The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py contains logic to compile a C-language shim at runtime.
It writes an embedded C source string to a temporary file.
It executes gcc to compile the source into a shared object (.so).
It uses the LD_PRELOAD environment variable to load the shim when running LibreOffice, ensuring functionality in restrictive sandbox environments.
[COMMAND_EXECUTION]: Several scripts utilize subprocess.run() to interface with system-level tools.
scripts/accept_changes.py and scripts/office/soffice.py execute soffice (LibreOffice).
scripts/office/validators/redlining.py executes git diff for comparing document text.
scripts/office/soffice.py executes gcc for the shim compilation mentioned above.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document files which could theoretically contain malicious instructions.
Ingestion points: The skill unzips and parses .docx and .dotx files provided by the user or external parties (referenced in SKILL.md and scripts/merge_runs.py).
Boundary markers: The instructions in SKILL.md explicitly recommend running find unpacked -type l -delete to remove potentially malicious symlinks from untrusted archives.
Capability inventory: The skill has the capability to write to the file system and execute shell commands (soffice, git, gcc).
Sanitization: The scripts use the defusedxml library throughout (e.g., in scripts/merge_runs.py and scripts/office/validators/base.py) to prevent XML External Entity (XXE) and other XML-based attacks.

docx