The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill offers to download a pretrained model (nvidia/NV-KERMT-70M-v2) from Hugging Face if no checkpoint is provided. This is managed via a consent gate requiring a user prompt or an explicit opt-in flag.
[COMMAND_EXECUTION]: The skill executes several shell commands and Python scripts within a Docker container environment via a helper script (kermt_container.sh). These commands are used for system probes, data validation, and running the embedding extraction process. The commands are scoped to the project repository ($KERMT_REPO) and use relative paths for internal scripts.
[DATA_EXPOSURE_&_EXFILTRATION]: While the skill involves network operations (Hugging Face download), these are directed to well-known infrastructure for the purpose of fetching necessary model assets. There are no patterns suggesting the exfiltration of sensitive user data.
[PROMPT_INJECTION]: No prompt injection patterns or attempts to override AI safety guidelines were detected in the instructions.
[DYNAMIC_EXECUTION]: The skill generates a RUN_DIR path using the date command and constructs a replay command stored in a JSON manifest, which is standard for workflow reproducibility.

kermt-embed