The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/office/soffice.py performs runtime compilation and library injection. It writes a C source string to a temporary file, compiles it using gcc into a shared library (lo_socket_shim.so), and injects it into the soffice process using the LD_PRELOAD environment variable. This technique is used to intercept and redirect system socket calls at a low level, which is a high-risk dynamic execution behavior.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute multiple external command-line binaries, including pandoc, soffice, pdftoppm, gcc, and git. While these are utilized for document conversion and processing, the capability to execute system commands is a sensitive vector that requires containment.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the docx library from the npm registry, as specified in the configuration steps within SKILL.md. This is an external dependency from a well-known package registry.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8).
Ingestion points: The skill processes external content from .docx files through pandoc text extraction and XML unpacking (SKILL.md, scripts/office/unpack.py).
Boundary markers: Absent. There are no delimiters or specific instructions to the agent to ignore or isolate embedded commands found within the document content.
Capability inventory: The skill possesses extensive system capabilities, including arbitrary subprocess execution, file system writes, and runtime code compilation (scripts/office/soffice.py, scripts/accept_changes.py).
Sanitization: While the skill uses defusedxml to mitigate XML-based attacks during parsing, it does not sanitize the natural language content extracted for agent processing.

docx