The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture is designed to process untrusted external data for Retrieval-Augmented Generation.
Ingestion points: Untrusted data enters the agent context through the ingest_file_task (defined in celery.md) and the LlamaIndexParser base class (in llamaindex.md).
Boundary markers: The provided implementation code lacks explicit prompt delimiters or 'ignore instructions' directives for isolating ingested text from agent instructions.
Capability inventory: The skill's capabilities are focused on parsing, embedding, and storing data. Analysis of the scripts shows no presence of subprocess calls, arbitrary code execution (exec/eval), or unauthorized network exfiltration logic.
Sanitization: The skill includes a _clean_metadata function in chromadb.md that ensures all metadata values are primitive types (str, int, float, bool) and serializes nested objects to JSON, providing protection against schema confusion and metadata-based injection attacks.

rag-skills