豆包 Seed 又更新了：一张永远“最新”的模型卡

这次豆包推出的不是一个过段时间就会落后的固定版本，而是 Doubao-Seed-Evolving一个专门面向 Coding 和 Agent 场景的 latest 分支。Evolving 的主要就是按周持续更新。者不用隔一段时间追一次新模型，使用同一个模型入口。背后的能力会因为每周迭代继续提高。可以把它理解成一张永远保持“最新”的模型卡。

使用者痛点：每次升级都要手动迁移、检查依赖、重新设置环境；旧模型在实际项目中很快失效，导致生产力下降。

主要升级点

这次升级最直观的两个点，是 1M 上下文和更强的长程任务能力。1M 上下文让模型一次能接住更大的代码仓库、更长的需求和更多历史信息；长程能力则让它在多轮规划、工具调用、代码修改和测试验证中不容易中途丢掉目标。对于 Codex 这类 Coding Agent 能把一个复杂任务持续做下去，比只写好某一段代码更关键。老实说，

实测准备

好的。今天我们就用 Codex 配置好这个模型，实测一下模型效果如何。

怎么接入和配置模型，我不在这里重复贴命令。怎么说呢，火山引擎已经整理了 Codex CLI、Codex 桌面客户端等配置步骤。需要的同学直接按照 操作即可。

图中红框是官方教程的章节目录。按照自己使用的客户端选择对应部分即可，这篇文章不再重复配置过程。

桌面脑洞

每天开机以后Chrome、Edge、微信、VS Code 这些软件图标都安安静静地待在 Windows 桌面上。它们的工作只有一个：等我双击。

使用者痛点：工作时桌面被一堆死板图标占满，视觉疲劳且缺乏互动；下班后桌面仍然毫无变化，让人感觉电脑“闲置”而浪费了潜在的趣味空间。

我突然想到一个有点无聊、但越想越有画面的事情：如果我下班了桌面图标也下班了它们会不会自己出来活动？浏览器之间可能互相看不顺眼。聊天软件可能凑在一起说悄悄话，开发工具也许会组队踢球。

后来这个脑洞真的变成了一个 Windows 小工具——《桌面下班后》。有的图标长出手脚满桌散步，有的躲到一边聊天两个图标还会突然打起来。过一会儿，整桌软件又会自动分成两队踢球。整个过程无需作，我最多在打架时点一下“劝架”。

使用者痛点：担心此类动画会影响程序性能或误触真实快捷方式。

实现思路概览

下面是实际运行时的画面：左边是真实桌面快捷方式，一枚都没动；右边那些长出手脚、满桌跑的，是程序绘制出的角色分身。

我最初只说了一句话：“让图标可行性。再给出“不要移动原生图标，改画角色分身”的关键方案。红色标签标出了我的最初想法和模型给出的关键方案：原生桌面图标不动，对了生成角色分身。

这个点子差点又被我做复杂

“会打架、会踢球”，听起来很像游戏。 第一轮方案里自然出现了点击角色、拖动角色、指定聊天对象、主动发起足球比赛这些玩法。

使用者痛点：如果加入交互式游戏元素。会让本来只想欣赏动画的使用者感到困扰，还可能导致误操作打开软件。

于是我把边界说死了：

• 不能点角色，不能拖角色；
• 不能指定谁聊天、谁打架；
• 不能手动开球，也不能控制任何一队；
• 没有积分、成就和输赢；
• 桌面舞台永远鼠标穿透；
• 能做的只有换上、换下图标，还有打架时劝一下。

If you don't intervene,conflicts resolve mselves after a few seconds. The “mediation” button is just a light entry point,not mandatory.

先别写代码，先把三个方案砍一遍

方案	看起来怎么样	实际问题	结论
直接移动 Windows 原生图标	最有噱头，看起来像真的动画 但坐标/DPI/Explorer 状态出错会乱掉使用者布局 双击意外打开软件等风险高	破坏使用者桌面布局 触发真实快捷方式导致误打开程序	不采用
读取 .lnk 图标再画分身	真实图标不动。又能使用本机图标 需要处理透明窗口、穿透和层级	实现难度适中，但能保留代入感	采用此方案

项目第一条规矩不是“动画要好看”，而是：

The program scans user's desktop and public desktop for top‑level .lnk,extracts name & icon,n draws a copy in a transparent overlay window. The real shortcuts stay untouched on left side .

第一脚先踢到 Windows 的铁板

透明舞台不能挡住鼠标

A moving image in WPF is easy;making whole window invisible to mouse events is hard.

• The naïve transparent window still captures clicks → users can't click real icons.
• A top‑most window would also obscure or apps.
• The solution: set Win32 extended styles to make window click‑through,tool‑window and non‑activating.

var current = GetWindowLongPtr.ToInt64;SetWindowLongPtr(
handle。GwlExStyle,new IntPtr);

This makes mouse pass through to underlying desktop,prevents keyboard focus stealing。and hides overlay from Alt+Tab.

The overlay only appears when Explorer’s desktop is foreground;opening or apps hides it automatically.

一点“隐藏面板”，图标也全没了

The first prototype worked—until I clicked “Start Show”. The control panel disappeared and so did all characters.

• User Pain: Unexpected disappearance of visual feedback makes users think tool crashed.
• The bug wasn't WPF closing stage but my overly strict “desktop foreground” check. Background programs like SpeedBall were mistakenly treated as normal windows.

I refined logic into three categories:

• Explorer Desktop:: show overlay.
• This tool & recognized desktop companions:: show overlay.
• Browsers / IDEs / or apps:: hide & pause overlay.

个图标一起上。我自己都看不下去

The overlay initially displayed every valid shortcut – resulting in a chaotic “desktop rush hour”. Icons overlapped,chats were unreadable and finding a specific icon became impossible.

• User Pain:
• - Overcrowded screen leads to visual noise. - Users lose control over which icons are animated. - Performance degrades with too many sprites.

I asked model to limit initial roster. The new default shows only first N=8?，valid shortcuts;users can later swap icons manually. Existing users who previously selected “all icons” keep that setting unless y upgrade from an old version where all icons were shown – n we auto‑shrink back to N=8?.

public const int DefaultInitialCount = 8;return valid
.Take)
.ToArray;

会走了不代表有东西可看

自动事件怎么选

The earliest version only moved icons randomly – boring after two minutes. Later we added state,stamina,personality。relationships and cooldowns so events feel purposeful.

• A scheduler checks which feature is enabled,respects individual cooldowns and n rolls a random number to decide which event fires next.

var roll = _random.Next;if ) return true;if ) return true;if ) return true;if
return roll 

• The UI now clearly indicates who is chatting,who’s fighting,and which icons are currently playing football .
• No emojis – plain text labels avoid ambiguity.
• The same icon remains recognizable – Edge stays Edge,WeChat stays WeChat – only hands/feet are added for animation.

足球也改过最早并不是全员参加

全员参加。奇数也不落单

The original spec required exactly 10 players split evenly – leaving odd numbers idle on sidelines looked weird. I changed it to “all active characters participate”. If total count is odd,left team gets one extra player / 2)) so nobody stays out.

var selected = _characters
.Where
.ToArray;话说回来，if return false;Shuffle，selected = selected.OrderByDescending.ToArray;话说回来，var leftTeamSize = / 2;var rightTeamSize = selected.Length / 2;...

• User Pain:
• - Inconsistent team sizes broke immersion. - Users expected every icon on screen to be involved in events.
\

The scoreboard,team colors and countdown are all generated automatically – users just watch.

After football we added group dance and parade events with identical participation rules.

Performance tip: when many characters are active we cap rendering at 30 FPS and reuse drawing objects – sufficient for a desktop pet without draining CPU.

Doubao-Seed-Evolving 在这个项目里到底做了什么

1. I described brain‑storming idea → model split it into product shapes & risks.
2. I rejected game‑like interactions → model rewrote specs into a pure‑watch tool.
3. I asked for requirement & development plans → model produced two docs which I cross‑checked for errors.
4. A minimal transparent stage prototype was built first → validated mouse‑through & foreground switching.
5. I n implemented desktop scanning。character state machine and auto‑event scheduler.
6. Eeach real‑world run produced screenshots → bugs fed back to model for iterative fixes.
7. \

The tech stack stayed simple: C#,.NET 6+,WPF + Win32 interop + JSON + xUnit – no Unity/Electron or heavy AI SDKs.

The codebase split into two layers:

• Core – manages characters,relationships,cooldowns and event logic.
• App – handles WPF UI,desktop scanning,transparent overlay,tray icon and local persistence.

  This separation let us write unit tests for “all players join football?” or “mediation should not double‑count?” without watching animations.

  Real‑world issues like hidden panels wiping out icons or stale UI strings never showed up in compile‑time checks – y only appeared during live runs and were fixed by feeding concrete feedback back into Doubao‑Seed-Evolving.

  现在这版能做到什么

---