The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implementation involves running commands for applying changes and verifying steps as part of its core plan-execution functionality. Evidence: The apply_change and run_verification methods in the PlanExecutor class in SKILL.md are designed to execute operations defined within the plan's steps.
[INDIRECT_PROMPT_INJECTION]: The skill processes external plan data which serves as a vulnerability surface if the plan content originates from an untrusted source. 1. Ingestion points: The PlanExecutor class in SKILL.md accepts a plan object as input. 2. Boundary markers: The skill documents the use of human checkpoints (pause_for_human_review) and verification commands as controls. 3. Capability inventory: The skill possesses file writing (apply_change, restore_file) and shell execution (run_verification) capabilities. 4. Sanitization: No programmatic filtering of command strings is explicitly shown, with the process relying on human-in-the-loop oversight.

executing-plans