The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the trackio CLI to manage projects, retrieve metric snapshots, and list training alerts. This provides a direct interface for an agent to interact with the local experiment database.
[EXTERNAL_DOWNLOADS]: The skill requires installing the trackio Python package. This package is managed by Hugging Face, which is a recognized trusted organization.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where an agent monitors training by reading alerts and metrics, which represents an external data ingestion surface.
Ingestion points: Data is ingested through CLI commands such as trackio list alerts, trackio get metric, and trackio get snapshot as documented in references/retrieving_metrics.md.
Boundary markers: The instructions encourage using the --json flag to ensure structured data parsing, which helps delimit content but does not sanitize the values within the JSON.
Capability inventory: The agent has capabilities to execute shell commands (trackio CLI), sync data to Hugging Face Spaces, and potentially modify training scripts to insert alert calls.
Sanitization: There is no explicit sanitization or filtering logic provided to handle potentially adversarial text contained within the title or text fields of training alerts.

hugging-face-trackio