Agent task coverage audit
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
调研方向
Start by reading docs/tools.md, docs/recipes.md, src/tools.ts, and the native client, then compare the documented task sequences with the current code and the referenced issues and pull requests. The audit says its ratings were not checked live and points to #42 and npm run live-check for pending verification. Done means updating the task coverage assessment with evidence for each rating and recording any remaining gaps.
由索引模型根据 Issue 内容生成。
描述
How well desktop-tools covers everyday agent tasks on a Mac, mapped from docs/tools.md, docs/recipes.md, src/tools.ts and the native client.
Refreshed 2026-10-17 (second pass) against main at aefabc2 (0.11.0 with #107 and #109), plus open #110 (web accessibility for desktop_wait) and #111 (incomplete-read retry). The first refresh was at 6783905 (0.8.0); the original audit was at acae444. Rows changed since then say which PR changed them. None of this was run live: the screen is locked, and the live checks for each PR are in #42. Ratings come from the code, and the ones marked (unverified) depend on how an app or the OS exposes its UI.
- Gap: there is no reasonable tool sequence.
- Awkward: it works, but takes many calls, has an unverifiable result, or relies on a fragile workaround.
- Works: there is a direct sequence.
Common shorthand: inspect = desktop_inspect (it returns the snapshot that input uses). Each completed action returns a fresh inspection, so chains don't need a separate inspect.
Counts: 5 gaps (10 originally), 11 awkward (13 at the first refresh, 16 originally), 29 works (26 at the first refresh, 18 originally).
Gaps (5)
| # | Task | Tool sequence | Why it's a gap |
|---|---|---|---|
| 2 | Enter a password: a website login, joining Wi-Fi, or an admin prompt (SecurityAgent) | inspect → desktop_type / desktop_insert / desktop_dialog_answer field |
All three refuse secure fields, and so does #82's web typing. Left for the user's decision. |
| 3 | Install an app from a DMG when it replaces an existing app, or change a locked System Settings pane | desktop_open dmg → inspect → desktop_drag app → Applications |
The copy works, but admin authentication hits gap 2. Touch ID can't be done at all. |
| 4 | Connect to a secured Wi-Fi network | desktop_menubar_click Wi-Fi → inspect → desktop_click network → password |
Opening the list and choosing the network work, and an open network works. The password field hits gap 2. |
| 8 | Move a window to another Space | — | Documented as not supported: since macOS 14.5 only the Dock can do it. The Mission Control drag workaround is always unknown. |
| 10 | Select non-editable text in a PDF (to copy or highlight) | inspect → desktop_drag |
Web pages are done (row g10 under Works). A PDF view has no selected text range, and #88's WebKit text markers apply to web areas only, so a PDF needs a drag-select, which is unverifiable. desktop_read covers copying. |
Works but awkward (11)
| # | Task | Tool sequence | Why awkward |
|---|---|---|---|
| 2 | Use Spotlight to open something | desktop_system_key Cmd-Space with expect window_appears → inspect → desktop_insert → desktop_key enter |
Many steps. The system key now completes when Spotlight's window appears (expect, #89), but is still unknown otherwise. #84 has a recipe. desktop_open / desktop_launch is usually better. |
| 4 | Print to PDF | desktop_menu File > Print… → … → desktop_dialog_answer |
About 6 supported steps. #84 has a recipe, including the Export as PDF shortcut. |
| 7 | Drag a file into a web upload area | Finder inspect + browser inspect → desktop_drag with to_snapshot |
Cross-app drags are foreground and unknown unless an expect condition (#89), such as the file name appearing, holds afterwards. Prefer the upload button and desktop_dialog_answer path. |
| 8 | Install an app from a DMG (no replace) | desktop_open dmg → desktop_wait → inspect → desktop_drag → desktop_dialog Gatekeeper |
Many steps with waits. #84 has a recipe. |
| 9 | Use the share sheet | desktop_menu File > Share > , or the toolbar button → popover |
The menu path works. The popover may live in another process (unverified). #84 has a recipe. |
| 11 | Pick a color in the color panel | desktop_menu Format > Show Colors → … → desktop_type hex → desktop_key enter |
Many steps, and the panel layout varies (unverified). #84 has a recipe. |
| 12 | Choose a font | desktop_menu Format > Font > Show Fonts → … |
The lists are long. #86's table read can list them. #84 has a recipe. |
| 14 | Copy files between Finder folders | desktop_clipboard_write files → inspect target → desktop_key Cmd-V |
Verifiable but indirect. #99's desktop_insert format: "files" pastes into text areas only; a Finder folder paste confirmed by the copied names is a #3 follow-up. #84 has a recipe. Shell cp is simpler. |
| 15 | Respond to a notification | desktop_notifications → desktop_notification_action |
Built in #74 against macOS 11–15. Pending a live check on 27 (#42). |
| 16 | Use a context menu on a Dock item | desktop_dock right-click + item |
Moves the real pointer. |
| 7g | Zoom or rotate in Maps, Photos or Preview (was gap 7) | desktop_menus → desktop_menu View > Zoom In / Rotate, or desktop_key Cmd-= / Cmd-- (with repeat) |
Use the menu or keys (#87 recipe). There is no pinch or rotate gesture by decision: undocumented events, GPL prior art, and no Chrome support. Canvas views may not report the zoom level. |
Works (29)
| # | Task | Tool sequence |
|---|---|---|
| 1 | Rename a file in Finder | inspect → desktop_click file → desktop_key enter → desktop_type → desktop_key enter |
| 2 | Attach a file in Mail | desktop_menu File > Attach Files… → desktop_dialog_answer path → button Choose File |
| 3 | Change a System Settings toggle | desktop_open settings URL → desktop_wait element_appears → desktop_set_value with readback |
| 4 | Accept a cookie banner | inspect (with scope if truncated) → desktop_click Accept |
| 5 | Use a context menu | desktop_click kind right or AXShowMenu → AXPick |
| 6 | Scroll a list until an item is visible | AXScrollToVisible, or desktop_scroll + desktop_wait element_appears |
| 7 | Switch tabs in Safari or Terminal | desktop_click tab, desktop_menu, or desktop_key Cmd-Shift-] |
| 8 | Answer a permission prompt (button) | desktop_dialog → desktop_dialog_answer button |
| 9 | Choose from a popup button or web <select> |
desktop_click → AXPick |
| 10 | Handle a modal sheet or alert | desktop_wait dialog_appears → desktop_dialog → desktop_dialog_answer |
| 11 | Search the web | desktop_open URL → desktop_wait text_contains → desktop_read. In Chrome and other Chromium browsers, the page's text appears once web accessibility is on (#107 for inspect and read, #110 for the wait). |
| 12 | Read a long article or mail thread | desktop_read |
| 13 | Save a document to a given path | desktop_menu File > Save… → desktop_dialog_answer path → Save |
| 14 | Quit an app with unsaved changes | desktop_quit → desktop_dialog → desktop_dialog_answer button |
| 15 | Copy and paste between apps | Cmd-C → desktop_clipboard_read, or desktop_clipboard_write → desktop_insert |
| 16 | Run a Terminal command and read its output | desktop_insert → Enter → desktop_wait text_changes → desktop_read |
| 17 | Arrange windows: display, size, full screen | desktop_spaces → desktop_move / desktop_resize / desktop_maximize, and desktop_fullscreen with an explicit enabled (#78) |
| 18 | Reorder items by dragging within one window | desktop_drag (background, exact window) |
| g1 | Select several files or messages (Cmd-click, Shift-click) (was gap 1) | desktop_click with modifiers command/shift/option, through the foreground modifier route (#77). The result is unknown unless an expect condition holds afterwards (#89); otherwise check the selection by inspecting. |
| g5 | Set a date or time in a date picker (was gap 5) | desktop_set_value with an ISO 8601 string on AXDateTimeArea/AXDateField/AXTimeField, read back. The keyboard route is given when a picker can't be set (#79). Web <input type=date> is (unverified). |
| g6 | Take a screenshot to a file (was gap 6) | desktop_screen or desktop_inspect with save_path, region and overwrite, writing the full-resolution PNG (#81) |
| g9 | Volume, brightness or play/pause by key (was gap 9) | desktop_system_key volume_up, mute, play_pause, brightness_up… The result is unknown (#80), and these keys have no observable condition for expect. For a set level, use the Control Center slider. |
| a1 | Keyboard-only navigation (was awkward 1) | desktop_key with keys (up to 16 chords) or repeat (up to 32) against one snapshot, focus re-checked before each key (#77) |
| a3 | Fill a web form's text fields (was awkward 3) | desktop_click the field → desktop_type, which types with key events so page scripts see input, then reads the field back (#82, recipe in #84). In Chromium and Electron apps (Chrome, Slack, VS Code), the fields are listed once #107 turns on the app's web accessibility at the first inspect. |
| a5 | Read a table: Numbers, Finder list, web table (was awkward 5) | desktop_read with table: true and a scope → columns, rows of cells, outline depth, and row IDs for desktop_click (#86) |
| a13 | Adjust volume or brightness to a level (was awkward 13) | desktop_menubar_click Control Center → desktop_set_value slider with readback (recipe in #84), or the keys in g9 |
| a6 | Edit a spreadsheet or table cell (was awkward 6) | desktop_read with table: true → desktop_set_cell by row and column, confirmed by reading the table again (#92). Whether Numbers and Excel cells expose a control is (unverified), listed in #42; a cell without one is refused with CELL_NOT_EDITABLE. |
| a10 | Switch the input source (was awkward 10) | desktop_input_source with select, read back through the host, covered by pause and the activity list (#90, #101) |
| g10 | Select non-editable text on a web page (was part of gap 10) | desktop_select on an element inside a web area, through WebKit text markers, read back (#88). Chromium pages are (unverified). |
Top remaining gaps and awkward items
Password entry (gap 2, and gaps 3 and 4 behind it) is left for the user's decision. The first refresh's top three are done: table cells (#92), input source by ID (#90, #101) and expect for foreground input (#89).
- Run the pending live checks in #42. None of these ratings were run live, and most rows since 0.5.0 rest on code alone.
npm run live-check(#102) automates part of them. - Background paste into web content (#5 part a). It affects a3 and Ace's composer when the app isn't frontmost. A live spike, listed in #42, must answer its three questions before any patch.
- Paste files into a Finder folder (awkward 14, a #3 follow-up). Confirm the paste by the copied names appearing in the folder, so the copy is
completedrather than indirect. - PDF text selection (gap 10). First find out what Preview's PDF view exposes to Accessibility.
- 主要语言
- Swift
- 星标
- 0
- 派生
- 0
- 平均合并
- 15 分钟
- 30 天内合并 PR
- 94
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
githubnext/desktop-tools 的其他 Issue
-
ace documentation
难度 2/5 1-3 小时 新手友好度 79/100
githubnext/desktop-tools#104 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 55/100
githubnext/desktop-tools#42 · 29 条评论 ·
维护者通常 1 天内回复
-
enhancement
难度 5/5 一周以上 新手友好度 15/100
githubnext/desktop-tools#27 · 3 条评论 ·
维护者通常 1 天内回复
-
question
难度 5/5 一周以上 新手友好度 8/100
githubnext/desktop-tools#24 ·
维护者通常 1 天内回复
-
live-check-only
难度 4/5 3-5 天 新手友好度 25/100
githubnext/desktop-tools#23 · 1 条评论 ·
维护者通常 1 天内回复
查看 githubnext/desktop-tools 的全部 Issue
相似的 Issue
-
bug
难度 2/5 1-3 小时 新手友好度 92/100
维护者通常 1 天内回复
-
test(BrainBar): inject synthetic observabilityCadence in BackupsHealth/one-page fixtures (DCI-1)未关闭
难度 2/5 1-3 小时 新手友好度 74/100
EtanHey/brainlayer#1367 ·
维护者通常 1 天内回复
-
cmux-next: new workspaces show no name instead of workspace-N可能已有人在做 关联的 PR 仍在进行中或已合并。 未关闭area: workspaces
难度 2/5 1-3 小时 新手友好度 82/100
manaflow-ai/cmux#21357 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 4 天内回复
-
难度 2/5 1-3 小时 新手友好度 74/100