The Agent Skills Directory

Script Generation and Execution: The skill is designed to accumulate code into a Python script and then execute it using the local system's Python interpreter. While this is the intended functionality, the ability to generate and run arbitrary code is a significant capability that warrants user awareness.
Evidence: Section 7 in SKILL.md describes writing the accumulated script to a file and executing it via a command like /tmp/meridian_eval_cache/bin/python3 model_build/my_model.py.
Processing External Data: The skill ingests untrusted data from user-provided CSV files to automatically determine column mappings and model configurations. This introduces a surface for indirect prompt injection if the file content is designed to influence the agent's behavior.
Ingestion Point: Reading CSV headers and data in Step 2.
Boundary Markers: No explicit delimiters are used in the generated code to isolate data from logic.
Capability Inventory: The skill can write files and execute shell commands.
Sanitization: The skill implements 'Data Quality Guardrails' described in references/csv_format_reference.md to validate types and values.
Interactive Safety Controls: A key safety mechanism is the inclusion of 'CRITICAL INTERACTIVE CHECKPOINT' rules throughout the workflow. These instructions require the agent to pause and obtain explicit user approval before proceeding with configuration or execution, which helps mitigate the risk of autonomous unintended actions.
Evidence: The 'Interactivity Checkpoint Rule' and multiple checkpoints listed in steps 1 through 7 of SKILL.md.

meridian-model-building