The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on variant definitions and input data provided during the benchmarking loop. If this data is sourced from untrusted external inputs, it creates a surface for indirect prompt injection. The skill includes a high-level instruction to reject variants that fail safety checks, which serves as a manual boundary marker for the agent.
[DYNAMIC_EXECUTION]: The core functionality of the skill involves generating multiple variants of a command or script and executing them to measure performance metrics. This runtime generation and execution is a form of dynamic execution. The risk is managed by the requirement for a 'Correctness Gate' and 'Safety' validation before any variant is promoted or codified.
[COMMAND_EXECUTION]: The skill uses the Bash tool to run measurement commands (e.g., 'npm run job') as part of the optimization process. This is the intended behavior for performance profiling and includes instructions to ensure reproducibility and explainability of the execution results.

benchmark-optimization-loop