The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/run_state.py utilizes subprocess.run and subprocess.Popen to interact with the local Git repository for retrieving metadata, tree structures, and file blobs. These interactions are performed without a shell and utilize argument delimiters to prevent command injection via manipulated file paths.
[INDIRECT_PROMPT_INJECTION]: The skill's snapshotting mechanism creates a surface for ingesting untrusted project data.
Ingestion points: The run_state.py script recursively walks the project directory and reads the content of scoped files to calculate SHA256 hashes (scripts/run_state.py).
Boundary markers: The skill treats file contents as opaque binary data for hashing rather than natural language, so no specific prompt delimiters are applied to the file contents themselves.
Capability inventory: The script can execute Git commands and perform atomic file writes to the .daodan state directory within the project root.
Sanitization: Path integrity is strictly enforced by the no_links function, which checks every component of a path for symbolic links or junctions to prevent directory traversal and unauthorized file access.
[SAFE]: The skill exhibits secure design principles, including the use of cross-platform file locking (fcntl or msvcrt), atomic file operations with UUID-based temporary files, and a dependency-free implementation relying solely on the Python standard library (specifically using tomllib for configuration).

project-protocol