The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a 'fan out' architecture where multiple agent candidates are executed in parallel and their outputs are subsequently read 'end to end' to be selected and 'grafted' into a final result. This pattern creates a surface where a malicious or adversarial output from a candidate could attempt to influence the lead agent's logic during the synthesis phases.
Ingestion points: Candidate rationales and artifacts are read from the filesystem (specifically /tmp/arena-<slug>/candidate-<n>/) during Phases B, D, and E of the workflow.
Boundary markers: The instructions do not specify the use of XML tags, delimiters, or 'ignore embedded instructions' warnings when the lead agent processes candidate outputs.
Capability inventory: The skill utilizes the herdr platform tool to spawn and manage sub-agents and has the capability to write synthesized artifacts to the local filesystem.
Sanitization: There are no instructions to sanitize, filter, or validate the content produced by the parallel candidates before it is integrated into the final synthesized output.

arena