mirror of
https://github.com/musistudio/claude-code-router.git
synced 2026-04-28 02:09:27 +00:00
update sponsors
This commit is contained in:
musistudio
parent
c73fe0d49d
commit
a4254eeb68
5 changed files with 157 additions and 5 deletions
|
|
@ -7,9 +7,11 @@
|
|||
<hr>
|
||||
|
||||
|
||||
> This project is sponsored by Z.ai, supporting us with their GLM CODING PLAN.
|
||||
> GLM CODING PLAN is a subscription service designed for AI coding, starting at just $3/month. It provides access to their flagship GLM-4.7 model across 10+ popular AI coding tools (Claude Code, Cline, Roo Code, etc.), offering developers top-tier, fast, and stable coding experiences.
|
||||
> Get 10% OFF GLM CODING PLAN：https://z.ai/subscribe?ic=8JVLJQFSKB
|
||||
> This project is sponsored by Z.ai, supporting us with their GLM CODING PLAN.
|
||||
|
||||
> GLM CODING PLAN is a subscription service designed for AI coding, starting at just $10/month. It provides access to their flagship GLM-4.7 & （GLM-5 Only Available for Pro Users）model across 10+ popular AI coding tools (Claude Code, Cline, Roo Code, etc.), offering developers top-tier, fast, and stable coding experiences.
|
||||
|
||||
> Get 10% OFF GLM CODING PLAN：https://z.ai/subscribe?ic=8JVLJQFSKB
|
||||
|
||||
> [Progressive Disclosure of Agent Tools from the Perspective of CLI Tool Style](/blog/en/progressive-disclosure-of-agent-tools-from-the-perspective-of-cli-tool-style.md)
|
||||
|
||||
|
|
|
|||
|
|
@ -7,8 +7,9 @@
|
|||
<hr>
|
||||
|
||||
|
||||
> GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元，即可在十余款主流AI编码工具如Claude Code、中畅享智谱旗舰模型GLM-4.7，为开发者提供顶尖的编码体验。
|
||||
> 智谱AI为本软件提供了特别优惠，使用以下链接购买可以享受九折优惠：https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII
|
||||
> 本项目由 Z智谱 提供赞助, 他们通过 GLM CODING PLAN 对本项目提供技术支持。
|
||||
> GLM CODING PLAN 是专为AI编码打造的订阅套餐，每月最低仅需20元，即可在十余款主流AI编码工具如 Claude Code、Cline、Roo Code 中畅享智谱旗舰模型GLM-4.7（受限于算力，目前仅限Pro用户开放），为开发者提供顶尖的编码体验。
|
||||
> 智谱AI为本产品提供了特别优惠，使用以下链接购买可以享受九折优惠：https://www.bigmodel.cn/claude-code?ic=RRVJPB5SII
|
||||
|
||||
> [从CLI工具风格看工具渐进式披露](/blog/zh/从CLI工具风格看工具渐进式披露.md)
|
||||
|
||||
|
|
|
|||
Binary file not shown.
|
Before Width: | Height: | Size: 343 KiB After Width: | Height: | Size: 107 KiB |
Binary file not shown.
|
Before Width: | Height: | Size: 396 KiB After Width: | Height: | Size: 1.1 MiB |
|
|
@ -10,9 +10,158 @@
|
|||
|
||||
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：
|
||||
> 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hropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fa3bca2763d7892982a59c28aa4df7993aaae55ae-2292x673.jpg&w=3840&q=75)
|
||||
> 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
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue