The Agent Skills Directory

[COMMAND_EXECUTION]: All provided Python helper scripts are non-executing by design. For example, scripts/plan_batch_command.py generates command-line arguments (argv) for MATLAB or Octave but does not invoke the binaries or spawn any subprocesses, returning the plan as data for user review.
[DATA_EXFILTRATION]: The skill is restricted to local file operations. The shared utility in scripts/_common.py includes a URL_LIKE regex that explicitly rejects any path inputs containing protocol signatures (e.g., http://), and the Python scripts do not import network-capable libraries like requests or urllib.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface involving the ingestion of untrusted MATLAB source files, MAT binaries, and JSON manifests. This is mitigated by several layers of defense: first, the logic is limited to static text scanning and metadata inventory; second, the _common.py library implements _reject_symlink_chain and relative_to(root) checks to prevent the agent from accessing files outside the designated project directory; third, the skill's instructions mandate a manual review workflow where all generated plans must be shown to the user before any external execution occurs.
[DYNAMIC_EXECUTION]: While the skill helps prepare code for MATLAB, which supports dynamic execution, the skill's own logic does not use eval() or exec() on any processed data. Furthermore, scripts/scan_m_code.py is specifically designed to flag dynamic execution risks (such as eval, feval, and assignin) within the MATLAB files being reviewed.

matlab