本文来源:https://mp.weixin.qq.com/s/PNn0rJw0I_xRhwFx5-lIkw
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
聊聊背景

这台 Quote/0 是去年买的。当时我很喜欢它的硬件形态:小、薄、安静,放在桌面上像一张永远不会熄的纸。开箱后的新鲜感也确实维持了一阵子。
在之前的周刊里面有提到
i-vox,公众号:i-voxVol.05 奔赴热爱 | 极光与冰雪星露谷物语;Google与AirDrop;喜人与朝圣之路
后来它逐步成了摆件。偶尔放一句话、换一张图、看一眼天气,更多时候只是静静地立在表明器旁边。它的内容能力没有难题,难题在于这些内容和我的一天缺少联系:工作节奏变了,显示屏上还是那张昨天就看腻的卡片。
直到最近听说 MindReset 给 Dot. / Quote/0 做了官方 Skill,我才重新把它拿出来。这个 Skill 同时给出本地 MCP Server 和 Open API,Agent 可以列设备、读状态、查看 Loop 内容、切换内容,或者把文字、图片和 Canvas 直接推到显示屏上。
其中最让我在意的是 Canvas API。它允许 Agent 按当前状态生成结构化界面:字多一点还是少一点,黑底还是白底,时间放大还是任务放大,都可以在发送时决定。
于是我给自己出了一道题:一块已经吃灰一年的墨水屏,能不能变成 AI Agent 的桌面输出端?
答案是做一个 Attention OS。

微信、日历、CI、代码评审、各种 Agent,都在争夺注意力。每个工具都觉得自己很关键,最后的结果是:显示屏越来越多,注意力越来越碎。
这块显示屏不负责展示一整天的待办、所有指标和全部通知。它每次只选一条当前信号,决定该表明什么。告警浮现时给出行动项;会议临近时给出下一场会议;需要深度工作时保留当前唯一目标;任务完成后安静地留下一个 checkpoint。

这个方向让我觉得很有意思。桌面上的小屏终于有了一个更合适的角色:它不再抢占视线,它帮我筛掉不值得抢占视线的东西。
官方 Skill 到底给了什么
官方接口把能力分成了三层:设备、内容和 Canvas。
data、windowData、layoutFull |
Text API 合适一句提醒;Image API 合适固定海报或已有图片;Canvas API 合适我这次的情形,因为它把内容和布局拆开了。数据变化时,卡片跟着变化;优先级变化时,整张卡的视觉语言也能一起切换。
底层请求落在 POST /api/authV2/open/device/:deviceId/canvas。完整字段和设备设置说明可以在 Dot Skill / Open API 官方文档 里查看。

一次只表明一个信号的强规则
所谓 Attention OS,核心不在于信息源有多多元,而在于取舍足够明确。我给它设了一条很朴素的优先级:
CI 失败 / 构建异常 100设备离线 97Agent 卡住、需要输入 94 / 8830 分钟内的下一场会议 92两小时内的下一场会议 78当前深度工作目标 65任务完成后的 checkpoint 58安静待机 0日历、GitHub、Agent 状态、当前主任务都会产生候选信号。策略层只保留优先级最高的一条,再把它转换成合适 296×152 的 Canvas。
日历 / CI / Agent / 当前主任务 ↓ Attention Policy ↓ Canvas API ↓ Quote/0这点很关键。小屏的价值来自克制:同一时刻只有一个结论,且这个结论要能立刻理解、立刻行动。
它是怎么跑起来的
我没有让模型每隔几分钟随意生成一张图。整套流程被拆成了几个稳定步骤:
收集信号 → 排优先级 → 挑选卡片形态 → 生成 Canvas → 校验 → 发送 → 读取实机渲染结果第一步是 Content Studio。要先在 Quote/0 的 Loop 里创建一个 Canvas API 内容项。它相当于设备上预留的“远程画布槽位”。后续每次发送都带着这个内容项对应的 taskKey,设备才知道要更新哪一项内容。
第二步是信号收集。这个版本已经接入了当前主任务和 Agent 状态,也预留了日历、GitHub Actions 的输入口。实际接入时,它们都会被规整成同一种数据:标题、详情、优先级、来源与可选跳转链接。
第三步是策略。它只做一件事:从候选信号中挑出最高优先级的一条。这里我没有让大模型直接裁决,因为“会议还剩多久”“CI 是否失败”“Agent 是否卡住”都有明确规则,用确定性策略更可控,也更容易排查。
第四步才轮到 Canvas。发送给设备的内容由三部分组成:
{"data": { "headline": "Infra Weekly", "detail": "21:35 · 28 分钟后" },"windowData": { "default": ["结构化 UI 树"] },"layoutFull": { "tw": "p-0 bg-white" }}data 是这次要表明的内容,windowData 是布局树,layoutFull 是全局布局规格。换句话说,显示屏并不接收一张预先做好的 PNG;它接收一份可渲染的结构化描述,再由服务端生成适配墨水屏的图片。
最后是两个容易被忽略的保护措施。
四种状态,四种显示屏语言
假设 Android CI 有两个失败用例,显示屏会切到 ATTENTION。标题足够大,原因一句话说清,底部保留行动提示。

这张卡并不试图解释所有日志。完整信息仍然在 GitHub、终端和 IDE 里;墨水屏只负责把“该处理了”送到视线里。
会议进入 30 分钟窗口后,显示屏切成 NEXT UP。时间、会议名、倒计时和会前准备集中在同一块区域。

(当然这只是 showcase 并还未真的在公司用起来)
我很喜欢这种提醒方式。手机推送很容易被划掉,电脑角落的通知也容易被其他窗口盖住。墨水屏没有声音,没有弹窗,它只是在你抬头时把关键信息放在那里。
当天没有会议压力、没有构建异常、Agent 也没有卡住,显示屏进入 FOCUS MODE,只表明当前最关键的一件事。

“写完这篇公众号文章”这种提醒看上去很简单,却很合适放在桌面上。它给人的感觉更接近一张便签:持续存在,又不会像通知一样要求你立刻回应。
哪些是真实的,哪些是演示的
这次实验的设备、Canvas API 调用、内容切换和四张显示屏图片都来自同一台 Quote/0。每次发送后,我都借助设备状态接口取回服务端渲染出的图片,再把它们放进本文。
为了覆盖告警、会议、专注三类视觉状态,截图里的 CI 失败、会议和写文章任务使用了演示信号。它们的作用是验证优先级和布局是否正确,显示屏渲染本身完全走真实链路。
一块好显示屏,应该帮你留下什么
我以前总想把更多数据塞进桌面屏:天气、日历、股票、服务器、待办、使用量……后来发现,信息越完整,越容易变成背景噪音。
这次实验给了我一个更喜欢的答案:桌面屏应该保留经过判断后的结果。
它可以在构建失败时提醒我处理;在会议前留出准备时间;在一切正常时守住手头的任务;在工作完成后告诉我可以收尾。显示屏很小,反而逼着平台做出挑选。
如果把 AI 当作一个会执行任务的协作者,Quote/0 很合适做它的物理输出端。它留在桌面上,安静、可见、低干扰。下一步我会接入真实日历、GitHub Actions 和 Codex 任务状态,让这块显示屏真正跟着一天的工作节奏变化。
对我来说,这才是 AI 和硬件结合最舒服的一种形态:少一点打扰,多一点恰到好处的浮现。
最后
如果你也对此感兴趣,也可以自己上手 DIY 一下:
设备:Quote/0(296×152 墨水屏)接口:MindReset Dot Skill / Canvas API项目:Attention OS,按优先级把实时工作信号投射到桌面显示屏。
Source: https://mp.weixin.qq.com/s/PNn0rJw0I_xRhwFx5-lIkw. Contact for copyright: 13951431913#139.com or comment.
Background
I bought Quote/0 last year. I loved its hardware — small, thin, quiet — like a never-fading piece of paper on my desk. The initial excitement lasted a while.
As mentioned in previous issues
i-vox, WeChat ID: i-voxVol.05 — Aurora, ice, Stardew; Google, AirDrop; Pilgrimage, joy.
But soon it became an ornament. Occasionally I’d add a line, change a picture, check weather — mostly just sitting there next to my clock. Its content capability wasn’t the problem; the problem was that the content didn’t connect with my daily rhythm. Work shifted, but the screen still showed yesterday’s stale card.
Recently, I heard MindReset created an official Skill for Dot. / Quote/0. It offers both local MCP server and open API, letting agents list devices, read states, view Loop content, switch content, or push text/image/canvas directly to the screen.
What excites me most is the Canvas API. It allows the agent to generate structured displays based on current state — more text or less, dark or light background, enlarged time or tasks — all determined at send-time.
So I posed a question to myself: can a year-old dusty e-reader become the desktop output for an AI agent?
The answer is to build an Attention OS.
WeChat, calendar, CI, code reviews, various agents — all vying for attention. Each claims importance, but the result is fragmented focus and cluttered screens.
This little screen doesn’t display everything: all tasks, metrics, notifications. It selects one signal at a time — alerting when action’s needed, reminding of next meeting, preserving the current goal during deep work, leaving a checkpoint after completion.
This role feels right: it doesn’t steal attention; it filters out what doesn’t deserve it.
What the official Skill actually provides
The official interface breaks down into three layers: device, content, and canvas.
Layer
Capabilities
My use in this project
Device
List devices, read state, R/W settings, list timezones
Verify device online, Wi-Fi/power status before startup
Content
View Loop tasks, switch next, push text/image
Locate pre-created Canvas API content item
Canvas
Validate, send data, windowData, layoutFull
Generate four layouts based on live work signals
The Text API suits single reminders; Image API fits static posters; Canvas API suits my case — separating content from layout. When data changes, the card adapts; when priority shifts, the entire visual language changes with it.
The underlying request is POST /api/authV2/open/device/:deviceId/canvas. Full field descriptions and device settings are in the Dot Skill / Open API docs.
Communicate Only One Signal at a Time
The core of Attention OS is not about how diverse the information sources are, but about making sufficiently clear choices. I set a very simple priority rule for it:
Candidates for signals include calendar, GitHub, agent status, and the current main task. The strategy layer only keeps the highest-priority one and then converts it into a suitable 296×152 canvas.
This is critical. The value of a small screen lies in restraint: there is only one conclusion at any given time, and that conclusion must be immediately understandable and actionable.
How Does It Run
I didn’t ask the model to randomly generate a plot every few minutes. The entire process was broken down into several stable steps:
The first step is Content Studio. You must first create a Canvas API content item in the Loop of Quote/0. This acts as a reserved “remote canvas slot” on the device. Subsequent sends include the taskKey of this content item, so the device knows what to update.
The second step is signal collection. This version already integrates the current main task and agent status, and has reserved inputs for calendar and GitHub Actions. In actual implementation, all of them will be normalized into a single data format: title, details, priority, source, and optional jump link.
The third step is strategy. It does one thing: pick the highest-priority signal from the candidates. I didn’t let a large model make the decision here because things like “how much time is left in the meeting” and “did the CI fail” have clear rules; using a deterministic strategy is more controllable and easier to debug.
The fourth step is Canvas. The content sent to the device consists of three parts:
data — the content to display; windowData — the layout tree; layoutFull — the global layout specification. In other words, the display screen does not receive a pre-made PNG; it receives a renderable, structured description which the server then converts into an image optimized for the e-ink display.
Finally, there are two protective measures:
Four States, Four Display Modes
Suppose Android CI has two failed cases — the screen switches to ATTENTION. The title is large enough, the reason is stated in one line, and the bottom retains an action prompt.
This card doesn’t try to explain all the logs. The full information remains in GitHub, the terminal, and the IDE; the e-ink screen only serves to push “you need to handle this” into view.
Once a meeting enters the 30-minute window, the screen switches to NEXT UP. The time, meeting name, countdown, and pre-meeting preparation are all concentrated in one area.
(This is just a showcase; it hasn’t actually been used in the company yet.)
I like this reminder method. Mobile notifications are easily dismissed, and computer corner notifications are often covered by other windows. The e-ink screen makes no noise, has no pop-ups — it just puts the critical information there when you look up.
On days without meeting pressure, build failures, or stuck agents, the screen enters FOCUS MODE, displaying only the most critical thing at hand.
Reminders like “Finish this blog post” seem simple, but are perfect for the desktop. They feel more like a physical note: persistent, yet not demanding immediate response like a notification.
What’s Real, What’s Demo
In this experiment, the device, Canvas API calls, content switching, and the four display images all come from the same Quote/0. After each send, I retrieve the server-rendered image via the device status interface and include it in this article.
To cover alarm, meeting, and focus visuals, the CI failure, meeting, and writing-article tasks in the screenshots use demo signals. Their purpose is to verify priority and layout correctness — the rendering itself is fully on the real pipeline.
What a Good Display Should Leave You With
I used to want to cram more data into the desktop: weather, calendar, stocks, servers, todo lists, usage stats... but I found that the more information, the more it becomes background noise.
This experiment gives me a better answer: the desktop screen should hold the result of a judgment.
It should remind me to handle a build failure, leave me time before a meeting, keep me locked onto the current task when everything is normal, and tell me I’m done when the work is finished. The small size of the screen forces the platform to make selections.
If we think of AI as a collaborator that executes tasks, Quote/0 is a perfect physical output terminal. It sits on the desk, quiet, visible, low-interference. The next step is to connect it to real calendars, GitHub Actions, and Codex task statuses, letting the display truly follow the rhythm of the day.
To me, this is the most comfortable form of AI and hardware integration: less distraction, more timely emergence.
最后
If you're interested in this, you can also try DIYing it yourself:
Device: Quote/0 (296×152 E-ink screen) Interface: MindReset Dot Skill / Canvas API Project: Attention OS, project real-time work signals to the desktop display screen according to priority.