Claude Code权限六模式：安全到极速全实战

严格 ←————————————————————————→ 自由
default → acceptEdits → plan → auto → dontAsk → bypassPermissions

default → acceptEdits → plan → auto → dontAsk → (循环回 default)

claude                          # 默认 default 模式
claude --permission-mode auto   # 直接启动 auto 模式
claude -p "任务" --permission-mode dontAsk  # 非交互模式指定权限

Claude wants to edit src/index.ts:
  ┌─────────────────────────────────────┐
  │  Allow Edit to src/index.ts? [y/n]  │
  └─────────────────────────────────────┘

# 方式1：Shift+Tab切换到acceptEdits
# 方式2：命令行参数
claude --permission-mode acceptEdits

claude --permission-mode auto

# CI/CD中的典型用法
claude -p "运行测试并修复所有失败的用例" --permission-mode dontAsk

{
  "deny": [
    "Bash(rm -rf *)",
    "Bash(git push *)",
    "Bash(git reset --hard *)",
    "Read(.env*)",
    "Edit(.env*)"
  ]
}

claude --dangerously-skip-permissions

# Docker中的安全用法
docker run -it --rm \
  -v $(pwd):/workspace \
  node:20 \
  bash -c "npm install -g @anthropic-ai/claude-code && \
           cd /workspace && \
           claude --dangerously-skip-permissions -p '任务描述'"

企业管理策略 (managed)
    ↓ 覆盖
用户全局配置 (~/.claude/settings.json)
    ↓ 覆盖
项目团队配置 (.claude/settings.json)
    ↓ 覆盖
项目本地配置 (.claude/settings.local.json)
    ↓ 覆盖
命令行参数 (--allow / --deny)

{
  "allow": [
    "Bash(npm test)",
    "Bash(npm run *)",
    "Bash(git status)",
    "Bash(git diff *)",
    "Bash(git log *)",
    "Bash(ls *)",
    "Bash(cat *)",
    "Bash(python *)",
    "Bash(make *)"
  ],
  "deny": [
    "Bash(rm -rf *)",
    "Bash(git push --force *)",
    "Bash(git reset --hard *)",
    "Bash(sudo *)",
    "Bash(curl * | bash)",
    "Bash(chmod 777 *)"
  ]
}

{
  "allow": [
    "Edit(src/**)",
    "Edit(tests/**)",
    "Read(**)"
  ],
  "deny": [
    "Edit(.env*)",
    "Edit(*.secret*)",
    "Edit(production.*)",
    "Write(/etc/*)",
    "Read(/etc/shadow)"
  ]
}

1. 检查 deny 规则 → 匹配则拒绝
2. 检查 allow 规则 → 匹配则通过
3. 都不匹配 → 根据当前权限模式决定（提示/通过/拒绝）

flowchart TD
    A[Claude Code请求操作] --> B{匹配deny规则?}
    B -->|是| C[拒绝执行]
    B -->|否| D{匹配allow规则?}
    D -->|是| E[自动通过]
    D -->|否| F{当前权限模式?}
    F -->|default/plan| G[弹出确认提示]
    F -->|acceptEdits| H[编辑自动通过<br/>命令弹出提示]
    F -->|auto| I[AI分类器判断]
    F -->|dontAsk| J[自动通过]
    F -->|bypassPermissions| K[跳过所有检查]

{
  "allow": [
    "Bash(npm *)",
    "Bash(npx *)",
    "Bash(git status)",
    "Bash(git diff *)",
    "Bash(git add *)",
    "Bash(git commit *)",
    "Bash(git log *)",
    "Bash(node *)",
    "Bash(tsc *)",
    "Bash(eslint *)",
    "Bash(prettier *)",
    "Edit(src/**)",
    "Edit(tests/**)",
    "Edit(docs/**)",
    "Write(src/**)",
    "Write(tests/**)"
  ],
  "deny": [
    "Bash(rm -rf *)",
    "Bash(git push --force *)",
    "Bash(sudo *)",
    "Bash(npm publish *)",
    "Edit(package.json)",
    "Edit(tsconfig.json)",
    "Edit(.github/**)",
    "Edit(.env*)"
  ]
}

claude --permission-mode acceptEdits

{
  "allow": [
    "Bash(npm run lint)",
    "Bash(npm run test*)",
    "Bash(npm run build*)",
    "Bash(npm run format)",
    "Bash(git status)",
    "Bash(git diff *)",
    "Bash(git log *)",
    "Bash(git branch *)",
    "Bash(git checkout *)",
    "Bash(npx jest *)",
    "Bash(npx tsc --noEmit)",
    "Edit(src/**)",
    "Edit(tests/**)",
    "Edit(*.md)",
    "Read(**)"
  ],
  "deny": [
    "Bash(rm *)",
    "Bash(git push *)",
    "Bash(git merge *)",
    "Bash(git rebase *)",
    "Bash(sudo *)",
    "Bash(docker *)",
    "Bash(kubectl *)",
    "Edit(.env*)",
    "Edit(terraform/**)",
    "Edit(.github/workflows/**)",
    "Edit(docker-compose.*)",
    "Edit(Dockerfile*)",
    "Edit(package.json)",
    "Edit(package-lock.json)",
    "Edit(.claude/**)",
    "Write(.env*)",
    "Write(.github/**)"
  ]
}

{
  "allow": [
    "Bash(npm run dev*)",
    "Bash(docker compose up*)",
    "Bash(git push origin *)"
  ]
}

flowchart LR
    subgraph 团队共享配置
        A[.claude/settings.json]
    end
    subgraph 个人本地配置
        B[.claude/settings.local.json]
    end
    subgraph 运行时
        C[default模式<br/>+ allow/deny引擎]
    end
    A --> C
    B --> C
    C --> D{Claude Code请求操作}
    D --> E[deny规则检查]
    E --> F[allow规则检查]
    F --> G[模式行为决策]

name: Claude Auto Fix
on:
  issues:
    types: [labeled]
  workflow_dispatch:

jobs:
  auto-fix:
    runs-on: ubuntu-latest
    permissions:
      contents: write
      pull-requests: write

    steps:
      - uses: actions/checkout@v4
        with:
          token: ${{ secrets.GITHUB_TOKEN }}

      - name: Setup Claude Code
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: npm install -g @anthropic-ai/claude-code

      - name: Run Claude Code
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          claude -p "读取issue #${{ github.event.issue.number }}的描述，修复相关问题，创建新分支并提交PR" \
            --permission-mode dontAsk \
            --allowedTools "Edit(src/**)" "Edit(tests/**)" "Bash(npm run *)" "Bash(git *)" "Read(**)" \
            --deny "Bash(git push --force *)" "Bash(rm *)"

{
  "hooks": {
    "preToolUse": [
      {
        "matcher": "Bash",
        "command": "python3 scripts/check-dangerous-commands.py"
      }
    ],
    "postToolUse": [
      {
        "matcher": "Edit",
        "command": "python3 scripts/log-edits.py"
      }
    ]
  }
}

flowchart TD
    A[操作请求] --> B{deny规则匹配?}
    B -->|是| C[拒绝]
    B -->|否| D{allow规则匹配?}
    D -->|是| E{preToolUse Hook}
    E -->|通过| F[执行操作]
    E -->|阻止| G[Hook拒绝]
    D -->|否| H{权限模式决策}
    H --> I{preToolUse Hook}
    I -->|通过| F
    I -->|阻止| G
    F --> J[postToolUse Hook]
    J --> K[记录日志/后处理]

# 允许特定操作
claude --allow "Bash(npm test)" --allow "Edit(src/**)"

# 拒绝特定操作
claude --deny "Bash(rm *)" --deny "Edit(.env*)"

# 组合使用
claude --permission-mode acceptEdits \
  --allow "Bash(npm run *)" \
  --deny "Bash(npm publish)" \
  --deny "Edit(.env*)"

claude -p "修复所有TypeScript编译错误" \
  --permission-mode dontAsk \
  --output-format stream-json \
  --deny "Bash(rm *)" "Bash(git push --force *)"

flowchart TD
    Start[开始Claude Code会话] --> Q1{项目类型?}
    Q1 -->|个人项目/学习| Q2{是否熟悉Claude Code?}
    Q1 -->|团队项目| Q3{是否CI/CD环境?}
    Q1 -->|生产环境/基础设施| Plan[plan模式<br/>+ 严格deny规则]

    Q2 -->|不熟悉| Default[default模式<br/>观察AI行为]
    Q2 -->|熟悉| AcceptEdits[acceptEdits模式<br/>+ 基础deny规则]

    Q3 -->|是| DontAsk[dontAsk模式<br/>+ 精确allow规则<br/>+ 完整deny规则]
    Q3 -->|否| Q4{信任等级?}
    Q4 -->|高信任| Auto[auto模式<br/>+ 关键deny规则]
    Q4 -->|低信任| Default2[default模式<br/>+ 完整deny规则]

    Default --> Config1[配置deny规则:<br/>rm -rf, git push --force, sudo]
    AcceptEdits --> Config2[配置deny规则:<br/>rm -rf, .env*, package.json]
    DontAsk --> Config3[配置allow规则:<br/>仅允许必要操作]
    Plan --> Config4[配置deny规则:<br/>所有写操作和Shell命令]
    Auto --> Config5[配置deny规则:<br/>生产环境文件, 敏感操作]
    Default2 --> Config6[配置完整deny列表]

    Config1 --> Run[开始编码]
    Config2 --> Run
    Config3 --> Run
    Config4 --> Run
    Config5 --> Run
    Config6 --> Run