pstack-claude：将 Cursor 的严谨 Agent 工作流移植至 Claude Code、Codex、Pi 等平台

Lauren Tan 的 pstack 是一套具有明确方法论的 Cursor 技能栈，可改善 Agent 的任务执行结果。本项目将其移植至 Claude Code、Codex、Pi 及其他 Agent 运行环境。它会跟踪上游版本，同时维护具名策略分支；每个分支均在 tools/forks.json 中声明。

向 poteto-mode 描述你的目标，它将针对任务调用正确的工作流。它旨在让代码保持精简、简单并经过验证。

对于测试无法触及的并发缺陷和不变量问题，请参阅独立的 agent-formal-verify 插件；该插件加入了 TLA+ 模型检查和 Lean 证明。

在 Claude Code 中运行：

/plugin marketplace add michael-denyer/pstack-claude
/plugin install pstack@pstack-claude

在终端中运行：

codex plugin marketplace add michael-denyer/pstack-claude
codex plugin add pstack@pstack-claude

在终端中运行：

pi install git:github.com/michael-denyer/pstack-claude

该软件包会加载技能以及 pstack Pi 扩展。该扩展提供技能所使用的子 Agent、提问和唤醒工具，以及 /loop 和路由指令。可通过 /skill:<name> 调用技能。

运行 setup-pstack 可更改默认模型、为每个角色设置推理强度（例如 arena runners: opus @xhigh, fable @max；Claude Code 会通过插件的 pstack:effort-<level> 或 pstack:poteto-agent-<level> Agent 进行调度；未指定强度等级的角色将保留当前会话的推理强度，除非配置表中的 default effort 行指定了某个等级），或关闭自动路由。该插件会在 Claude Code 和 Codex 中安装路由钩子；Codex 会在执行前通过 /hooks 要求你信任该钩子。在 Pi 中，扩展会注入相同的路由指令。在 Claude Code 中，请使用 /pstack:setup-pstack。

如需在 Prime Agent、OpenCode、Gemini CLI 中安装，或为任意运行环境仅安装技能，请参阅共享安装说明。

Use poteto-mode to fix the search filter resetting when I change pages.

对于缺陷，它会复现失败情况，使用 how 和 why 进行调查，将修复任务委派出去，然后重新运行失败用例。如果修复跨越函数边界，它会在实现之前引入 architect。你将获得修复结果，以及失败与通过的验证证据。

其他工作手册涵盖规划、功能开发、重构、性能问题、调查、原型、PR 维护、发布和长期项目。

pstack 没有服务器或遥测机制。其技能要求你的 Agent 读取的任何内容，包括会话记录，都会发送至你的模型提供商。脚本在本地运行，PR 工具使用你的 GitHub CLI 登录凭据。

感谢你帮助改进这一移植版本。欢迎提交 Bug 报告、文档修复和运行时改进。请参阅 CONTRIBUTING.md，了解检查项及你的改动应归属的位置。请按照 SECURITY.md 所述方式私下报告安全漏洞。

本移植版本及其修改和新增内容同样采用 MIT 许可证，© 2026 Michael Denyer。原始 pstack © 2026 Lauren Tan；导入的 cursor-team-kit 技能 © 2026 Cursor。请参阅 LICENSE-cursor-team-kit 和 NOTICE.md。

Lauren Tan's pstack is an opinionated Cursor skill stack that improves agent outcomes. This is a port for Claude Code, Codex, Pi and other agent harnesses. It tracks upstream and also carries named policy forks, each declared in tools/forks.json.

Tell poteto-mode your goal and it will invoke the correct workflow for the task. It keeps your code concise, simple and verified.

For concurrency bugs and invariants that tests cannot reach, see the separate agent-formal-verify plugin, which adds TLA+ model checking and Lean proofs.

Run in Claude Code:

/plugin marketplace add michael-denyer/pstack-claude
/plugin install pstack@pstack-claude

Run in your terminal:

codex plugin marketplace add michael-denyer/pstack-claude
codex plugin add pstack@pstack-claude

Run in your terminal:

pi install git:github.com/michael-denyer/pstack-claude

The package loads the skills and the pstack Pi extension, which adds the subagent, question, and wake-up tools the skills use, plus /loop and the routing instruction. Invoke a skill with /skill:<name>.

Run setup-pstack to change model defaults, set a reasoning effort per role (for example arena runners: opus @xhigh, fable @max, which Claude Code dispatches through the plugin's pstack:effort-<level> or pstack:poteto-agent-<level> agents; roles without a level keep the session's effort unless the sheet's default effort line names one), or turn automatic routing off. The plugin installs the routing hook on Claude Code and Codex; Codex asks you to trust it through /hooks before it runs. On Pi the extension injects the same routing instruction. In Claude Code, use /pstack:setup-pstack.

For Prime Agent, OpenCode, Gemini CLI, or skills-only installs for any harness, see shared installation.

Use poteto-mode to fix the search filter resetting when I change pages.

For a bug, it reproduces the failure, uses how and why to investigate, delegates the fix, then reruns the failing case. If the fix crosses a function boundary, it brings in architect before implementation. You receive the fix and the failing and passing evidence.

Other playbooks cover planning, features, refactoring, performance issues, investigations, prototypes, PR maintenance, shipping, and longer projects.

pstack has no server or telemetry. Anything its skills ask your agent to read, including session transcripts, goes to your model provider. Scripts run locally, and PR tools use your GitHub CLI login.

Thanks for helping make this port better. Bug reports, documentation fixes, and runtime improvements are welcome. See CONTRIBUTING.md for the checks and where your change belongs. Report vulnerabilities privately as described in SECURITY.md.

This port, including its modifications and additions, is also MIT-licensed, © 2026 Michael Denyer. Original pstack © 2026 Lauren Tan; imported cursor-team-kit skills © 2026 Cursor. See LICENSE-cursor-team-kit and NOTICE.md.