The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and compare artifacts generated by multiple isolated workers (candidates). This creates an attack surface where a candidate could include instructions meant to influence the parent agent's synthesis. Evidence: Phase C (Cross-judge) and Phase D (Pick a base) involve reading these untrusted artifacts from output paths like /tmp1-<slug>/. Boundary markers are established via isolation and the 'prompt as contract' principle. Capabilities include local file writing to worktrees and spawning sub-workers. Sanitization is implemented through mandatory human-like reasoning (rationales) and a structured rubric-based evaluation.
[SAFE]: The skill defines a logical, procedural framework for model orchestration without including executable scripts or suspicious network operations. It follows least-privilege principles by using isolated worktrees/directories for candidate outputs.

arena