Claude Code自主智能体：从/goal到守护进程全实战

graph TB
    subgraph 自主运行三层次
        G["/goal<br/>目标驱动层<br/>定义完成条件"]
        M["Monitor工具<br/>事件驱动层<br/>实时反应变化"]
        D["守护进程<br/>持久运行层<br/>跨会话存活"]
    end

    G -->|"每轮评估条件"| M
    M -->|"事件触发动作"| D
    D -->|"后台持续执行"| G

    style G fill:#4A90D9,color:#fff
    style M fill:#50C878,color:#fff
    style D fill:#FF6B6B,color:#fff

sequenceDiagram
    participant U as 用户
    participant C as Claude Code
    participant E as 评估模型(Haiku)

    U->>C: /goal all tests pass and lint is clean
    C->>C: 开始第一轮工作
    C->>E: 条件满足了吗？
    E-->>C: 否，还有3个测试失败
    C->>C: 开始第二轮（修复测试）
    C->>E: 条件满足了吗？
    E-->>C: 否，还有1个lint错误
    C->>C: 开始第三轮（修复lint）
    C->>E: 条件满足了吗？
    E-->>C: 是的！所有测试通过，lint干净
    C-->>U: ◎ goal achieved

# 在交互模式中设定目标
claude
> /goal all tests in test/auth pass and the lint step is clean

# 查看当前目标状态
> /goal
# 输出：条件、运行时长、已执行轮数、Token消耗、评估模型的最近判断

# 清除目标（提前终止）
> /goal clear

claude -p "Fix all failing tests in the auth module" \
  --goal "all tests in test/auth pass and the lint step is clean" \
  --allowedTools "Bash,Read,Edit,Glob,Grep"

# 好的条件：可衡量、可验证、有约束
/goal npm test exits 0 and no other test file is modified, stop after 20 turns

# 差的条件：模糊、无法验证
/goal make the code better

claude --autoacceptEdits \
  --append-system-prompt "You are fixing CI failures for a Node.js project. Focus on test failures and lint errors." \
  -p "Check the current CI status with 'gh run list --limit 1' and fix any failures"

> /autoaccept
> /goal gh run list --limit 1 shows the latest run as "success", stop after 15 turns

flowchart LR
    subgraph 后台进程
        P1["npm run build"]
    end

    P1 -->|"stdout逐行"| M["Monitor工具"]
    M -->|"ERROR行触发"| C["Claude反应"]
    M -->|"成功行"| C2["Claude记录"]

    style M fill:#50C878,color:#fff
    style C fill:#FF6B6B,color:#fff

> Watch the build output. If any errors appear, fix them automatically.

Claude会使用Monitor工具：
- 启动 npm run build 作为后台进程
- 监视输出流
- 发现ERROR行时自动诊断并修复
- 不需要你持续盯着终端

> Run the test suite and monitor it. Fix any failing tests you find.

> /loop Monitor the deployment: run 'kubectl rollout status deployment/web-app' and 
> if it reports an error, roll back and investigate.

graph TB
    subgraph 传统模式
        T1["终端1<br/>Claude Code"] -->|"关闭终端"| X["进程终止"]
    end

    subgraph 守护进程模式
        D["守护进程<br/>本地Socket监听"]
        T2["终端1"] -->|"连接"| D
        T3["终端2"] -->|"连接"| D
        VS["VS Code"] -->|"连接"| D
        RC["手机Remote Control"] -->|"连接"| D
        D -->|"独立运行"| W["后台持续工作"]
    end

    style D fill:#FF6B6B,color:#fff
    style W fill:#50C878,color:#fff

graph LR
    subgraph Agent Fleet
        CO["协调者<br/>Coordinator"]
        A1["代码审查代理"]
        A2["测试代理"]
        A3["文档代理"]
    end

    subgraph CoWork Memory
        M1["共享发现"]
        M2["任务状态"]
        M3["上下文快照"]
    end

    CO --> A1
    CO --> A2
    CO --> A3
    A1 --> M1
    A2 --> M2
    A3 --> M3
    M1 --> A2
    M2 --> A3

# 方法一：通过环境变量启动
export CLAUDE_BG_BACKEND=daemon
claude

# 方法二：在settings.json中配置
# ~/.claude/settings.json
{
  "env": {
    "CLAUDE_BG_BACKEND": "daemon"
  }
}

# 终端1：启动守护进程并开始工作
claude
> /goal all tests pass

# 终端2：attach到同一个会话（使用--resume）
claude --resume
# 选择正在运行的会话

# 手机上：通过Remote Control发送指令
# Claude Code会在守护进程上继续执行

flowchart TB
    subgraph 触发层
        WH["Webhook/GitHub Actions"]
        CH["Channel通知"]
        SCH["定时调度"]
    end

    subgraph 执行层
        G["/goal<br/>定义完成条件"]
        AU["Auto模式<br/>自动批准工具"]
        MO["Monitor<br/>事件驱动反应"]
    end

    subgraph 持久层
        DA["守护进程<br/>跨会话存活"]
        FL["Agent Fleet<br/>多代理协作"]
        CM["CoWork Memory<br/>共享记忆"]
    end

    WH --> G
    CH --> G
    SCH --> G
    G --> AU
    AU --> MO
    MO --> DA
    DA --> FL
    FL --> CM

    style G fill:#4A90D9,color:#fff
    style MO fill:#50C878,color:#fff
    style DA fill:#FF6B6B,color:#fff

mkdir ~/ci-auto-fixer && cd ~/ci-auto-fixer

{
  "permissions": {
    "allow": [
      "Bash(git *)",
      "Bash(npm *)",
      "Bash(npx *)",
      "Bash(gh run *)",
      "Bash(gh pr *)",
      "Bash(node *)",
      "Read",
      "Edit",
      "Glob",
      "Grep"
    ]
  },
  "hooks": {
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "echo 'Session ended at $(date)' >> ~/ci-auto-fixer.log"
          }
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Edit",
        "hooks": [
          {
            "type": "command",
            "command": "echo '[EDIT] $(jq -r '.tool_input.file_path') at $(date)' >> ~/ci-auto-fixer.log"
          }
        ]
      }
    ]
  }
}

# CI Auto-Fixer 巡检指令

检查当前分支的CI状态：
1. 运行 `gh run list --limit 1 --json status,conclusion,name` 检查最新运行
2. 如果状态是 "failure"：
   a. 获取失败日志：`gh run view --log-failed`
   b. 分析根因
   c. 修复代码
   d. 运行本地测试验证：`npm test`
   e. 如果本地通过，提交并推送
3. 如果状态是 "success"：
   检查是否有新的review comments需要处理
4. 如果没有进行中的工作，报告一行状态

#!/bin/bash
# start-fixer.sh - CI自动修复启动脚本

BRANCH=${1:-$(git branch --show-current)}
LOG_FILE=~/ci-auto-fixer.log

echo "=== CI Auto-Fixer Started ===" >> "$LOG_FILE"
echo "Branch: $BRANCH" >> "$LOG_FILE"
echo "Time: $(date)" >> "$LOG_FILE"

# 启动守护进程模式 + /loop
claude --bare \
  -p "You are a CI auto-fixer agent. Monitor CI for branch $BRANCH." \
  --append-system-prompt "Always check CI status first. Fix errors methodically. Run tests locally before pushing." \
  --allowedTools "Bash(gh *),Bash(git *),Bash(npm *),Bash(npx *),Bash(node *),Read,Edit,Glob,Grep" \
  --permission-mode acceptEdits \
  --output-format json

# 安装Telegram插件
claude
> /plugin install telegram@claude-plugins-official
> /reload-plugins
> /telegram:configure YOUR_BOT_TOKEN

# 启动时开启通道
claude --channels plugin:telegram@claude-plugins-official
> /loop 10m Check CI status and fix failures. Send a Telegram message when done.

# 目标：将所有Class组件迁移为函数组件
claude --bare \
  -p "Migrate all React class components to functional components in src/components/" \
  --goal "all files in src/components/ are functional components (no 'extends Component'), npm test exits 0, and eslint reports no errors. Stop after 50 turns." \
  --permission-mode acceptEdits \
  --allowedTools "Bash,Read,Edit,Write,Glob,Grep" \
  --output-format stream-json \
  --verbose

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "if": "Bash(rm -rf *)",
            "command": ".claude/hooks/block-destructive.sh"
          }
        ]
      }
    ]
  }
}

#!/bin/bash
# .claude/hooks/block-destructive.sh
COMMAND=$(jq -r '.tool_input.command')

if echo "$COMMAND" | grep -qE 'rm -rf|git push --force|git reset --hard'; then
  jq -n '{
    hookSpecificOutput: {
      hookEventName: "PreToolUse",
      permissionDecision: "deny",
      permissionDecisionReason: "Destructive command blocked by safety hook"
    }
  }'
else
  exit 0
fi

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "FILE=$(jq -r '.tool_input.file_path'); npx prettier --write \"$FILE\" 2>/dev/null; exit 0"
          }
        ]
      }
    ]
  }
}

> /goal All tests pass and the build succeeds. Stop after 30 turns.
> 
> Also monitor 'npm run build' in the background. If you see any errors 
> in the output, fix them immediately.

# 步骤1：启动守护进程+Channel
claude --channels plugin:telegram@claude-plugins-official

# 步骤2：设定目标
> /autoaccept
> /goal gh run list --limit 1 shows status "completed" and conclusion "success"

# 步骤3：Claude自主运行，你关闭终端
# 当目标达成或遇到问题，通过Telegram通知你

> I need to fix the failing CI. Use the explore agent to analyze the test failures,
> the plan agent to design fixes, and then implement the fixes yourself.
> Set a goal: all tests pass, stop after 20 turns.

# 周末代码健康巡检

按以下顺序执行：

1. 检查当前分支CI状态。如果失败，诊断并修复
2. 运行 `npm run lint` 修复可自动修复的lint问题
3. 搜索代码中的 TODO 和 FIXME，创建issue
4. 检查依赖是否有安全漏洞：`npm audit`
5. 如果有中高风险漏洞，尝试升级对应依赖
6. 每完成一步，记录到 ~/health-patrol.log
7. 完成所有步骤后，总结一行报告

# 周五下班前启动
nohup claude --bare \
  -p "Run the weekend code health patrol for the main branch" \
  --append-system-prompt "Work methodically through each step. Commit fixes separately. Never force-push." \
  --permission-mode acceptEdits \
  --allowedTools "Bash(git *),Bash(npm *),Bash(npx *),Bash(gh *),Read,Edit,Write,Glob,Grep" \
  --output-format json \
  > ~/health-patrol-output.json 2>&1 &

# 编辑 crontab
crontab -e

# 添加
0 2 * * * cd /path/to/project && claude --bare -p "Run the daily health patrol" --permission-mode acceptEdits --allowedTools "Bash(git *),Bash(npm *),Bash(npx *),Bash(gh *),Read,Edit,Write,Glob,Grep" --output-format json >> ~/daily-patrol.json 2>&1

> /tasks
# 显示所有后台代理的ID、状态、Token使用

> /goal
# 显示当前目标的条件、运行时长、已执行轮数、评估结果

# 策略一：限制轮数
/goal ... stop after 15 turns

# 策略二：使用--model指定低成本模型
claude --model haiku -p "Quick lint fix" --goal "..."

# 策略三：Bare模式减少上下文开销
claude --bare -p "..."  # 跳过hooks/skills/plugins/MCP的加载

# 策略四：输出JSON跟踪成本
claude -p "..." --output-format json | jq '.total_cost_usd'

flowchart TD
    A["自主运行启动"] --> B{"PreToolUse Hook"}
    B -->|"阻止破坏性命令"| C["安全检查通过"]
    B -->|"允许"| D["执行操作"]
    C --> D
    D --> E{"PostToolUse Hook"}
    E -->|"记录审计日志"| F["继续"]
    E -->|"格式化代码"| F
    F --> G{"Goal评估"}
    G -->|"条件满足"| H["停止 ✓"]
    G -->|"超过轮数限制"| H
    G -->|"继续"| A

flowchart TB
    START["启动自主智能体"] --> CONFIG["加载配置<br/>settings.json + CLAUDE.md"]
    CONFIG --> MODE{"运行模式？"}

    MODE -->|"/goal"| GOAL["设定完成条件<br/>评估模型就绪"]
    MODE -->|"/loop"| LOOP["设定调度间隔<br/>加载loop.md"]
    MODE -->|"守护进程"| DAEMON["启动Socket监听<br/>Agent Fleet就绪"]

    GOAL --> WORK["执行一轮工作"]
    LOOP --> WORK
    DAEMON --> WORK

    WORK --> TOOLS["调用工具<br/>Read/Edit/Bash/Grep..."]
    TOOLS --> HOOK1{"PreToolUse Hook"}
    HOOK1 -->|"允许"| EXEC["执行"]
    HOOK1 -->|"阻止"| BLOCK["记录阻止原因"]
    EXEC --> HOOK2["PostToolUse Hook<br/>审计 + 格式化"]
    BLOCK --> EVAL
    HOOK2 --> EVAL{"评估条件"}

    EVAL -->|"条件未满足"| WORK
    EVAL -->|"条件满足"| DONE["◎ Goal Achieved"]
    EVAL -->|"超过轮数"| DONE
    EVAL -->|"时间间隔到"| WAIT["等待下一轮"]
    WAIT --> WORK

    DONE --> NOTIFY["Channel通知<br/>Telegram/Discord"]
    NOTIFY --> LOG["记录日志<br/>审计回放"]

    style START fill:#4A90D9,color:#fff
    style DONE fill:#50C878,color:#fff
    style BLOCK fill:#FF6B6B,color:#fff
    style HOOK1 fill:#FFA500,color:#fff