The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/write_outputs.py uses importlib.util to dynamically load and execute a Python module (write_explore_bundle.py) from a path computed at runtime based on the script's location on the filesystem. This dynamic loading mechanism executes code from a computed path, which could be exploited if the filesystem structure is manipulated.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input through a JSON specification file, which creates a surface for indirect prompt injection.
Ingestion points: The scripts/plan_variants.py script reads an external JSON file provided via the --spec-json argument.
Boundary markers: The instructions in SKILL.md and references/execution-policy.md mandate experiment isolation and explicit authorization, but the scripts lack programmatic enforcement or delimiters to prevent embedded instructions in the JSON from influencing logic.
Capability inventory: The skill manages the construction of experiment variants and influences the base_command used by downstream execution tools such as minimal-run-and-audit or run-train.
Sanitization: While the script performs basic type validation for numerical scores (e.g., safe_float), it does not sanitize or validate the base_command or other structural fields that are passed into the experiment planning data structure.

explore-run