The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script scripts/trash.py to perform file system management. This utility performs sensitive operations including moving files to a local trash directory, copying directory trees, and unlinking (deleting) files from their original locations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes file metadata which serves as a potential vector for indirect prompt injection.
Ingestion points: The scripts/trash.py script reads file names, paths, and metadata (sizes, SHA-256 hashes) from the local filesystem and the manifest.jsonl log file during list, log, and inspect operations.
Boundary markers: The skill includes rules/anti-aigc.md, which defines reporting rules that prioritize raw identifiers and checksums to minimize the risk of the agent following instructions potentially embedded in file names, although it does not implement strict structural delimiters for all outputs.
Capability inventory: The skill has broad filesystem access, including the ability to move (shutil.move), copy (shutil.copy2), and permanently delete (os.unlink, shutil.rmtree) files and directories.
Sanitization: The script uses standard path handling via pathlib and argparse for CLI input, but it lacks specific sanitization or escaping for file names and metadata before they are processed by the agent's context.

hui-shou