claudit

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Claudit: Claude Code Configuration Audit

Claudit：Claude Code配置审计

You are the Claudit orchestrator. When the user runs

/claudit

, execute this 5-phase audit workflow. Follow each phase in order. Do not skip phases.

你是Claudit编排器。当用户运行

/claudit

时，执行以下5阶段审计工作流。按顺序执行每个阶段，不得跳过。

Phase 0: Environment Detection & Configuration Map

阶段0：环境检测与配置映射

Step 1: Environment Detection

步骤1：环境检测

PROJECT_ROOT: Run
```
git rev-parse --show-toplevel 2>/dev/null
```
via Bash. If this fails (not in a git repo), set PROJECT_ROOT to empty.
HOME_DIR: Run
```
echo $HOME
```
via Bash.

PROJECT_ROOT：通过Bash执行
```
git rev-parse --show-toplevel 2>/dev/null
```
。如果执行失败（当前不在Git仓库中），则将PROJECT_ROOT设为空。
HOME_DIR：通过Bash执行
```
echo $HOME
```
。

Step 2: Scope Detection

步骤2：范围检测

If
```
PROJECT_ROOT
```
is found → comprehensive (global + project)
If
```
PROJECT_ROOT
```
is empty → global only

如果找到
```
PROJECT_ROOT
```
→ 全面审计（全局+项目）
如果
```
PROJECT_ROOT
```
为空 → 仅全局审计

Step 3: Comprehensive Configuration Scan

步骤3：全面配置扫描

Run parallel Glob calls to discover every Claude-related file. Cap at 50 total files — if a project has more, report the cap and proceed with the 50 most recently modified.

Project-level (if comprehensive):

Category	Glob Pattern	Notes
Instructions	`{PROJECT_ROOT}/**/CLAUDE.md`	Exclude node_modules, .git, vendor, dist, build via pattern
Local instructions	`{PROJECT_ROOT}/CLAUDE.local.md`	Personal/gitignored
Rules	`{PROJECT_ROOT}/.claude/rules/*/.md`	Modular rules with optional path frontmatter
Settings (shared)	`{PROJECT_ROOT}/.claude/settings.json`	Team settings
Settings (local)	`{PROJECT_ROOT}/.claude/settings.local.json`	Personal project settings
Skills	`{PROJECT_ROOT}/.claude/skills/*/SKILL.md`	Project skills
Agents	`{PROJECT_ROOT}/.claude/agents/*.md`	Project subagents
Memory	`{PROJECT_ROOT}/.claude/MEMORY.md`	Project memory
MCP	`{PROJECT_ROOT}/.mcp.json`	Project MCP servers
Plugin hooks	`{PROJECT_ROOT}/.claude/plugins/*/hooks/hooks.json`	Plugin-level hooks

For the Instructions glob, exclude common vendor directories. Use Glob with pattern

**/CLAUDE.md

rooted at PROJECT_ROOT, then filter out paths containing

node_modules

.git

vendor

dist

, or

build

Global-level (always):

Category	Path	Notes
Settings	`~/.claude/settings.json`	Global settings
Instructions	`~/.claude/CLAUDE.md`	Global instructions (check `~/CLAUDE.md` too as legacy)
Rules	`~/.claude/rules/*/.md`	Personal modular rules
Memory	`~/.claude/MEMORY.md`	Global memory
MCP	`~/.claude/.mcp.json`	Global MCP servers
Plugins	`~/.claude/plugins/installed_plugins.json`	Installed plugins
Marketplaces	`~/.claude/plugins/known_marketplaces.json`	Registered marketplaces
Managed policy (macOS)	`/Library/Application Support/ClaudeCode/CLAUDE.md`	macOS managed policy
Managed policy (Linux)	`/etc/claude-code/CLAUDE.md`	Linux/WSL managed policy

For each file found, get its line count via

wc -l

(batch multiple files in a single Bash call for efficiency). Quote paths containing spaces (e.g.,

/Library/Application Support/...

) in any Bash commands.

并行执行Glob调用，发现所有与Claude相关的文件。最多保留50个文件——如果项目中文件超过50个，报告数量上限并选取最近修改的50个文件。

项目级（仅全面审计时）：

类别	Glob模式	说明
指令文件	`{PROJECT_ROOT}/**/CLAUDE.md`	排除node_modules、.git、vendor、dist、build目录
本地指令文件	`{PROJECT_ROOT}/CLAUDE.local.md`	个人/已忽略Git的文件
规则文件	`{PROJECT_ROOT}/.claude/rules/*/.md`	带可选路径前置信息的模块化规则
共享设置	`{PROJECT_ROOT}/.claude/settings.json`	团队设置
本地设置	`{PROJECT_ROOT}/.claude/settings.local.json`	个人项目设置
技能文件	`{PROJECT_ROOT}/.claude/skills/*/SKILL.md`	项目技能
代理文件	`{PROJECT_ROOT}/.claude/agents/*.md`	项目子代理
记忆文件	`{PROJECT_ROOT}/.claude/MEMORY.md`	项目记忆
MCP配置	`{PROJECT_ROOT}/.mcp.json`	项目MCP服务器
插件钩子	`{PROJECT_ROOT}/.claude/plugins/*/hooks/hooks.json`	插件级钩子

对于指令文件的Glob模式，排除常见的第三方依赖目录。以PROJECT_ROOT为根目录使用

**/CLAUDE.md

模式，然后过滤掉包含

node_modules

、

.git

、

vendor

、

dist

或

build

的路径。

全局级（始终执行）：

类别	路径	说明
全局设置	`~/.claude/settings.json`	全局设置
全局指令	`~/.claude/CLAUDE.md`	全局指令（同时检查 `~/CLAUDE.md` 作为遗留路径）
全局规则	`~/.claude/rules/*/.md`	个人模块化规则
全局记忆	`~/.claude/MEMORY.md`	全局记忆
全局MCP	`~/.claude/.mcp.json`	全局MCP服务器
已安装插件	`~/.claude/plugins/installed_plugins.json`	已安装的插件
已注册市场	`~/.claude/plugins/known_marketplaces.json`	已注册的插件市场
macOS托管策略	`/Library/Application Support/ClaudeCode/CLAUDE.md`	macOS平台的托管策略
Linux托管策略	`/etc/claude-code/CLAUDE.md`	Linux/WSL平台的托管策略

对于每个找到的文件，通过

wc -l

获取行数（为提高效率，可在单个Bash调用中批量处理多个文件）。对包含空格的路径（如

/Library/Application Support/...

）在Bash命令中添加引号。

Step 4: Build and Present the Configuration Map

步骤4：构建并展示配置映射

Build a structured manifest grouping files by category with line counts. Present it to the user:

=== CONFIGURATION MAP ===
Scope: Comprehensive (project + global)

PROJECT: {PROJECT_ROOT}
  Instructions (N files, ~N tokens):
    CLAUDE.md                        45 lines
    src/api/CLAUDE.md                30 lines
    CLAUDE.local.md                  10 lines
    .claude/rules/testing.md         15 lines
  Settings (N files):
    .claude/settings.json            exists
    .claude/settings.local.json      exists
  Skills (N): [list]
  Agents (N): [list]
  Memory: .claude/MEMORY.md          30 lines
  MCP: .mcp.json                     N servers configured

GLOBAL: ~/.claude/
  Instructions: ~/.claude/CLAUDE.md  20 lines
  Rules: [list or "none"]
  Settings: ~/.claude/settings.json  exists
  Memory: ~/.claude/MEMORY.md        15 lines
  MCP: ~/.claude/.mcp.json           N servers configured
  Plugins: N installed

MANAGED POLICY: [found (N lines) / not found]
=== END MAP ===

Estimate tokens for instruction files as

(total_lines * 40) / 4

(rough estimate: ~10 words per line, ~4 chars per word, divided by 4 chars per token). This line-based estimate is for the config map display only. Audit agents use

chars/4

for more precise per-file token counts after reading file contents. Show the aggregate token estimate for instruction files.

After presenting the map, tell the user:

Phase 1: Building expert context from official Anthropic documentation...

构建结构化清单，按类别分组文件并显示行数。向用户展示如下：

=== 配置映射 ===
范围：全面审计（项目+全局）

项目：{PROJECT_ROOT}
  指令文件（共N个文件，约N个token）：
    CLAUDE.md                        45行
    src/api/CLAUDE.md                30行
    CLAUDE.local.md                  10行
    .claude/rules/testing.md         15行
  设置文件（共N个）：
    .claude/settings.json            已存在
    .claude/settings.local.json      已存在
  技能文件（共N个）：[列表]
  代理文件（共N个）：[列表]
  记忆文件：.claude/MEMORY.md          30行
  MCP配置：.mcp.json                     已配置N个服务器

全局：~/.claude/
  指令文件：~/.claude/CLAUDE.md  20行
  规则文件：[列表或“无”]
  设置文件：~/.claude/settings.json  已存在
  记忆文件：~/.claude/MEMORY.md        15行
  MCP配置：~/.claude/.mcp.json           已配置N个服务器
  插件：已安装N个

托管策略：[已找到（N行）/未找到]
=== 映射结束 ===

指令文件的token估算公式为

(总行数 * 40) / 4

（大致估算：每行约10个单词，每个单词约4个字符，每个token约4个字符）。此基于行数的估算仅用于配置映射展示。审计代理在读取文件内容后，会使用

字符数/4

来获得更精确的单文件token统计。显示指令文件的总token估算值。

展示完映射后，告知用户：

阶段1：从官方Anthropic文档构建专家上下文...

Phase 1: Build Expert Context

阶段1：构建专家上下文

Dispatch 3 research subagents in parallel using the Task tool. All must be foreground (do NOT use

run_in_background

使用Task工具并行调度3个研究子代理。所有代理均需在前台运行（请勿使用

run_in_background

）。

Dispatch All Three Simultaneously

同时调度三个代理

In a single message, dispatch all 3 Task tool calls:

Research Core:

```
description
```
: "Research core config docs"
```
subagent_type
```
: "claudit:research-core"
```
prompt
```
: "Build expert knowledge on Claude Code core configuration. Read the baseline from ${CLAUDE_PLUGIN_ROOT}/skills/claudit/references/known-settings.md first, then fetch official Anthropic documentation for settings, permissions, CLAUDE.md, and memory. Return structured expert knowledge."

Research Ecosystem:

```
description
```
: "Research ecosystem docs"
```
subagent_type
```
: "claudit:research-ecosystem"
```
prompt
```
: "Build expert knowledge on Claude Code ecosystem features. Fetch official Anthropic documentation for MCP servers, hooks, skills, sub-agents, and plugins. Return structured expert knowledge."

Research Optimization:

```
description
```
: "Research optimization docs"
```
subagent_type
```
: "claudit:research-optimization"
```
prompt
```
: "Build expert knowledge on Claude Code performance and over-engineering patterns. Fetch official Anthropic documentation for model configuration, CLI reference, and best practices. Search for context optimization and over-engineering anti-patterns. Return structured expert knowledge."

在一条消息中调度所有3个Task工具调用：

核心配置研究：

```
description
```
: "研究核心配置文档"
```
subagent_type
```
: "claudit:research-core"
```
prompt
```
: "构建关于Claude Code核心配置的专家知识。首先读取${CLAUDE_PLUGIN_ROOT}/skills/claudit/references/known-settings.md中的基准内容，然后获取关于设置、权限、CLAUDE.md和记忆的官方Anthropic文档。返回结构化的专家知识。"

生态系统研究：

```
description
```
: "研究生态系统文档"
```
subagent_type
```
: "claudit:research-ecosystem"
```
prompt
```
: "构建关于Claude Code生态系统功能的专家知识。获取关于MCP服务器、钩子、技能、子代理和插件的官方Anthropic文档。返回结构化的专家知识。"

优化研究：

```
description
```
: "研究优化文档"
```
subagent_type
```
: "claudit:research-optimization"
```
prompt
```
: "构建关于Claude Code性能与过度设计模式的专家知识。获取关于模型配置、CLI参考和最佳实践的官方Anthropic文档。搜索上下文优化和过度设计反模式的相关内容。返回结构化的专家知识。"

Assemble Expert Context

整合专家上下文

Once all 3 return, combine their results into a single Expert Context block:

=== EXPERT CONTEXT ===

当三个代理均返回结果后，将它们的结果合并为一个专家上下文块：

=== 专家上下文 ===

Core Configuration Knowledge

核心配置知识

[Results from research-core]

[来自research-core的结果]

Ecosystem Knowledge

生态系统知识

[Results from research-ecosystem]

[来自research-ecosystem的结果]

Optimization & Over-Engineering Knowledge

优化与过度设计知识

[Results from research-optimization]

=== END EXPERT CONTEXT ===


Tell the user:

Expert context assembled. Proceeding to configuration analysis...

Phase 2: Analyzing your configuration against expert knowledge...

---

[来自research-optimization的结果]

=== 上下文结束 ===


告知用户：

专家上下文已整合。开始配置分析...

阶段2：对照专家知识分析你的配置...

---

Phase 2: Expert-Informed Audit

阶段2：基于专家知识的审计

Dispatch audit subagents using the Task tool. Each agent receives the Expert Context from Phase 1 plus only its relevant slice of the configuration map.

使用Task工具调度审计子代理。每个代理将接收阶段1的专家上下文以及其相关的配置映射片段。

Build Agent Dispatch Prompts

构建代理调度提示词

For
audit-global
, include:

Full Expert Context
Global slice of config map: global instructions, global rules, global settings, global memory, global MCP, plugins, managed policy paths
If comprehensive: also include the content of the project's root CLAUDE.md (read it and paste the first 200 lines) so the agent can detect cross-scope redundancy. Cap at 200 lines to avoid bloating the agent prompt for very large files.

For
audit-project
(comprehensive only), include:

Full Expert Context
Project slice of config map: all project instructions (with full paths), rules, settings, skills, agents, memory

For
audit-ecosystem
, include:

Full Expert Context
Ecosystem slice: all MCP config paths (global + project as applicable), plugins path, plugin hooks paths, paths to all settings files (agent reads them to check for hooks)

对于
audit-global
，需包含：

完整的专家上下文
配置映射的全局片段：全局指令、全局规则、全局设置、全局记忆、全局MCP、插件、托管策略路径
如果是全面审计：还需包含项目根目录下CLAUDE.md的内容（读取并粘贴前200行），以便代理检测跨范围冗余。对于超大文件，最多粘贴200行以避免代理提示词过于冗长。

对于
audit-project
（仅全面审计时），需包含：

完整的专家上下文
配置映射的项目片段：所有项目指令（含完整路径）、规则、设置、技能、代理、记忆

对于
audit-ecosystem
，需包含：

完整的专家上下文
生态系统片段：所有MCP配置路径（全局+项目，如适用）、插件路径、插件钩子路径、所有设置文件的路径（代理将读取这些文件以检查钩子配置）

Dispatch Based on Scope

根据审计范围调度代理

Global only → dispatch

audit-global

audit-ecosystem

in parallel (2 agents) Comprehensive → dispatch all three in parallel (3 agents)

Use these agent types:

```
subagent_type
```
: "claudit:audit-global"
```
subagent_type
```
: "claudit:audit-project"
```
subagent_type
```
: "claudit:audit-ecosystem"

仅全局审计 → 并行调度

audit-global

audit-ecosystem

（2个代理） 全面审计 → 并行调度所有三个代理（3个代理）

使用以下代理类型：

```
subagent_type
```
: "claudit:audit-global"
```
subagent_type
```
: "claudit:audit-project"
```
subagent_type
```
: "claudit:audit-ecosystem"

Phase 3: Scoring & Synthesis

阶段3：评分与综合

Once all audit agents return, read the scoring rubric:

Read

${CLAUDE_PLUGIN_ROOT}/skills/claudit/references/scoring-rubric.md

当所有审计代理返回结果后，读取评分规则：

读取

${CLAUDE_PLUGIN_ROOT}/skills/claudit/references/scoring-rubric.md

Score Each Category

为每个类别评分

Apply the rubric to the audit findings. For each of the 6 categories:

Start at base score of 100
Apply matching deductions from the rubric based on audit findings
Apply matching bonuses from the rubric based on audit findings
Clamp to 0-100 range

Categories and their weights:

Category	Weight	Primary Audit Source
Over-Engineering Detection	20%	audit-project (CLAUDE.md analysis) + audit-ecosystem (hook/MCP sprawl)
CLAUDE.md Quality	20%	audit-project (structure, sections, references, multi-file)
Security Posture	15%	audit-project (permissions) + audit-global (settings)
MCP Configuration	15%	audit-ecosystem (server health, sprawl)
Plugin Health	15%	audit-ecosystem (plugin structure) + audit-global (installed plugins)
Context Efficiency	15%	All audits (token cost estimates, aggregate instruction size)

Scope-aware scoring:

Global only: Exclude CLAUDE.md Quality from scoring (no project to evaluate). Renormalize the remaining 5 category weights proportionally: Over-Engineering = 20/80 = 25%, Security = 15/80 = 18.75%, MCP = 18.75%, Plugin = 18.75%, Context = 18.75%. Note "CLAUDE.md Quality: skipped (no project detected)" in the report.
Comprehensive: Score all 6 categories normally.

根据审计结果应用评分规则。对以下6个类别分别评分：

初始基准分100
根据审计结果应用匹配的扣分规则
根据审计结果应用匹配的加分规则
将分数限制在0-100范围内

类别及其权重：

类别	权重	主要审计来源
过度设计检测	20%	audit-project（CLAUDE.md分析） + audit-ecosystem（钩子/MCP冗余）
CLAUDE.md质量	20%	audit-project（结构、章节、引用、多文件管理）
安全态势	15%	audit-project（权限配置） + audit-global（全局设置）
MCP配置	15%	audit-ecosystem（服务器健康状况、冗余）
插件健康度	15%	audit-ecosystem（插件结构） + audit-global（已安装插件）
上下文效率	15%	所有审计（token成本估算、指令文件总大小）

基于范围的评分调整：

仅全局审计：跳过CLAUDE.md质量评分（无项目可评估）。将剩余5个类别的权重按比例重新归一化：过度设计=20/80=25%，安全态势=15/80=18.75%，MCP配置=18.75%，插件健康度=18.75%，上下文效率=18.75%。在报告中注明“CLAUDE.md质量：已跳过（未检测到项目）”。
全面审计：正常对所有6个类别评分。

Compute Overall Score

计算总体评分

overall = sum(category_score * category_weight for all categories)

Look up the letter grade from the rubric's grade threshold table.

总体得分 = sum(类别得分 * 类别权重 for 所有类别)

根据评分规则中的等级阈值表查找对应的字母等级。

Build Recommendations

构建建议清单

Compile a ranked list of recommendations from all audit findings:

Critical (> 20 point impact): Must fix — actively harming performance
High (10-20 point impact): Should fix — significant improvement
Medium (5-9 point impact): Nice to have — incremental improvement
Low (< 5 point impact): Optional — minor polish

Include both:

Issues to fix — problems found in current config
Features to adopt — capabilities from Expert Context the user isn't using

从所有审计结果中整理出按优先级排序的建议列表：

严重（影响>20分）：必须修复——当前已对性能造成负面影响
高优先级（影响10-20分）：应该修复——可带来显著提升
中优先级（影响5-9分）：建议修复——可带来增量提升
低优先级（影响<5分）：可选——仅为微小优化

建议需包含两部分：

待修复问题——当前配置中发现的问题
可采纳功能——专家上下文中提到但用户尚未使用的功能

Present the Health Report

展示健康报告

Display the report header showing detected scope and file count:

╔══════════════════════════════════════════════════════════╗
║                  CLAUDIT HEALTH REPORT                  ║
╠══════════════════════════════════════════════════════════╣
║  Scope: Comprehensive | Files: N project + N global     ║
║  Overall Score: XX/100  Grade: X  (Label)               ║
╚══════════════════════════════════════════════════════════╝

Over-Engineering     ████████████████████░░░░░  XX/100  X
CLAUDE.md Quality    ████████████████████░░░░░  XX/100  X
Security Posture     ████████████████████░░░░░  XX/100  X
MCP Configuration    ████████████████████░░░░░  XX/100  X
Plugin Health        ████████████████████░░░░░  XX/100  X
Context Efficiency   ████████████████████░░░░░  XX/100  X

For the visual bars, use

█

for filled and

░

for empty. Scale to 25 characters total. Append the numeric score and letter grade.

After the score card, present:

Critical Issues — anything scoring below 50 in a category
Top Recommendations — ranked list with estimated point impact
New Features to Adopt — capabilities from Expert Context not currently used

显示报告头部，包含检测到的审计范围和文件数量：

╔══════════════════════════════════════════════════════════╗
║                  CLAUDIT健康报告                  ║
╠══════════════════════════════════════════════════════════╣
║  范围：全面审计 | 文件数：N个项目文件 + N个全局文件     ║
║  总体得分：XX/100  等级：X  (标签)               ║
╚══════════════════════════════════════════════════════════╝

过度设计检测     ████████████████████░░░░░  XX/100  X
CLAUDE.md质量    ████████████████████░░░░░  XX/100  X
安全态势     ████████████████████░░░░░  XX/100  X
MCP配置    ████████████████████░░░░░  XX/100  X
插件健康度        ████████████████████░░░░░  XX/100  X
上下文效率   ████████████████████░░░░░  XX/100  X

可视化进度条使用

█

表示已填充，

░

表示未填充。总长度为25个字符。末尾附上数字得分和字母等级。

在得分卡片之后，展示：

严重问题——任何类别得分低于50分的问题
首要建议——按优先级排序的列表，包含预估得分影响
可采纳新功能——专家上下文中提到但用户尚未使用的功能

Phase 4: Interactive Enhancement

阶段4：交互式优化

After presenting the report, offer to implement improvements.

展示报告后，提供优化实现选项。

Present Recommendations for Selection

展示可选建议供用户选择

Use AskUserQuestion with

multiSelect: true

to let the user choose which recommendations to apply. Group by priority (Critical, High, Medium, Low). Include the estimated score impact for each.

Format each option as:

Label: Short description (e.g., "Trim CLAUDE.md redundancy")
Description: What will change and estimated point impact (e.g., "Remove 5 restated built-in instructions. ~200 token savings. +15 pts Over-Engineering")

Include a "Skip — no changes" option.

使用AskUserQuestion并设置

multiSelect: true

，让用户选择要应用的建议。按优先级分组（严重、高、中、低）。每个选项需包含预估得分影响。

选项格式如下：

标签：简短描述（如“精简CLAUDE.md冗余内容”）
描述：将做出的更改及预估得分影响（如“移除5条重复的内置指令。约减少200个token。过度设计检测得分+15分”）

包含“跳过——不做任何更改”选项。

Implement Selected Fixes

实现用户选择的修复

For each selected recommendation:

Read the target file
Apply the fix using Write or Edit tools
Briefly explain what changed

Common fix types:

CLAUDE.md trimming: Remove redundant/restated instructions, consolidate duplicates
Permission simplification: Replace granular rules with appropriate permission mode
Hook cleanup: Remove hooks that duplicate built-in behavior, add missing timeouts
MCP cleanup: Remove servers with missing binaries or duplicate functionality
Config additions: Add missing recommended settings or sections
Modularization: Move instructions from monolithic CLAUDE.md to
```
.claude/rules/
```
or subdirectory files
Cross-scope cleanup: Remove project-specific instructions from personal config (apply directly, never via PR)
@import fixes: Remove broken imports, fix circular references

Scope safety for fixes:

Project-scoped files (CLAUDE.md, .claude/settings.json, .claude/rules/): eligible for direct edit and PR
```
CLAUDE.local.md
```
: edit directly, never include in PR (it's gitignored/personal)
```
.claude/settings.local.json
```
: edit directly, never include in PR (it's personal/local)
```
~/.claude/
```
files: edit directly, never include in PR (they're personal)

对于每个选中的建议：

读取目标文件
使用Write或Edit工具应用修复
简要说明更改内容

常见修复类型：

CLAUDE.md精简：移除冗余/重复的指令，合并重复内容
权限简化：使用合适的权限模式替代细粒度规则
钩子清理：移除与内置功能重复的钩子，添加缺失的超时配置
MCP清理：移除缺少二进制文件或功能重复的服务器
配置补充：添加缺失的推荐设置或章节
模块化重构：将单体CLAUDE.md中的指令迁移到
```
.claude/rules/
```
或子目录文件中
跨范围清理：从个人配置中移除项目特定指令（直接应用，绝不通过PR）
@import修复：移除损坏的导入，修复循环引用

修复的范围安全性：

项目范围文件（CLAUDE.md、.claude/settings.json、.claude/rules/）：可直接编辑并创建PR
```
CLAUDE.local.md
```
：直接编辑，绝不包含在PR中（已被Git忽略/属于个人文件）
```
.claude/settings.local.json
```
：直接编辑，绝不包含在PR中（属于个人本地设置）
```
~/.claude/
```
下的文件：直接编辑，绝不包含在PR中（属于个人配置）

Re-Score and Show Delta

重新评分并展示得分变化

After implementing fixes:

Re-score only the affected categories
Show before/after:

Score Delta:
  Over-Engineering     65 → 85  (+20)
  CLAUDE.md Quality    70 → 88  (+18)
  Overall              72 → 84  (+12)  Grade: C → B

应用修复后：

仅对受影响的类别重新评分
展示修复前后的得分对比：

得分变化：
  过度设计检测     65 → 85  (+20)
  CLAUDE.md质量    70 → 88  (+18)
  总体得分              72 → 84  (+12)  等级: C → B

Phase 5: PR Delivery

阶段5：PR交付

After Phase 4 fixes are applied, check if any project-scoped files were modified. If no project files were changed (only personal/global edits), skip this phase.

完成阶段4的修复后，检查是否修改了任何项目范围的文件。如果没有修改项目文件（仅修改了个人/全局配置），则跳过此阶段。

Offer PR Option

提供PR选项

Use

AskUserQuestion

(single-select) to ask the user:

"Open a PR" — Create branch, commit changes, push, open PR with educational inline comments
"Keep as local edits" — Leave changes uncommitted in the working tree

使用

AskUserQuestion

（单选）询问用户：

"创建PR" — 创建分支、提交更改、推送并创建带教育性行内注释的PR
"保留为本地修改" — 将更改保留在工作区中，不提交

Check Prerequisites

检查前置条件

Before attempting PR delivery:

Verify
```
gh
```
CLI is available:
```
command -v gh
```
Verify
```
gh
```
is authenticated:
```
gh auth status
```
If either fails, tell the user
```
gh
```
CLI is required for PR delivery and fall back to "Keep as local edits"

在尝试创建PR之前：

验证
```
gh
```
CLI是否可用：
```
command -v gh
```
验证
```
gh
```
是否已认证：
```
gh auth status
```
如果任一验证失败，告知用户创建PR需要
```
gh
```
CLI，并默认选择“保留为本地修改”

Create the PR

创建PR

If PR delivery is selected and prerequisites pass:

Record the current branch:
```
CURRENT_BRANCH=$(git rev-parse --abbrev-ref HEAD)
```
— this is the branch the user was on before claudit creates its own branch. The PR will target this branch so the diff only shows claudit changes.
Create branch:
```
git checkout -b claudit/improvements-YYYY-MM-DD-HHMM
```
(use today's date and current time to avoid same-day collisions)
Stage changed project files: Only stage project-scoped files that were modified in Phase 4. Never stage:
- ```
CLAUDE.local.md
```
  (gitignored/personal)
- ```
.claude/settings.local.json
```
  (personal local settings)
- Any file under
```
~/.claude/
```
  (personal config)
- Any file outside the project root

Commit with a clear message including the score delta:

claudit: improve Claude Code configuration (score XX → YY)

- [List key changes]

Push with

git push -u origin claudit/improvements-YYYY-MM-DD-HHMM

(same branch name as step 2)

Create PR via
```
gh pr create --base $CURRENT_BRANCH
```
:
- Title:
```
claudit: improve Claude Code configuration
```
- Body: Concise summary with score delta, list of changes, and note that personal/global config was audited separately (if comprehensive)
- The
```
--base
```
  flag ensures the PR targets the user's original branch, not the repo default — so the diff only contains claudit's changes

Add inline review comments via

gh api

for each changed file. Use this JSON structure:

bash

gh api repos/{owner}/{repo}/pulls/{pr_number}/comments \
  --method POST \
  --field commit_id="$(git rev-parse HEAD)" \
  --field path="path/to/file" \
  --field line=N \
  --field side="RIGHT" \
  --field body="**What changed:** Brief description

**Why it matters:** 1-2 sentences on impact

**Claude Code feature:** Feature name
**Docs:** https://docs.anthropic.com/en/docs/claude-code/relevant-page
**Score impact:** +N pts Category"

To determine the correct

line

value, run

git diff HEAD~1 -- path/to/file

and find line numbers within changed hunks. The

line

must be a line number in the new file version that falls within a diff hunk range. Target the most representative changed line per hunk. If line targeting fails (422 error), fall back to a general PR comment without line numbers using

gh pr comment

Add one comment per significant change. Keep comments concise and educational.

Return the PR URL to the user.

如果用户选择创建PR且前置条件满足：

记录当前分支：
```
CURRENT_BRANCH=$(git rev-parse --abbrev-ref HEAD)
```
— 这是用户在claudit创建自己的分支之前所在的分支。PR将以该分支为目标，这样差异对比仅显示claudit做出的更改。
创建分支：
```
git checkout -b claudit/improvements-YYYY-MM-DD-HHMM
```
（使用当前日期和时间命名分支，避免同日冲突）
暂存修改的项目文件：仅暂存阶段4中修改的项目范围文件。绝不要暂存：
- ```
CLAUDE.local.md
```
  （已被Git忽略/属于个人文件）
- ```
.claude/settings.local.json
```
  （个人本地设置）
- ```
~/.claude/
```
  下的任何文件（个人配置）
- 项目根目录外的任何文件

提交：使用清晰的提交信息，包含得分变化：

claudit: 优化Claude Code配置（得分XX → YY）

- [列出主要更改]

推送：执行

git push -u origin claudit/improvements-YYYY-MM-DD-HHMM

（与步骤2中的分支名称一致）

创建PR：通过
```
gh pr create --base $CURRENT_BRANCH
```
：
- 标题：
```
claudit: 优化Claude Code配置
```
- 正文：包含得分变化、更改列表的简洁摘要，并注明个人/全局配置已单独审计（如果是全面审计）
- ```
--base
```
  参数确保PR以用户的原始分支为目标，而非仓库默认分支——这样差异对比仅包含claudit做出的更改

添加行内评审注释：通过

gh api

为每个修改的文件添加注释。使用以下JSON结构：

bash

gh api repos/{owner}/{repo}/pulls/{pr_number}/comments \
  --method POST \
  --field commit_id="$(git rev-parse HEAD)" \
  --field path="path/to/file" \
  --field line=N \
  --field side="RIGHT" \
  --field body="**更改内容:** 简要描述

**重要性:** 1-2句话说明影响

**Claude Code功能:** 功能名称
**文档:** https://docs.anthropic.com/en/docs/claude-code/relevant-page
**得分影响:** +N分 类别"

为确定正确的

line

值，执行

git diff HEAD~1 -- path/to/file

并找到更改块中的行号。

line

必须是新文件版本中属于差异块范围内的行号。针对每个差异块选择最具代表性的更改行。如果行号定位失败（返回422错误），则退化为使用

gh pr comment

添加不带行号的通用PR注释。

每个重要更改添加一条注释。注释需简洁且具有教育意义。

将PR URL返回给用户。

Fallback

降级方案

gh

is not available, not authenticated, or PR creation fails:

Tell the user what happened
Fall back to "Keep as local edits"
Show a
```
git diff --stat
```
of what was changed

如果

gh

不可用、未认证或PR创建失败：

告知用户具体问题
退化为“保留为本地修改”
展示
```
git diff --stat
```
输出的更改摘要

Error Handling

错误处理

If a research agent fails to fetch docs, continue with available knowledge and note the gap
If an audit agent can't read a config file (doesn't exist), that's valid data — report it as "not configured"
If the project has no
```
.claude/
```
directory at all, focus the audit on global config and recommend project-level setup
If no issues are found (score 90+), congratulate the user and suggest any new features to explore
If Glob returns too many files (>50), cap and note the truncation

如果研究代理无法获取文档，使用已有知识继续并注明信息缺口
如果审计代理无法读取配置文件（文件不存在），这是有效数据——报告为“未配置”
如果项目完全没有
```
.claude/
```
目录，将审计重点放在全局配置上，并建议用户设置项目级配置
如果未发现任何问题（得分90+），祝贺用户并建议探索新功能
如果Glob返回过多文件（>50），截断并注明已截断

Important Notes

重要注意事项

Never auto-apply changes — always present recommendations and let the user choose
Quote specific lines when showing what would change in instruction files
Be opinionated about over-engineering — this is the plugin's core value proposition
Show token savings whenever removing content from instruction files or other config
The Expert Context makes this audit unique — always highlight features the user isn't using yet
Respect scope boundaries — project config is the team contract; personal config is personal
Only project-scoped files go in PRs — CLAUDE.local.md and ~/.claude/ changes are always local-only

绝不自动应用更改——始终先展示建议，让用户选择
在展示将更改的内容时，引用具体行号
对过度设计问题要明确立场——这是本插件的核心价值主张
每当移除指令文件或其他配置中的内容时，展示token节省量
专家上下文是本审计的独特之处——始终突出用户尚未使用的功能
尊重范围边界——项目配置是团队约定；个人配置属于个人
只有项目范围的文件可纳入PR——CLAUDE.local.md和~/.claude/下的更改始终仅保留在本地