Claude Code 最厉害的工具。一行代码都不写

我当时觉得这可能是模型性格差异——有的模型比较谨慎，有的比较奔放。但后来我把几个主流 Coding Agent的工具定义全拆了一遍。才发现这不是性格问题，是工具设计问题。

Claude Code 的工具集里藏着一批不写代码的工具——它们不改文件、不跑命令，做的事情是约束 Agent 自己的行为。有的让 Agent 在执行前必须停下来出方案。有的强制它改代码前必须先引用旧内容，有的给无所不能的 Shell 套上权限锁链。

这批工具，才是 Claude Code 和其他 Agent 拉开差距的地方。

至于两类工具。做事的 vs 管事的

Claude Code 的工具列表按功能分，就是一堆"读、写、执行"。但换一个角度——按作用对象分——会看到完全不同结构：

Pi only has a handful of tools,all on left side – giving model free play.Popencode writes "You MUST iterate and keep going until problem is solved" in its prompt – no pause mechanism.PClaude gives model brakes: some actions must be read-first,some require approval.

一、PlanMode 不是"建议"。是断路器

The tool CreatePlanMode: does nothing but put agent into a read‑only state until it produces a plan that is approved by a human.

• User Pain Point: When reviewing code diffs that are huge and messy,it's hard to see what changed at all.
• The prompt “plan first” is just advice;once context grows too long or tasks get complex it gets ignored.
• A Tool call like CreatePlanMode,however,is explicit and can be intercepted by harness – no execution until approval.
• This design leaves *when* to decide up to model but *how* to enforce caution up to harness.
• No or coding agent harness implements such an explicit circuit breaker.

二、Edit 的 old_string：不是查找替换。而是安全机制

• Model Pain Point: A model may generate an entire new file with only one line of change—leading to massive diffs that are impossible to review cleanly.
• Design Insight: Edit requires you provide both old and new strings – forcing a read before any write.
• Result: If old_string mismatches,Edit aborts instead of silently overwriting—preventing accidental data loss.
• Why not Bash/>shell sed?:  - Bash can run any command without knowing context. Claude forces Edit first so it must read before writing—this is an optimistic lock on content.
• Comparison:  - Pi trusts model with bulk edits .  - Codex allows arbitrary patch text .  - Claude demands explicit read → reference → edit chain—most conservative approach.

  三、Bash：能力无限，所以必须受限

  • •User Pain Point:  Without Bash you cannot run tests or invoke build tools—essential for any realistic coding workflow.
  • •Bash gives full OS access but needs constraints!
  • Categorized commands: common read commands auto‑allowed while write/unknown ones trigger approval.
  The safety net relies on model judgment + user vigilance + harness enforcement—no single layer is enough alone. Thus Claude offers a powerful yet controlled shell while still allowing deep system interactions when necessary.

  工具设计救不了的问题

  You might think 娱乐ter tools = controllable agents—but y only enforce structure . They can’t fix poor decision making inside those constraints.

These are limits of reasoning capability rar than tool design.

The bottom line:A robust tool architecture sets an upper bound on control。but actual reliability depends on how well underlying model follows those constraints.

完整工具速查表

层次	工具名称	主要功能
感知	Glob,Grep,Read,LSP	按方法/内容/语义检索，无副作用
修改	Edit,NotebookEdit,Bash,Monitor	修改文件、输入命令、监听进程控制
约束Agent行为	AskUserQuestion,EnterPlanMode,ExitPlanMode,Skill,Agent…Task，TodoWrite,…话说回来，TaskStop/Output…子Agent委托等隔离与管理
工作区与资源隔离	EnterWorktree,ExitWorktree…SendMessage,ListAgents…Artifact等外部交付和审计通道…Cron/ReportFindings等定时与报告功能…EndConversation 等会话控制 …，​ ​​ ​ ​​ ​          ​

参考资料与进一步阅读：

• Mega-Reference Guide of Claude Code Tools – 官方文档链接

LI> LII) GitHub Repository of ClaudeCode - 包含所有示例脚本及其使用方式

LIII) Phistory — Coding Agent Prompt Snapshots – 查看不同版本对比如何演变。

• ) Tech Blog Post from Anthropic about Harness Safeguards —
• ) OpenAI Cookbook for Coding Agents – 比较不同网站对比实现细节。