The Agent Skills Directory

[DATA_EXFILTRATION]: The skill facilitates consensus by sending project artifacts, live files, and proposals to external models (including Gemini and Claude) using platform-specific tmux tools (codex-tmux, agy-tmux). This represents a data flow where local repository content is shared with well-known model providers for the purpose of cross-model review.
[PROMPT_INJECTION]: The skill implements an adversarial review loop where a fresh model reads untrusted proposals or diffs, creating a surface for indirect prompt injection.
Ingestion points: The reviewer model is directed to read repo paths, artifact paths, and specific proposed diffs/claims provided in the consensus round prompt (SKILL.md).
Boundary markers: The round-prompt skeleton provides structural instructions (e.g., writing to a specific result.json path and restricting further delegation) which offer some process isolation, though it lacks explicit instructions for the reviewer to disregard commands contained within the data being audited.
Capability inventory: The skill utilizes codex-tmux, agy-tmux, and tmux-agent-dialogue to dispatch tasks to other models and writes consensus reports to a scratchpad file.
Sanitization: No specific content filtering or sanitization of the input diffs or artifacts is mentioned before they are presented to the reviewer model.

second-model-consensus