The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents the use of autowrap and ufuncify in references/code-generation-printing.md. These functions generate C or Fortran source code from symbolic expressions, compile them at runtime (using Cython or f2py), and execute the resulting binary. This pattern represents dynamic code generation and execution.
[DYNAMIC_EXECUTION]: The documentation includes instructions for using pickle.load() in references/code-generation-printing.md to load saved SymPy expressions. Python's pickle module is known for being vulnerable to arbitrary code execution if used with untrusted data files.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a surface for indirect prompt injection by documenting the parse_expr() function in references/code-generation-printing.md. This function is a wrapper for Python's eval() and can be exploited if an agent processes untrusted strings containing mathematical expressions.
Ingestion points: Untrusted mathematical strings are processed via parse_expr(), parse_latex(), and parse_mathematica() in references/code-generation-printing.md.
Boundary markers: No explicit boundary markers or 'ignore embedded instructions' prompts are suggested for the data ingestion phase.
Capability inventory: The skill has access to codegen for file writing, autowrap for shell-level compilation, and eval via the parsing functions, all documented in references/code-generation-printing.md.
Sanitization: The skill provides a parse_trusted_expr example in references/code-generation-printing.md that uses a regex-based whitelist (re.search(r'__|import|=')) to filter some malicious Python patterns, though this is not a comprehensive sandbox.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the sympy library and its dependencies (NumPy, SciPy, Matplotlib) from the official Python Package Index (PyPI).

sympy