The Agent Skills Directory

[COMMAND_EXECUTION]: Several Python scripts within the skill (scripts/config_audit.py, scripts/dataset_audit.py, and scripts/inspect_upstream.py) utilize subprocess.run to call system utilities such as git and ffprobe. These operations are limited to inspecting local repository state and media file metadata, which are routine tasks for AI model training and auditing.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of data from potentially untrusted local sources, such as dataset captions and repository files (e.g., README.md, version.py). While this presents a theoretically exploitable surface if these files contain malicious content, the skill's scripts focus strictly on structural validation and data integrity.
Ingestion points: scripts/dataset_audit.py (sidecar captions), scripts/inspect_upstream.py (toolkit metadata files).
Boundary markers: The skill does not explicitly utilize delimiters for these external files, but instructs the agent to maintain a cautious and diagnostic engineering posture.
Capability inventory: File reading, metadata parsing, and restricted execution of CLI utilities (git, ffprobe).
Sanitization: Employs regex pattern matching and standard string cleaning (e.g., strip(), casefold()) to process text data.
[DYNAMIC_EXECUTION]: The documentation in references/upstream-and-versioning.md includes a shell snippet that uses Python's exec() function to parse a local version.py file. This is intended as a manual verification method for the user to confirm the toolkit version and does not constitute hidden or automated execution by the agent.

ai-toolkit-training