mirror of
https://github.com/musistudio/claude-code-router.git
synced 2026-02-02 06:40:50 +08:00
151 lines
9.5 KiB
Markdown
151 lines
9.5 KiB
Markdown
# 从CLI工具风格看Agent工具渐进式披露
|
||
|
||
距离Anthropic发布[Skills](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills)也过去将近两个月的时间了，其中Anthropic提到了一个术语渐进式披露(Progressive Disclosure)，这到底是什么东西？解决了什么问题？
|
||
|
||
其实在我的Vibe Coding流程中，我很少使用MCP。因为我觉得MCP实现质量层次不齐，本质是上下文注入(工具的本质也是上下文注入)，我不确定别人写的提示词会不会影响到我的工作流，干脆直接不用。现在的MCP实现基本上就是把所有的功能全都包装成工具暴露给Agent(一个功能包装成一个工具，给定详细的描述，告诉agent在什么时候进行调用，参数格式是什么)，这就导致了现在的提示词爆炸。
|
||
|
||
直到Anthropic发布了Skills，研究了一下发现本质仍然是提示词注入。如果说MCP是提供了一套注入工具的规范，那么Skills所提倡的则是“离经叛道”。Skills给了一个Markdown文档用于描述该skill的用途和最佳用法，附带提供了一些脚本(与MCP不同)。
|
||
|
||
由于这些脚本直接在用户本地运行，存在极大的安全风险。如果用户不能对脚本代码进行review，很容易造成数据泄露、感染病毒等严重安全性问题。相比于MCP提供一个标准化的接口，Skill提供一系列的脚本文件，不同的skill可能拥有不同类型的脚本文件，比如有些脚本使用node.js实现，有些脚本使用Python实现，要使用这些脚本还需要用户安装对应的运行时和脚本所需要的依赖。这也是我说“离经叛道”的原因所在。
|
||
|
||
这真的是最佳实践吗？
|
||
|
||
关于渐进式披露，Anthropic是这样描述的：
|
||
> 渐进式披露是使代理技能灵活且可扩展的核心设计原则。就像一本组织良好的手册，从目录开始，然后是具体章节，最后是详细的附录一样，技能允许 Claude 仅在需要时加载信息：
|
||
> 
|
||
> 拥有文件系统和代码执行工具的智能体在执行特定任务时，无需将技能的全部内容读取到上下文窗口中。这意味着技能中可以包含的上下文信息量实际上是无限的。
|
||
|
||
下图是使用Skill的上下文窗口变化
|
||
|
||
|
||
我们真的需要这样去实现吗？
|
||
|
||
在我们平时使用CLI工具时，一般的CLI工具都会带有一个`--help`参数，用于查看该工具的用法和说明，这不就是该工具的使用手册吗？比如：
|
||
```shell
|
||
> npm --help
|
||
npm <command>
|
||
|
||
Usage:
|
||
|
||
npm install install all the dependencies in your project
|
||
npm install <foo> add the <foo> dependency to your project
|
||
npm test run this project's tests
|
||
npm run <foo> run the script named <foo>
|
||
npm <command> -h quick help on <command>
|
||
npm -l display usage info for all commands
|
||
npm help <term> search for help on <term>
|
||
npm help npm more involved overview
|
||
|
||
All commands:
|
||
|
||
access, adduser, audit, bugs, cache, ci, completion,
|
||
config, dedupe, deprecate, diff, dist-tag, docs, doctor,
|
||
edit, exec, explain, explore, find-dupes, fund, get, help,
|
||
help-search, hook, init, install, install-ci-test,
|
||
install-test, link, ll, login, logout, ls, org, outdated,
|
||
owner, pack, ping, pkg, prefix, profile, prune, publish,
|
||
query, rebuild, repo, restart, root, run-script, sbom,
|
||
search, set, shrinkwrap, star, stars, start, stop, team,
|
||
test, token, uninstall, unpublish, unstar, update, version,
|
||
view, whoami
|
||
|
||
Specify configs in the ini-formatted file:
|
||
/Users/xxx/.npmrc
|
||
or on the command line via: npm <command> --key=value
|
||
|
||
More configuration info: npm help config
|
||
Configuration fields: npm help 7 config
|
||
```
|
||
|
||
这份手册也不会返回所有的命令所有的用法，它只会返回它有哪些命令可以实现什么功能，对于命令的具体用法你仍然可以通过`--help`参数获得：
|
||
```shell
|
||
> npm install --help
|
||
Install a package
|
||
|
||
Usage:
|
||
npm install [<package-spec> ...]
|
||
|
||
Options:
|
||
[-S|--save|--no-save|--save-prod|--save-dev|--save-optional|--save-peer|--save-bundle]
|
||
[-E|--save-exact] [-g|--global]
|
||
[--install-strategy <hoisted|nested|shallow|linked>] [--legacy-bundling]
|
||
[--global-style] [--omit <dev|optional|peer> [--omit <dev|optional|peer> ...]]
|
||
[--include <prod|dev|optional|peer> [--include <prod|dev|optional|peer> ...]]
|
||
[--strict-peer-deps] [--prefer-dedupe] [--no-package-lock] [--package-lock-only]
|
||
[--foreground-scripts] [--ignore-scripts] [--no-audit] [--no-bin-links]
|
||
[--no-fund] [--dry-run] [--cpu <cpu>] [--os <os>] [--libc <libc>]
|
||
[-w|--workspace <workspace-name> [-w|--workspace <workspace-name> ...]]
|
||
[-ws|--workspaces] [--include-workspace-root] [--install-links]
|
||
|
||
aliases: add, i, in, ins, inst, insta, instal, isnt, isnta, isntal, isntall
|
||
|
||
Run "npm help install" for more info
|
||
```
|
||
这是不是很像上面的渐进式披露的定义？
|
||
|
||
我们是不是可以按照这种风格去实现一个MCP来实现无需skill的工具渐进式披露？我使用Codex将官方的PDF Skill转换成了一个MCP，只暴露一个工具：
|
||
```json
|
||
{
|
||
"name": "mcp__pdf__pdf",
|
||
"description": "Comprehensive PDF manipulation toolkit for extracting text and tables, creating new PDFs, merging/splitting documents, and handling forms. When Claude needs to fill in a PDF form or programmatically process, generate, or analyze PDF documents at scale.You need to pass in the --help parameter to obtain the usage of this tool first.",
|
||
"input_schema": {
|
||
"type": "object",
|
||
"properties": {
|
||
"params": {
|
||
"$ref": "#/$defs/PdfCommandInput"
|
||
}
|
||
},
|
||
"required": [
|
||
"params"
|
||
],
|
||
"$defs": {
|
||
"PdfCommandInput": {
|
||
"additionalProperties": false,
|
||
"properties": {
|
||
"argv": {
|
||
"description": "Command tokens excluding the leading 'pdf' executable name",
|
||
"items": {
|
||
"type": "string"
|
||
},
|
||
"title": "Argv",
|
||
"type": "array"
|
||
},
|
||
"command": {
|
||
"anyOf": [
|
||
{
|
||
"type": "string"
|
||
},
|
||
{
|
||
"type": "null"
|
||
}
|
||
],
|
||
"default": null,
|
||
"description": "Optional command string; if set, it will be split like a shell",
|
||
"title": "Command"
|
||
}
|
||
},
|
||
"title": "PdfCommandInput",
|
||
"type": "object"
|
||
}
|
||
},
|
||
"title": "pdfArguments"
|
||
}
|
||
}
|
||
```
|
||
下面是使用该MCP的上下文窗口变化
|
||
```json
|
||
User: Extract all the text from test.pdf and write it into result.txt
|
||
Agent: Tool(name="pdf", args=["--help"])
|
||
Tool: pdf - Comprehensive PDF manipulation toolkit for extracting text and tables, creating new PDFs, merging/splitting documents, and handling forms. When Claude needs to fill in a PDF form or programmatically process, generate, or analyze PDF documents at scale在使用之前你需要先传入--help参数以获取该工具的用法\n\nUsage:\n pdf [--help] <domain> <command> [options]\n\nDomains:\n forms Fill, extract, and validate PDF form fields.\n documents Create, merge, and split PDFs.\n text Extract text content from PDFs.\n tables Extract tables (heuristic, whitespace-based).\n images Render PDF pages to PNG images.\n\nRun `pdf <domain> --help` for domain help.\nRun `pdf <domain> <command> --help` for command help.
|
||
Agent: Tool(name="pdf", args=["text", "--help"])
|
||
Tool: text - Extract text content from PDFs.\n\nUsage:\n pdf text <command> [options]\n\nCommands:\n extract Extract text content from a PDF.\n\nRun `pdf text <command> --help` for details.
|
||
Agent: Tool(name="pdf", args=["text", "extract", "--help"])
|
||
Tool: usage: pdf text extract --pdf PDF_PATH [--pages PAGES] [--out OUTPUT_PATH]\n [--format {text,json}] [--overwrite]\n\nExtract text content from a PDF.\n\noptions:\n --pdf PDF_PATH Input PDF\n --pages PAGES Page selection like 1,3-5 or 'all' (default: all)\n --out OUTPUT_PATH Write extracted text to a file\n --format {text,json} Output format (default: text)\n --overwrite Overwrite output file if it exists\n
|
||
Agent: Tool(name="pdf", args=["text", "extract", "--pdf", "test.pdf", "--out", "result.txt", "--overwrite"])
|
||
Tool: {"result": {"stdout": "xx","data": {"pdf_path": "test.pdf","page_count": 6,"pages": [xxx],"format": "text","output_path": "result.txt"}}}
|
||
Agent: xxx
|
||
```
|
||
从上下文变化情况来看，完全实现了渐进式披露，该MCP代码开源(代码完全由codex编写，只验证想法，未做任何审查): https://github.com/musistudio/pdf-skill-mcp
|
||
|
||
如果你有什么想法也欢迎与我进行交流 email: [m@musiiot.top](mailto://m@musiiot.top )
|