The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided content (Markdown and HWPX files) and performs document assembly, which represents a potential injection surface.
Ingestion points: md_parser.py ingests user-supplied Markdown; analyze_template.py and text_extract.py ingest user-supplied HWPX files.
Boundary markers: Instructions require the use of specific placeholders (e.g., ) and delimiters when inserting content.
Capability inventory: The skill has extensive file-writing capabilities through build_hwpx.py, xml_writer.py, and zip_surgery.py.
Sanitization: Content is passed through an xml_escape function in md_parser.py and xml_writer.py before being interpolated into HWPX XML fragments to prevent XML structural corruption.
[DYNAMIC_EXECUTION]: Several scripts use the importlib module to load peer scripts dynamically.
md_merger.py loads md_parser.py via importlib.util.spec_from_file_location.
section_transplant.py loads zip_surgery.py via importlib.import_module.
office/pack.py loads cell_writer.py via importlib.util.spec_from_file_location.
In all cases, the target paths are computed relative to the script directory using Path(__file__), which keeps the execution scope limited to the skill's own code.
[COMMAND_EXECUTION]: The validate.py script and various test files use subprocess.run() to execute shell commands.
validate.py executes proofread.py using sys.executable and a local path to perform quality checks.
Tests (e.g., test_e2e_pipeline.py) use subprocess.run() to execute the internal build pipeline for verification.
These executions are restricted to calling local Python scripts provided within the skill package.

hwpx-core