The Agent Skills Directory

[DYNAMIC_EXECUTION]: Runtime compilation of C source code and library injection via LD_PRELOAD.
File: scripts/office/soffice.py
The script defines a C shim designed to redirect system socket calls to work around environment-specific constraints. At runtime, it writes this source to a temporary file, compiles it using gcc, and sets the LD_PRELOAD environment variable to inject the shared library into the LibreOffice (soffice) process. While the C source is hardcoded and serves a functional purpose, this technique represents a high-privilege execution pattern.
[DYNAMIC_EXECUTION]: Automated generation and execution of LibreOffice macros.
File: scripts/accept_changes.py
The skill automates document cleaning by writing a LibreOffice Basic macro to the local profile directory and executing it via a special URI (vnd.sun.star.script:...). This allows the agent to perform complex internal document operations through the LibreOffice engine.
[INDIRECT_PROMPT_INJECTION]: Large attack surface for untrusted external data.
File: SKILL.md, scripts/merge_runs.py
The skill is explicitly designed to process Word documents from external parties. Such documents can contain malicious formatting or text content designed to influence the agent's logic when extracted into markdown or text for processing.
The skill includes proactive security controls, such as the safe_extract helper and manual removal of symlinks (find unpacked -type l -delete), which successfully mitigate common path traversal and ZipSlip attacks during the unpacking process.
[COMMAND_EXECUTION]: Extensive use of sub-processes for document manipulation.
Files: scripts/office/soffice.py, scripts/office/validate.py, scripts/office/validators/redlining.py
The skill orchestrates several external binaries including soffice, gcc, pandoc, pdftoppm, and git to perform its tasks. These commands are executed with structured arguments, though they operate on the contents of the untrusted document directory.

docx