Skip to main content

​
What Are Skills?

Skills are reusable diagnostic playbooks that extend Siclaw’s capabilities. Each skill is a directory containing a description and optional executable scripts. 
skills/core/k8s-pod-diagnostics/
├── SKILL.md        ← Description, parameters, usage
└── scripts/
    └── check-pod.sh  ← Executable diagnostic script

​
Skill Tiers

skills/
├── core/             ← Built-in, read-only, ships with Siclaw
├── team/             ← Team-managed via Web UI, admin-published
├── user/<userId>/    ← Personal skills, per-user
└── extension/        ← Optional overlay builds
Loading priority (highest wins): personal > team > core

​
Creating a Skill

​
1. Write the SKILL.md

# GPU Health Check

Check NVIDIA GPU health status across cluster nodes.

## Parameters

- `namespace` (optional): Target namespace. Default: all namespaces.
- `node` (optional): Specific node to check.

## Usage

Check all GPUs: `run_skill gpu-health-check`
Check specific node: `run_skill gpu-health-check --node gpu-worker-01`

​
2. Add Scripts (Optional)

#!/bin/bash
# scripts/check-gpu.sh
# Arguments passed as positional parameters

kubectl get nodes -l nvidia.com/gpu.present=true -o wide
kubectl exec -n gpu-operator -- nvidia-smi --query-gpu=gpu_name,temperature.gpu,utilization.gpu,memory.used --format=csv

​
Security Review Gate

Scripts follow a mandatory review workflow before execution: 
draft → request review → pending → AI + static analysis → approved/rejected

​
Static Analysis

Danger patterns are checked automatically:
SeverityExamples
Criticalrm -rf, mkfs, dd if=, fork bomb
Highchmod 777, `curlsh, eval, > /dev/sda`
Mediumkill -9, pkill, reboot, shutdown

​
AI Review

An LLM reviews the script semantics with a mandatory rule: “Skills MUST be strictly read-only.”

​
Human Approval

A user with skill_reviewer role must approve before the skill can be executed.

​
Script Execution

When a skill script is approved and executed via run_skill:
  • Interpreter: bash for .sh, python3 for .py (detected automatically)
  • Timeout: Default 180s, max 300s
  • Arguments: Passed as array to spawn() — no shell interpolation (injection-safe)
  • Max output: 10 MB combined stdout + stderr
  • Injected environment: KUBECONFIG, SICLAW_DEBUG_IMAGE, SICLAW_CREDENTIALS_DIR
Skill scripts are exempt from the binary allowlist. This is the only way to run otherwise-blocked binaries (like sed or awk) — the security review gate is the safety mechanism.