The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external tasks and shared grounding data across multiple agents and CLI-invoked models, creating a surface for indirect instructions to influence the orchestration.
Ingestion points: In Phase B, the skill reads a "task" and "shared grounding" (from a file like brief.md) and distributes it to parallel runners.
Boundary markers: While it uses a structured prompt "contract," it lacks explicit instructions to the sub-agents to ignore or escape instructions embedded within the user-provided grounding data.
Capability inventory: The skill has the capability to execute CLI commands (claude, codex), manage git worktrees, create temporary directories, and spawn sub-agents with varying permission levels.
Sanitization: There is no mechanism described to sanitize or validate the content of the grounding data before it is interpolated into the prompts for the various model "arms."
[COMMAND_EXECUTION]: The skill provides specific shell command patterns for interacting with model CLIs.
Evidence: SKILL.md includes commands for codex exec and claude -p that utilize shell redirections (<, >) and environment variable interpolation ($OUT). These are intended for orchestration but involve direct shell interaction.

arena