Cisco Defenseclaw Deskside Setup

Cisco DefenseClaw Deskside Setup

Prerequisites

Workflow Overview

┌───────────┐   ┌───────────────┐   ┌───────────────┐   ┌─────────────────┐
│ Preflight │ → │ Render/review │ → │ Apply/handoff │ → │ Validate evidence │
└───────────┘   └───────────────┘   └───────────────┘   └─────────────────┘

When to Activate

• Deploying, upgrading, configuring, or validating DefenseClaw on an AMD Ryzen AI or other Lemonade-backed deskside.
• Preview and review the cisco defenseclaw deskside setup workflow before any live apply phase.
• Diagnose failed prerequisites, generated assets, configuration, or validation evidence.

Scope

Follow the documented read-only or render-first path whenever it is available. This skill does not imply permission to mutate live systems. Require explicit apply flags, protected credentials, and operator review for state changes.

Examples

Inspect the supported setup modes before selecting one:

bash skills/cisco-defenseclaw-deskside-setup/scripts/setup.sh --help

Expected output: usage, supported modes, and required arguments are displayed without changing the target environment.

Inspect validation modes before running completion checks:

bash skills/cisco-defenseclaw-deskside-setup/scripts/validate.sh --help

Expected output: offline, live, and completion options are displayed when the skill supports them; help exits without mutation.

Troubleshooting

Use this skill for the native, signed Cisco release. The public Cisco release does not publish an OCI runtime image; do not substitute a mutable third-party container tag for the release artifacts.

Safety Contract

• Require a pre-pinned SSH host key. The scripts use StrictHostKeyChecking=yes and never use TOFU.
• Keep the Lemonade and DefenseClaw inference endpoints on loopback. Do not add an API key to a local provider that does not require one.
• Use Lemonade's stable server API on port 13305. Backend ports such as 8001 or 8002 are assigned dynamically and must not be persisted in DefenseClaw.
• Require --replace-active-model before replacing an already loaded Lemonade model. Model loading can interrupt active requests.
• Begin DefenseClaw in observe mode. Require an explicit --guardrail-mode action apply after a clean Codex canary before blocking.
• Do not place credentials in argv, chat, or generated files.

Workflow

Run a non-mutating preflight first:

bash skills/cisco-defenseclaw-deskside-setup/scripts/setup.sh \
  --host "<deskside-host>" --user cisco

Apply the current Cisco release and switch Lemonade to the desired model:

bash skills/cisco-defenseclaw-deskside-setup/scripts/setup.sh \
  --host "<deskside-host>" --user cisco \
  --release latest \
  --model Qwen3.6-27B-GGUF \
  --replace-active-model --apply

The apply path requires an explicit release selection. Prefer a concrete version for repeatable deployments; --release latest resolves it once and then uses that immutable release tag. The vendor installer verifies release checksums. The script waits for Lemonade's asynchronous model download, configures the credential-free lm_studio compatibility adapter against http://127.0.0.1:13305/api/v1, initializes Cisco's scanners and default rule pack, and wires Codex hooks, native OTel, and notify telemetry. It pins Codex to the Lemonade registration after connector setup so a managed config rewrite cannot restore a different cached model. Qwen3.6-27B-GGUF is the registration used by the stable API; do not persist a dynamically assigned backend port or GGUF filename in Codex configuration.

Before changing the DefenseClaw LLM config, the script creates a timestamped copy beside ~/.defenseclaw/config.yaml. If a later configuration or gateway step fails, it restores that config and reloads the previous model when it had replaced one. Downloaded model files remain cached by Lemonade.

Validate the model endpoint, configured judge, and gateway afterward:

bash skills/cisco-defenseclaw-deskside-setup/scripts/validate.sh \
  --host "<deskside-host>" --user cisco --live --check-inference \
  --expect-mode observe

After a clean observe-mode Codex canary, explicitly enable enforcement:

bash skills/cisco-defenseclaw-deskside-setup/scripts/setup.sh \
  --host "<deskside-host>" --user cisco \
  --release 0.8.3 --model Qwen3.6-27B-GGUF \
  --replace-active-model --guardrail-mode action --apply

bash skills/cisco-defenseclaw-deskside-setup/scripts/validate.sh \
  --host "<deskside-host>" --user cisco --live --check-inference \
  --expect-mode action

Codex is a hook connector: model traffic remains direct to Lemonade while DefenseClaw inspects prompts and lifecycle/tool events through Codex hooks. In action mode, supported policy hits return a native deny verdict. A disabled OpenClaw fleet uplink and closed proxy port are expected for this standalone Codex architecture; require the sidecar API, Codex connector, guardrail, local judge, scanners, and rule pack to be healthy instead.