The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a process where it reads and incorporates output from multiple sub-agents ('candidates') and judges. This creates a surface for indirect prompt injection, as malicious instructions embedded in a candidate's output could influence the parent agent's behavior during the 'Pick' and 'Graft' phases. • Ingestion points: Output files produced by sub-agents located in /tmp directories or git worktrees, and recommendations from a cross-judge model. • Boundary markers: The instructions do not specify the use of delimiters or 'ignore instructions' markers when reading candidate output. • Capability inventory: The skill utilizes t3_thread_launch, delegate_task, and various file system operations. • Sanitization: There are no instructions for validating or sanitizing the content generated by sub-agents before it is processed by the main agent.
[COMMAND_EXECUTION]: The skill dynamically constructs and executes tasks using the t3_thread_launch and delegate_task tools. It reads model configurations and provider settings from local configuration files (e.g., ~/.config/pstack/pstack-models.md) to parameterize these executions.

arena