The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (LOW): The skill requires installing trackio via pip. As the skill is authored by Hugging Face (a trusted organization) and points to a legitimate repository (gradio-app/trackio), the download risk is significantly mitigated per the trust-scope rules.
[PROMPT_INJECTION] (LOW): The skill is susceptible to indirect prompt injection (Category 8) because it retrieves and processes training metrics and configurations that could be controlled by an external actor or malicious dataset.
Ingestion points: Metric data and run configurations are retrieved via trackio get metric --json and trackio get run --json (documented in references/retrieving_metrics.md).
Boundary markers: None identified in the provided documentation to delimit metric data from agent instructions.
Capability inventory: The agent can execute CLI commands and potentially use the retrieved JSON data to make subsequent decisions or generate code.
Sanitization: No explicit sanitization or validation of the metric values or config strings is mentioned before processing.
[DATA_EXFILTRATION] (SAFE): The skill documentation describes syncing data to Hugging Face Spaces (trackio sync). This is a core, transparent feature for remote experiment monitoring and does not constitute unauthorized exfiltration.
[COMMAND_EXECUTION] (SAFE): The CLI commands used (trackio list, trackio get) are restricted to the library's specific database querying functionality and do not appear to allow arbitrary shell execution.

hugging-face-trackio