dev-pipeline-skill

dev-pipeline

一个 Claude Code skill：多代理开发流水线。主会话（Fable）负责设计和写 spec，三个不同厂商的 agent 分别写测试、实现、评审，主会话最后亲自验证并提交。

English version below.

为什么这样分工

• 评审者和实现者不是同一个模型。同一个模型对模糊 spec 的误读高度相关，换模型才能把这类错误暴露出来。
• 测试从 spec 独立写出，实现方看得到测试但不能改，测试和实现的分歧由主会话对照 spec 裁决。
• 额度分摊。评审最吃额度，放额度最多的订阅；写测试便宜，放额度少的；实现放最强的编码 agent。
• 主会话上下文干净。重活都在独立进程里跑，主会话只做判断。

默认分工

任务先分三档：琐碎改动主会话直接做；小改动走轻量档（短 spec → codex 实现 → 主会话验证）；其余走完整流水线。测试只在有明确输入输出逻辑的任务里写，视觉、文案、配置类跳过。

前置条件

• Claude Code 作为主会话。作者只在 Fable 上跑过；主会话要做设计、裁决和最终把关，模型弱了流水线会散。
• codex CLI（实现）。
• kimi CLI（测试，评审替补）。
• opencode + 智谱 GLM Coding Plan（评审，测试替补）。
• 以上任何一个都可以换成你自己的订阅，改 SKILL.md 里的角色表和对应的三条命令即可。

安装

git clone https://github.com/Abreto/dev-pipeline-skill.git ~/.claude/skills/dev-pipeline

或者放进某个项目的 .claude/skills/dev-pipeline/ 只对该项目生效。装好后在 Claude Code 里说「走流水线做 X」「让 codex 实现，GLM review」即可触发。

自定义

• 模型分配：改 SKILL.md 的角色表和 opencode / kimi / codex 三条命令。
• 本机信息：在同目录建 LOCAL.md（已在 .gitignore），放 key 位置、额度状况、撞限报错文案、事故记录。skill 开工前会先读它。
• 没装 codex、或暂时不想用它：SKILL.md 里有用 Claude Code 自己的 opus 子代理当实现 agent 的写法（含退路表怎么跟着变），在 LOCAL.md 写一行即可切换。
• examples/spec-template.md：spec 模板。examples/prompts.md：三个角色的 prompt 模板。

已知的坑

都写在 SKILL.md 对应 CLI 的小节里。最常踩的三条：三个 CLI 的非交互模式都要 < /dev/null，否则等 stdin 挂住；opencode 的模型前缀必须是 zhipuai-coding-plan/，写成 zhipuai/ 会打到按量接口；Claude Code 的 settings.json env 优先级高于 shell 环境变量，嵌套 claude -p 换端点要用 CLAUDE_CONFIG_DIR。

许可

MIT

dev-pipeline (English)

A Claude Code skill that runs a multi-agent development pipeline. The main session (Fable) does design and writes the spec; three agents from three different vendors write tests, implement, and review; the main session verifies and commits.

Why split it this way

• Reviewer and implementer are never the same model. A single model misreads an ambiguous spec the same way twice. A different model surfaces the disagreement.
• Tests are written from the spec, independently. The implementer can see them but cannot edit them. Conflicts go back to the main session, which rules against the spec.
• Quota is spread out. Review burns the most tokens, so it goes on the subscription with the most headroom. Test writing is cheap, so it goes on the smallest one. Implementation goes to the strongest coding agent.
• The main session stays lean. Heavy work runs in separate processes; the main session only judges.

Default roles

Tasks are triaged into three tiers first: trivial edits are done directly by the main session; small changes take the light path (short spec → codex → verify); everything else takes the full pipeline. Tests are only written for tasks with clear input/output logic; visual, copy, and config changes skip them.

Prerequisites

• Claude Code as the main session. The author has only run this on Fable; the main session designs, arbitrates, and gates the result, so a weaker model makes the pipeline fall apart.
• codex CLI for implementation.
• kimi CLI for tests and as review fallback.
• opencode with a Zhipu GLM Coding Plan for review and as test fallback.
• Any of these can be swapped for your own subscriptions by editing the role table and the three commands in SKILL.md.

Install

git clone https://github.com/Abreto/dev-pipeline-skill.git ~/.claude/skills/dev-pipeline

Or drop it into a project's .claude/skills/dev-pipeline/ to scope it to that project. Then tell Claude Code to "run the pipeline for X" or "have codex implement it and GLM review it".

Customize

• Model assignment: edit the role table and the opencode / kimi / codex commands in SKILL.md.
• Machine-specific notes: create LOCAL.md next to it (already gitignored) for key locations, quota status, observed rate-limit messages, and incident notes. The skill reads it before starting.
• No codex, or not using it for now: SKILL.md documents running a Claude Code opus subagent as the implementation agent (including how the fallback chains shift); one line in LOCAL.md switches it.
• examples/spec-template.md is a spec skeleton; examples/prompts.md has prompt templates for all three roles.

Known pitfalls

All recorded under each CLI's section in SKILL.md. The three most common: every CLI's non-interactive mode needs < /dev/null or it hangs waiting on stdin; opencode's model prefix must be zhipuai-coding-plan/, since zhipuai/ hits the pay-as-you-go endpoint; Claude Code's settings.json env overrides shell env, so a nested claude -p pointed at another endpoint needs CLAUDE_CONFIG_DIR.

License

MIT