Downloads, uploads & file dialogs
Move files between apps, the web and disk through download bars, upload pickers and native Open/Save dialogs (save as, location, name, format).
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- Chrome · Safari · Edge
- Class GUI-need label
G3· GUI-only
Task proposals · 20
Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.
R01.10.01.02-T01 Download a statement PDF from one portal and upload it to another
Original instruction
Signed in (test accounts) to the sandbox bank portal and the sandbox accounting portal in Chrome, download the October 2026 statement PDF for account `...4471` and upload it to the accounting portal's `Bank statements > 2026-10` slot using the portal's file picker. Do not rename the file.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicknav· Pan/zoom/scrollmulti· Cross-app- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Accounting DB has one upload in slot 2026-10 whose SHA-256 equals the bank's October PDF for ...4471 (September and another account's October are decoys); filename unchanged. Both portals issue no API token and accept requests only from browser sessions (CSRF form token + HttpOnly session cookie); direct HTTP uploads are rejected.
Proposed partial credit
40% correct file downloaded; 50% uploaded to correct slot; 10% filename preserved
R01.10.01.02-T02 Save three downloads into different folders through the native Save dialog
Original instruction
In Chrome, turn on 'Ask where to save each file before downloading'. Then from `http://files.local/` download `contract.pdf` into `~/Documents/Legal` renamed `2026-10_contract.pdf`, `logo.svg` into `~/Pictures/Brand`, and `data.csv` into a new folder `~/Documents/Q4` created from inside the Save dialog.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Save dialog
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsnav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Files exist at the three exact paths with server hashes; Q4 folder created during the run (birth time after start); Chrome prompt_for_download true; ~/Downloads has none of the three. Download API/extensions disabled.
Proposed partial credit
15% setting; 30% renamed contract; 25% logo; 30% new folder + csv
R01.10.01.02-T03 Stop Safari from auto-opening downloads and keep a zip intact
Original instruction
In Safari settings, turn off 'Open safe files after downloading', set the download location to `~/Downloads/Web`, and then download `http://files.local/site_backup.zip` so it stays zipped there.
- Platforms
mac· macOSweb· Any OS (web)- Applications
- Safari
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicknav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
com.apple.Safari AutoOpenSafeDownloads false and DownloadsPath ~/Downloads/Web; site_backup.zip present with server hash; no extracted folder site_backup/ exists anywhere in ~/Downloads. The agent shell lacks Full Disk Access (Safari's prefs container is not accessible to defaults) and eslogger detects curl/wget fetches of the zip, which void the download milestone.
Proposed partial credit
35% auto-open off; 25% download path; 40% zip saved intact
R01.10.01.02-T04 Upload a profile photo by dragging it from the file manager
Original instruction
On the sandbox HR portal profile page, the upload area only accepts drag-and-drop (no Browse button). Drag `~/Pictures/headshot_2026.jpg` from Finder or File Explorer into the drop zone, crop it to the square guide shown, and save.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Finder · File Explorer
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & dropmulti· Cross-app- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Portal DB: avatar uploaded, source hash equals headshot_2026.jpg (two similarly named decoys in the folder); saved crop rectangle IoU >= 0.8 with the guide; drop event originated from OS drag (DataTransfer.files from native drag, recorded by page). Remote debugging is disabled (no CDP synthetic drops) and the portal accepts the file only from the browser session.
Proposed partial credit
50% correct file dropped; 30% crop IoU; 20% saved
R01.10.01.02-T05 Multi-select five files in a crowded native Open dialog for upload
Original instruction
On the sandbox claims portal, click `Add documents` and, in the native Open dialog, select exactly the 5 files listed in `~/Desktop/claim_docs.txt` from `~/Documents/Scans` (200 files with similar names), then upload them in one batch.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Open dialog
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clickkey· Shortcutsnav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Portal DB: one upload batch with exactly 5 files whose hashes match the list (precision/recall); single batch id. Typing paths into the File name box is allowed; JS injection into input[type=file] is not possible (DevTools disabled).
Proposed partial credit
80% F1 over the 5 files; 20% single batch
R01.10.01.02-T06 Upload a hidden dotfile through the macOS Open panel
Original instruction
The sandbox config-sync website asks for `~/.config/devtool/settings.json`. In Chrome's file chooser (macOS Open panel), navigate to that hidden folder (show hidden files or Go to Folder) and upload it. Then confirm the site's preview shows key `theme`.
- Platforms
mac· macOS- Applications
- Chrome · Open panel
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
key· Shortcutstype· Text & formsnav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Server received a file with SHA-256 of the seeded settings.json (a decoy settings.json exists in ~/Documents); page preview state 'theme' present; upload performed via the chooser (page records change event from input element). Remote debugging is disabled (no CDP setInputFiles) and the site accepts uploads only from the browser session.
Proposed partial credit
70% correct file uploaded; 30% preview confirmed
R01.10.01.02-T07 Print a web page to PDF with backgrounds and a page range
Original instruction
In Chrome, open `http://fixtures.local/itinerary.html` and use Print > Save as PDF to save pages 2-3 only, landscape, with 'Background graphics' enabled and headers/footers off, to `~/Documents/itinerary_p2-3.pdf`.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Print dialog
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & forms- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
PDF has exactly 2 pages with landscape mediabox; text of pages equals source pages 2-3; rendered background color band present (pixel sample within ΔE 5 of #E8F0FE); no URL/date header text. The OS process audit detects headless Chrome (--print-to-pdf) or remote-debugging launches, which void the task.
Proposed partial credit
30% page range; 20% landscape; 30% backgrounds; 20% no headers/footers and right path
R01.10.01.02-T08 Keep a blocked insecure download from an internal test server
Original instruction
Chrome blocks `http://build.lab.test/artifacts/tool-1.4.2.zip` as an insecure download. The ticket in `~/Desktop/ticket.txt` confirms the server is the lab's internal build host and lists the expected SHA-256. Keep the download, verify the hash matches before using it, and record `match` or `mismatch` in `~/Desktop/hash_check.txt`.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Terminal
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
click· Point & clickhover· Hover & contextcode· Code & commands- Work product
WP09· Files & folders- Conditions
risk· Irreversible/risky- Proposed difficulty
- 3 / 5
Proposed verifier
~/Downloads/tool-1.4.2.zip exists with expected hash; hash_check.txt 'match'; Chrome History download state COMPLETE with danger type accepted by user. A variant seed serves a tampered file where 'mismatch' and deleting it is the correct outcome. curl/wget fetches are detected by the process audit and cannot create the user-accepted danger record.
Proposed partial credit
40% download kept via browser UI; 40% correct hash verdict; 20% correct handling of file
R01.10.01.02-T09 Upload a whole folder with its structure to a sandbox drive
Original instruction
In the sandbox cloud-drive web app, upload the folder `~/Projects/site_assets` (3 levels, 46 files) into `/Clients/Acme/` using the app's folder-upload option so the directory structure is preserved. Skip the `.DS_Store`/`Thumbs.db` files if the app asks.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Folder picker
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicknav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Drive DB: tree under /Clients/Acme/site_assets equals the local tree (paths + hashes) excluding system junk files; no flattened duplicates at the root. The drive issues no API token and accepts uploads only from browser sessions (CSRF + session cookie).
Proposed partial credit
60% files present with paths (scaled); 25% structure exact; 15% junk files excluded
R01.10.01.02-T10 Resume a large download that the server interrupts
Original instruction
Download `http://files.local/dataset_2gb.tar` in Chrome. The sandbox server drops the connection at about 60%. Resume it from Chrome's Downloads page (do not restart from zero) until it completes, then confirm it against `dataset_2gb.sha256`.
- Platforms
mac· macOSwin· Windowsweb· Any OS (web)- Applications
- Chrome
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
CL· CLI-sufficient- Interaction primitives
click· Point & clicknav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
interrupt· Interruptions- Proposed difficulty
- 4 / 5
Proposed verifier
Final file hash matches; server access log shows a Range request starting near the drop offset (resume, not full restart); exactly one complete file in Downloads (no ' (1)' duplicates). The OS process audit detects curl/wget fetches, so the resume must come from Chrome's Downloads page.
Proposed partial credit
30% interruption noticed; 40% resumed with Range; 30% complete file verified
R01.10.01.02-T11 Shrink an image to meet an upload size limit, then upload it
Original instruction
The sandbox visa portal rejects photos over 500 KB or not 600x600 px. `~/Pictures/passport_photo.heic` is 3.1 MB and 3024x3024. Use Preview (macOS) or Photos/Paint (Windows) to export a JPEG at 600x600 under 500 KB, then upload it through the portal's photo field.
- Platforms
mac· macOSwin· Windows- Applications
- Preview · Paint · Chrome
- Interfaces
desk· Desktop GUIbrow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsmulti· Cross-app- Work product
WP09· Files & folders- Conditions
multiapp· Multi-app- Proposed difficulty
- 4 / 5
Proposed verifier
Portal stored image: JPEG, 600x600, <= 500 KB, SSIM >= 0.9 against a reference downscale of the original; portal status 'photo accepted'. Images not derived from the original fail SSIM. The portal accepts uploads only from the browser session (CSRF token), and the process audit detects sips/ImageMagick/ffmpeg conversions, which void the export milestone.
Proposed partial credit
40% correct dimensions/size; 30% SSIM to source; 30% accepted by portal
R01.10.01.02-T12 Download a paginated set of invoices and rename them on save
Original instruction
The sandbox supplier portal lists 37 invoices over 4 pages. Download every invoice dated in Q3 2026 (12 of them) and save each as `INV-<invoice number>.pdf` in `~/Documents/Invoices/Q3`, using the Save dialog to rename. Skip credit notes.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Save dialog
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clicktype· Text & formsnav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
long· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Target folder contains exactly the 12 expected names, each with the hash of the matching invoice; no credit notes; server-generated filenames are opaque (doc_8f2c.pdf) so renaming is required. The portal serves PDFs only to the browser session (no API token) and the process audit detects curl/wget or scripted renames (mv/ren), which void the task.
Proposed partial credit
70% correct files with correct names (scaled over 12); 20% no extras; 10% folder
R01.10.01.02-T13 Save a chat attachment to a network share from an Electron app
Original instruction
In the sandbox team-chat desktop app (Electron), find the file `floorplan_v3.pdf` posted in channel `#facilities` yesterday and save it with the app's Save As dialog to the mapped share `Z:\Facilities\Plans` (Windows) or `/Volumes/Facilities/Plans` (macOS).
- Platforms
mac· macOSwin· Windows- Applications
- Team chat (Electron) · Save dialog
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
click· Point & clicknav· Pan/zoom/scrolltype· Text & forms- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
File at the share path with hash of the v3 attachment (v2 and v3-draft are decoys in the same channel); share server SMB log shows write from the client. The chat app has DevTools/remote debugging disabled and no API token for the agent; the process audit detects copy/robocopy/cp writes to the share, which fail.
Proposed partial credit
50% right attachment; 50% saved to share path
R01.10.01.02-T14 Upload a cloud-only OneDrive placeholder file through a browser picker
Original instruction
On Windows, `C:\Users\agent\OneDrive\Reports\audit_2025.pdf` is a cloud-only Files On-Demand placeholder (sandbox tenant). Upload it to the sandbox compliance portal. The first attempt fails because the placeholder is not hydrated; make the file available locally and complete the upload.
- Platforms
win· Windows- Applications
- Chrome · OneDrive · File Explorer
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clickhover· Hover & contextmulti· Cross-app- Work product
WP09· Files & folders- Conditions
hidden· Hidden/stale state- Proposed difficulty
- 4 / 5
Proposed verifier
Portal DB upload hash equals the real file; zero-byte or error uploads are recorded as failed attempts; file attribute shows pinned/hydrated afterwards. The portal accepts uploads only from browser sessions (CSRF + session cookie); attrib/PowerShell hydration or curl uploads are detected by Sysmon and void the matching milestone.
Proposed partial credit
30% hydration done; 60% successful upload with correct hash; 10% no stray failed uploads left attached
R01.10.01.02-T15 Download a form on Android Chrome and attach it through the document picker
Original instruction
On the Android emulator, open the sandbox school portal in Chrome, download `consent_form.pdf`, then on the `Submit documents` page attach that file using the Android document picker (navigate to Downloads), plus the photo `IMG_2041.jpg` from the Pictures folder, and submit.
- Platforms
and· Android- Applications
- Chrome (Android) · Files
- Interfaces
mob· Mobile touchbrow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
touch· Touch gesturesnav· Pan/zoom/scroll- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Portal DB submission has two files with hashes of consent_form.pdf and IMG_2041.jpg (decoy IMG_2014.jpg exists); /sdcard/Download contains consent_form.pdf; adb is not exposed to the agent.
Proposed partial credit
25% download; 35% PDF attached; 25% photo attached; 15% submitted
R01.10.01.02-T16 Upload a photo at full resolution from the iOS photo picker
Original instruction
On the iOS Simulator in Safari, open the sandbox print-shop upload page and upload the photo taken at `Lighthouse` (album `Trip`) at actual size, not the default reduced size. Use the Photos picker's options to choose 'Actual Size' or equivalent.
- Platforms
ios· iOS/iPadOS- Applications
- Safari (iOS) · Photos picker
- Interfaces
mob· Mobile touchbrow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
touch· Touch gesturesnav· Pan/zoom/scrollvis· Visual judgment- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Server received a file whose pixel dimensions equal the original (4032x3024) and perceptual hash matches the Lighthouse photo (album has 9 similar seascapes); reduced-size uploads fail the size milestone. Host-side simctl addmedia/openurl or curl uploads are detected in the host process audit, and the server accepts uploads only from the simulator Safari session.
Proposed partial credit
50% correct photo; 50% full resolution
R01.10.01.02-T17 Paste a window screenshot directly into a web issue form
Original instruction
Take a screenshot of only the `Calculator` window showing the error state, copy it to the clipboard (do not save to disk), and paste it into the description field of a new issue in the sandbox issue tracker, titled `Calculator overflow on 1e308*10`. Submit the issue.
- Platforms
mac· macOSwin· Windows- Applications
- Screenshot tool · Chrome
- Interfaces
desk· Desktop GUIbrow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
key· Shortcutsclick· Point & clickmulti· Cross-app- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Tracker DB: issue with exact title; attachment image whose size ≈ Calculator window bounds (±10 px) and SSIM >= 0.85 vs a grader capture of that window; no new image files on Desktop/Downloads. The tracker issues no API token and accepts attachments only via the browser session; shell-launched screencapture or PowerShell capture commands are detected and void the screenshot milestone.
Proposed partial credit
40% window-only screenshot; 40% pasted into issue; 20% title and no file saved
R01.10.01.02-T18 Control pair of T01: transfer the statement with curl and a session cookie
Original instruction
Control pair of T01. Using the sandbox portals' documented test REST endpoints and the API tokens in `~/Desktop/tokens.env`, download the same October 2026 statement for account ...4471 and upload it to the same accounting slot with curl only.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- curl
- Interfaces
api· API/SDKterm· Terminal- GUI need · task setting
AP· API-sufficient- GUI need · ordinary access
AP· API-sufficient- Interaction primitives
code· Code & commands- Work product
WP09· Files & folders- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Same DB checks as T01; request log shows API-token auth (no browser session).
Proposed partial credit
40% correct download; 60% correct upload
R01.10.01.02-T19 Drag a batch of files into a web uploader and set per-file metadata
Original instruction
In the sandbox digital-asset manager, drag the 8 images listed in `~/Desktop/batch.txt` from Finder/File Explorer into the upload queue, reorder the queue by dragging to match the list order, and set each file's `Credit` and `Alt text` fields from the list before starting the upload.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Finder · File Explorer
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
drag· Drag & droptype· Text & formsmulti· Cross-app- Work product
WP09· Files & folders- Conditions
multiapp· Multi-app- Proposed difficulty
- 4 / 5
Proposed verifier
DAM DB: 8 assets with hashes of listed files; upload sequence numbers follow list order (Kendall tau); Credit and Alt fields exact per asset. The DAM issues no API token and accepts uploads only from browser sessions (CSRF + session cookie); remote debugging is disabled, so files must be dropped in the page.
Proposed partial credit
30% correct 8 files; 25% queue order; 45% metadata (scaled)
R01.10.01.02-T20 Assemble and upload a sandbox visa application document set
Original instruction
For the sandbox e-visa portal: download the three blank forms (DS-A, DS-B, consent), fill each PDF with data from `~/Desktop/applicant.json` in Preview or Acrobat Reader, save them as `<form>_<surname>.pdf`, and upload each into its matching slot with the passport scan from `~/Documents/ID/`. Stop at the 'Submit application' confirmation page.
- Platforms
mac· macOSwin· Windows- Applications
- Chrome · Preview · Acrobat Reader
- Interfaces
brow· Browserdesk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
type· Text & formsclick· Point & clickmulti· Cross-app- Work product
WP09· Files & folders- Conditions
multiapp· Multi-applong· Long-horizonrisk· Irreversible/risky- Proposed difficulty
- 5 / 5
Proposed verifier
Portal DB: 4 slots filled; each uploaded PDF's AcroForm fields equal applicant.json; filenames follow the pattern; passport scan hash correct; application status 'pending_confirmation', not submitted. The portal issues no API token and accepts uploads only from browser sessions; the process audit detects scripted PDF filling (pdftk, Python) or curl uploads, which void those milestones.
Proposed partial credit
20% forms downloaded; 40% fields filled (scaled); 25% correct slots and names; 15% stopped before submit
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.