The Agent Skills Directory

[COMMAND_EXECUTION]: The skill dynamically assembles and executes shell commands using the herdr tool (e.g., herdr agent start <nome> --kind <kind> --pane <pane-id>). While the skill provides regex constraints for the agent names ([a-z][a-z0-9_-]{0,31}), the assembly of these commands from user-provided intake data represents a potential command injection surface if the parameters are not strictly sanitized by the execution environment.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a mediator, ingesting and processing text files (turnos/*.md) generated by independent sub-agents. This creates a surface for indirect prompt injection where a sub-agent could attempt to influence the mediator's behavior.
Ingestion points: The mediator reads the output of sub-agents using herdr agent read or by directly accessing files in the turnos/ directory.
Boundary markers: The skill uses strict Markdown templates (e.g., templates/turno.md) and a shared agenda to delimit responses.
Capability inventory: The skill possesses shell execution capabilities through the herdr tool and file system write access for managing artifacts.
Sanitization: A mandatory 'Gate of Quality' (qualidade.md) is used to validate each turn, checking for school adherence, citation accuracy, and formatting before the data is integrated into the synthesis.
[DYNAMIC_EXECUTION]: The skill generates and manages external agent processes at runtime based on the intake configuration. It orchestrates a multi-pane layout and issues commands to sub-agents, relying on dynamic logic to control the debate flow and handle retries.

debate