作者分享其 Claude Code 模組的完整基準測試，該模組讓 Jev 為每次 subagent 呼叫挑選 agent 與 model，三輪測試中前兩輪失敗，貼文附上部落格連結供討論。
I wrote up the full benchmark for my Claude Code mod that lets Jev pick the agent and model for every subagent call. Three rounds, and the first two lost. If you're routing subagents to cheaper models yourself, or you think my 0.8 threshold is wrong, reply here. This is the thread the post links to for discussion. https://benenewton.com/blog/jev-agent-routing-claude-code?utm_source=twitter&utm_medium=social&utm_campaign=jev-routing-post-discussion-tweet
貼文為作者 Ben Newton 對於其 Claude Code 模組的說明，該模組讓 Jev 在每次 subagent 呼叫時選擇 agent 與 model。作者公布三輪完整基準測試結果，前兩輪表現不及基準，並邀請正在自行將 subagent 路由到較便宜模型，或對其 0.8 門檻有疑問的人回覆討論。內文附上個人部落格文章連結，未提供程式碼倉庫、安裝步驟或示範媒體。
安装方法请查看原始来源。
I wrote up the full benchmark for my Claude Code mod that lets Jev pick the agent and model for every subagent call. Three rounds, and the first two lost.
If you're routing subagents to cheaper models yourself, or you think my 0.8 threshold is wrong, reply here. This is the thread the post links to for discussion.
https://benenewton.com/blog/jev-agent-routing-claude-code?utm_source=twitter&utm_medium=social&utm_campaign=jev-routing-post-discussion-tweet
将工作阶段作为 skill 使用的浏览器自动化：Playwright + RVF 认知容器 + ruvector 轨迹 + AgentDB 选择器记忆 + AIDefence PII/注入守门。作为 mod（ADR-445 模式）：只收紧的工具守门、本地斜杠命令和控制台显示的状态文件。
为 Claude 提供 register_assumption 工具，并在提示列上方显示本回合的假设、决定和考虑过但未完成的事项。
通过 SSH RemoteForward 隧道，将 REMOTE 主机上的 Claude Code 会话上下文发布到 Mac 上的 localvoxtral。命令钩子运行随附的 POSIX-sh shim，通过 curl POST；远程主机只需 sh 和 curl，缺少任一项时会静默放行。