The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system-level commands to perform its core functions.
Evidence: scripts/scan.py uses subprocess.run to call du, diskutil, sw_vers, and uname for gathering disk usage and system metadata.
Evidence: scripts/server.py uses subprocess.run to execute osascript for trashing files on macOS and open or explorer for revealing files in the OS file manager.
[DYNAMIC_EXECUTION]: The skill implements a local HTTP server that performs file system modifications based on requests from the web-based report.
Evidence: scripts/server.py starts a ThreadingHTTPServer on 127.0.0.1. It processes POST requests to /action to perform file deletions or moves.
Mitigations: The server uses a random session token, validates the Host header to prevent DNS rebinding, and strictly validates all target paths against an allowlist generated during the initial scan. It also enforces that operations are scoped to the user's home directory or the /Applications folder.
[INDIRECT_PROMPT_INJECTION]: The skill processes uncontrolled file and directory names from the local disk, which could be exploited via indirect prompt injection.
Ingestion points: scripts/scan.py reads file and directory names from the disk using os.listdir and os.scandir.
Boundary markers: The agent is provided instructions to interpret the resulting JSON output and categorize items into priority tiers.
Capability inventory: scripts/server.py provides the capability to delete, trash, or open files within $HOME based on the agent's categorization.
Sanitization: The scripts/server.py utility uses realpath resolution and a session-based allowlist to ensure only items identified during the scan can be manipulated.

storage-analyzer-khazix