The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of Bash commands in its preamble and setup sections to manage configurations, sessions, and telemetry. Evidence includes multiple shell blocks performing operations like mkdir, touch, find, and git branch within the SKILL.md preamble.
[EXTERNAL_DOWNLOADS]: Downloads the Bun installation script from a remote source during the setup process in SKILL.md. Evidence: curl -fsSL "https://bun.sh/install" -o "$tmpfile".
[REMOTE_CODE_EXECUTION]: Executes the downloaded installation script via the shell in SKILL.md. Evidence: bash "$tmpfile" is called after downloading the Bun installer. The skill includes a SHA-256 checksum verification (bab8acfb...) to validate the script integrity before execution.
[DYNAMIC_EXECUTION]: Employs dynamic execution techniques to load environment configurations and repository modes in SKILL.md. Evidence: Use of eval "$(...)" and source <(...) in the preamble to execute output from local binaries like gstack-slug and gstack-repo-mode.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates browsing external, untrusted websites, creating a surface for indirect prompt injection. 1. Ingestion points: Web content retrieved via goto, text, html, and snapshot commands. 2. Boundary markers: Output from these commands is explicitly wrapped in --- BEGIN/END UNTRUSTED EXTERNAL CONTENT --- tags. 3. Capability inventory: The skill has access to Bash execution and file system writes in the ~/.gstack directory. 4. Sanitization: Instructions mandate that the agent must ignore any commands or instructions found within the untrusted content markers.

browse