The Agent Skills Directory

[PERSISTENCE]: The skill modifies shell configuration files to ensure that tool paths and environment managers are initialized in future sessions.
Evidence: scripts/lib/path_check.sh contains functions add_to_shell_rc and add_shell_hook which append export statements and initialization logic (e.g., `eval "$(rbenv init
bash)") to files like ~~/.bashrc,~~ /.zshrc, and ~/.profile`.
[PRIVILEGE_ESCALATION]: Several scripts utilize sudo to perform administrative tasks, such as system-wide package updates and writing to restricted directories.
Evidence: scripts/auto_update.sh executes sudo apt-get upgrade -y and sudo snap refresh. scripts/install_composer.sh uses sudo cp to install the composer binary to /usr/local/bin.
[COMMAND_EXECUTION]: The skill relies extensively on shell commands to audit the environment, detect project types, and manage tools.
Evidence: scripts/check_environment.sh runs various diagnostic commands like which, type -a, and --version across multiple binaries. scripts/detect_project_type.sh executes grep on local files to identify frameworks.
[EXTERNAL_DOWNLOADS]: The skill fetches binaries and scripts from various external sources to facilitate tool installation.
Evidence: scripts/install_composer.sh downloads from https://getcomposer.org. scripts/installers/hashicorp_zip.sh downloads from https://releases.hashicorp.com. Multiple installers fetch releases from GitHub repositories.
[REMOTE_CODE_EXECUTION]: Binaries and scripts downloaded from external sources are executed on the host system.
Evidence: scripts/installers/aws_installer.sh downloads the AWS CLI zip, unzips it, and runs the ./aws/install script. scripts/installers/github_release_binary.sh downloads binaries, applies chmod +x, and executes them to verify installation.
[DYNAMIC_EXECUTION]: The skill executes shell commands constructed from tool metadata (catalog entries).
Evidence: scripts/installers/package_manager.sh uses bash -c within the safe_version_check function to run version commands retrieved from JSON catalog files. While it includes an allowlist-based validator, this still facilitates dynamic command assembly and execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from local project files that could potentially influence agent behavior or script execution.
Ingestion points: scripts/detect_project_type.sh (reading project manifest files), scripts/check_environment.sh (capturing tool output).
Boundary markers: Absent.
Capability inventory: Full file-write via shell profiles, network operations (curl/wget), and privileged execution (sudo).
Sanitization: A command allowlist exists for metadata execution, but project file content is processed via standard grep/sed without specific injection sanitization.

cli-tools