The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses sudo and system management commands such as apt, yum, and systemctl. These are standard and necessary operations for installing, configuring, and managing a system-level monitoring agent.
Evidence: sudo apt install -y observe-agent, sudo systemctl restart observe-agent, and sudo rm -rf /etc/observe-agent/ (the latter is used correctly in the uninstall flow).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted output from the host (such as logs or agent status). It includes a robust defense mechanism to prevent these inputs from being interpreted as instructions.
Ingestion points: Host command outputs (e.g., observe-agent status) and user-pasted log data.
Boundary markers: Implements a wrap helper that uses random nonces to encapsulate untrusted data within <untrusted-data> tags.
Capability inventory: Limited to the observe CLI and standard Linux system management via user-run commands.
Sanitization: The wrap helper uses sed and tr to strip ANSI escape sequences and non-printable control characters, preventing terminal-based UI redressing attacks.
[CREDENTIALS_SAFE]: The skill demonstrates safe credential management by ensuring the OBSERVE_TOKEN remains in the user's local shell environment and is never passed into the LLM conversation context.
Evidence: Explicit instructions in SKILL.md to use "$OBSERVE_TOKEN" for interpolation by the local shell rather than substituting literal values.

setup-linux-host-collection