Compound Engineering 3.31.0:越用越顺手的编码 Agent 技能集
编码 Agent 技能集 Compound Engineering 3.31.0 发布,新增 Antigravity 跨模型审查并修复 /lfg 多项问题。
来源摘录
Compound Engineering 3.31.0 is out. The highlights: - `/ce-simplify-code` now prunes useless tests and comments, not just code. Try it with "/ce-simplify-code find useless comments across the repo" or "/ce-simplify-code audit the tests in the repo" - Added support for Antigravity (agy) (incl Gemini 4 Argon) as a Gemini peer for cross-model review, so a different model family can check your work. - Fixed a bug where the cross-model review resulots may not show up in results when it's run unattended (via `/lfg`). This is so you can tell whether the second opinion actually ran - `/lfg` fixed a bug where in some cases it was stalling partway through https://t.co/YmqTzRGvgC
深度解析
是什么:Compound Engineering 是 Every 推出的一套编码 Agent 技能集(v3.31.0),核心理念是“每个工作单元都应让下一个更轻松”,通过 brainstorm、plan、work、simplify、review、compound 六步循环,把传统开发中“每次变更成本递增”的曲线反过来。它以插件形式安装在 Claude Code、Cursor、Codex、Pi、oh-my-pi、Antigravity CLI 等数十个 Agent 宿主上。
核心设计:精力主要前置到规划与审查,而非写码本身;compound 步骤把本次学到的经验写入 docs/solutions/,让后续无关的计划能检索复用,形成知识复利。技能目录覆盖核心循环(/ce-plan、/ce-work、/ce-compound)、Git 工作流、自主流水线 /lfg 等 30 余项。
本次更新:/ce-simplify-code 从只简化代码扩展到清理无用测试与注释;引入 Antigravity(agy,含 Gemini 4 Argon)作为 Gemini 对等端参与跨模型审查,让不同模型家族互查工作;修复 /lfg 无人值守时跨模型审查结果不显示、以及部分情况下中途停滞两个 bug。
关键洞察:一是跨模型审查用异构模型家族提供“第二意见”,降低同源盲区;二是无人值守场景的可观测性——必须能确认第二意见真的运行过;三是经验落盘是整套方法复利的核心来源。
边界与风险:本条为版本发布说明,以变更列表为主,无视觉材料佐证,实际效果需结合其仓库与文档进一步验证。