The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from an engineering knowledge graph (ForgeRAG).
Ingestion points: Data is ingested through tools like ask_engineering_question and find_relevant_chunks which query a local service endpoint (localhost:8200).
Boundary markers: The skill assumes the underlying LLM platform handles prompt synthesis and isolation.
Capability inventory: The handler can write image files to the local workspace (writeFileBuffer) and persist records to a database (prisma.generatedImage.create).
Sanitization: The get_page_image tool implements regex validation for document hashes and number checks for page indices to prevent SSRF-like behavior against the internal service.
[EXTERNAL_DOWNLOADS]: The skill connects to a local service (localhost:8200 or configured via environment variables) to fetch engineering documentation. This is a standard architectural pattern for RAG skills and does not involve downloading untrusted third-party code or execution.
[DATA_EXFILTRATION]: No evidence of data exfiltration was found. The skill uses standard environment variables (FORGERAG_ENDPOINT, FORGERAG_API_TOKEN) for internal service authentication and configuration. Images and text are retrieved from a local/trusted service and kept within the agent's workspace and chat context.

knowledge-graph