The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the bash tool to execute local Python scripts for discovering and invoking various model runners (e.g., discover_runners.py, run_codex.py, run_gemini.py, run_kimi.py, run_glm.py). These executions are part of the core functionality to facilitate communication between the orchestrator and the individual model seats.
[DATA_EXFILTRATION]: No exfiltration patterns detected. Analysis confirms the skill does not access sensitive local file paths like SSH keys or AWS credentials. It provides optional tool profiles for repository reading or web research, but explicitly restricts these to read-only operations and prohibits write or execution capabilities.
[PROMPT_INJECTION]: No evidence of malicious prompt injection, behavior overrides, or safety filter bypasses was found. The instructions include specific Hard Rules that enforce unbiased moderation and prevent the orchestrator from influencing or framing the seats' answers.
[EXTERNAL_DOWNLOADS]: The skill relies on local dependencies expected to be present in the .agents/skills/ directory. It does not perform any unverified remote downloads or execute code directly from external URLs.
[SAFE]: The skill follows least-privilege principles by enforcing read-only access and disabling tool execution for all involved sub-agents (organizers, judges, and synthesizers).

models-roundtable