The Agent Skills Directory

[SAFE]: The skill implementation follows security best practices by defining a clear 'Safety boundary' that explicitly prohibits unauthorized data access, credential storage, or modification of system startup files. No malicious patterns, obfuscation, or unauthorized network operations were detected.
[INDIRECT_PROMPT_INJECTION]: The skill has a small attack surface for indirect prompt injection as it processes command output from the host environment to detect tool availability. However, this risk is thoroughly mitigated by the skill's design.
Ingestion points: Output from environmental discovery commands such as uname and command -v in SKILL.md.
Boundary markers: Includes specific instructions to ignore embedded instructions in data and requires the agent to show exact proposed commands to the user before execution.
Capability inventory: Limited to terminal multiplexer management (tmux) and infrastructure assessment. It explicitly forbids using dangerous flags like --yolo or --allow-all.
Sanitization: Instructions mandate that no credentials, session links, or terminal transcripts are included in the final handoff record.
[COMMAND_EXECUTION]: The skill utilizes standard terminal commands for session persistence. Security is maintained by requiring explicit user confirmation before any software installation or session creation, and by prohibiting elevated privileges without separate approval.

keep-agent-running