# Compound Engineering v3
**作者**: Trevin Chow
**日期**: 2026-04-22T21:32:47.000Z
**来源**: [https://x.com/trevin/status/2047066108763770998](https://x.com/trevin/status/2047066108763770998)
---
v3 is here! The naming cleanup we've been teasing for 3 weeks finally lands, our brainstorm and plan created artifacts now give requirements a paper trail from idea to commit, every harness outside Claude Code gets first-class install and runtime support, and our review loops stop letting you rubber-stamp your way through findings — all in Compound Engineering 3.0.0.
## One consistent namespace, no more collisions
Every skill and agent now lives under a single ce- hyphen prefix. ce:work becomes ce-work, git-commit becomes ce-commit, setup becomes ce-setup, and so on.
The old mix of prefixed, unprefixed, and colon-prefixed names collided with similarly named skills from other plugins across harnesses — a persistent source of ambiguity any time you ran more than one plugin side by side. The colon in ce:work also required filesystem sanitization on Windows. After v3, everything CE ships sits under its own unambiguous namespace, so our skills don't fight with anyone else's. 2 small renames went along for the ride to help clarity and simplification: ce-review became ce-code-review, and ce-document-review became ce-doc-review.
This is a breaking change. Existing chains, references, and muscle memory need to move to the new names... but it's worth it! 😀
## A paper trail from brainstorm to commit, for you and your agents
Before v3, a requirement you wrote during ce-brainstorm didn't have a stable identity. By the time it reached ce-plan it was prose; by the time it was implemented it was vibes. When a test failed or a reviewer asked "what was this covering?", the answer lived in somebody's head — and the agents running ce-work or ce-code-review had no better way to recover intent than you did.
v3 rewrites the ce-brainstorm and ce-plan product docs to carry much better structure rather than just prose. The brainstorm requirements doc now has 1st class sections for Actors, Key Flows, Acceptance Examples, and Requirements, each with stable IDs. ce-plan pulls those IDs forward into a Requirements Trace section and assigns plan-local implementation unit IDs that survive plan deepening and reorders. Test scenarios reference the acceptance examples they cover. ce-work recognizes those unit IDs in blockers, verification, and task labels.
The obvious payoff is provenance for you and your agents. Trace a failing test back to the acceptance example it was supposed to cover, a plan unit back to the flow that motivated it, and a commit back to the brainstorm entry that started it all. The less obvious and arguably bigger payoff is that the agents can follow those same threads. A reviewer can check whether an implementation actually satisfies the acceptance example it claims to cover. A debugger can see which flow a failing test was exercising. A work agent hitting a blocker on U3 can go read the origin criteria instead of guessing at intent.
The biggest change this also support is improved greenfield product-tier brainstorms. They also get structural support they didn't have before prior to v3. The template used to optimize for incremental changes inside an existing product, so deep-scope work on something new had thin guidance and forced the agent to improvise a shape.
Better structure upstream produces better outcomes downstream...and not just for the human in the loop.
## First-class across every harness
Claude Code was where CE started, and for a long time everything else felt like a second-class port. v3 changes that across the board.
- Codex users can install the native way: codex plugin marketplace add plus the /plugins TUI. Skills and MCP servers register directly with no Bun converter step. Codex's plugin spec doesn't cover custom agents yet, so a small converter fills the gap by writing CE's agents as Codex TOML under ~/.codex/agents/.
- Pi gets an even bigger shift: we stopped shipping our own compatibility layer which collided with the most popular community subagent extension and serialized parallel dispatch. We now delegate to 2 community extensions, @nicopreme's pi-subagents for native parallel sessions and @edlzsh's pi-ask-user, for a blocking-question UX. Pi users get real parallelism and proper question overlays without us competing with community maintainers on infrastructure they already do better.
- Copilot on the CLI and in VSCode both get first class support with the native plugin install too!
Cross-harness cleanup landed alongside the big lifts:
- question-tool no longer silently skips blocking questions on Codex (where request_user_input is only exposed in Plan mode) or on Claude Code (where the tool is deferred and its schema isn't loaded at session start), fixed with a unified instruction applied across 37 SKILL.md files.
- Skill descriptions are now capped at the harness limit with a frontmatter test enforcing it.
## Reviews that make you decide, not rubber-stamp
Our review skills all used to ask bucket-level policy questions which results in one decision covering many findings. In practice that degraded into rubber-stamping or wholesale deferral, and reviewers ended up re-litigating the same premise in different wordings across a dozen findings.
v3 reshapes the whole review family around per-finding engagement:
- ce-code-review Interactive mode walks you through findings one at a time with Apply/Defer/Skip/"LFG the rest" options and a plain-English framing of each proposed fix. Finally, easier way to understand review feedback so you can make better and faster decisions. If you choose to defer the findins, we now route to your real issue tracker instead of a deprecated internal todo store that CE managed.
- ce-doc-review adds three-tier autofix classification and premise-dependency chain grouping — collapse findings that all hinge on one root into a single decision with cascading dependents. Typical engagement dropped from 14+ findings per run to roughly 4-6 real decisions.
- ce-resolve-pr-feedback got tightened clustering (cross-round evidence is required before it activates, so unrelated single-round feedback no longer triggers holistic refactors) with bot-wrapper noise filtered out.
- A new ce-swift-ios-reviewer persona joins the reviewer cohort for Swift/iOS stacks, catching SwiftUI state-wrapper misuse, Combine retain cycles, and Core Data context threading. Thanks to @jcjvm for contribution.
## Debugging that diagnoses, not guesses
ce-debug already had a 4-phase methodology, but in practice it defaulted to print-debugging, skipped environment sanity before deep tracing, and sometimes declared heisenbugs "fixed" when instrumentation perturbed them out of sight. v3 tightens all of that:
- An early environment sanity check before deep tracing: branch, deps, runtime, env vars, stale artifacts.
- An assumption audit at hypothesis time to help agent steering. Many wrong hypotheses are correct hypotheses tested against a wrong assumption.
- A parallel read-only subagent dispatch option for broad searches.
- A new technique reference covering boundary instrumentation, test-order pollution, repro minimization, stepping debuggers versus instrumentation, heisenbugs, and a bug-class checklist (timezone, encoding, float, overflow, cache, auth).
## Less friction, fewer sharp edges
Small quality-of-life upgrades to existing skills:
- ce-proof now triggers on direct user requests to share or iterate on a markdown file. "Share this to Proof so we can iterate" works without needing an upstream caller like ce-brainstorm or ce-plan. You can try it with any markdown down in your repo for example with something like "/ce-proof iterate with me on <doc>"
- ce-demo-reel adds local save as an alternative to catbox upload, so you can keep demo assets on disk without manually fishing them out of temp directories. Thanks to Lucas Henn for this contribution.
- ce-setup checks for ast-grep and the corresponding agent skill, with a new Skills section in the health-check output matching the existing Tools section.
And a round of correctness and safety fixes:
- ce-plan no longer misclassifies non-software research as only factual lookup, silently substitutes generic web search for user-named resources, or skips the blocking post-plan handoff menu. You can get real thoughtful agent assisted planning for your non-software problems whether it's your kids math lessons, that roadtrip to Portland or the home improvement project you've been trying to pin down.
- ce-work stopped inventing human-time estimates and multi-day session breakdowns — framings that don't apply to agent execution.
- ce-update derives its cache directory from the plugin root parent with a shape check, so malformed env values can't collapse into a dangerous rm -rf target. Thanks to @andrewlook for raising this issue.
- ce-compound YAML frontmatter now survives strict parsers and Cowork's plugin validator.
What's next: more work on product and strategy skills, sharper verification and shipping support, and configurable preferences so things like the compounded-docs output path don't have to live hardcoded as docs/solutions. That last one was close to ready for v3 but needed more time to bake.
v3 is as much a foundation as a release....more landing soon!
## 相关链接
- [Trevin Chow](https://x.com/trevin)
- [@trevin](https://x.com/trevin)
- [82K](https://x.com/trevin/status/2047066108763770998/analytics)
- [Compound Engineering 3.0.0](https://github.com/everyinc/compound-engineering-plugin)
- [@nicopreme](https://x.com/@nicopreme)
- [pi-subagents](https://github.com/nicobailon/pi-subagents)
- [@edlzsh](https://x.com/@edlzsh)
- [pi-ask-user](https://github.com/edlsh/pi-ask-user)
- [@jcjvm](https://x.com/jcjvm)
- [Lucas Henn](https://github.com/Isaccseven)
- [ast-grep](https://github.com/ast-grep/ast-grep)
- [@andrewlook](https://x.com/andrewlook)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [5:32 AM · Apr 23, 2026](https://x.com/trevin/status/2047066108763770998)
- [82.9K Views](https://x.com/trevin/status/2047066108763770998/analytics)
- [View quotes](https://x.com/trevin/status/2047066108763770998/quotes)
---
*导出时间: 2026/4/23 15:05:24*
---
## 中文翻译
# 复合工程 v3
**作者**：Trevin Chow
**日期**：2026-04-22T21:32:47.000Z
**来源**：[https://x.com/trevin/status/2047066108763770998](https://x.com/trevin/status/2047066108763770998)
---
v3 来了！我们预告了 3 周的命名清理终于落地，头脑风暴和计划创建的产物现在为需求提供了从想法到提交的完整记录链，Claude Code 之外的每个 harness 都获得了一流的安装与运行时支持，而且我们的审查循环不再让你对发现的问题敷衍盖章——这些全部包含在 Compound Engineering 3.0.0 中。
## 统一的命名空间，不再有冲突
现在每个 skill 和 agent 都统一使用 ce- 连字符前缀。ce:work 变成 ce-work，git-commit 变成 ce-commit，setup 变成 ce-setup，以此类推。
过去这种前缀、无前缀和冒号前缀混杂的命名方式，与其他插件在各个 harness 上的同名 skill 产生冲突——只要你同时运行多个插件，就会不断遇到这种歧义。ce:work 中的冒号在 Windows 上还需要进行文件系统清理。v3 之后，CE 发布的所有内容都位于自己无歧义的命名空间下，我们的 skill 不会再与任何其他人的冲突。为了清晰和简化，还顺带进行了 2 个小的重命名：ce-review 改为 ce-code-review，ce-document-review 改为 ce-doc-review。
这是一项破坏性变更。现有的链、引用和肌肉记忆都需要迁移到新名称……但这是值得的！😀
## 从头脑风暴到提交的完整记录链，为你和你的智能体服务
在 v3 之前，你在 ce-brainstorm 中写下的需求没有稳定的身份标识。当它到达 ce-plan 时已经变成了散文；等它被实现时就成了感觉。当测试失败或审查者问“这覆盖了什么？"，答案只存在于某个人的脑子里——而运行 ce-work 或 ce-code-review 的智能体恢复意图的能力并不比你强。
v3 重写了 ce-brainstorm 和 ce-plan 的产品文档，使其承载更好的结构而不仅仅是散文。头脑风暴需求文档现在有一等公民级的章节：参与者、关键流程、验收示例和需求，每个都有稳定的 ID。ce-plan 将这些 ID 提取到需求追踪章节，并分配计划局部的实现单元 ID，这些 ID 在计划深化和重排时保持不变。测试场景引用它们所覆盖的验收示例。ce-work 能在阻塞项、验证和任务标签中识别这些单元 ID。
最直接的好处是为你和你的智能体提供了可追溯性。你可以把一个失败的测试追溯到它本应覆盖的验收示例，把一个计划单元追溯到激发它的流程，把一次提交追溯到引发这一切的头脑风暴条目。不太明显但可以说更重要的好处是，智能体也能遵循同样的线索。审查者可以检查一个实现是否真正满足它声称覆盖的验收示例。调试者可以看到失败的测试在执行哪个流程。一个在 U3 上遇到阻塞的工作智能体可以去阅读原始标准，而不是猜测意图。
这也支持的最大变化是改进了绿地产品级头脑风暴。它们在 v3 之前也获得了以前没有的结构化支持。旧模板是为现有产品内的增量变更而优化的，因此对新事物的深度范围工作只有很薄弱的指导，迫使智能体即兴发挥一个结构。
上游更好的结构会带来下游更好的结果……而且不仅仅是针对流程中的人类。
## 在每个 harness 上都是一等公民
Claude Code 是 CE 的起点，很长一段时间里其他一切都像二等移植。v3 全面改变了这一点。
- Codex 用户可以用原生方式安装：codex plugin marketplace add 加上 /plugins TUI。Skill 和 MCP 服务器直接注册，无需 Bun 转换步骤。Codex 的插件规范尚未覆盖自定义 agent，因此一个小型转换器会填补这个空缺，将 CE 的 agent 写入 ~/.codex/agents/ 下的 Codex TOML。
- Pi 迎来了更大的转变：我们停止发布自己的兼容层，它与最受欢迎的社区 subagent 扩展冲突，并将并行分发串行化。我们现在委托给 2 个社区扩展：@nicopreme 的 pi-subagents 用于原生并行会话，@edlzsh 的 pi-ask-user 用于阻塞性提问的 UX。Pi 用户获得真正的并行和合适的提问叠加层，而我们不必在社区维护者已经做得更好的基础设施上与之竞争。
- CLI 和 VSCode 上的 Copilot 也通过原生插件安装获得了一流支持！
跨 harness 的清理工作也随着这些大改动一起落地：
- question-tool 不再在 Codex（其 request_user_input 仅在 Plan 模式下暴露）或 Claude Code（该工具被延迟，且其 schema 在会话启动时未加载）上静默跳过阻塞性问题，通过一个应用于 37 个 SKILL.md 文件的统一指令修复。
- Skill 描述现在被限制在 harness 上限内，并由 frontmatter 测试强制执行。
## 让你做出决定的审查，而不是敷衍盖章
我们的审查 skill 过去都询问桶级别的政策问题，导致一个决定覆盖许多发现。在实践中，这退化为敷衍盖章或整体搁置，审查者最终不得不在不同措辞的十几个发现中反复争论同一个前提。
v3 围绕逐项发现参与重塑了整个审查系列：
- ce-code-review 交互模式会逐条引导你查看发现的问题，提供“应用/推迟/跳过/‘一路放行其余问题’”等选项，并以通俗易懂的语言解释每项建议的修复。终于有了一种更简单的方式来理解审查反馈，让你能更好更快地做出决策。如果你选择推迟处理某些发现的问题，我们现在会路由到你真正的问题跟踪器，而不是 CE 所管理的那个已废弃的内部待办事项存储。
- ce-doc-review 增加了三层自动修复分类和前提依赖链分组——将所有依赖于同一根问题的发现折叠为单个决策，并附带级联的依赖项。典型的交互量从每次运行的 14 个以上发现降低到大约 4-6 个真正的决策。
- ce-resolve-pr-feedback 采用了更严格的聚类（需要跨轮次的证据才会激活，因此无关的单轮反馈不再触发整体性重构），并过滤掉了机器人包装器产生的噪音。
- 一个新的 ce-swift-ios-reviewer 角色加入了针对 Swift/iOS 技术栈的审查者行列，能够捕获 SwiftUI 状态包装器误用、Combine 循环引用以及 Core Data 上下文线程问题。感谢 @jcjvm 的贡献。
## 诊断而非猜测的调试
ce-debug 本来就有一套 4 阶段方法论，但在实践中它默认使用打印调试，在深入追踪之前跳过环境健全性检查，有时还会在插桩干扰让 heisenbug（海森 bug）暂时消失时就宣布其“已修复”。v3 收紧了所有这些方面：
- 在深入追踪之前进行早期环境健全性检查：分支、依赖、运行时、环境变量、陈旧的构建产物。
- 在提出假设时进行假设审计，以帮助引导 agent。许多错误的假设其实是针对错误假设进行测试的正确假设。
- 一个用于广泛搜索的并行只读子 agent 调度选项。
- 一份新技术参考，涵盖边界插桩、测试顺序污染、复现最小化、单步调试器与插桩对比、heisenbug，以及一个 bug 类别清单（时区、编码、浮点数、溢出、缓存、认证）。
## 更少的摩擦，更少的隐患
对现有技能进行了一些提升使用体验的小升级：
- ce-proof 现在会在用户直接请求分享或迭代 markdown 文件时触发。“Share this to Proof so we can iterate”无需 ce-brainstorm 或 ce-plan 等上游调用者即可使用。你可以在仓库中的任意 markdown 文件上试用，例如使用类似 "/ce-proof iterate with me on <doc>" 的命令。
- ce-demo-reel 增加了本地保存作为 catbox 上传的替代方案，因此你可以将演示素材保留在磁盘上，而无需手动从临时目录中翻找。感谢 Lucas Henn 的这一贡献。
- ce-setup 会检查 ast-grep 及相应的 agent 技能，健康检查输出中新增了与现有 Tools 部分对应的 Skills 部分。
以及一轮正确性和安全性修复：
- ce-plan 不再将非软件研究误分类为单纯的事实查询，不再静默地用通用网络搜索替代用户指定的资源，也不再跳过计划完成后的阻塞式交接菜单。无论非软件问题是你孩子的数学课、去波特兰的自驾游，还是你一直想敲定的家居装修项目，你都可以获得真正深思熟虑的 agent 辅助规划。
- ce-work 不再凭空编造人类时间估算和多天会话分解——这些框架并不适用于 agent 执行。
- ce-update 现在从插件根目录的父目录派生其缓存目录，并带有形状检查，因此格式错误的环境变量值不会坍缩成危险的 rm -rf 目标。感谢 @andrewlook 提出此问题。
- ce-compound 的 YAML frontmatter 现在可以通过严格的解析器和 Cowork 的插件验证器。
下一步计划：继续完善产品和策略技能，提供更精准的验证和发布支持，并加入可配置的偏好设置，使诸如复合文档输出路径之类的配置不再需要硬编码为 docs/solutions。最后一项在 v3 中已接近就绪，但还需要更多时间来打磨。
v3 既是一次发布，也是一个基础……更多内容即将上线！
## 相关链接
- [Trevin Chow](https://x.com/trevin)
- [@trevin](https://x.com/trevin)
- [82K](https://x.com/trevin/status/2047066108763770998/analytics)
- [Compound Engineering 3.0.0](https://github.com/everyinc/compound-engineering-plugin)
- [@nicopreme](https://x.com/@nicopreme)
- [pi-subagents](https://github.com/nicobailon/pi-subagents)
- [@edlzsh](https://x.com/@edlzsh)
- [pi-ask-user](https://github.com/edlsh/pi-ask-user)
- [@jcjvm](https://x.com/jcjvm)
- [Lucas Henn](https://github.com/Isaccseven)
- [ast-grep](https://github.com/ast-grep/ast-grep)
- [@andrewlook](https://x.com/andrewlook)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [2026年4月23日 上午5:32](https://x.com/trevin/status/2047066108763770998)
- [82.9K 次浏览](https://x.com/trevin/status/2047066108763770998/analytics)
- [查看引用](https://x.com/trevin/status/2047066108763770998/quotes)
---
*导出时间: 2026/4/23 15:05:24*
