The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses docker run to execute image processing tasks. Invocations include high-privilege flags such as --gpus all, --ipc=host, and -v $WORKSPACE:$WORKSPACE for host filesystem access. This is standard and required for the skill's GPU-accelerated data science purpose.
[EXTERNAL_DOWNLOADS]: The workflow pulls the tao_toolkit.data_services image (resolved from a local versions.yaml) and the official alpine image from Docker Hub. These downloads are from trusted or well-known sources and are necessary for the workflow.
[DATA_EXPOSURE]: The hooks/mining-package.sh script automates the collection of Claude Code session logs (.jsonl) and local settings (settings.json) to include them in the final mining report package. This data remains on the local filesystem and is intended for auditability of the mining iteration.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes untrusted data from .parquet files and tool outputs via automation hooks.
Ingestion points: target_embeddings.parquet, source_embeddings.parquet, mined.parquet, Mining_Report.md, and tool-use stdout/stderr.
Boundary markers: Absent; the skill does not use explicit delimiters to isolate data from instructions in these inputs.
Capability inventory: Extensive capabilities including docker run (root access), host file writes, and arbitrary bash execution.
Sanitization: The skill performs basic validation via regex and schema checks in the hooks but lacks robust sanitization for embedded instructions in processed data.

tao-mine-aoi-images