The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied dataset URLs and configuration variables. Ingestion points: Dataset URL (storage_url, dataset), cookbook name, and custom --set-string values. Boundary markers: Instruction Gate 6 in SKILL.md explicitly tells the agent to validate these inputs. Capability inventory: osmo workflow submit, kubectl commands, and bash scripts for file/network operations. Sanitization: Instruction Gate 6 mandates agent-side sanitization, and worker scripts use safe patterns like NUL-delimited arrays to prevent injection.\n- [COMMAND_EXECUTION]: Orchestrates infrastructure setup and workflow submission using osmo and kubectl commands. This behavior is consistent with the skill's primary purpose as a workflow orchestrator.\n- [DYNAMIC_EXECUTION]: Employs python -c to perform lightweight data processing and YAML parsing. The scripts use safe data-passing patterns, such as environment variables and command-line arguments, rather than direct string interpolation of untrusted data into Python code.\n- [EXTERNAL_DOWNLOADS]: Performs network probes to nvcr.io and user-specified model endpoints using curl. These operations are limited to health checks and registry access validation, which are standard for the intended use case.

physical-ai-image-attribute-augmentation