The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Python scripts to manage agent roles and capabilities. Evidence: Phase A uses python3 <pstack-runtime>/scripts/roles.py show ... to resolve model seats.
[INDIRECT_PROMPT_INJECTION]: The skill processes outputs from multiple parallel 'candidate' agents (Phase C and D) and instructs the main agent to review and 'graft' components from them. This creates a surface for indirect prompt injection where a candidate could provide adversarial instructions disguised as artifact content to influence the main agent's behavior during the synthesis phase.
Ingestion points: Phase C and D read outputs from parallel tasks launched via delegate_task.
Boundary markers: None. The skill uses rubric criteria for evaluation but lacks specific delimiters or 'ignore embedded instructions' warnings for the candidate-generated content.
Capability inventory: The skill possesses capabilities for shell command execution (python3), file writing (git worktree and /tmp), and further task delegation (delegate_task).
Sanitization: No automated sanitization or filtering is described; the skill relies on manual agent review ('graft each... by hand') as a mitigation.
[EXTERNAL_DOWNLOADS]: The skill references and depends on an external directory (../pstack-runtime/) for its core logic, runtime rules, and scripts, which are not contained within the skill's own directory structure.

arena