The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/office/soffice.py dynamically writes C source code to a temporary file and compiles it using gcc into a shared library at runtime. This library is then injected into the soffice process using the LD_PRELOAD environment variable to intercept and simulate Unix socket system calls (socket, listen, accept). This approach bypasses system-level socket restrictions but introduces high-risk dynamic execution behavior and binary interception.
[PROMPT_INJECTION]: The skill extracts content from untrusted .pptx files for processing by the agent, creating a surface for indirect prompt injection. * Ingestion points: PowerPoint slide text is extracted using the markitdown tool as described in SKILL.md. * Boundary markers: No delimiters or behavioral instructions are used to distinguish untrusted slide content from agent instructions in the extraction flow. * Capability inventory: Arbitrary command execution via gcc, soffice, and node processes, along with broad file system access. * Sanitization: No input filtering or content sanitization is applied to extracted text before it enters the agent's context.
[EXTERNAL_DOWNLOADS]: The skill depends on well-known packages from established organizations, including Microsoft's markitdown and the widely used pptxgenjs, Pillow, and react-icons libraries. These are established tools and are treated as safe sources.

pptx