The Agent Skills Directory

Filesystem Tool Capability: The skill documentation describes how to enable and use filesystem tools including ls, read_file, write_file, edit_file, glob, and grep. These tools allow the agent to interact directly with the storage backends, which can include the local disk via FilesystemBackend. The skill appropriately highlights the use of virtual_mode to restrict access to specific directories and recommends human-in-the-loop checkpoints (interrupt_on) for write operations.
Indirect Prompt Injection Surface: By facilitating the reading of external data through filesystem tools, the skill creates a surface where an agent might process and potentially follow instructions embedded within the files it reads.
Ingestion points: File reading tools such as read_file, glob, and grep described in SKILL.md introduce data into the agent's context.
Boundary markers: While the skill recommends virtual_mode for path safety, it does not explicitly define prompt-level delimiters or instructions to prevent the agent from executing commands found within processed file content.
Capability inventory: The skill provides a full suite of file manipulation capabilities: ls, read_file, write_file, edit_file, glob, and grep across Python and TypeScript implementations.
Sanitization: The skill documents path-level sanitization using the virtual_mode=True parameter to prevent path traversal attempts.

deep-agents-memory