The Agent Skills Directory

[COMMAND_EXECUTION]: Spawns system utilities for document processing and visual conversion.
The skill executes soffice, pdftoppm, and git via subprocess.run calls in scripts/thumbnail.py, scripts/office/soffice.py, and scripts/office/validators/redlining.py to handle presentation files and version comparisons.
[REMOTE_CODE_EXECUTION]: Performs runtime compilation and process injection.
Runtime Compilation: The file scripts/office/soffice.py dynamically writes C source code to a temporary directory and invokes gcc at runtime to compile a shared library (lo_socket_shim.so).
Library Injection: The script utilizes the LD_PRELOAD environment variable to inject the dynamically compiled shared object into the soffice process. This is used to shim network/socket calls that may be blocked in sandboxed environments.
[PROMPT_INJECTION]: Surface for indirect prompt injection via document processing.
Ingestion points: Untrusted .pptx files are ingested and parsed for text extraction (markitdown) and visual analysis (thumbnail.py).
Boundary markers: None identified in the XML manipulation or thumbnail generation scripts to distinguish extracted content from agent instructions.
Capability inventory: The skill has capabilities for file system read/write, zip/unzip operations, and execution of subprocesses (soffice, pdftoppm, git).
Sanitization: Content extracted from slide XML is processed without explicit sanitization or filtering of potential adversarial instructions.

pptx