agentsclimarketplace

Pdlc loop run

Skill kanfu-panda/pdlc-skills/skills/pdlc-loop-run

36 slash commands that turn Claude Code into a real PDLC workflow — PRD → TDD → implement → review → ship. Hard contracts force AI to persist artifacts, write failing tests first, and run self-checks. No more "looks done" in chat.

Install
npx -y skills add kanfu-panda/pdlc-skills --skill pdlc-loop-run

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 9 stars9 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

收敛循环引擎(自动推进 tdd→implement→review 到 review_done 或 blocked)

SKILL.md

3.6 KB, as published. Nobody here has run it

收敛循环引擎(loop-run)

把机械收敛段 tdd → implement → review 烧成一个自主循环:从 current_stage 出发逐阶段自动推进,直到 review_done 或被 blocked。这是 /pdlc-loop-next 之上的高层引擎——用户不必自己写 bash 循环。

<!-- @include templates/prompts/noninteractive.md -->

终态即 review_done(设计如此):本引擎只覆盖机械收敛段。到达 review_done成功停机,交人工决定是否 /pdlc-ship绝不自动进入 pdlc-ship / pdlc-deploy——发布/部署是不可逆·外发操作,永远留人(--autonomous 对它们无效)。prd/design 的关键取舍同样不在本引擎范围内。

两种运行形态

  • 默认推荐 · 外部 Runbook(真进程隔离):长跑 / 过夜 / 多 feature 并行时,用独立进程逐轮跑,每轮全新进程 = 真 fresh context 最可信。本命令可打印该 Runbook 脚本(见 usage-guide「自主循环 Runbook」)。
  • 便捷 · 插件内 Task 版(本命令默认行为):适合短收敛。每个 stage 派发给一个 fresh Task subagent(context 相对隔离),返回后读状态机决定推进 / 停 / block。

循环算法(Task 版)

  1. $ARGUMENTS 取功能ID;--max-steps 取迭代上限(缺省 4)。读 docs/.pdlc-state/<功能ID>.json
  2. 循环,每轮:
    1. /pdlc-loop-next 的映射判定下一条命令:pdlc-tdd / pdlc-implement / pdlc-review / done / blocked
    2. done成功停机,输出 <<<PDLC done stage=review>>>,提示交人工 /pdlc-ship
    3. blocked → 停机交还人类,输出 <<<PDLC blocked reason="...">>>
    4. 否则用 Task 工具派发该命令到一个 fresh subagent,--autonomous,模型取目标 skill frontmatter 的 recommended_model(无则继承)。
    5. subagent 返回后重新读状态机 last_phase_result
      • ok=falsefail-stop:停机、不重跑同一 stage,输出 blocked 哨兵。
      • current_stage 未推进(违反 IRON LAW 第 6 条)→ stuck-stop:停机报错。
      • ok=true 且已推进 → 记一步,继续下一轮。
    6. 步数 > --max-steps上限停机(防病态空转烧 token)。
  3. 输出循环小结(跑了几步、终态、每步 last_phase_result 摘要)。

护栏(不可协商)

  • 迭代上限:收敛段线性前进 tdd→implement→review(3 段),默认上限 4 = 3 段 + 1 段容错余量(供「首轮 block、人工修好后从原 stage 续跑一次」)。不给更多余量。
  • 只前进不回炉:循环只在不同 stage 间前进,永不原地重跑同一 stage(与 IRON LAW「修复单次不递归」一致——单 stage 内的修复仍是单次)。
  • fail-stop / stuck-stop:任一 stage ok=false 或状态未推进即停,绝不空转。
  • 预算:外部 Runbook 形态必须配 --max-budget-usd(见 usage-guide)。自主循环持续烧 token,护栏是硬要求。

与 IRON LAW 的关系

本引擎不违反「修复单次不递归」:单个 stage 的自检-修复仍单次;引擎做的是跨 stage 前进 + 失败即停 + 上限,不是「重试同一步」。

功能ID: $ARGUMENTS

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.