The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses runtime compilation and library injection to enable Office document processing in restricted environments.
scripts/office/soffice.py writes C source code to a temporary file and compiles it into a shared library using gcc via a subprocess.run call.
The resulting library is loaded using the LD_PRELOAD environment variable to intercept socket-related system calls, a technique used to bypass sandbox restrictions on AF_UNIX sockets.
scripts/accept_changes.py dynamically generates and writes a StarBasic macro (ACCEPT_CHANGES_MACRO) to the LibreOffice profile directory to be executed at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document content and XML metadata, providing a surface for indirect instructions to influence agent behavior.
Ingestion points: scripts/office/unpack.py extracts XML files from untrusted .docx, .pptx, or .xlsx files using zipfile.extractall. pandoc is also used for text extraction from documents.
Boundary markers: Absent. There are no explicit delimiters or instructions to the agent to ignore potentially adversarial instructions embedded within the document content or XML tags.
Capability inventory: The skill possesses high-privilege capabilities including arbitrary command execution via subprocess.run (in soffice.py, accept_changes.py, and redlining.py), runtime compilation (gcc), and extensive file system write access.
Sanitization: Absent. Document content and metadata are passed directly into the agent's context without filtering, escaping, or instruction-trapping.
[COMMAND_EXECUTION]: The skill executes several external CLI tools via the shell.
In addition to the gcc compiler, the skill invokes git diff with the --word-diff flag in scripts/office/validators/redlining.py to compare document versions.
It heavily relies on soffice (LibreOffice) for format conversion and change management.

docx