ICDYCT / I Can Do, You Can Too

 Preface Previously, I implemented a Telegram keyword monitoring system . It runs well. https://t.me/FuckGFWNewbieNeedHelp Today I feel that all keyword messages are tiled together without distinction of importance, which is worth optimizing.

 前言 之前, 我实现过 一个telegram关键字监控系统 . 运行效果良好. https://t.me/FuckGFWNewbieNeedHelp 今天我觉得 所有关键字消息平铺在一起, 没有重要程度的区分, 值得优化一下.

Preface Got a Dedirock early on, storage is only 15G, installed Hermes, played around with it, disk space is down to just over 1G. This isn't the first time disk space has been running low. Previously I just had Hermes analyze which folders could be cleaned, which caches could be deleted, and which logs could have their size limited.  Today it's really not working anymore, I think I'll just reinstall.

前言 很早入的一台 Dedirock, 存储只有15G, 装了Hermes, 玩来玩去, 磁盘空间只剩1个多G了. 这不是第一次硬盘空间告急了, 之前就是让Hermes自己分析哪些文件夹能清理, 哪些缓存能删, 哪些日志能限制大小.  今天实在是不行了, 我想干脆重装吧.

Before Modification Requesting Modifications from Muse Analyze https://github.com/crazypeace/xkcd-password-generator Add the following toggle options to the page: All lowercase letters / Capitalize first letter of each segment No separator / - separator No numbers and symbols / Append a 4-digit number at the end, which is the alphabetical position of the last 2 letters of the original password (a=01, z=26, 4 digits total) After Modification Demo https://git.io/xkcdpw Github https://github.com/crazypeace/xkcd-password-generator

修改前 向muse提修改要求 分析 https://github.com/crazypeace/xkcd-password-generator 在页面中增加以下开关项: 全小写字母 / 每小节首字母大写 无分隔 / -分隔 无数字和符号 / 尾部添加一节4位数字, 为原密码最后2个字母的排序数字（a=01, z=26，共4位） 修改后 Demo https://git.io/xkcdpw Github https://github.com/crazypeace/xkcd-password-generator

Introduction A few months ago, I wrote an article urging beginners to use Agent to deploy projects . It was still a text-and-image tutorial based on the project's README. A few days ago, while discussing with a fellow group member how to use Agent to set up a 3x-ui system , they said that the Youtube video tutorial they followed worked as expected. At the time, I also said that I would be confident Agent could follow a text-and-image tutorial and succeed, but I could not guarantee the outcome with a video tutorial. Today, I wanted to see what would happen if I had Agent follow a video tutorial and set it up. Over the past couple of days, I feel space-bunny-alpha has become less capable. So I'll use muse instead.

前言 几个月前, 我写过 呼吁小白们都使用Agent部署项目 , 那还是基于项目的README 这种图文教程. 几天前, 我在 和一个群友交流使用Agent搭3x-ui系统 时, TA说看的Youtube视频教程 操作下来结果是正常使用的, 我当时还说, 要是图文教程我肯定有信心让Agent照着操作成功, 视频教程我不敢保证效果. 我今天想, 试试看让Agent照着视频教程搭建会是什么结果? 这两天 space-bunny-alpha 我感觉降智了. 那就用muse来吧.

Introduction The currently popular muse is an Agent. Moreover, it comes "fully furnished": utilities and the network are connected, and the furniture and appliances are all in place. You don't need to install a system or configure a model; you can start chatting right away. All the examples in my blog about developing for Agents and putting Agents to work can be copied over to muse and tried there.

前言 最近很火的muse就是Agent. 而且是"精装修"的, 水电网络接好了, 家具家电都摆好了. 你不需要安装系统, 不需要配置模型, 直接开始对话就行了. 我的博客中, 所有 面向Agent开发 , 使唤Agent干活 的例子, 你都可以照搬到 muse 上试试.

Introduction Previously, I showed muse bypassing Turnstile human verification .  I have also seen many spam bots in other groups pass nmbot's Turnstile human verification and join groups all day long. I thought I would test in my own group whether advertisers could solve Turnstile human verification. * I do not use group-management bots deployed by other people in my own group . Then I'll develop my own Telegram group-join verification bot  using  Turnstile for human verification 

前言 前面展示了 muse破Turnstile人机验证 .  我在其它群也见识过一天到晚很多广告机器人都过了nmbot的Turnstile人机验证加群. 我在想自己的群试一下发广告的人能不能解决Turnstile人机验证. * 我是 不在自己的群使用别人部署的群管机器人 的. 那么就自己开发一个 Turnstile作人机验证 的电报入群验证机器人 

Introduction muse has been very popular lately. Has everyone gotten one and started playing with it? Today, we're testing muse's ability to operate its built-in browser to pass Cloudflare's free Turnstile CAPTCHA.

前言 最近 muse 热度很高啊, 大家都搞了一个到手开始玩了吗? 今天来测试一下 muse 操作自带浏览器 通过 Cloudflare 免费版 Turnstile 人机验证 的能力.

Preface A few months ago, I wrote  Configuring a Cloudflare API Token for Your Hermes Agent Although the overall process is already streamlined, it still isn't particularly seamless. Either you have to follow the tutorial and operate the website step by step yourself, or your Agent needs strong browser automation capabilities. But now, Cloudflare is enthusiastically embracing AI and has introduced the Cloudflare CLI cf tool Let's take a look at the entire OAuth process

前言 几个月前我写过  Cloudflare 设置 API token 为你的Hermes Agent 虽然整个过程已经很精简了, 但还算不上特别丝滑. 要么, 需要你自己照着教程一步一步操作网页; 要么, 需要你的Agent有较强的操作浏览器的能力. 而现在, Cloudflare 热烈地拥抱AI, 推出了 Cloudflare CLI cf工具 让我们来看看整个OAuth过程

I was originally going to have stealth/space-bunny-alpha try writing an image-and-text tutorial based on a Youtube video. Then, quite by accident, I saw this: Huh? Where did this proxy come from? I opened a new session and asked. It came from a proxy pool. But where did the information about that proxy pool come from? It turned out to be the model's own "knowledge."

本来我是想让stealth/space-bunny-alpha试试根据Youtube视频写图文教程的任务. 然后一不小心我看到了这么一条 唉? 这是哪里来的代理? 新开一个session问了一下. 来自一个代理池. 进一步的, 那代理池的信息是哪里来的呢? 原来是自来模型自己的"知识"

Preface A few years ago, I wrote an article titled How to Ask Questions Properly and Bad Examples of Asking for Help Recently, the world has entered the era of AI and Agents. In various chat groups, I have seen many cases of people failing to get Agents to do what they want. Some are indeed due to the limitations of AI and Agents, but many more are caused by giving them the wrong instructions.

前言 几年前, 我写过一篇  如何正确提问及错误求助反面实例 最近, 世界来到了AI 和 Agent的时代, 我在各个群里看到了很多使唤Agent失败的案例. 有一些确实是AI和Agent能力不足, 但是还有很多是使唤方法不对.

Preface I have always enjoyed using Telegram as a platform for interacting with Hermes . Furthermore, I like opening a group forum with topics in a new window, so I can continue browsing other groups in the main window without interruption. Friends who often  develop with an Agent  and  put an Agent to work  know that searching for keywords in session history is a feature they use frequently. So, the problem is this: in a group forum with topics opened in a new window, if you select a topic and click the search button, it automatically switches to All. For someone like me, who has countless topics open, it amounts to searching for keywords across all sessions, including historical sessions.

前言 我一直都喜欢 用Telegram作为与Hermes交流的平台 . 进一步的, 我喜欢用新窗口打开 带Topic的群组Forum, 这样不影响我在主窗口中去其它的群组中闲逛. 经常  用Agent开发  和  使唤Agent干活  的朋友们都知道, 在session历史中搜索关键字是经常使用的功能. 那么, 问题来了, 在新窗口打开 带Topic的群组Forum 中如果选中了topic再按搜索按钮, 会自动跳到 All 对于我这种开了无数个topic的人来说, 相当于是在所有的session(包括历史session)中搜索关键字了.

Introduction We already have a specially modified version of cpa-for-agentrouter-d4f to address the issue encountered by deepseek-v4-flash on agentrouter: HTTP 400: The content[].thinking However, the  request-format: anthropic  parameter added by this modified version is not supported by the web management interface. So, what is the most elegant way to use it?

前言 我们已经有了一份 特殊修改版 cpa-for-agentrouter-d4f 针对 agentrouter 的 deepseek-v4-flash 遇到 HTTP 400: The content[].thinking问题 但是, 这份修改版增加的  request-format: anthropic  参数不被web管理界面支持. 那么, 我们应该怎样优雅地使用呢?

Introduction Earlier, to use the nonprofit agentrouter site 's deepseek-v4-flash model for free in Hermes, we created a modified version of cpa-for-agentrouter-d4f Today, I registered muse.ai using the method from cloud.browser-use.com  * If you have just finished registering for muse.ai, you can redeem the invitation code  XKGZJU  in settings to receive 1B tokens I wanted muse to reproduce the issue and then analyze it independently.

前言 前面我们为了白嫖 公益站agentrouter 的deepseek-v4-flash 给 Hermes 用, 修改了一版 cpa-for-agentrouter-d4f 今天用 cloud.browser-use.com 的方法注册了 muse.ai  * 如果你刚完成注册 muse.ai 可以在设置中验证邀请码  XKGZJU  得 1B token 想让muse复现一下问题, 再独立分析一下问题.

Introduction Previously, we set up CPA cliproxyapi configured User-Agent to access the free agentrouter service Recently, I always encounter this error when using agentrouter 's deepseek-v4-flash HTTP 400: The content[].thinking in the thinking mode must be passed back to the API.

前言 之前, 我们 搭建了CPA cliproxyapi 设置了User-Agent来白嫖公益站agentrouter 最近在使用 agentrouter 的deepseek-v4-flash时, 总是遇到这样的报错 HTTP 400: The content[].thinking in the thinking mode must be passed back to the API.

Preface I listen to some audio channels on Telegram. If I close Telegram halfway through listening, the next time I open Telegram, I have to find that channel again, recall where I left off, and then start playing again. I find this a bit inconvenient. So I thought, could forkgram/tdesktop remember the last played audio and automatically load it at the top on the next startup?

前言 我在Telegram上面会播放一些 音频的频道 . 如果听到一半, 关闭Telegram, 下次再打开Telegram, 就要再找到这个频道, 回忆自己上次听到哪里了, 再开始播放. 我觉得有点不方便. 那么我就想, 能不能让 forkgram/tdesktop 记住上次播放的音频, 下次启动后自动加载到顶部?

Preface Previously, I successfully compiled v6.9.4 with github action Today I wanted to develop a new feature, but compiling v7.2.8 failed. And the Agent even said it couldn't learn any successful experience from the previous successful compilation records. I was so mad.

前言 之前 成功用github action编译了 v6.9.4 今天想开发个新需求, 结果 编译 v7.2.8 不成功. 而且Agent还说从以前编译成功的记录中学不到成功经验. 气死我了.

Preface Previously  improved the Blogger editor and added pre and code function buttons Also improved the blog template to better display blockquote, pre, and code Then I felt that the display effect inside the editor was a bit ugly. Mainly because the editor originally did not have pre and code functions, so there was no suitable display effect.

前言 前面  改进了 blogger编辑器, 加入了 pre 和 code 的功能按钮 也 改进了 博客的模板, 能更好的显示 blockquote, pre, code 然后就觉得, 编辑器里面的显示效果有点丑. 主要是原本编辑器就没有 pre 和 code 的功能, 没有合适的显示效果.

Preface Regarding syntax highlighting in my blog, I initially used modifying the blockquote CSS to achieve the display effect for code snippets Later I experimented with introducing Google's code-prettify to implement syntax highlighting and converting blockquote to pre code Finally, because syntax highlighting would erase the text background colors I set in code snippets , I switched back to the approach of modifying blockquote CSS to render code snippets Now that I've made the Blogger editor accept pre and code tags , I use blockquote to actually represent quotation — quoting the natural-language prompts I send to Agents, which is indeed a scenario I've used very frequently in the last few months. I also used an Agent to operate the Blogger API to help me convert blockquote to pre in my historical posts. Now I'm reconsidering the syntax-highlighting issue.

前言 关于博客中的代码高亮, 最早我是  修改 blockquote 的CSS 来实现 代码段 的显示效果 后来实验   引入google的code-prettify实现代码高亮 把blockquote转换为pre code 最后 因为代码高亮会把我在代码段中设置的文字底色清除 , 换回了  修改 blockquote 的CSS 来实现 代码段  的效果 现在我让  blogger编辑器可以输入pre和code了 , 所以我用blockquote真的去表示引用, 引用我发送给Agent的自然语言, 这也确实是我最近几个月使用非常多的场景. 我也通过Agent 操作Blogger API 帮助我把历史博文中的 blockquote 换成了 pre 现在我重新考虑代码高亮的问题.

Preface Previously, I implemented a simple anti-advertising bot for personal use in Python , implementing the speaking verification feature. Recently my server expired, and a thought suddenly crossed my mind — can it be converted to a Cloudflare Worker?

前言 之前 用python实现了自用的简易防广告机器人 , 实现发言验证的功能. 最近服务器到期了, 突然心里动了一个念头, 能转换成cloudflare worker吗?

Preface Previously I implemented a simple anti-advertising bot for personal use in Python , featuring a join-group verification function. Recently my server expired, and suddenly an idea crossed my mind — can I convert it to a Cloudflare Worker?

前言 之前 用python实现了自用的简易防广告机器人 , 实现加群验证的功能. 最近服务器到期了, 突然心里动了一个念头, 能转换成cloudflare worker吗?

Implementation Results What people who join the group see What other group members see

实施效果 加群的人看到的 其它群友看到的

Preface Previously I had applied text-autospace.js to add spaces between Chinese and English , and have been using it until now. The display effect looks great. Today I wanted AI to modify it so that it only takes effect within the blog post content; in other parts of the page, such as widgets, it should not take effect. After AI analysis, I was told that this script adds the hanla tag to the entire HTML body element. What? (I was really surprised — I had never thought about what this script does under the hood; I only looked at the final display effect.)

前言 之前 应用过text-autospace.js给中英文之间 添加空格 , 一直用到现在, 显示效果不错的. 今天想让AI改改, 只在博文正文中生效; 在页面的其它部分, 比如小工具之类, 不要生效. AI分析告诉我, 这个脚本是对整个HTML的 body 元素全部添加 hanla 标签. 什么? (我很吃惊, 我从来没想过这个脚本在底层干了什么, 我只知道眼睛看最终的显示效果)

Preface When I use some streaming sites , domain-based traffic splitting doesn't work well. So I simply use it in Edge, and I set Edge to use no proxy, a completely direct connection. I already have a Tampermonkey script that lets videos go fullscreen on the webpage , but I don't want to install too many things in Edge. So I thought: what if I convert this Tampermonkey script into a browser extension?

前言 我在使用一些 影视网站 的时候, 域名分流 并不能好好工作. 所以我干脆在edge中使用, 我把edge设置为不用代理, 完全直连. 我已经有了一个 让视频可以网页全屏的油猴脚本 , 但我的edge并不想安装太多东西. 所以我想, 如果把这个油猴脚本转换为浏览器扩展呢?

Preface We already have a Blogger/blogspot blank blog template The next step is to make it look good. Let's start with a simple task, using Moonlight Blog as an example.

前言 我们已经有了一份 Blogger/blogspot 空白博客模板 接下来就是怎么弄漂亮. 先从简单的任务开始吧, 以 月光博客 为例.

Together with  oc/x-preview-f-free , stealth/ox-alpha , I created a blank Blogger/blogspot template. No CSS included.

我和  oc/x-preview-f-free , stealth/ox-alpha 一起, 制作了一份 Blogger/blogspot 空白博客模板. 不含CSS.

Blogger/blogspot's Blog Search Widget  Set the title to empty Prompt: Required field cannot be left blank If you modify the xml template file and forcibly change the title to empty, then, when importing the template file, it will prompt that the required field cannot be left blank Conclusion The Blog Search widget of Blogger/blogspot cannot be set with an empty title If you don't want the page to display the title of the Blog Search widget, you can only hide it via CSS.

Blogger/blogspot 的 博客搜索 小工具  设置标题为 空 提示 必填字段不能留空

Importing a template file into Blogger/blogspot Theme - the dropdown menu to the right of Customize - Restore Select the .xml file * I prepared a test file that contains all the widgets (except the Followers widget), and none of the widgets use the h2 tag https://github.com/crazypeace/blogspot-template-xml/blob/main/theme-ReportAbuse-testcase.xml

向 Blogger/blogspot 导入模板文件 主题背景 - 自定义右边下拉菜单 - 还原 选择 .xml 文件 * 我准备了一份测试文件, 里面包含了全部的小工具widget (除了关注者 小工具 Followers widget), 并且所有小工具都没有使用 h2 标签 https://github.com/crazypeace/blogspot-template-xml/blob/main/theme-ReportAbuse-testcase.xml

Import a Template File into Blogger/blogspot Theme - Customize right-side dropdown menu - Revert Select the .xml file * I have prepared a test file that contains all the widgets, and none of the widgets use the h2 tag https://github.com/crazypeace/blogspot-template-xml/blob/main/theme-Followers-testcase.xml

向 Blogger/blogspot 导入模板文件 主题背景 - 自定义右边下拉菜单 - 还原 选择 .xml 文件 * 我准备了一份测试文件, 里面包含了全部的小工具widget, 并且所有小工具都没有使用 h2 标签 https://github.com/crazypeace/blogspot-template-xml/blob/main/theme-Followers-testcase.xml

Preface Many people like Blogger/blogspot, the server and domain are both from Google, no need to worry about being attacked . However, Blogger's editor is not friendly for writing posts with code; there are no code or pre buttons. To deal with this problem, I've seen all kinds of solutions. Some hand-write pre code in HTML mode , some use plugins to assist in HTML mode , I came up with several workaround methods on my own. One is: using the editor's built-in blockquote block container, and restyling its appearance Another is: while writing the post, use the editor's built-in blockquote block container, and use JS to convert blockquote into pre code when displaying Yet another is: write plain Markdown text directly in HTML mode, and use JS to convert it to HTML when displaying With AI being so powerful these days, I got the itch again today to let AI analyze how the original editor works and see if there's a chance to enhance the editor.

前言 Blogger/blogspot 有不少人喜欢, 服务器和域名都是google的, 不怕别人打 . 但是 Blogger 的编辑器对写代码的文章不友好, 没有 code 和 pre 按钮. 面对这个问题, 我看到过各种各样的方案. 有 在HTML模式下手写 pre code 的 , 有 在HTML模式下用插件辅助的 , 我自己研究出了几种绕过的方案 一种是: 利用编辑器自带的 引用blockquote 块容器, 改造显示样式 另一种是: 在编写博文时, 使用 编辑器自带的 引用blockquote 块容器, 在显示时用JS 把 blockquote转换为pre code 还有一种是: 直接在HTML模式下写纯Markdown语法的文本, 在显示时用JS转换为HTML显示 现在AI能力这么强, 我今天又动了念头想, 让AI来分析一下原本的编辑器是怎么工作的, 看有没有机会对编辑器进行增强.

Preface I don't use the audio/video features on Telegram, so I thought, is it possible to disable the audio/video buttons?

前言 我不在Telegram上使用 音频/视频 功能, 于是想, 能不能把 音频/视频 按钮禁用.

Preface We have already implemented pinning groups with @me messages to help us "quickly reach the battlefield" We also implemented jump to the last message that @me   Furthermore, we optimized it into jump to the previous or next message that @me But, if there are many @me messages in a group, then the efficiency of navigating previous and next is also too low.

前言 我们之前已经实现了 有@我消息的群 置顶 方便我们"快速到达战场" 又实现了 跳转到最后一条@我的消息   进一步地, 优化成了  跳转到前一条后一条@我的消息 但, 如果群里面有很多条@我的消息, 那么前一条后一条的导航效率也太低了吧.

Preface I posted on the forum  Modifying forkgram/tdesktop to implement jumping to the previous/next message that @mentions me A forum member replied saying "Actually, just searching #m allows you to view messages that mention and reply to you"

前言 我在论坛发布了  修改 forkgram/tdesktop 实现 跳转到前一条后一条@我的消息 有坛友回复说 "其实搜索 #m 就可以查看提及和回复自己的消息" 我试了一下, 确实是这个效果. 我想知道底层的原理, 所以就问了问各个模型 * 我这种穷鬼 只能用各个模型的网页版免费账户.

Preface Recently a public welfare site agentrouter popped up offering free claude-opus-4-8, claude-opus-5, gpt-5.6-sol However, when you connect this to Hermes, you'll find that it doesn't work. Inspiration Came across this big shot's solution of modifying the UA while browsing online https://blog.notett.com/post/2026/08/202608-agentrouter-proxy/ Practice After some exploration, I found that CPA cliproxyapi has the function of setting UA. The content to fill in is just copied from the big shot User-Agent Kilo-Code/7.3.50 ai-sdk/provider-utils/4.0.27 runtime/bun/1.3.14 Then you'll find that the test passes in CPA. Then connect it to Hermes and use it, gpt-5.6-sol works. But claude-opus-4-8 and claude-opus-5 still don't work. ======== Postscript I used it to run some slightly "heavy" development work , and sometimes encountered refusal to work. The model provider failed after retries. I kept raw provider details out of chat; check gateway logs for diagnost...

前言 最近冒出来一个 公益站agentrouter 提供免费的claude-opus-4-8, claude-opus-5, gpt-5.6-sol 但是, 当你把这个接到 Hermes 里面后, 会发现, 不能使用

Preface We previously implemented pinning groups with @mention messages to help us "quickly get to the battlefield" We went further and implemented  jumping to the last @mention message But what if there are more than one new @mention messages in the group?

前言 我们之前已经实现了 有@我消息的群 置顶 方便我们"快速到达战场" 又进一步实现了  跳转到最后一条@我的消息 但是, 如果群组里面新 @我的消息多于一条呢? 

Preface We have previously implemented pinning groups with @me messages to help us "quickly get to the battlefield". However, if we scroll up and down in this group, or jump to other groups/channels to find "paper reference materials", and then return to the previous group, we may not be able to find the message that just @-ed me. If you like to "intensely communicate with others" across multiple groups, you may need this feature: Find the previous message that @-ed me

前言 我们之前已经实现了  有@我消息的群 置顶  方便我们"快速到达战场" 但是, 如果我们在这个群组里上下滚动了一下, 或者跳到其它群/频道 去找 "论文引用材料", 再回到刚才的群组, 就有可能出现找不到 刚才@我的那条消息 的情况. 如果你喜欢在多个群里面"剧烈地与人交流", 说不定你需要这样的功能: 找到上一条@我的消息

My last share  on GPT-Oriented Development examples was still back in January Most of those projects were small in scale, the main workflow was: describe requirements in a chatbot interface, then a human "carries" the AI's output code into the project, the human does testing, sends the results back to GPT, and then GPT outputs the revised product. Loop and repeat ... Now, GPT-Oriented Development is all Agent-Oriented Development.

上次分享  面向GPT开发实例集合 还是1月份 大部分开发规模很小, 主要形式是在 chatbot 界面描述需求, 人再把AI输出的代码"搬"到项目里, 人做测试, 把结果发给 GPT, 然后 GPT再输出修改后的产物. 如此循环 ... 现在的面向GPT开发, 都是面向Agent.

Preface A group member asked whether there's a feature on the phone to mute all of an entire folder I thought about it, and iOS or Android apps are too troublesome. We can use the web version of Telegram. I tried Version A and Version K , and found that Version K's folder long-press has a pop-up menu, which is more suitable for further development

前言 有群友问, 手机上有没有 mute all 整个 folder 的功能 我想了一下, ios 或 android APP 太麻烦了, 我们可以利用 web版 telegram. 我试用了一下 A版 和 K版 , 发现K版的 folder 长按是有弹出菜单的, 比较适合进一步开发

Preface Sometimes, you want the Agent to help you push the project to Github, and then use Github's action to run the build. For example, https://zelikk.blogspot.com/2026/06/forkgram-tdesktop-github-action-win10-amd64.html Sometimes, you want the Agent to save a copy of the proxy node subscription text to a Github gist All of the above operations require a prerequisite: you have set up a Github API token for the Agent to use. Github 's API token can be set with repo, workflow, gist permissions. Generally that's enough. * The following practical solutions were all completed under the condition of Hermes connecting to tencent/hy3:free Practice 1 If your Agent runs on your own computer.  You can open the Chrome browser, log into your Github account. Then tell your Agent, The current page in the browser is already logged into a Github account, please continue the operation and create an API token with repo, workflow, gist permissions Then you can save this token separately. Yo...