The Agent Skills Directory

[DATA_EXPOSURE]: The hooks/rca-package.sh script captures the execution environment for reproducibility by copying configuration files from the .claude/ directory, including settings.json and settings.local.json, into the timestamped output folder. This ensures that the exact state of the analysis environment is preserved alongside the results.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources including inference.csv, train.csv, val.csv, experiment.yaml, and status.json to drive its investigation phases.
Ingestion points: Data is read from experiment result files such as inference/inference.csv and training configuration files like experiment.yaml.
Boundary markers: The workflow does not explicitly define delimiters when passing data extracted from these files to parallel subagents.
Capability inventory: The skill leverages the Bash tool for script execution, the Read tool for visual inspection, and the Agent tool for distributed analysis.
Sanitization: Standard parsing libraries (JSON, CSV) are used for data ingestion, though no specialized filtering is applied to the content used for downstream reasoning.
[DYNAMIC_EXECUTION]: The skill workflow extensively uses dynamically generated Python scripts to perform data analysis tasks. This includes calculating pixel intensity for golden image audits, performing component-type clustering, and mining metadata for correlations between scores and hardware positions.

tao-analyze-changenet-rca