本文是 《qca-wifi-host-cmn objmgr 解析》 的续篇。objmgr 的“骨架”梳理完之后，读代码时最容易卡住的反而是几个更基础的问题：

• HDD Adapter 到底是什么？
• PSOC / PDEV / VDEV / PEER 分别是什么？
• Link Info 是什么？
• Link Info 和 VDEV 是什么关系？
• Adapter 和 VDEV 是什么关系？

这一篇就把它们一次讲透：从 Linux 网络设备讲起，一层层往下剥，目标是初学者也能看懂。

This article is a sequel to qca-wifi-host-cmn objmgr Analysis. With the objmgr “skeleton” laid out, what actually trips people up while reading the code turns out to be a few more basic questions:

• What exactly is the HDD Adapter?
• What are PSOC / PDEV / VDEV / PEER, respectively?
• What is Link Info?
• How do Link Info and VDEV relate to each other?
• How do Adapter and VDEV relate to each other?

This post answers them all in one go: starting from Linux network devices, we peel off the layers one by one, with the goal of making it understandable even for beginners.

在使用 FakeHTTP 绕过运营商网络限速📤一文中，我记录了运营商会对上传流量进行限速，并借助 FakeHTTP 规避这一限制。

起初，我以为上传限速对普通用户的影响并不大：大多数人的上传需求明显少于下载需求，而且运营商通常还会通过白名单让主流应用免受影响。直到最近，我尝试用天翼云盘分享文件，才发现实际上传速度慢得难以接受。我本想继续使用 FakeHTTP，却发现原项目没有提供 Windows 版本。

于是，趁 Copilot 还有一些可用额度，我让 GPT-5.6 Luna 协助实现了一个 Windows 版本。实际运行效果如下：

In Using FakeHTTP to Bypass ISP Network Throttling 📤, I documented how ISPs throttle upload traffic and used FakeHTTP to work around the restriction.

At first, I assumed upload throttling would not matter much to ordinary users: most people upload far less than they download, and ISPs often exempt mainstream applications through allowlists. That changed when I recently tried to share a file through Tianyi Cloud Drive. The upload speed was painfully slow. I wanted to use FakeHTTP again, only to find that the original project did not provide a Windows version.

So, while I still had some Copilot quota available, I asked GPT-5.6 Luna to help implement a Windows version. Here is what it looked like in practice:

事件从哪里开始

近日，重庆长江国际办公大楼楼下发生了一件颇为超现实的事情。如果你对粉丝文化一无所知，那么我们先来简单介绍一下背景。长江国际一直是TF娱乐旗下偶像练习生们来来往往、接受训练的地方。TFBOYS 和 Teens in Times都是这家公司一手打造的，所以多年来一直有人在公司大楼外露营拍照，等待偶像出现。但最近，这里发生了一件大事。一边是大量女性偶像粉丝，另一边是一些男性直播主播和围观者，双方发生争执，后来甚至出现了泼水追逐和其他肢体冲突。

Where the Story Begins

Recently, at Chongqing’s Changjiang International office building something rather surreal happened downstairs. If you know absolutely nothing about fan culture then let’s first briefly explain the background. Changjiang International has long been where idol trainees under TF Entertainment come and go and receive training. TFBOYS and Teens in Times were both created by this company so people have been camping outside the building for years taking photos and waiting for idols. But recently, something big happened here. On one side were large numbers of female idol fans and on the other were some male livestreamers and onlookers the two sides argued, and later there was even water-throwing chasing, and other physical clashes.

本文基于 deepseek-harness 仓库当前检出的源码（0.1.0-rc.5）。文中的文件路径以这份源码为准；项目仍在快速演化，内部 API 可能变化。

DeepSeek Harness（下面简称 dsh）是一套 Agent 运行时。它做的事情并不神秘：接收用户输入，请求模型，执行模型要求的工具，把工具结果交回模型，直到得到最终回答。真正值得读源码的地方，是它怎样把这条流程变成可以恢复、回放和替换的程序。

本文只跟踪一个小任务：用户要求运行 echo hello。先把这个任务跑通，再解释它经过的 agent-loop、Session 日志、工具调度器、LLM 适配器和 Cordis 插件系统。这样读到后面的类型和事件时，都能知道它们在解决哪个具体问题。

1
2
3
4
5
6
7
8
9

用户输入
  -> inbox
  -> ReactLoopAgent 的 turn / step
  -> Session 投影成模型历史
  -> DeepSeek 适配器发起请求
  -> 模型返回 bash 工具调用
  -> 工具执行并写入 tool/result
  -> 新 step 再次请求模型
  -> 最终文本

This article is based on the checked-out source of deepseek-harness (0.1.0-rc.5). File paths refer to this checkout; the project is still evolving quickly, so internal APIs may change.

DeepSeek Harness, abbreviated as dsh below, is an Agent runtime. Its job is straightforward: accept user input, call the model, run the tools requested by the model, feed the results back, and stop when a final answer is available. The interesting part in the source is how this ordinary loop becomes recoverable, replayable, and replaceable.

This article follows one small task from beginning to end: the user asks the Agent to run echo hello. We first make that task concrete, then explain the agent-loop, Session log, tool scheduler, LLM adapter, and Cordis plugin system it passes through. That way, every type and event later in the article has a concrete problem to solve.

1
2
3
4
5
6
7
8
9

user input
  -> inbox
  -> turn / step in ReactLoopAgent
  -> Session projects the model history
  -> DeepSeek adapter sends a request
  -> the model returns a bash tool call
  -> the tool runs and writes tool/result
  -> a new step calls the model again
  -> final text

本文翻译自: https://arxiv.org/pdf/1808.03156

摘要

Apple Wireless Direct Link (AWDL) 是一个专有的、未公开的、基于 IEEE 802.11 的自组网（ad hoc）协议。Apple 于 2014 年左右首次引入 AWDL，此后将其集成到整个产品线中，包括 iPhone 和 Mac。我们发现 AWDL 驱动着超过十亿终端设备上的流行应用，如 AirPlay 和 AirDrop。然而，该协议本身以及潜在的安全和 Wi-Fi 共存问题尚未被研究。在本文中，我们通过二进制分析和运行时分析，揭示了该协议的工作原理。简而言之，每个 AWDL 节点会宣布一系列可用性窗口（Availability Windows, AWs），表明其与其他 AWDL 节点通信的准备状态。一个被选举出的主节点（master）负责同步这些序列。在 AW 之外，节点可以将其 Wi-Fi 无线电调谐到不同信道以与接入点通信，或关闭无线电以节省能源。基于我们的分析，我们进行了实验，研究了主节点选举过程、同步精度、跳频动态和可实现的吞吐量。我们还进行了初步安全评估，并发布了一个开源的 AWDL Wireshark 解析器，以促进未来的研究。

关键词： AWDL, 逆向工程, 自组网, IEEE 802.11, 专有协议, Apple, macOS, iOS

This article is translated from: https://arxiv.org/pdf/1808.03156

Abstract

Apple Wireless Direct Link (AWDL) is a proprietary and undocumented IEEE 802.11-based ad hoc protocol. Apple first introduced AWDL around 2014 and has since integrated it into its entire product line, including iPhone and Mac. While we have found that AWDL drives popular applications such as AirPlay and AirDrop on more than one billion end-user devices, neither the protocol itself nor potential security and Wi-Fi coexistence issues have been studied. In this paper, we present the operation of the protocol as the result of binary and runtime analysis. In short, each AWDL node announces a sequence of Availability Windows (AWs) indicating its readiness to communicate with other AWDL nodes. An elected master node synchronizes these sequences. Outside the AWs, nodes can tune their Wi-Fi radio to a different channel to communicate with an access point, or could turn it off to save energy. Based on our analysis, we conduct experiments to study the master election process, synchronization accuracy, channel hopping dynamics, and achievable throughput. We conduct a preliminary security assessment and publish an open source Wireshark dissector for AWDL to nourish future work.

Keywords: AWDL, Reverse engineering, Ad hoc networks, IEEE 802.11, Proprietary protocol, Apple, macOS, iOS

2026 年 6 月，Github Copilot 的计费方式从原来的按请求次数收费变成了按 Token 消耗量收费。并且不再提供任何免费模型。

这一计费方式的改变使得 Github Copilot 的使用成本大幅增加，对于重度使用者经常可以一天就消耗掉一个月的使用额度。

Token 焦虑时代正式来临。

本文介绍如何在 VS Code 中使用 Github Copilot 对接现有的免费模型 API。

In June 2026, GitHub Copilot’s billing model changed from charging per request to charging by token consumption. It also stopped offering any free models.

This billing change dramatically increased the cost of using GitHub Copilot — heavy users can easily burn through a month’s quota in a single day.

The era of token anxiety has officially arrived.

This article explains how to connect GitHub Copilot to existing free model APIs in VS Code.

最近在折腾博客的时候，偶然发现自己复制的 Cloudflare 账户 ID 在粘贴的时候就变样了，即使这个账户 ID 藏在一段话的中间，也会被篡改！

全程火绒安全形同虚设，完全没有任何提示和拦截。

演示视频：

While tinkering with my blog recently, I happened to notice that the Cloudflare account ID I copied got altered when pasted — even when the ID was buried in the middle of a paragraph, it was still tampered with!

Huorong Security was completely useless throughout this whole ordeal, without any alerts or blocks whatsoever.

Demo video:

大黑山

Big Black Mountain

太炸裂了！！！

4 月 30 日下班后刷短视频，发现了这个超级劲爆的安全漏洞：CVE-2026-31431，又名 Copy Fail。
它只需 10 行 Python 代码，就能完成 Linux 本地权限提升，直接获得 root 权限。而且这个漏洞从 2017 年就存在，几乎影响所有主流 Linux 发行版。

This is absolutely insane!!!

While scrolling through short videos after work on April 30th, I discovered this absolutely explosive security vulnerability: CVE-2026-31431, also known as Copy Fail.
It only takes 10 lines of Python to achieve Linux local privilege escalation and directly obtain root privileges. What’s more, this vulnerability has existed since 2017 and affects almost all mainstream Linux distributions.