The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external RAG traces from traces.jsonl, which contains untrusted text fields such as queries, answers, and context documents.
Ingestion points: The agent reads trace data from traces.jsonl (referenced in SKILL.md and scripts/rag_diagnostic.py) to perform evaluation and passes it to the scripts/rag_diagnostic.py script and openjudge graders.
Boundary markers: The instructions define a data schema but do not specify delimiters or "ignore" instructions to prevent the model from obeying text contained within the trace fields during the evaluation process.
Capability inventory: The skill performs local file reading via a bundled script and refers to the openjudge library, which typically uses language models (potentially via network APIs) to perform grading.
Sanitization: There is no mention of sanitizing or escaping the text content of the traces before they are processed by the agent or the graders.

rag-eval