The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/tor_respond.py contains a hardcoded local filesystem path C:\Users\AGasser\OneDrive\5 ANE CLAUDE work folder assigned as a default to the WORK_FOLDER_ROOT variable. This exposes the author's local directory structure and username.
[DYNAMIC_EXECUTION]: The skill performs dynamic path manipulation to resolve local dependencies. The _bootstrap_ane_package function in scripts/tor_docx.py and logic in scripts/tor_respond.py modify sys.path at runtime to import the ane_package vendor resource from locations determined by environment variables or directory traversal.
[COMMAND_EXECUTION]: Instructions in the build mode of SKILL.md direct the agent to generate new Python scripts (e.g., gen_tor_v01_docx.py) by populating a scaffold and then executing those scripts using the python interpreter to render document deliverables.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of reviewer comments extracted from .docx files, which could potentially contain instructions intended to influence the agent's behavior.
Ingestion points: The load_threads function in scripts/tor_respond.py uses the officeops library to extract comment threads from external Word documents.
Boundary markers: While the skill uses templates and instructions for factual reliability, it lacks explicit security delimiters or "ignore instructions" wrappers for the processed comment text.
Capability inventory: The skill has capabilities to write files to the local system and execute subprocesses via Python script generation.
Sanitization: Minimal sanitization is implemented in _escape_cell to escape markdown characters for table formatting, but no security-specific filtering of the input content is present.

tor-procurement