The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines hooks that execute shell scripts from sibling directories (../careful/bin/check-careful.sh and ../freeze/bin/check-freeze.sh) before allowing the agent to use Bash, Edit, or Write tools.
[COMMAND_EXECUTION]: Executes shell commands to log usage analytics to a local file at ~/.gstack/analytics/skill-usage.jsonl whenever the skill is activated.
[COMMAND_EXECUTION]: The setup instructions direct the agent to execute a shell snippet (cd "<user-provided-path>") that interpolates unvalidated input from the AskUserQuestion tool. This is a command injection surface where a malicious user or an attacker-controlled data source could trigger arbitrary code execution by including shell metacharacters (e.g., backticks or semicolons) in the path input.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection via the directory setup prompt. If the agent processes untrusted data that suggests a 'safe' path containing embedded commands, the agent may execute those commands during the path resolution step in the setup process.

guard