DeskMind 0.4.1

DeskMind for Mac 0.4.1 fixes what the first week turned up: setting up on a new Mac, a document that kept growing, keyboard layouts other than US, and how far an approval reaches.

Fixes

• Screen Recording is recognised as soon as you turn it on. On a Mac setting DeskMind up for the first time, the setup could keep saying Screen Recording was off after it was turned on, and quitting and reopening DeskMind didn't help. macOS's own "Quit & Reopen" had started a copy of DeskMind Hands whose permission didn't count. DeskMind now always runs the right copy, and restarts it by itself once the permission is granted.
• Writing into a document no longer loops. After writing and saving the right text, the model could choose to "replace" part of it with the whole text, and DeskMind spliced that in, so the document nested inside itself and every later step wrote a longer copy. Such a step now rewrites the field; when the field already says that, nothing changes, and the task finishes. On our benchmark task for this (G04), 3 of 3 runs now finish in 4 steps. Before, all 6 runs we have stayed in that loop until the 20-step limit (5 of them happened to stop on the right text). The whole diag suite passes 39 of 39 (diag-v27, run on the same runtime as 0.4.1-rc.2).
• An approval covers one step. When DeskMind asks before a step that may not be undoable (deleting, sending, paying, publishing, sharing), your yes covers that step and the confirmation it opens. Before, it could also cover another step of the same kind a step or two later in the same app: approving one deletion could let a second, different item be deleted without asking.
• DeskMind acts only on what it offered the model. Every answer is checked against the options DeskMind actually offered. An answer outside them stops the run with the reason, instead of being turned into one of the options.
• On a French (AZERTY) or Dvorak keyboard, select-all and paste press the right keys. In apps read from the screen, DeskMind used the key positions of a US keyboard: on AZERTY, cmd+A could arrive as cmd+Q and quit the app; on Dvorak, cmd+V was cmd+K. The keys now come from your keyboard layout. Checked on a Mac with the French layout.
• No crash when the helper is stopped while quitting. DeskMind Hands could crash if macOS stopped it at the moment it was quitting, and leave the model server running. It now stops cleanly and takes the model server with it.

New

• A Restart button for the background helper, in the setup.
• Save Full Log… next to "Report on GitHub" after a failed run. It saves a .zip on your Mac with the run's steps, a screenshot of each step, the model servers' logs and the error. It holds what was on your screen, file names and your account name, so look through it before you share it; DeskMind sends nothing.
• Reports open a GitHub form already filled in: what you asked, what happened (with the steps and a folded diagnostics section), the app version and macOS. Check it and submit it yourself.
• View › Smooth Recordings (30 fps, Slows Tasks About 10%). Recording a task now takes a picture ten times a second instead of running a capture stream, which makes each decision about 3% slower instead of about 10%. Turn this on for footage where smooth motion matters.

Changes

• The app follows your system language unless you've chosen one in the language menu. It used to start in English everywhere.
• The whole setup fits on screen: the examples appear once setup is done, and long descriptions are no longer cut off.
• A mistake in the apps file (~/.config/deskmind/apps.yaml, which you write only to configure apps by hand) is reported by name.
• The models are the same as in 0.4.0, and models already downloaded are not downloaded again.

Known

• In Notes, DeskMind can't yet select a note in the list, so tasks on a single note don't work there (deskmind#49).

Requirements

A Mac with Apple Silicon and macOS 15 or later.

---

中文说明

DeskMind for Mac 0.4.1 修复了第一周发现的问题：新 Mac 上的设置、越写越长的文档、非美式键盘布局，以及一次批准能管多远。

修复

• 打开「屏幕录制」后马上生效。 在第一次设置 DeskMind 的 Mac 上，打开屏幕录制后，设置里可能仍显示未开启，退出重开 DeskMind 也没用：macOS 自带的「退出并重新打开」启动了一个权限不算数的 DeskMind Hands 副本。现在 DeskMind 总是运行正确的那一份，并在权限打开后自动重启它。
• 往文档里写内容不再死循环。 以前写好并保存正确内容后，模型可能选择把其中一部分「替换」成整段文字，DeskMind 照做，文档就嵌套进了自己，之后每一步都写一份更长的副本。现在这种操作按「整段改写」处理；内容已经是这样时不做改动，任务随之完成。在专测这个问题的评测任务（G04）上，现在 3 次运行都在 4 步内完成；以前我们记录的 6 次运行都陷在这个循环里，直到用满 20 步（其中 5 次碰巧停在正确内容上）。整套 diag 评测 39 题全部通过（diag-v27，与 0.4.1-rc.2 相同的运行时）。
• 一次批准只管一步。 DeskMind 在可能无法撤销的步骤（删除、发送、付款、发布、分享）之前会先问你。你的同意只覆盖这一步，以及这一步自己弹出的确认框。以前，同一个 App 里隔一两步的另一个同类步骤也可能沿用这次批准：批准删除一项后，另一项可能不经询问就被删掉。
• DeskMind 只执行它给模型提供过的选项。 每个回答都会和实际提供的选项核对。超出范围的回答会让这次运行停下并说明原因，不会被硬套成某个选项。
• 使用法语（AZERTY）或 Dvorak 键盘时，全选和粘贴会按对键。 对靠屏幕识别的应用，DeskMind 以前按美式键盘的键位发送：在 AZERTY 上 ⌘A 可能变成 ⌘Q，直接退出应用；在 Dvorak 上 ⌘V 变成了 ⌘K。现在按你的键盘布局找对应的键。已在法语键盘布局的 Mac 上实测。
• 后台助手在退出时被系统停止，不再崩溃。 以前 DeskMind Hands 正在退出时如果被 macOS 停止，可能崩溃，并留下没关掉的模型服务。现在会正常退出，并把模型服务一起关掉。

新功能

• 设置里的「后台助手」一行新增「重启」按钮。
• 运行失败后，「在 GitHub 上反馈」旁边新增「保存完整日志…」。 在你的 Mac 上存一个 .zip，里面有这次运行的步骤、每一步的截图、模型服务的日志和错误原文。它包含你屏幕上的内容、文件名和账户名，分享前请先检查；DeskMind 不会发送任何内容。
• 反馈会打开已填好的 GitHub 表单： 你让它做什么、实际发生了什么（含步骤和折叠的诊断信息）、App 版本和 macOS 版本。你检查后自己提交。
• 「显示 › 流畅录制（每秒 30 帧，任务约慢 10%）」。 录制任务现在每秒截 10 张图，不再开持续录屏流，每次决策约慢 3%（录屏流约 10%）。需要画面流畅的录像时再打开这个选项。

其他

• App 默认跟随系统语言，除非你在语言菜单里选过。以前在任何系统上都默认英文。
• 设置页能完整显示： 示例在设置完成后才出现，较长的说明也不再被截断。
• ~/.config/deskmind/apps.yaml（只在手动配置应用时才需要写）里写错的字段，现在会直接指出来。
• 模型与 0.4.0 相同，已经下载的模型不会重复下载。

已知问题

• 在「备忘录」里，DeskMind 暂时还不能在列表里选中某条备忘录，所以针对单条备忘录的任务做不了（deskmind#49）。

系统要求：Apple 芯片的 Mac，macOS 15 或更新版本。

DeskMind 0.4.1-rc.3

Pre-release of 0.4.1, for testing before it becomes the latest version. 0.4.0 stays the latest until then.

Since rc.2

• An approval covers one step. When DeskMind asks before a step that may not be undoable (deleting, sending, paying, publishing, sharing), your yes covers that step and the confirmation it opens. Before, it could also cover another step of the same kind a step or two later in the same app: approving one deletion could let a second, different item be deleted without asking.
• DeskMind acts only on what it offered the model. It now checks every answer against the options it actually offered. An answer outside them stops the run with the reason, instead of being turned into one of the options.

Everything in rc.2 is included; see the 0.4.1-rc.2 notes.

---

中文说明

0.4.1 的预发布版，转为正式版之前供测试。在此之前，正式版仍是 0.4.0。

相比 rc.2

• 一次批准只管一步。 DeskMind 在可能无法撤销的步骤（删除、发送、付款、发布、分享）之前会先问你。你的同意只覆盖这一步，以及这一步自己弹出的确认框。以前，同一个 App 里隔一两步的另一个同类步骤也可能沿用这次批准：批准删除一项后，另一项可能不经询问就被删掉。
• DeskMind 只执行它给模型提供过的选项。 现在每个回答都会和实际提供的选项核对。超出范围的回答会让这次运行停下并说明原因，不会被硬套成某个选项。

rc.2 的全部内容都包含在内，见 0.4.1-rc.2 的说明。

DeskMind 0.4.1-rc.2

Pre-release of 0.4.1, for testing before it becomes the latest version. 0.4.0 stays the latest until then.

Since rc.1

• Screen Recording is recognised as soon as you turn it on. On a Mac setting DeskMind up for the first time, the setup could keep saying Screen Recording was off after it was turned on, and quitting and reopening DeskMind didn't help. macOS's own "Quit & Reopen" had started a copy of DeskMind Hands whose permission didn't count. DeskMind now always runs the right copy, and restarts it by itself once the permission is granted.
• The background helper has a Restart button in the setup.
• The app follows your system language unless you've chosen one in the language menu. It used to start in English everywhere.
• The whole setup fits: the examples appear once setup is done, so the last setup step is no longer pushed out of sight, and long descriptions are no longer cut off.

Everything in rc.1 is included; see the 0.4.1-rc.1 notes.

---

中文说明

0.4.1 的预发布版，转为正式版之前供测试。在此之前，正式版仍是 0.4.0。

相比 rc.1

• 打开「屏幕录制」后马上生效。 在第一次设置 DeskMind 的 Mac 上，打开屏幕录制权限后，设置里可能仍显示未开启，退出重开 DeskMind 也没用：macOS 自带的「退出并重新打开」启动了一个权限不算数的 DeskMind Hands 副本。现在 DeskMind 总是运行正确的那一份，并在权限打开后自动重启它。
• 设置里的「后台助手」一行新增了「重启」按钮。
• App 默认跟随系统语言，除非你在语言菜单里选过。以前在任何系统上都默认英文。
• 设置能完整显示： 示例在设置完成后才出现，最后一步设置不会再被挤到看不见的地方，较长的说明也不再被截断。

rc.1 的全部内容都包含在内，见 0.4.1-rc.1 的说明。

DeskMind 0.4.1-rc.1

Pre-release of 0.4.1, for testing before it becomes the latest version. 0.4.0 stays the latest until then.

Fixes

• Chinese text written into a document no longer loops. After writing and saving the right text, the planner could choose to "replace" one number in it with the whole text, and DeskMind spliced that in, so the document nested inside itself and every later step rewrote a longer copy. Now such a step counts as rewriting the field; when the field already says that, nothing changes, and the task finishes. On our benchmark task for this (G04), 3 of 3 runs now finish in 4 steps. Before, all 6 runs we have stayed in that loop until the 20-step limit (5 of them happened to stop on the right text). The whole diag suite passes 39 of 39 (diag-v27, run on the same runtime as this pre-release's successor rc.2).
• On a French (AZERTY) or Dvorak keyboard, select-all and paste press the right keys. In apps read from the screen, DeskMind pastes with cmd+A and cmd+V, sent as the key positions of a US keyboard. On AZERTY that position types Q, so cmd+A could quit the app; on Dvorak, cmd+V was cmd+K. The keys now come from your keyboard layout.
• No crash when the helper is stopped while quitting. DeskMind Hands could crash if macOS stopped it at the moment it was quitting, and leave the model server running. It now stops cleanly and takes the model server with it.

New

• Save Full Log… next to "Report on GitHub" after a failed run. It saves a .zip on your Mac with the run's steps, a screenshot of each step, the model servers' logs and the error. It holds what was on your screen, file names and your account name, so look through it before you share it; DeskMind sends nothing.
• Reports open a GitHub form already filled in: what you asked, what happened (with the steps and a folded diagnostics section), the app version and macOS. Check it and submit it yourself.
• View › Smooth Recordings (30 fps, Slows Tasks About 10%). Recording a task now takes a picture ten times a second instead of running a capture stream. On 40 real decisions of the local model, recording made each decision about 3% slower, against about 10% with the stream. Turn this on for footage where smooth motion matters.

Changes

• A mistake in the apps file (~/.config/deskmind/apps.yaml, which you write only to configure apps by hand) is now reported by name ("unknown keys: groundings…"), instead of a generic error.
• "It guessed instead of asking" reports link the issue that collects those cases.
• For developers running DeskMind Hands from a terminal: runs no longer leave Dock tiles behind.
• The models are the same as in 0.4.0, and models already downloaded are not downloaded again.

Requirements

A Mac with Apple Silicon and macOS 15 or later.

---

中文说明

0.4.1 的预发布版，转为正式版之前供测试。在此之前，正式版仍是 0.4.0。

修复

• 往文档里写中文不再死循环。 以前写好并保存正确内容后，模型可能选择把其中一个数字「替换」成整段文字，DeskMind 照做，文档就嵌套进了自己，之后每一步都改写一份更长的副本。现在这种操作按「整段改写」处理；内容已经是这样时不做改动，任务随之完成。在专测这个问题的评测任务（G04）上，现在 3 次运行都在 4 步内完成；以前我们记录的 6 次运行都陷在这个循环里，直到用满 20 步（其中 5 次碰巧停在正确内容上）。整套 diag 评测 39 题全部通过（diag-v27，与 rc.2 相同的运行时）。
• 使用法语（AZERTY）或 Dvorak 键盘时，全选和粘贴会按对键。 对靠屏幕识别的应用，DeskMind 用 cmd+A、cmd+V 粘贴，以前按的是美式键盘上的键位。在 AZERTY 上那个键位是 Q，cmd+A 可能直接退出应用；在 Dvorak 上 cmd+V 变成了 cmd+K。现在按你的键盘布局找对应的键。
• 后台助手在退出时被系统停止，不再崩溃。 以前 DeskMind Hands 正在退出时如果被 macOS 停止，可能崩溃，并留下没关掉的模型服务。现在会正常退出，并把模型服务一起关掉。

新功能

• 运行失败后，「在 GitHub 上反馈」旁边新增「保存完整日志…」。 在你的 Mac 上存一个 .zip，里面有这次运行的步骤、每一步的截图、模型服务的日志和错误原文。它包含你屏幕上的内容、文件名和账户名，分享前请先检查；DeskMind 不会发送任何内容。
• 反馈会打开已填好的 GitHub 表单： 你让它做什么、实际发生了什么（含步骤和折叠的诊断信息）、App 版本和 macOS 版本。你检查后自己提交。
• 「显示 › 流畅录制（每秒 30 帧，任务约慢 10%）」。 录制任务现在每秒截 10 张图，不再开持续录屏流。在本地模型的 40 次真实决策上实测，开着录制时每次决策慢约 3%（录屏流约 10%）。需要画面流畅的录像时再打开这个选项。

其他

• 应用配置文件（~/.config/deskmind/apps.yaml，只有手动配置应用时才会写）里有写错的字段时，现在会直接指出是哪个字段，而不是笼统地报错。
• 「它没问就猜了」的反馈会链接到收集这类情况的 issue。
• 面向开发者：在终端里运行 DeskMind Hands 时，不再在 Dock 里留下图标。
• 模型与 0.4.0 相同，已下载的模型不会重新下载。

运行要求：Apple Silicon 芯片的 Mac，macOS 15 或更高版本。

DeskMind 0.4.0

DeskMind for Mac 0.4.0 shows you the window a task is working in, live, while it works in the background.

New: a live view of the task

• See what DeskMind is doing, wherever the window is. A small card in a corner of the screen shows the window the task is working in, live, even when that window is behind your other windows. Under the picture is the step it is taking, in words, and a dot shows where each step acted. In the title bar, 小方 (DeskMind’s little frame) watches where DeskMind is working, looks up when it needs you, and smiles when it’s done.
• It stays out of your way. The card moves to another corner when it would cover the window being worked in, never takes the keyboard, and doesn't come to the front. Over an app that fills the screen it sits in the corner farthest from where DeskMind is clicking and lets clicks pass through; rest the pointer on it for a moment to use its buttons. Drag it to any corner and it stays there for later tasks. Double-click it to make it larger, or collapse it to a small capsule.
• It tells you when it needs you. When the task asks a question or wants your approval, the card says so; click it to answer in DeskMind. Stop is on the card too. When the task ends, the card says how it ended, then goes.
• Only the task's window. The card shows the window the task works in and nothing else. If that window is minimized or closed, the card says so; it never shows another window of the same app instead. The card is an ordinary window: your own screenshots and screen sharing show it. The models never see it, and task recordings leave it out (recording the whole screen keeps it).
• Turn it off in View › Show a Live View of the Task.

Also new

• Tell us when it gets it wrong. After a task, 👎 asks what went wrong ("It guessed instead of asking", "Wrong result", "It got stuck") and opens a GitHub issue filled in with your instruction, the result and the steps in words. It includes only your instruction, how the task ended, the number and kinds of steps, and the versions, nothing read from your screen. You check it and submit it in your browser; DeskMind sends nothing itself. A failed run has "Report on GitHub" too, and Help › Report an Issue is always there.
• You see the result. When a task gives an answer, or your first task finishes, the DeskMind window comes back with it (without taking the keyboard). A notification also comes when the window is out of the way.
• All three examples are visible under the instruction box, one per line.

Fixes

• Organising a large folder works. On a real Downloads folder, "整理目录" stopped at once: the model was handed every file's list of every subfolder, more than it could take ("didn't answer in time", or out of memory). It is now shown a bounded part at a time, and gets up to 2 minutes a step on slower Macs.
• Report on GitHub includes diagnostics (the error, step times, versions, the Mac's model, chip and memory), folded at the end, with no screen content, file names or paths.
• Organising a folder with subfolders no longer fails at once. "整理目录" on a Downloads folder with subfolders ended before its first step, saying the local model "didn't answer in time": one question was too large for the model. Questions now stay within its limits, and when the model does refuse a step DeskMind says so, and why.
• A rename in your own folder waits for your approval, as deleting and sending do (the sample folder is not affected). An app the instruction names that isn't on this Mac is named before the run starts. The file example always names a file the folder really has. In Self-test, the window comes back after a task, picking another shows Start, and task names follow the app's language.
• A question no longer catches what you type elsewhere. If you were typing in another app when DeskMind asked a question, your keys (Return included) could go into the answer and send it. Now the window comes back without taking the keyboard while you type.
• A task in an app that isn’t open, or whose window is closed, no longer ends at once. DeskMind saw no window, and the task could end as done after one step, having done nothing (seen with NetEase Cloud Music). Now the app is opened, or its window brought back, before the task starts, without bringing it to the front.
• The file example names the file the sample folder actually holds, after a language switch too.
• The vision model is described as optional on the confirmation sheet, and the "Don't ask again" box no longer shows where it had no effect.
• Clearer words: "Didn't finish" in Chinese reads 没做完; the island says "Starting" instead of "Hands"; "Try examples" (the six diagnostics) is now "Self-test".
• Recorded runs no longer show the purple "being shared" mark in place of the task windows' traffic lights, and recordings no longer include the menu bar's status icons.
• The GitHub line after your first finished task now links to the DeskMind repository, waits until you have seen the result, and comes back after a 👍 until you have followed it once. Home has a GitHub link at the bottom.

Changes

• The models are the same as in 0.3.x (DeskMind Brain G18b, routing threshold 0.96; Eyes-4B at de8e13b), so the published benchmark numbers still apply, and models already downloaded are not downloaded again.
• The app's source now lives in deskmind-ai/deskmind (app/). This release was built and notarized by CI from the source commits pinned in app/deps.lock.

Requirements

A Mac with Apple Silicon and macOS 15 or later. The planner models (about 5.3 GB) are downloaded on first launch, from Hugging Face or, if that is slow, from ModelScope. The optional vision model needs another 3.3 GB. The live view uses the Screen Recording permission DeskMind already has; macOS shows its screen-recording indicator in the menu bar while a task runs with the live view on.

Known limitations

• Apps with no accessibility tree are read from the screen, and are briefly brought to the front while a step acts. The vision model as the app runs it (4-bit, up to 2 megapixels) scores 50.9 on ScreenSpot-Pro (67.7 on a GPU at full resolution), and small icons are its weak spot.
• Chrome's address bar is not recognised as the browser's own field; the guard that keeps table rows out of it works in Safari.
• Copying more than about four rows in a given order is less reliable than short tables.

---

中文说明

DeskMind for Mac 0.4.0：任务在后台运行时，可以实时看到它正在操作的窗口。

新功能：任务实时画面

• 无论窗口在哪，都能看到 DeskMind 在做什么。 屏幕角落的小卡片会实时显示任务正在操作的窗口，即使这个窗口被你的其他窗口挡住也一样。画面下方用文字写出当前这一步，画面里的小圆点标出每一步落在哪里。标题栏里的小方会看着 DeskMind 正在操作的位置，需要你时抬头看你，完成时眯眼笑。
• 不挡你的事。 如果卡片会挡住正在操作的窗口，它会自动换到另一个角落；它不抢键盘，也不会跑到最前面。应用占满屏幕时，卡片停在离 DeskMind 点击位置最远的角落，点击会穿透过去；鼠标在卡片上停一下，就能用它的按钮。把它拖到任意角落，之后的任务都会留在那里。双击可以放大，也可以收起成一个小胶囊。
• 需要你时会告诉你。 任务提问或需要你批准时，卡片会提示；点一下卡片就能在 DeskMind 里回答。卡片上也有停止按钮。任务结束时，卡片会显示结果，然后消失。
• 只显示任务的窗口。 卡片只显示任务正在操作的那个窗口。窗口被最小化或关闭时，卡片会说明，绝不会改为显示同一应用的其他窗口。卡片是普通窗口：你自己截图、共享屏幕时能看到它；模型看不到它，任务录像也不包含它（录制整个屏幕时会包含）。
• 可以在「显示 › 显示任务窗口的实时画面」里关闭。

其他新功能

• 做错时告诉我们。 任务结束后点 👎，会问哪里不对（「它没问就猜了」「结果不对」「卡住了」），然后打开一个已填好的 GitHub issue 页面（只含你的指令、任务结果、步骤数量和类型、版本号，不含任何屏幕内容），由你在浏览器里检查后提交；DeskMind 自己不发送任何内容。运行出错时也有「在 GitHub 上反馈」，菜单「帮助 › 反馈问题」随时可用。
• 能看到结果。 任务给出答案，或你的第一个任务完成时，DeskMind 窗口会带着结果回来（不抢键盘）；窗口被移开时也会发通知。
• 三个示例都能看到，在输入框下方各占一行。

修复

• 整理大文件夹不再失败。 以前在真实的「下载」里说「整理目录」会马上停下：模型一次收到了每个文件对应的全部子文件夹列表，超出了它能处理的量（提示「没及时回答」或内存不足）。现在每次只给它有限的一部分，慢一些的 Mac 上每步最多等 2 分钟。
• 在 GitHub 上反馈时会附带诊断信息（报错、每步耗时、版本、Mac 型号、芯片和内存），折叠在 issue 末尾，不含屏幕内容、文件名或路径。
• 整理带子文件夹的目录不会再一开始就失败。 以前在有子文件夹的「下载」里说「整理目录」，第一步之前就结束，还提示本地模型「没及时回答」：其实是一道题太大，被模型拒绝了。现在题目不会超出模型的限制；模型真的处理不了某一步时，会如实说明原因。
• 在你自己的文件夹里改名，会先征得你同意，和删除、发送一样（示例文件夹不受影响）。指令里提到的应用没安装时，开始前就会告诉你。文件示例只写文件夹里真实存在的文件。自检里一个任务跑完后窗口会回来，选别的任务时按钮显示「开始」，任务名跟随 App 语言。
• 提问不会再吃掉你在别处打的字。 以前 DeskMind 提问时如果你正在别的应用里打字，按键（包括回车）可能落进答案并直接发送。现在你打字时窗口回来但不抢键盘。
• 应用没打开、或窗口被关掉时，任务不会再一步就「完成」。 以前 DeskMind 看不到窗口，可能走一步就判定完成，实际什么都没做（网易云音乐上出现过）。现在开始前会先打开应用或找回它的窗口，且不会把它切到前台。
• 切换语言后，文件示例也会写示例文件夹里真实存在的文件名。
• 确认页把视觉模型写明为可选；不起作用的「不再询问」勾选框不再出现。
• 用词更清楚：中文「Didn't finish」改为「没做完」；顶部状态不再显示「Hands」而是「启动中」；「试试示例」（6 个自检任务）改名为「自检」。
• 录制的任务里，窗口左上角不再出现紫色的「正在共享」标记，录像里也不再出现菜单栏的状态图标。
• 第一次有任务完成后出现的 GitHub 那一行，现在链接到 DeskMind 仓库，要等你看到结果才出现；点 👍 后也会出现，直到你点过一次。首页底部加了 GitHub 链接。

其他

• 模型与 0.3.x 相同（DeskMind Brain G18b，路由门槛 0.96；Eyes-4B de8e13b），已公布的评测成绩依然适用，已下载的模型不会重新下载。
• App 的源码现在位于 deskmind-ai/deskmind 的 app/ 目录。本版本由 CI 按 app/deps.lock 固定的源码提交构建并公证。

运行要求：Apple Silicon 芯片的 Mac，macOS 15 或更高版本。首次启动会下载约 5.3 GB 的规划模型（Hugging Face 慢时自动改用 ModelScope）；可选的视觉模型另需 3.3 GB。实时画面使用 DeskMind 已有的「屏幕录制」权限；开启实时画面运行任务时，菜单栏会显示系统的屏幕录制指示。

DeskMind 0.4.0-rc.3

Pre-release of 0.4.0, for testing before it becomes the latest version. 0.3.1 stays the latest until then.

Since rc.2

• Organising a large folder no longer fails. On a real Downloads folder, "整理目录" still stopped at once: the local model was handed every file's list of every subfolder and ran out of memory ("metal::malloc … greater than the maximum allowed buffer size"). DeskMind now shows it a bounded part at a time, so a folder of 150 files and 12 subfolders is sorted step by step (about 18 s a step on an M4 Pro).
• The local model gets up to 2 minutes for a step, instead of 1, on slower Macs.
• Report on GitHub now includes diagnostics, folded at the end of the issue: the error, how long each step took, the shape of a request that failed, the versions, and the Mac's model, chip and memory. It has no screen content, file names or paths, and you can read and edit it before you submit.

Since rc.1 (in rc.2)

• Organising a folder with subfolders no longer fails at once. "整理目录" on a Downloads folder with subfolders ended before its first step, saying the local model "didn't answer in time". Each file's "move to" list of every subfolder made one question too large for the model, and the request was refused. Questions now stay within the model's limits, and moves into subfolders work.
• A rename in your own folder waits for your approval, as deleting and sending do. It cannot be undone, and an unclear instruction could otherwise rename files to something you didn't want. The sample folder is not affected.
• An app that isn't on this Mac is named before the run starts ("WeChat isn't installed on this Mac…"), instead of the run trying without it.
• The file example always names a file the folder really has. If the attached folder has no file it could name, the example isn't shown.
• Self-test: after a task finishes, the window comes back, and picking another task shows Start. Task names follow the app's language.
• When the local model refuses a step, DeskMind now says that, and why, instead of saying it didn't answer in time.

DeskMind for Mac 0.4.0 shows you the window a task is working in, live, while it works in the background.

New: a live view of the task

• See what DeskMind is doing, wherever the window is. A small card in a corner of the screen shows the window the task is working in, live, even when that window is behind your other windows. Under the picture is the step it is taking, in words, and a dot shows where each step acted. In the title bar, 小方 (DeskMind’s little frame) watches where DeskMind is working, looks up when it needs you, and smiles when it’s done.
• It stays out of your way. The card moves to another corner when it would cover the window being worked in, never takes the keyboard, and doesn't come to the front. Over an app that fills the screen it sits in the corner farthest from where DeskMind is clicking and lets clicks pass through; rest the pointer on it for a moment to use its buttons. Drag it to any corner and it stays there for later tasks. Double-click it to make it larger, or collapse it to a small capsule.
• It tells you when it needs you. When the task asks a question or wants your approval, the card says so; click it to answer in DeskMind. Stop is on the card too. When the task ends, the card says how it ended, then goes.
• Only the task's window. The card shows the window the task works in and nothing else. If that window is minimized or closed, the card says so; it never shows another window of the same app instead. The card is an ordinary window: your own screenshots and screen sharing show it. The models never see it, and task recordings leave it out (recording the whole screen keeps it).
• Turn it off in View › Show a Live View of the Task.

Also new

• Tell us when it gets it wrong. After a task, 👎 asks what went wrong ("It guessed instead of asking", "Wrong result", "It got stuck") and opens a GitHub issue filled in with your instruction, the result and the steps in words. It includes only your instruction, how the task ended, the number and kinds of steps, and the versions, nothing read from your screen. You check it and submit it in your browser; DeskMind sends nothing itself. A failed run has "Report on GitHub" too, and Help › Report an Issue is always there.
• You see the result. When a task gives an answer, or your first task finishes, the DeskMind window comes back with it (without taking the keyboard). A notification also comes when the window is out of the way.
• All three examples are visible under the instruction box, one per line.

Fixes

• A question no longer catches what you type elsewhere. If you were typing in another app when DeskMind asked a question, your keys (Return included) could go into the answer and send it. Now the window comes back without taking the keyboard while you type.
• A task in an app that isn’t open, or whose window is closed, no longer ends at once. DeskMind saw no window, and the task could end as done after one step, having done nothing (seen with NetEase Cloud Music). Now the app is opened, or its window brought back, before the task starts, without bringing it to the front.
• The file example names the file the sample folder actually holds, after a language switch too.
• The vision model is described as optional on the confirmation sheet, and the "Don't ask again" box no longer shows where it had no effect.
• Clearer words: "Didn't finish" in Chinese reads 没做完; the island says "Starting" instead of "Hands"; "Try examples" (the six diagnostics) is now "Self-test".
• Recorded runs no longer show the purple "being shared" mark in place of the task windows' traffic lights, and recordings no longer include the menu bar's status icons.
• The GitHub line after your first finished task now links to the DeskMind repository, waits until you have seen the result, and comes back after a 👍 until you have followed it once. Home has a GitHub link at the bottom.

Changes

• The models are the same as in 0.3.x (DeskMind Brain G18b, routing threshold 0.96; Eyes-4B at de8e13b), so the published benchmark numbers still apply, and models already downloaded are not downloaded again.
• The app's source now lives in deskmind-ai/deskmind (app/). This release was built and notarized by CI from the source commits pinned in app/deps.lock.

Requirements

A Mac with Apple Silicon and macOS 15 or later. The planner models (about 5.3 GB) are downloaded on first launch, from Hugging Face or, if that is slow, from ModelScope. The optional vision model needs another 3.3 GB. The live view uses the Screen Recording permission DeskMind already has; macOS shows its screen-recording indicator in the menu bar while a task runs with the live view on.

Known limitations

• Apps with no accessibility tree are read from the screen, and are briefly brought to the front while a step acts. The vision model as the app runs it (4-bit, up to 2 megapixels) scores 50.9 on ScreenSpot-Pro (67.7 on a GPU at full resolution), and small icons are its weak spot.
• Chrome's address bar is not recognised as the browser's own field; the guard that keeps table rows out of it works in Safari.
• Copying more than about four rows in a given order is less reliable than short tables.

---

中文说明

DeskMind for Mac 0.4.0：任务在后台运行时，可以实时看到它正在操作的窗口。

新功能：任务实时画面

• 无论窗口在哪，都能看到 DeskMind 在做什么。 屏幕角落的小卡片会实时显示任务正在操作的窗口，即使这个窗口被你的其他窗口挡住也一样。画面下方用文字写出当前这一步，画面里的小圆点标出每一步落在哪里。标题栏里的小方会看着 DeskMind 正在操作的位置，需要你时抬头看你，完成时眯眼笑。
• 不挡你的事。 如果卡片会挡住正在操作的窗口，它会自动换到另一个角落；它不抢键盘，也不会跑到最前面。应用占满屏幕时，卡片停在离 DeskMind 点击位置最远的角落，点击会穿透过去；鼠标在卡片上停一下，就能用它的按钮。把它拖到任意角落，之后的任务都会留在那里。双击可以放大，也可以收起成一个小胶囊。
• 需要你时会告诉你。 任务提问或需要你批准时，卡片会提示；点一下卡片就能在 DeskMind 里回答。卡片上也有停止按钮。任务结束时，卡片会显示结果，然后消失。
• 只显示任务的窗口。 卡片只显示任务正在操作的那个窗口。窗口被最小化或关闭时，卡片会说明，绝不会改为显示同一应用的其他窗口。卡片是普通窗口：你自己截图、共享屏幕时能看到它；模型看不到它，任务录像也不包含它（录制整个屏幕时会包含）。
• 可以在「显示 › 显示任务窗口的实时画面」里关闭。

其他新功能

• 做错时告诉我们。 任务结束后点 👎，会问哪里不对（「它没问就猜了」「结果不对」「卡住了」），然后打开一个已填好的 GitHub issue 页面（只含你的指令、任务结果、步骤数量和类型、版本号，不含任何屏幕内容），由你在浏览器里检查后提交；DeskMind 自己不发送任何内容。运行出错时也有「在 GitHub 上反馈」，菜单「帮助 › 反馈问题」随时可用。
• 能看到结果。 任务给出答案，或你的第一个任务完成时，DeskMind 窗口会带着结果回来（不抢键盘）；窗口被移开时也会发通知。
• 三个示例都能看到，在输入框下方各占一行。

修复

• 整理带子文件夹的目录不会再一开始就失败。 以前在有子文件夹的「下载」里说「整理目录」，第一步之前就结束，还提示本地模型「没及时回答」：其实是一道题太大，被模型拒绝了。现在题目不会超出模型的限制；模型真的处理不了某一步时，会如实说明原因。
• 在你自己的文件夹里改名，会先征得你同意，和删除、发送一样（示例文件夹不受影响）。指令里提到的应用没安装时，开始前就会告诉你。文件示例只写文件夹里真实存在的文件。自检里一个任务跑完后窗口会回来，选别的任务时按钮显示「开始」，任务名跟随 App 语言。
• 提问不会再吃掉你在别处打的字。 以前 DeskMind 提问时如果你正在别的应用里打字，按键（包括回车）可能落进答案并直接发送。现在你打字时窗口回来但不抢键盘。
• 应用没打开、或窗口被关掉时，任务不会再一步就「完成」。 以前 DeskMind 看不到窗口，可能走一步就判定完成，实际什么都没做（网易云音乐上出现过）。现在开始前会先打开应用或找回它的窗口，且不会把它切到前台。
• 切换语言后，文件示例也会写示例文件夹里真实存在的文件名。
• 确认页把视觉模型写明为可选；不起作用的「不再询问」勾选框不再出现。
• 用词更清楚：中文「Didn't finish」改为「没做完」；顶部状态不再显示「Hands」而是「启动中」；「试试示例」（6 个自检任务）改名为「自检」。
• 录制的任务里，窗口左上角不再出现紫色的「正在共享」标记，录像里也不再出现菜单栏的状态图标。
• 第一次有任务完成后出现的 GitHub 那一行，现在链接到 DeskMind 仓库，要等你看到结果才出现；点 👍 后也会出现，直到你点过一次。首页底部加了 GitHub 链接。

其他

• 模型与 0.3.x 相同（DeskMind Brain G18b，路由门槛 0.96；Eyes-4B de8e13b），已公布的评测成绩依然适用，已下载的模型不会重新下载。
• App 的源码现在位于 deskmind-ai/deskmind 的 app/ 目录。本版本由 CI 按 app/deps.lock 固定的源码提交构建并公证。

运行要求：Apple Silicon 芯片的 Mac，macOS 15 或更高版本。首次启动会下载约 5.3 GB 的规划模型（Hugging Face 慢时自动改用 ModelScope）；可选的视觉模型另需 3.3 GB。实时画面使用 DeskMind 已有的「屏幕录制」权限；开启实时画面运行任务时，菜单栏会显示系统的屏幕录制指示。

DeskMind 0.4.0-rc.2

Pre-release of 0.4.0, for testing before it becomes the latest version. 0.3.1 stays the latest until then.

Since rc.1

• Organising a folder with subfolders no longer fails at once. "整理目录" on a Downloads folder with subfolders ended before its first step, saying the local model "didn't answer in time". Each file's "move to" list of every subfolder made one question too large for the model, and the request was refused. Questions now stay within the model's limits, and moves into subfolders work.
• A rename in your own folder waits for your approval, as deleting and sending do. It cannot be undone, and an unclear instruction could otherwise rename files to something you didn't want. The sample folder is not affected.
• An app that isn't on this Mac is named before the run starts ("WeChat isn't installed on this Mac…"), instead of the run trying without it.
• The file example always names a file the folder really has. If the attached folder has no file it could name, the example isn't shown.
• Self-test: after a task finishes, the window comes back, and picking another task shows Start. Task names follow the app's language.
• When the local model refuses a step, DeskMind now says that, and why, instead of saying it didn't answer in time.

DeskMind for Mac 0.4.0 shows you the window a task is working in, live, while it works in the background.

New: a live view of the task

• See what DeskMind is doing, wherever the window is. A small card in a corner of the screen shows the window the task is working in, live, even when that window is behind your other windows. Under the picture is the step it is taking, in words, and a dot shows where each step acted. In the title bar, 小方 (DeskMind’s little frame) watches where DeskMind is working, looks up when it needs you, and smiles when it’s done.
• It stays out of your way. The card moves to another corner when it would cover the window being worked in, never takes the keyboard, and doesn't come to the front. Over an app that fills the screen it sits in the corner farthest from where DeskMind is clicking and lets clicks pass through; rest the pointer on it for a moment to use its buttons. Drag it to any corner and it stays there for later tasks. Double-click it to make it larger, or collapse it to a small capsule.
• It tells you when it needs you. When the task asks a question or wants your approval, the card says so; click it to answer in DeskMind. Stop is on the card too. When the task ends, the card says how it ended, then goes.
• Only the task's window. The card shows the window the task works in and nothing else. If that window is minimized or closed, the card says so; it never shows another window of the same app instead. The card is an ordinary window: your own screenshots and screen sharing show it. The models never see it, and task recordings leave it out (recording the whole screen keeps it).
• Turn it off in View › Show a Live View of the Task.

Also new

• Tell us when it gets it wrong. After a task, 👎 asks what went wrong ("It guessed instead of asking", "Wrong result", "It got stuck") and opens a GitHub issue filled in with your instruction, the result and the steps in words. It includes only your instruction, how the task ended, the number and kinds of steps, and the versions, nothing read from your screen. You check it and submit it in your browser; DeskMind sends nothing itself. A failed run has "Report on GitHub" too, and Help › Report an Issue is always there.
• You see the result. When a task gives an answer, or your first task finishes, the DeskMind window comes back with it (without taking the keyboard). A notification also comes when the window is out of the way.
• All three examples are visible under the instruction box, one per line.

Fixes

• A question no longer catches what you type elsewhere. If you were typing in another app when DeskMind asked a question, your keys (Return included) could go into the answer and send it. Now the window comes back without taking the keyboard while you type.
• A task in an app that isn’t open, or whose window is closed, no longer ends at once. DeskMind saw no window, and the task could end as done after one step, having done nothing (seen with NetEase Cloud Music). Now the app is opened, or its window brought back, before the task starts, without bringing it to the front.
• The file example names the file the sample folder actually holds, after a language switch too.
• The vision model is described as optional on the confirmation sheet, and the "Don't ask again" box no longer shows where it had no effect.
• Clearer words: "Didn't finish" in Chinese reads 没做完; the island says "Starting" instead of "Hands"; "Try examples" (the six diagnostics) is now "Self-test".
• Recorded runs no longer show the purple "being shared" mark in place of the task windows' traffic lights, and recordings no longer include the menu bar's status icons.
• The GitHub line after your first finished task now links to the DeskMind repository, waits until you have seen the result, and comes back after a 👍 until you have followed it once. Home has a GitHub link at the bottom.

Changes

• The models are the same as in 0.3.x (DeskMind Brain G18b, routing threshold 0.96; Eyes-4B at de8e13b), so the published benchmark numbers still apply, and models already downloaded are not downloaded again.
• The app's source now lives in deskmind-ai/deskmind (app/). This release was built and notarized by CI from the source commits pinned in app/deps.lock.

Requirements

A Mac with Apple Silicon and macOS 15 or later. The planner models (about 5.3 GB) are downloaded on first launch, from Hugging Face or, if that is slow, from ModelScope. The optional vision model needs another 3.3 GB. The live view uses the Screen Recording permission DeskMind already has; macOS shows its screen-recording indicator in the menu bar while a task runs with the live view on.

Known limitations

• Apps with no accessibility tree are read from the screen, and are briefly brought to the front while a step acts. The vision model as the app runs it (4-bit, up to 2 megapixels) scores 50.9 on ScreenSpot-Pro (67.7 on a GPU at full resolution), and small icons are its weak spot.
• Chrome's address bar is not recognised as the browser's own field; the guard that keeps table rows out of it works in Safari.
• Copying more than about four rows in a given order is less reliable than short tables.

---

中文说明

DeskMind for Mac 0.4.0：任务在后台运行时，可以实时看到它正在操作的窗口。

新功能：任务实时画面

• 无论窗口在哪，都能看到 DeskMind 在做什么。 屏幕角落的小卡片会实时显示任务正在操作的窗口，即使这个窗口被你的其他窗口挡住也一样。画面下方用文字写出当前这一步，画面里的小圆点标出每一步落在哪里。标题栏里的小方会看着 DeskMind 正在操作的位置，需要你时抬头看你，完成时眯眼笑。
• 不挡你的事。 如果卡片会挡住正在操作的窗口，它会自动换到另一个角落；它不抢键盘，也不会跑到最前面。应用占满屏幕时，卡片停在离 DeskMind 点击位置最远的角落，点击会穿透过去；鼠标在卡片上停一下，就能用它的按钮。把它拖到任意角落，之后的任务都会留在那里。双击可以放大，也可以收起成一个小胶囊。
• 需要你时会告诉你。 任务提问或需要你批准时，卡片会提示；点一下卡片就能在 DeskMind 里回答。卡片上也有停止按钮。任务结束时，卡片会显示结果，然后消失。
• 只显示任务的窗口。 卡片只显示任务正在操作的那个窗口。窗口被最小化或关闭时，卡片会说明，绝不会改为显示同一应用的其他窗口。卡片是普通窗口：你自己截图、共享屏幕时能看到它；模型看不到它，任务录像也不包含它（录制整个屏幕时会包含）。
• 可以在「显示 › 显示任务窗口的实时画面」里关闭。

其他新功能

• 做错时告诉我们。 任务结束后点 👎，会问哪里不对（「它没问就猜了」「结果不对」「卡住了」），然后打开一个已填好的 GitHub issue 页面（只含你的指令、任务结果、步骤数量和类型、版本号，不含任何屏幕内容），由你在浏览器里检查后提交；DeskMind 自己不发送任何内容。运行出错时也有「在 GitHub 上反馈」，菜单「帮助 › 反馈问题」随时可用。
• 能看到结果。 任务给出答案，或你的第一个任务完成时，DeskMind 窗口会带着结果回来（不抢键盘）；窗口被移开时也会发通知。
• 三个示例都能看到，在输入框下方各占一行。

修复

• 整理带子文件夹的目录不会再一开始就失败。 以前在有子文件夹的「下载」里说「整理目录」，第一步之前就结束，还提示本地模型「没及时回答」：其实是一道题太大，被模型拒绝了。现在题目不会超出模型的限制；模型真的处理不了某一步时，会如实说明原因。
• 在你自己的文件夹里改名，会先征得你同意，和删除、发送一样（示例文件夹不受影响）。指令里提到的应用没安装时，开始前就会告诉你。文件示例只写文件夹里真实存在的文件。自检里一个任务跑完后窗口会回来，选别的任务时按钮显示「开始」，任务名跟随 App 语言。
• 提问不会再吃掉你在别处打的字。 以前 DeskMind 提问时如果你正在别的应用里打字，按键（包括回车）可能落进答案并直接发送。现在你打字时窗口回来但不抢键盘。
• 应用没打开、或窗口被关掉时，任务不会再一步就「完成」。 以前 DeskMind 看不到窗口，可能走一步就判定完成，实际什么都没做（网易云音乐上出现过）。现在开始前会先打开应用或找回它的窗口，且不会把它切到前台。
• 切换语言后，文件示例也会写示例文件夹里真实存在的文件名。
• 确认页把视觉模型写明为可选；不起作用的「不再询问」勾选框不再出现。
• 用词更清楚：中文「Didn't finish」改为「没做完」；顶部状态不再显示「Hands」而是「启动中」；「试试示例」（6 个自检任务）改名为「自检」。
• 录制的任务里，窗口左上角不再出现紫色的「正在共享」标记，录像里也不再出现菜单栏的状态图标。
• 第一次有任务完成后出现的 GitHub 那一行，现在链接到 DeskMind 仓库，要等你看到结果才出现；点 👍 后也会出现，直到你点过一次。首页底部加了 GitHub 链接。

其他

• 模型与 0.3.x 相同（DeskMind Brain G18b，路由门槛 0.96；Eyes-4B de8e13b），已公布的评测成绩依然适用，已下载的模型不会重新下载。
• App 的源码现在位于 deskmind-ai/deskmind 的 app/ 目录。本版本由 CI 按 app/deps.lock 固定的源码提交构建并公证。

运行要求：Apple Silicon 芯片的 Mac，macOS 15 或更高版本。首次启动会下载约 5.3 GB 的规划模型（Hugging Face 慢时自动改用 ModelScope）；可选的视觉模型另需 3.3 GB。实时画面使用 DeskMind 已有的「屏幕录制」权限；开启实时画面运行任务时，菜单栏会显示系统的屏幕录制指示。

DeskMind 0.4.0-rc.1

Pre-release of 0.4.0, for testing before it becomes the latest version. 0.3.1 stays the latest until then.

DeskMind for Mac 0.4.0 shows you the window a task is working in, live, while it works in the background.

New: a live view of the task

• See what DeskMind is doing, wherever the window is. A small card in a corner of the screen shows the window the task is working in, live, even when that window is behind your other windows. Under the picture is the step it is taking, in words, and a dot shows where each step acted. In the title bar, 小方 (DeskMind’s little frame) watches where DeskMind is working, looks up when it needs you, and smiles when it’s done.
• It stays out of your way. The card moves to another corner when it would cover the window being worked in, never takes the keyboard, and doesn't come to the front. Over an app that fills the screen it sits in the corner farthest from where DeskMind is clicking and lets clicks pass through; rest the pointer on it for a moment to use its buttons. Drag it to any corner and it stays there for later tasks. Double-click it to make it larger, or collapse it to a small capsule.
• It tells you when it needs you. When the task asks a question or wants your approval, the card says so; click it to answer in DeskMind. Stop is on the card too. When the task ends, the card says how it ended, then goes.
• Only the task's window. The card shows the window the task works in and nothing else. If that window is minimized or closed, the card says so; it never shows another window of the same app instead. The card is an ordinary window: your own screenshots and screen sharing show it. The models never see it, and task recordings leave it out (recording the whole screen keeps it).
• Turn it off in View › Show a Live View of the Task.

Also new

• Tell us when it gets it wrong. After a task, 👎 asks what went wrong ("It guessed instead of asking", "Wrong result", "It got stuck") and opens a GitHub issue filled in with your instruction, the result and the steps in words. It includes only your instruction, how the task ended, the number and kinds of steps, and the versions, nothing read from your screen. You check it and submit it in your browser; DeskMind sends nothing itself. A failed run has "Report on GitHub" too, and Help › Report an Issue is always there.
• You see the result. When a task gives an answer, or your first task finishes, the DeskMind window comes back with it (without taking the keyboard). A notification also comes when the window is out of the way.
• All three examples are visible under the instruction box, one per line.

Fixes

• A question no longer catches what you type elsewhere. If you were typing in another app when DeskMind asked a question, your keys (Return included) could go into the answer and send it. Now the window comes back without taking the keyboard while you type.
• A task in an app that isn’t open, or whose window is closed, no longer ends at once. DeskMind saw no window, and the task could end as done after one step, having done nothing (seen with NetEase Cloud Music). Now the app is opened, or its window brought back, before the task starts, without bringing it to the front.
• The file example names the file the sample folder actually holds, after a language switch too.
• The vision model is described as optional on the confirmation sheet, and the "Don't ask again" box no longer shows where it had no effect.
• Clearer words: "Didn't finish" in Chinese reads 没做完; the island says "Starting" instead of "Hands"; "Try examples" (the six diagnostics) is now "Self-test".
• Recorded runs no longer show the purple "being shared" mark in place of the task windows' traffic lights, and recordings no longer include the menu bar's status icons.
• The GitHub line after your first finished task now links to the DeskMind repository, waits until you have seen the result, and comes back after a 👍 until you have followed it once. Home has a GitHub link at the bottom.

Changes

• The models are the same as in 0.3.x (DeskMind Brain G18b, routing threshold 0.96; Eyes-4B at de8e13b), so the published benchmark numbers still apply, and models already downloaded are not downloaded again.
• The app's source now lives in deskmind-ai/deskmind (app/). This release was built and notarized by CI from the source commits pinned in app/deps.lock.

Requirements

A Mac with Apple Silicon and macOS 15 or later. The planner models (about 5.3 GB) are downloaded on first launch, from Hugging Face or, if that is slow, from ModelScope. The optional vision model needs another 3.3 GB. The live view uses the Screen Recording permission DeskMind already has; macOS shows its screen-recording indicator in the menu bar while a task runs with the live view on.

Known limitations

• Apps with no accessibility tree are read from the screen, and are briefly brought to the front while a step acts. The vision model as the app runs it (4-bit, up to 2 megapixels) scores 50.9 on ScreenSpot-Pro (67.7 on a GPU at full resolution), and small icons are its weak spot.
• Chrome's address bar is not recognised as the browser's own field; the guard that keeps table rows out of it works in Safari.
• Copying more than about four rows in a given order is less reliable than short tables.

---

中文说明

DeskMind for Mac 0.4.0：任务在后台运行时，可以实时看到它正在操作的窗口。

新功能：任务实时画面

• 无论窗口在哪，都能看到 DeskMind 在做什么。 屏幕角落的小卡片会实时显示任务正在操作的窗口，即使这个窗口被你的其他窗口挡住也一样。画面下方用文字写出当前这一步，画面里的小圆点标出每一步落在哪里。标题栏里的小方会看着 DeskMind 正在操作的位置，需要你时抬头看你，完成时眯眼笑。
• 不挡你的事。 如果卡片会挡住正在操作的窗口，它会自动换到另一个角落；它不抢键盘，也不会跑到最前面。应用占满屏幕时，卡片停在离 DeskMind 点击位置最远的角落，点击会穿透过去；鼠标在卡片上停一下，就能用它的按钮。把它拖到任意角落，之后的任务都会留在那里。双击可以放大，也可以收起成一个小胶囊。
• 需要你时会告诉你。 任务提问或需要你批准时，卡片会提示；点一下卡片就能在 DeskMind 里回答。卡片上也有停止按钮。任务结束时，卡片会显示结果，然后消失。
• 只显示任务的窗口。 卡片只显示任务正在操作的那个窗口。窗口被最小化或关闭时，卡片会说明，绝不会改为显示同一应用的其他窗口。卡片是普通窗口：你自己截图、共享屏幕时能看到它；模型看不到它，任务录像也不包含它（录制整个屏幕时会包含）。
• 可以在「显示 › 显示任务窗口的实时画面」里关闭。

其他新功能

• 做错时告诉我们。 任务结束后点 👎，会问哪里不对（「它没问就猜了」「结果不对」「卡住了」），然后打开一个已填好的 GitHub issue 页面（只含你的指令、任务结果、步骤数量和类型、版本号，不含任何屏幕内容），由你在浏览器里检查后提交；DeskMind 自己不发送任何内容。运行出错时也有「在 GitHub 上反馈」，菜单「帮助 › 反馈问题」随时可用。
• 能看到结果。 任务给出答案，或你的第一个任务完成时，DeskMind 窗口会带着结果回来（不抢键盘）；窗口被移开时也会发通知。
• 三个示例都能看到，在输入框下方各占一行。

修复

• 提问不会再吃掉你在别处打的字。 以前 DeskMind 提问时如果你正在别的应用里打字，按键（包括回车）可能落进答案并直接发送。现在你打字时窗口回来但不抢键盘。
• 应用没打开、或窗口被关掉时，任务不会再一步就「完成」。 以前 DeskMind 看不到窗口，可能走一步就判定完成，实际什么都没做（网易云音乐上出现过）。现在开始前会先打开应用或找回它的窗口，且不会把它切到前台。
• 切换语言后，文件示例也会写示例文件夹里真实存在的文件名。
• 确认页把视觉模型写明为可选；不起作用的「不再询问」勾选框不再出现。
• 用词更清楚：中文「Didn't finish」改为「没做完」；顶部状态不再显示「Hands」而是「启动中」；「试试示例」（6 个自检任务）改名为「自检」。
• 录制的任务里，窗口左上角不再出现紫色的「正在共享」标记，录像里也不再出现菜单栏的状态图标。
• 第一次有任务完成后出现的 GitHub 那一行，现在链接到 DeskMind 仓库，要等你看到结果才出现；点 👍 后也会出现，直到你点过一次。首页底部加了 GitHub 链接。

其他

• 模型与 0.3.x 相同（DeskMind Brain G18b，路由门槛 0.96；Eyes-4B de8e13b），已公布的评测成绩依然适用，已下载的模型不会重新下载。
• App 的源码现在位于 deskmind-ai/deskmind 的 app/ 目录。本版本由 CI 按 app/deps.lock 固定的源码提交构建并公证。

运行要求：Apple Silicon 芯片的 Mac，macOS 15 或更高版本。首次启动会下载约 5.3 GB 的规划模型（Hugging Face 慢时自动改用 ModelScope）；可选的视觉模型另需 3.3 GB。实时画面使用 DeskMind 已有的「屏幕录制」权限；开启实时画面运行任务时，菜单栏会显示系统的屏幕录制指示。

DeskMind 0.3.1

DeskMind for Mac 0.3.1 fixes problems found in external reviews of 0.3.0. Update if you have tried the examples with unsaved TextEdit documents open.

Fixes

• Try examples no longer closes your TextEdit documents. In 0.3.0, the example "Type exact Chinese text and save" closed every open TextEdit document without saving before it started, so unsaved work was lost. Now only the example's own sandbox files are closed.
• An approval covers one step. Approving one "Send" covered every later "Send" in the same task, even in another app. Now each send, delete, pay, publish or share asks again; the confirmation dialog it opens right after counts as the same decision.
• The local model servers answer only to DeskMind. The planner and vision servers on 127.0.0.1 now require a random token made on this Mac, so another program can neither use them nor pass for them. A server left over from 0.3.0 is replaced automatically.
• Clear all deletes run data. Clear all in Recent now also deletes the runs' step screenshots and records, and your 👍/👎, from this Mac. Recordings in Movies › DeskMind stay. The README lists what is kept where, and for how long.
• Vision clicks near the screen's edge. The vision model's point at the very top-left of the screen could be read as the bottom-right corner. It is now read on the model's own 0–1000 scale, and an answer outside the screen is rejected.
• Two AppleScript strings are now escaped like the others, so a folder name with a quote or a backslash cannot break them.

Changes

• The models are the same as in 0.3.0 (DeskMind Brain G18b, routing threshold 0.96; Eyes-4B at de8e13b), so the published benchmark numbers still apply. A machine that already downloaded them for 0.3.0 does not download them again.

• In the Chinese interface, the product is called DeskMind.

• After your first task that finishes, one line offers the GitHub repository. It shows once, can be closed, and sends nothing.

• This release was built and notarized by CI in deskmind-ai/app, from the source commits pinned in prototype/deps.lock (now app/deps.lock in this repository). The app's source moved here on 10-04; the DMG is the same file, SHA-256 unchanged.

Download

DeskMind-0.3.1.dmg, signed with a Developer ID and notarized by Apple. SHA-256: 267edcf87e280e68226f8a05075b22d1914b18796d3962f52f306d3ff0987235

Requirements

A Mac with Apple Silicon and macOS 15 or later. The planner models (about 5.3 GB) are downloaded on first launch, from Hugging Face or, if that is slow, from ModelScope. The optional vision model needs another 3.3 GB.

Known limitations

• Apps with no accessibility tree are read from the screen, and are briefly brought to the front while a step acts. The vision model as the app runs it (4-bit, up to 2 megapixels) scores 50.9 on ScreenSpot-Pro (67.7 on a GPU at full resolution), and small icons are its weak spot.
• Chrome's address bar is not recognised as the browser's own field; the guard that keeps table rows out of it works in Safari.
• Copying more than about four rows in a given order is less reliable than short tables.

---

中文说明

DeskMind for Mac 0.3.1 修复了外部评审在 0.3.0 中发现的问题。如果你在打开着未保存的 TextEdit 文档时用过「试试示例」，建议更新。

• 「试试示例」不会再关掉你的 TextEdit 文档。 0.3.0 里，「中文精确写入并保存」这个示例开始前，会把所有打开的 TextEdit 文档不保存直接关掉，未保存的内容会丢失。现在只关闭示例自己的沙盒文件。
• 一次批准只对一步有效。 以前批准过一次「发送」，同一个任务里之后所有的「发送」都不再询问，换了应用也一样。现在每次发送、删除、付款、发布、分享都会重新询问；紧跟着弹出的确认框算同一次决定。
• 本地模型服务只响应 DeskMind。 127.0.0.1 上的规划服务和视觉服务，现在要求带上本机生成的随机令牌，其他程序既用不了，也冒充不了。0.3.0 留下的旧服务会被自动替换。
• 「全部清除」会删除运行数据。 在「最近的任务」里点「全部清除」，现在会同时从这台 Mac 上删除这些任务的步骤截图、运行记录，以及你的 👍/👎；录像保留在「影片 › DeskMind」中。README 里写明了每类数据存在哪里、保留多久。
• 视觉模型在屏幕边缘的定位。 视觉模型给出的屏幕左上角附近的点，可能被误读成右下角。现在按模型自己的 0–1000 尺度解读，超出屏幕的结果会被拒绝。
• 两处 AppleScript 字符串补上了转义，文件夹名里有引号或反斜杠也不会出错。
• 模型与 0.3.0 相同（DeskMind Brain G18b，路由门槛 0.96；Eyes-4B de8e13b），已公布的评测成绩依然适用；装过 0.3.0 的机器不会重新下载模型。
• 中文界面里的产品名改为 DeskMind；第一次有任务完成后，会出现一行可关闭的 GitHub 链接，只出现一次，不发送任何数据。
• 本版本由原 deskmind-ai/app 仓库的 CI 按 prototype/deps.lock（现为本仓库的 app/deps.lock）固定的源码提交构建并公证。App 源码已于 10-04 并入本仓库，DMG 是同一个文件，SHA-256 不变。

安装包：DeskMind-0.3.1.dmg，已用 Developer ID 签名并经 Apple 公证。SHA-256：267edcf87e280e68226f8a05075b22d1914b18796d3962f52f306d3ff0987235

运行要求：Apple Silicon 芯片的 Mac，macOS 15 或更高版本。首次启动会下载约 5.3 GB 的规划模型（Hugging Face 慢时自动改用 ModelScope）；可选的视觉模型另需 3.3 GB。

Peekaboo 4.3.1 (fabaadc56) for DeskMind builds

The Peekaboo CLI shipped in DeskMind builds (built from fabaadc56), and the licence files of what is linked into it. Used by CI; not an app release.