The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/generate_ts.py modifies the Python runtime environment by appending the current working directory to the module search path (sys.path.append(os.getcwd())). This creates a risk of module shadowing or path hijacking, where a malicious Python file placed in the execution directory could be imported and executed instead of legitimate system or project libraries.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill's setup instructions in SKILL.md direct the user to execute shell and Python scripts (bash install.sh and python download_models.py) that are not included in the provided file package. These constitute unverified remote code execution vectors. Additionally, the skill depends on the external reactot package, which is not a standard library.
[EXTERNAL_DOWNLOADS]: The skill requires downloading pre-trained model weights to the local cache directory (~/.cache/react-ot/) as part of its initialization, but the source URL and integrity of these weights cannot be verified from the provided files.
[INDIRECT_PROMPT_INJECTION]: The skill processes external chemical structure files (e.g., .xyz, .pdb, .cif) which may contain untrusted data that could influence the agent's behavior if parsed improperly by the underlying libraries.
Ingestion points: Reactant and product files loaded via scripts/generate_ts.py and passed to the reactot library.
Boundary markers: Absent; the script passes file paths directly to processing functions without validation or clear separators.
Capability inventory: The skill performs file system read/write operations and loads external Python modules.
Sanitization: None; the implementation assumes that input files are well-formatted chemical data and does not sanitize content before processing.

chem-react-ot