moa-x

Run a non-trivial planning task through a layered ensemble of frontier models from four different labs (proposers codex/gpt-5.4 xhigh + opencode/glm-5.2 + sonnet/4.6; refiners codex + kimi) before producing a final implementation plan. The configured proposers run in parallel, broadcast refiners (codex + kimi, each sees all proposals) verify and cross-check, then Opus 4.6 aggregates in place. Adapted from the 2024 Mixture-of-Agents paper (arXiv:2406.04692) for repo-grounded planning, not chat-answer ensembling. Use when: (1) the user invokes /mixture-of-agents, (2) the user pastes a substantial spec doc and asks for a "deeply considered plan" or "second opinion from another lab", (3) the user explicitly says "run MoA on this", (4) high-stakes architecture work where one model's blind spots could be expensive. Do NOT auto-activate for trivial tasks; this skill takes 6-12 minutes wall-clock and spends real quota (subscription or API-billed) across the external CLIs.