The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to ingest and process external data, specifically CI log files and ClickHouse database query results. This creates an attack surface where instructions embedded in logs could potentially influence agent behavior.
Ingestion points: External data enters the context through tools such as download_log_to_file_resource, run_clickhouse_query, and get_job_details_resource as described in SKILL.md.
Boundary markers: The instructions do not define explicit delimiters or instructions to ignore embedded prompts within the ingested data.
Capability inventory: The skill facilitates raw SQL execution, file system writes (log downloads), and network-based log retrieval.
Sanitization: While the documentation suggests using specific extraction tools (e.g., extract_log_patterns_resource) to limit context pollution, there is no mention of sanitizing or escaping the retrieved content before it is processed by the AI.
Raw Query Execution: The skill provides access to the clickhouse-mcp server, which allows for raw SQL execution via run_clickhouse_query.
This provides significant flexibility for debugging and optimization as intended. However, raw SQL access presents a potential consideration for unintended data modifications or exposure if the agent constructs queries based on untrusted inputs.
External Data Retrieval: The skill includes functionality to retrieve logs and artifacts from S3 via download_log_to_file_resource and get_artifacts_resource.
These operations facilitate the skill's core purpose of CI triage. As these resources typically originate from the well-known Amazon S3 service and are related to the official PyTorch organization infrastructure, they are treated as standard operational procedures within the skill's domain.

pytorch-clickhouse