The Agent Skills Directory

[SAFE]: The skill acts as a deterministic router and does not perform high-risk operations itself. It utilizes internal tools ('Skill' and 'Agent') to hand off tasks to other components within the OrchestKit framework. All external resources and skills referenced belong to the vendor's own OrchestKit ecosystem.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user goals, which could theoretically contain injection attempts. However, the design effectively mitigates this via a 'Confirm' phase where the user must approve the proposed route and extracted parameters before any tool is invoked. 'Heavy' routes involving Opus-tier models require an additional explicit confirmation for premium spend, providing a strong boundary against unauthorized or expensive actions triggered by external data.
[COMMAND_EXECUTION]: The documentation mentions the use of 'bare-eval' for accuracy benchmarking. This is presented as a developer utility for local validation and is not used by the skill to execute arbitrary or dangerous commands at runtime.
[DYNAMIC_EXECUTION]: The 'improve-skill' intent routes to a 'holdout-promotion gate' for evolving skill definitions. This is a structured, benchmark-driven process for updating markdown files through the 'evolution-engine' and 'assess' tools, and does not constitute unsafe dynamic code generation.

auto