The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses several local Python scripts to automate policy enforcement and configuration validation. Scripts such as validate_team_plan.py, validate_route_plan.py, and validate_team_ledger.py are executed by the lead agent to verify JSON configurations before dispatching subtasks to worker agents.
[DATA_EXFILTRATION]: The script scripts/model_preflight.py contains a semantic_probe function that utilizes urllib.request to perform network operations. This capability is strictly restricted to loopback addresses (localhost and 127.0.0.1) and is designed for internal verification of model reasoning capabilities using cryptographic nonces. It does not attempt to contact external or unauthorized domains.
[INDIRECT_PROMPT_INJECTION]: The skill architecture manages the ingestion of worker agent outputs into the lead agent's context for final integration and verification.
Ingestion points: Worker outputs are recorded in the output field of the task ledger (specified in references/audit-schema.json) and integrated by the lead agent.
Boundary markers: The skill uses strict role definitions in references/task-packet.md and explicitly prohibits worker agents from further delegating tasks or performing irreversible actions.
Capability inventory: The system relies on local validation scripts for configuration management and maintains execution logs in a dedicated agent_team/ directory.
Sanitization: All worker outputs are subjected to a final verification step by the lead agent against pre-defined completion standards.

codex-model-routing-team