The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/audit.py executes git config --get remote.origin.url to retrieve the repository's origin URL for metadata. While benign here, it demonstrates the skill's ability to invoke shell commands.
[COMMAND_EXECUTION]: The module _shared/eml_core/viz.py uses subprocess.run to call the Graphviz dot utility, piping generated tree data to its standard input for visualization.
[REMOTE_CODE_EXECUTION]: The functions compile_formula in _shared/eml_core/compile.py and verify_identity in _shared/eml_core/identity.py use sympy.parse_expr and sympy.lambdify to evaluate formulas. These functions can execute arbitrary Python code if provided with malicious strings, as they do not strictly sandbox the global environment or clear builtins.
[REMOTE_CODE_EXECUTION]: The script scripts/check.py dynamically loads scripts/audit.py using importlib.util. While structural, dynamic loading from computed paths is a significant capability.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection when processing formulas from untrusted external data.
Ingestion points: Formula strings and tree definitions passed to compile_formula and verify_identity.
Boundary markers: None present to isolate mathematical data from potential instructions.
Capability inventory: The skill can execute shell commands and evaluate Python code.
Sanitization: Input strings are not sufficiently validated or sandboxed before being passed to powerful evaluation functions.

eml-check