0 as an exec command can drop the session. Prefer reading short show output, or set paging inside an interactive session only.

5. Report the device's own verdict

Quote the summary line verbatim — Success rate is 100 percent (5/5), round-trip min/avg/max = 8/36/91 ms, %SYS-5-CONFIG_I, state: up. Do not paraphrase into your own numbers, and do not claim success without the device's line in the output.

6. Before a config change: plan, approval, checkpoint

Reading is cheap; writing is not. For anything that changes state:

  1. Read first. Pull the running config and the relevant show output before proposing a diff — never configure from memory of what the device "should" look like.
  2. Present the plan before touching the device, and stop there. Five blocks, in this order: (a) what changes and why, in one short reason; (b) the exact command list, verbatim; (c) the exact rollback command list; (d) the read-only command that will verify it afterwards plus the output/state you expect; (e) risk — which lines can cut you off, and why they do not here. A confirmation that has not seen all five is not approval. This applies even when the user asked for the change in the same message, and even for one line. If any line changes after approval, the old approval is void: re-present, re-ask.
  3. Save the pre-change running config to a file, and write the rollback script before applying. references/config-change-safety.md has the checkpoint → apply → save → verify → rollback loop and the per-vendor command profiles. Where the network-change-management skill is installed, run its rollback generator: rollback_gen <script> --existing-config <saved-config> drafts the rollback — read it line by line and fill its # TODO restore from snapshot: entries before use.
  4. Never send a command that can cut your own path back in: VTY/management interface, access-class, default route, the interface you are connected through.
  5. Verify with the command promised in (d) once the change lands, and quote the device's own output line. The device echoing the config line is not verification; reading the resulting state back is.

A change is not "done" because the device echoed the line. Re-read the running config and check the lines are actually present; a capture that shows no error is weaker evidence than a captured config containing the line.

7. Persistent interactive session (OpenSSH)

One-shot OpenSSH suits a single show. For a session that enters enable with a secret and then runs a dozen commands, hold one PTY open instead of spawning an ssh per command — the engine's own technique, no paramiko, no Python:

SSH_ASKPASS=/tmp/askpass SSH_ASKPASS_REQUIRE=force setsid \
  ssh -tt -o KexAlgorithms=+diffie-hellman-group14-sha1 \
      -o HostKeyAlgorithms=+ssh-rsa -o PubkeyAcceptedAlgorithms=+ssh-rsa \
      -o Ciphers=+aes128-cbc -o MACs=+hmac-sha1 admin@10.0.0.5 < plan
  • SSH_ASKPASS + setsid (no controlling tty) supplies the password without sshpass; SSH_ASKPASS_REQUIRE=force needs OpenSSH 8.4+.
  • -tt forces the PTY so the device prints prompts; feed a paced plan of delay<TAB>command pairs on stdin — pause after connect, then one line at a time. Pacing replaces read-until-idle: gear that pauses mid-output otherwise gets the next command typed into the middle of the previous one.
  • On a piped PTY the prompt is never a standalone line: the device prints it and the next command lands on the same line (R2>show clock). Parse prompts as a line prefix (^[A-Za-z0-9_.-]{1,63}[>#]), not a line suffix.
  • Disable paging first (terminal length 0); a stray --More-- cannot be answered from a pipe.
  • enable: send enable, pause, send the secret, then confirm the prompt ends in #. Report the prompt you actually saw rather than assuming the enable succeeded.
  • Keep the password out of argv — the askpass helper reads it from the environment.

scripts/ios_session.sh is this session shape as a runnable CLI (same flags as the old paramiko version: --host --user --password --secret --cmd --show-config --paging).

Credential parsing pitfall

Credentials arrive as run-together prose: user admin pass S3cr3t! secret en4ble! is user admin, password S3cr3t!, and en4ble! is the enable secret (privileged exec password) — not a passphrase with spaces. Try the whole string, then the single token, and report which one authenticated rather than silently using it. A Cisco banner plus a rejected multi-word password is the signal that you split the string wrong.

Supporting files

  • scripts/legacy-ssh.sh — wraps the flags, askpass, and hold-open stdin so a one-shot command works first try.
  • scripts/ios_session.sh — persistent OpenSSH PTY session (enable, paging off, read running config) as a runnable CLI.
  • references/config-change-safety.md — checkpoint → apply → save → verify → rollback loop, rollback-script rules, and per-vendor command profiles.
Files (ntnroute)
  • references
    • config-change-safety.md 5.9 KB
      Loading…
  • scripts
    • ios_session.sh 6.2 KB
      Loading…
    • legacy-ssh.sh 1.5 KB
      Loading…
  • SKILL.md 9 KB
    ---
    name: network-device-cli
    description: "Use when SSHing into a router, switch, or firewall CLI."
    version: 1.1.0
    author: Hermes Agent
    license: MIT
    metadata:
      hermes:
        tags: [ssh, cisco, ios, network, legacy-crypto, cli]
    ---
    
    # Network Device CLI (Cisco IOS and other appliances)
    
    ## When to Use
    
    - The user gives a host + credentials and asks for a command (ping, show) to be
      run **on** the device — router, switch, firewall, appliance.
    - An SSH attempt to network gear fails with `Permission denied` or
      `no acceptable kex algorithm`.
    - You need device-side reachability (can the router reach 8.8.8.8?) rather than
      host-side.
    - A request wants device config **changed** (interface, VLAN, route, ACL), not
      just read — see "Before a config change" below.
    - You are rebuilding a device-inventory / automation workflow that must survive
      session resets — the inventory (hostname, ip, port, credentials, config_list,
      checkpoint) belongs on disk, never in the conversation.
    
    Not for a Linux server over SSH — plain `ssh user@host` works there.
    
    Network gear is not a Linux box. It runs an exec-mode CLI with paging, restricted
    exec commands, and — on anything a few years old — SSH algorithms that modern
    OpenSSH disables by default. Follow this order; each step fails in a way that
    misdiagnoses as "wrong password" if you skip it.
    
    ## 1. Confirm reachability, then read the banner
    
    ```bash
    ping -c1 -W2 <host>
    ssh -v -o ConnectTimeout=10 -o StrictHostKeyChecking=no \
        -o UserKnownHostsFile=/dev/null <user>@<host> exit 2>&1 \
      | grep -iE "remote software|kex: algorithm|Authentications that can continue"
    ```
    
    `remote software version Cisco-1.25` plus `kex: algorithm: (no match)` means
    legacy crypto, NOT bad credentials. Solve the crypto before you touch the
    password — otherwise every auth attempt reports `Permission denied` and you burn
    time guessing passwords that were right all along.
    
    ## 2. Force legacy algorithms per connection
    
    Modern OpenSSH still *ships* these algorithms; it only stops offering them. So
    the fix is flags on the invocation, not an install and not a config edit:
    
    ```
    -o KexAlgorithms=+diffie-hellman-group14-sha1,diffie-hellman-group1-sha1,diffie-hellman-group-exchange-sha1 \
    -o HostKeyAlgorithms=+ssh-rsa -o PubkeyAcceptedAlgorithms=+ssh-rsa \
    -o Ciphers=+aes128-cbc,3des-cbc -o MACs=+hmac-sha1 \
    -o PreferredAuthentications=password -o NumberOfPasswordPrompts=1
    ```
    
    Do **not** edit `/etc/ssh/ssh_config.d/*.conf` for this: it needs root, and it
    weakens every future SSH connection instead of one. Keep the legacy flags inside
    the invocation (or a `Host` block in `~/.ssh/config` scoped to that IP).
    
    ## 3. Password auth without sshpass/expect
    
    `sshpass` is usually absent and installing it needs root. OpenSSH's own askpass
    hook works fine and keeps the password out of `argv` and out of logs:
    
    ```bash
    cat > /tmp/askpass.sh <<'EOF'
    #!/bin/bash
    printf '%s\n' "$SSH_PROBE_PW"
    EOF
    chmod +x /tmp/askpass.sh
    
    SSH_PROBE_PW="$PW" SSH_ASKPASS=/tmp/askpass.sh SSH_ASKPASS_REQUIRE=force DISPLAY=:0 \
      ssh -tt <legacy flags> <user>@<host> '<command>'
    ```
    
    `SSH_ASKPASS_REQUIRE=force` is what makes it work without `setsid`. Do not put
    the password inline in the script or on the ssh command line.
    
    ## 4. Issue commands the way the device accepts them
    
    - **One command per connection** on IOS. A single string with IOS arguments
      (`ping 8.8.8.8 repeat 5 timeout 2`) is rejected as
      `Line has invalid autocommand "…"` — exec mode takes the bare form only
      (`ping 8.8.8.8`, default 5 packets). For arguments, drive an interactive
      session instead of exec mode.
    - **Hold the channel open.** With `< /dev/null` the session closes the moment
      auth completes and you get a truncated reply (a lone `!` from a 5-packet
      ping). Pipe a sleep into it instead:
      ```bash
      { sleep 25; } | ssh -tt <flags> <user>@<host> 'ping 8.8.8.8' 2>&1 | tr -d '\r'
      ```
    - **`-tt`** allocates the PTY the CLI expects; without it IOS output arrives
      mangled or not at all. `tr -d '\r'` afterwards makes the output readable.
    - **Paging:** `terminal length 0` as an exec command can drop the session.
      Prefer reading short `show` output, or set paging inside an interactive
      session only.
    
    ## 5. Report the device's own verdict
    
    Quote the summary line verbatim — `Success rate is 100 percent (5/5),
    round-trip min/avg/max = 8/36/91 ms`, `%SYS-5-CONFIG_I`, `state: up`. Do not
    paraphrase into your own numbers, and do not claim success without the device's
    line in the output.
    
    ## 6. Before a config change: plan, approval, checkpoint
    
    Reading is cheap; writing is not. For anything that changes state:
    
    1. Read first. Pull the running config and the relevant `show` output before
       proposing a diff — never configure from memory of what the device "should"
       look like.
    2. Present the plan before touching the device, and stop there. Five blocks,
       in this order: (a) what changes and why, in one short reason; (b) the exact
       command list, verbatim; (c) the exact rollback command list; (d) the read-only
       command that will verify it afterwards plus the output/state you expect;
       (e) risk — which lines can cut you off, and why they do not here. A
       confirmation that has not seen all five is not approval. This applies even
       when the user asked for the change in the same message, and even for one line.
       If any line changes after approval, the old approval is void: re-present,
       re-ask.
    3. Save the pre-change running config to a file, and write the rollback script
       before applying. `references/config-change-safety.md` has the checkpoint →
       apply → save → verify → rollback loop and the per-vendor command profiles.
       Where the `network-change-management` skill is installed, run its rollback
       generator: `rollback_gen <script> --existing-config <saved-config>` drafts
       the rollback — read it line by line and fill its
       `# TODO restore from snapshot:` entries before use.
    4. Never send a command that can cut your own path back in: VTY/management
       interface, `access-class`, default route, the interface you are connected
       through.
    5. Verify with the command promised in (d) once the change lands, and quote the
       device's own output line. The device echoing the config line is not
       verification; reading the resulting state back is.
    
    A change is not "done" because the device echoed the line. Re-read the running
    config and check the lines are actually present; a capture that shows no error is
    weaker evidence than a captured config containing the line.
    
    ## 7. Persistent interactive session (OpenSSH)
    
    One-shot OpenSSH suits a single `show`. For a session that enters `enable` with a
    secret and then runs a dozen commands, hold one PTY open instead of spawning an
    ssh per command — the engine's own technique, no paramiko, no Python:
    
    ```bash
    SSH_ASKPASS=/tmp/askpass SSH_ASKPASS_REQUIRE=force setsid \
      ssh -tt -o KexAlgorithms=+diffie-hellman-group14-sha1 \
          -o HostKeyAlgorithms=+ssh-rsa -o PubkeyAcceptedAlgorithms=+ssh-rsa \
          -o Ciphers=+aes128-cbc -o MACs=+hmac-sha1 admin@10.0.0.5 < plan
    ```
    
    - `SSH_ASKPASS` + `setsid` (no controlling tty) supplies the password without
      sshpass; `SSH_ASKPASS_REQUIRE=force` needs OpenSSH 8.4+.
    - `-tt` forces the PTY so the device prints prompts; feed a paced plan of
      `delay<TAB>command` pairs on stdin — pause after connect, then one line at a
      time. Pacing replaces read-until-idle: gear that pauses mid-output otherwise
      gets the next command typed into the middle of the previous one.
    - On a piped PTY the prompt is never a standalone line: the device prints it and
      the next command lands on the same line (`R2>show clock`). Parse prompts as a
      line *prefix* (`^[A-Za-z0-9_.-]{1,63}[>#]`), not a line suffix.
    - Disable paging first (`terminal length 0`); a stray `--More--` cannot be
      answered from a pipe.
    - `enable`: send `enable`, pause, send the secret, then confirm the prompt ends
      in `#`. Report the prompt you actually saw rather than assuming the enable
      succeeded.
    - Keep the password out of `argv` — the askpass helper reads it from the
      environment.
    
    `scripts/ios_session.sh` is this session shape as a runnable CLI (same flags as
    the old paramiko version: `--host --user --password --secret --cmd --show-config
    --paging`).
    
    ## Credential parsing pitfall
    
    Credentials arrive as run-together prose: `user admin pass S3cr3t! secret en4ble!` is
    user `admin`, **password `S3cr3t!`**, and `en4ble!` is the *enable secret* (privileged
    exec password) — not a passphrase with spaces. Try the whole string, then the
    single token, and report which one authenticated rather than silently using it.
    A Cisco banner plus a rejected multi-word password is the signal that you split
    the string wrong.
    
    ## Supporting files
    
    - `scripts/legacy-ssh.sh` — wraps the flags, askpass, and hold-open stdin so a
      one-shot command works first try.
    - `scripts/ios_session.sh` — persistent OpenSSH PTY session (enable, paging off,
      read running config) as a runnable CLI.
    - `references/config-change-safety.md` — checkpoint → apply → save → verify →
      rollback loop, rollback-script rules, and per-vendor command profiles.
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.