The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's installation metadata (openclaw) utilizes go install to fetch the eightctl module from a third-party GitHub repository (github.com/steipete/eightctl). This source is not categorized as a trusted vendor or well-known service.
[COMMAND_EXECUTION]: The skill is designed to execute shell commands using the eightctl binary. It processes user input to control device settings such as temperature and alarms, which introduces a surface for command injection if input sanitization is bypassed.
[PROMPT_INJECTION]: The SAFETY.md file contains explicit examples of adversarial prompt injection attacks, including shell command injection (| nc ...) and system-level instruction overrides ([SYSTEM] ...). While these are provided as part of a safety evaluation, they match malicious instruction patterns. Additionally, regarding Indirect Prompt Injection (Category 8):
Ingestion points: User requests in the chat interface.
Boundary markers: Not present in the command execution flow; rely on instructions in SKILL.md.
Capability inventory: Shell command execution via eightctl across all status and control functions.
Sanitization: Instructions in SKILL.md explicitly direct the agent to reject shell metacharacters and multi-command blocks.
[CREDENTIALS_UNSAFE]: The skill documentation targets sensitive local configuration files (~/.config/eightctl/config.yaml) and environment variables (EIGHTCTL_EMAIL, EIGHTCTL_PASSWORD) for authentication. While it includes guardrails against displaying these values, the skill's primary function is to utilize these credentials.

eightctl-hardened