arch-doctor
诊断这台机器上的设置：hook、边界规则、构建 wrapper、java、项目来自哪个 .claude/，以及 compose 栈是否在运行。
arch-doctor，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。
## Diagnosis
!java "${CLAUDE_PROJECT_DIR:-.}/.claude/hooks/ArchHook.java" doctor 2>&1
## AI files
!find "${CLAUDE_PROJECT_DIR:-.}/.claude" -maxdepth 2 -type f | sort
Interpret the result above and tell the user, in two or three sentences, whether the setup is operational. If something is marked with ❌, give the concrete command that fixes it — don't describe the problem in general terms.
The Provenance line is the one exception to "give the command": when it lists files edited locally, say which ones and that an update would overwrite them. Don't offer to run the update — losing a hand-edited norm is the user's call, not a fix to apply.
Don't fix anything without the user asking.
## Why this is a manually-invoked skill
Form 2: the whole body is two shell injections plus how to read them, so it is a procedure (axis 5) fired by a person who wants a verdict (axis 2 — /arch-doctor). Form 1 was rejected because a diagnostic the model reaches on its own runs a JVM and a find on every session that mentions a hook; Form 3 was rejected because there is nothing to isolate — the output is the report, and it is what the user asked to see.
Runs on sonnet with effort: high: the report is computed, but naming the cause and the command that fixes it is judgment (@.claude/decisions/0080-procedural-skills-on-sonnet.md).
The name is not doctor: /doctor is a native command, and a skill folder with that name shadows it silently — @CLAUDE.md § Known pitfalls.
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
observer — the territory is skill_classes.observer in @.claude/schemas/extensions.json, and it is empty: this skill writes no file. It reports, and the fix is the user's call. ArchHook.java guard enforces it.
Reads
the output of ArchHook.java doctor (which folds in compose) and the listing of .claude/, both injected above. Nothing else: an injection that lied about the setup would be worse than no report.
Writes
nothing, ever. Leaves no audit report either — its class, observer in skill_classes of @.claude/schemas/extensions.json, declares audited: false, because an observer that records its own observation makes the trail mostly about reading it.
来源：SKILL.md § Contract（nerviz e4cae27）。
guard 如何处理它
observer 是这个 skill 的类别，左侧的列表就是它的领地。skill 在自己的提示上运行时阶段打开，ArchHook 对照这份列表判断每一次写入。右侧的文字是 hook 自己输出的。
arch-doctor 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。
> /arch-doctor … UserPromptSubmit guard prompt exit 0 # phase opened: arch-doctor (class observer) # write_allow: [] — this class writes no file
PreToolUse Write(pom.xml) guard write exit 2 ❌ `arch-doctor` is class `observer` — pom.xml is outside its territory. write_allow: (nothing — this class writes no file) Put the content in the spec as a code block, or finish the run and invoke the skill that owns this path. If the path is legitimately this skill's, widen skill_classes.observer in .claude/schemas/extensions.json — retrying will not help.