菜单
Toggle item-level multi-model fusion (MoE) for ai-harness fusion-proxy sessions (claude-moe / codex-moe). Use when the user says "moe", "MoE", "fusion mode", or asks to turn multi-model fusion on/off or check its status.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/yukimaru77/ai-harness --skill moe命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | moe |
| description | Toggle item-level multi-model fusion (MoE) for ai-harness fusion-proxy sessions (claude-moe / codex-moe). Use when the user says "moe", "MoE", "fusion mode", or asks to turn multi-model fusion on/off or check its status. |
Switch the ai-harness fusion proxy between "moe" (every step fans out to all candidate models in parallel and a synthesizer merges them) and "passthrough" (normal single model).
Run with the shell tool (argument: on / off / toggle / status; default toggle):
ai-harness-fusion <arg> codex-moe # in Codex
ai-harness-fusion <arg> claude-moe # in Claude Code
All Claude and Codex sessions run through the fusion proxies by default
(claude: ANTHROPIC_BASE_URL in ~/.claude/settings.json; codex: the fusion
model provider in ~/.codex/config.toml), so the new mode applies from the
next step. Passthrough mode is byte-transparent to the real provider.
Only sessions started BEFORE that wiring existed need a restart — for Claude
you can verify with echo "${ANTHROPIC_BASE_URL:-direct}" (8400 = proxied).
Report in 1-2 lines: the new mode and that it applies from the next step.
Diagnosis if anything misbehaves: ai-harness-fusion diag all.