nerviz

↑ ↓ 移动 · Enter 打开 · Esc 关闭点按一个结果

架构
skill · orchestrator 类 · sonnet

init-project

在主会话中完成项目首次生成的访谈，然后把整个生成交给一个 agent。

skill · orchestratorsonneteffort low不写任何文件留在 nerviz 中
§ skill 的定义

init-project，逐节说明

skill 自己的 SKILL.md 的各个章节，节选。悬停或点击一个章节，查看它引用的项目规则、模板和 references/ 文件。章节也是标签：点击即可跳转。

nerviz/.claude/skills/init-project/SKILL.md

## Available blueprints

!find "${CLAUDE_PROJECT_DIR:-.}/.claude/blueprints" -mindepth 2 -maxdepth 2 -name '*.yaml' 2>/dev/null | sort

## Next to this repository

!ls -a "${CLAUDE_PROJECT_DIR:-.}/.." | head -30

## Session

${CLAUDE_SESSION_ID}

## Arguments

$ARGUMENTS

## Procedure

  1. The project is generated next to this repository, at ../<artifactId>
  2. Interview — every question of the run, here, before delegating. The agent runs in the

    … 此处截断，其余见文件

    ## Why this is a manually-invoked skill

    Form 2: creating a project is a ritual a person starts (axis 2 — /init-project), and the interview it leads to cannot be triggered by inference from a sentence. Form 1 was rejected for exactly that: a skill that can scaffold a whole tree on the model's own judgment is a directory rewritten by accident. Form 3 was rejected here and used one level down instead — the isolation belongs to the generation, which is why project-initializer exists, and the interview belongs here, in the main session, the only place AskUserQuestion is guaranteed (@.claude/decisions/0123-lessons-learned-020-init-project-run.md).

    Runs on sonnet with effort: low: the interview reads its questions from the blueprints and from the two chained skills, and the generation is delegated to project-initializer, which declares its own model (@.claude/decisions/0081-skill-model-required-per-class.md).

    在 GitHub 上打开完整文件 →
    § 行为

    它读什么、写什么、拒绝什么

    skill 自己的契约，取自 SKILL.md，保留原文。

    Class

    orchestrator — the territory is skill_classes.orchestrator's override for this skill in @.claude/schemas/extensions.json, and it is empty: this skill writes no file itself. Everything is written by project-initializer, which the guard bypasses as an executor agent, and by project-bootstrap under it. ArchHook.java guard enforces it. The one write it causes is audit genesis replacing three placeholders in the generated project's GENESIS.md, outside this repository — done by the hook's code, never by a Write.

    Reads

    the blueprint listing and the listing of this repository's parent directory injected above, every blueprint YAML for the interview, the § 2 questions of transport-security-setup and sonarqube-setup, and the arguments.

    Owns

    the interview of a project's first generation. The questions about transport and Sonar stay owned by their skills; this one only asks them before the agent starts.

    Delegates

    to project-initializer — never generates a file inline, not even a single pom.xml. A partial generation from this thread is the failure mode the agent exists to prevent: the report would describe a tree nobody wrote in full.

    来源：SKILL.md § Contract（nerviz e4cae27）。

    § guard 与 hook

    guard 如何处理它

    orchestrator 是该 skill 的类，它的领地为空：自身不写任何文件。它打开的阶段正是阻止在主线程中生成项目的机制。右侧文字是 hook 自己的输出。

    init-project 类的 write_allow

    • 任何文件：领地为空

    拒绝以退出码 2 结束，消息回到模型。

    > /init-project …
UserPromptSubmit
  guard prompt                     exit 0
# phase opened: init-project (class orchestrator)
# write_allow: [] — this skill writes no file itself
    § 编排

    它在轨道上的位置

    每张卡片是 /init-project 的一步。访谈在主会话中进行；着色的步骤在 project-initializer 内运行；git-publish 在 audit genesis 之后回到主会话。虚线步骤仅在适用时运行。

    /init-project orchestrator · sonnet
    ← 返回链路