The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes external CLI agents (such as claude, codex, gemini, agy, cursor-agent, opencode, pi, and aider) to query different LLMs. The instructions in references/cli-runners.md provide shell templates and specifically warn the agent to avoid shell fragment interpolation and to use read-only modes.
[INDIRECT_PROMPT_INJECTION]: The skill aggregates output from various LLMs into a matrix for comparison. Ingestion points: Model outputs are collected and presented in a model-by-model answer matrix (defined in SKILL.md). Boundary markers: The skill recommends anonymizing answers and labeling synthesis as LLM interpretation when performing fusion. Capability inventory: The skill executes subprocesses to query runners and uses bin/detect-runners to discover models. Sanitization: Instructions mandate read-only or no-tools invocation and emphasize using stdin or quoted arguments to prevent injection.

tribunal