Skip to content

Arelis v0.2.7

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 06 Sep 04:01
· 575 commits to main since this release

Heads up: the Windows installer in this release does not work. It installs, but Arelis never opens. The installer build left out a graphics file the app needs at startup. I found this on Oct 4, 2026 by testing on a clean PC. I had only ever run Arelis on my own PC, from a developer version, which is different from the installed one, and this was the first time I installed it myself, so I had no idea. I tested 0.2.9 directly, and the installers from 0.2.4 on were built the same way. A fixed installer is coming in v0.3.0. If you downloaded this one, I'm sorry. Please wait for v0.3.0 or run Arelis from source, which is not affected.


Arelis v0.2.7

Version 0.2.6 got her through finishing research tasks properly. The last four days of work covered everything else — photos, the desk, mail, texts, calendar, and the remaining ways she was still doing tasks wrong. Here's what changed.

She does what you actually asked

Every request now gets evaluated against a clear, one-line goal instead of a vague match. A temperature question isn't treated as a weather lookup. "1960–2026" isn't misread as a subtraction problem. "Email it to me" means send it to you — not to a contact literally named "me." A finished text message that still contains leftover placeholder text like "I love her" won't get sent as-is. A Bloomberg captcha page isn't mistaken for the actual research report. And simple math stays in the chat — she won't launch a 32-step web crawl and generate a PDF just to answer basic algebra.

A screenshot isn't treated as a text message. Drop a file onto the tray and she'll read it directly. If a browser session crashes mid-task, your next message starts a fresh request rather than trying to resume the broken one. Starting a new topic opens a new chat, and the previous conversation stays safely in your history.

Pictures

Image generation runs on Comfy: you can choose square, 16:9, or portrait format, apply a style, and get four variations if you ask for them. If you hand her a photo and say "make this a watercolor," she edits that photo rather than inventing a new image. Cut-out and outpaint tools remain in the image generator; resize, crop, upscale, and rotate/flip stay in the image editor.

The desk now includes a proper image library — showing file names, prompts, square thumbnails, and left/right navigation. If you paste in a photo and ask "who is this?", she may identify a public figure. However, the live camera still won't attempt to identify people in the room.

Desk

The workspace no longer pops open empty every time she does research — a PDF or report she just finished writing now lands on the desk and opens automatically. Opening a file no longer leaves an awkward empty column next to the page.

Math written in a note using TeX notation (like $$ or \frac) now renders as properly formatted math, both in the saved file and in the chat bubble.

Permissions

Viewing images and files, screen access, and window control are now granted automatically. Sending mail or texts, deleting things, making payments, and running programs still require confirmation via a permission card. In Settings, you can manage this at two levels: turn off confirmation for an entire category, or hit "don't ask again" to stop being asked about that specific action going forward.

Notifications, texts, mail, and calendar

The inbox now works like an unread list — click any item and it opens directly: a text opens as a chat, a job opens its file, and mail opens in a reader you can reply from right there. Snooze options are 5 minutes, 15 minutes, 60 minutes, or tomorrow. Re-syncing your phone no longer re-surfaces last week's texts as if they were new.

Text threads now persist across restarts, messages show timestamps, and failed sends offer a retry option. Contacts also have a direct "Text" action.

Mail and phone connections are now managed in Settings → Notify — no more manually editing secrets.yaml. If a task needs mail access and it's not connected, she'll say so.

Calendar now works locally without needing Google. If your Google token expires, she'll still create the event locally and then prompt you to sign in — no terminal commands required. Duplicate events (same title and time) are merged instead of creating two entries. When a day has too many events to show, it displays a "+N more" indicator. A recurring authentication issue now shows as a small sign-in chip rather than a disruptive popup.

Her Chrome browser

Just tell her the task — if you grab the mouse while she's controlling her browser window, she'll pause instead of fighting you for control. Sign-ins on X are handled by you; she won't keep clicking "Continue with phone" on your behalf. A click that times out no longer causes the whole tab to close.

Phone

Opening the app for the first time each day starts a fresh chat, just like on desktop. Yesterday's conversations remain accessible in your chat history.

Reality mode (still in early access)

The 3D avatar, hand tracking, and Earth view still require checking out from source — Reality as a standalone room is still shipping as-is.

Earth is now a destination you enter after traveling there — say "take me to Tokyo" or give a street address, and she'll fly there. UI elements now follow your actual destination instead of lagging behind on the travel path. Double-clicking a plane or an earthquake marker takes you straight to it. Markers now look like what they represent — a plane, a ship, a station — instead of generic chevrons and rings. Streets are labeled with their real names, along with a compass and a distance readout. Entering a location turns on Live mode; leaving takes you back to the solar system view.

Unchanged from 0.2.6

"Deeply research" still triggers a full, long-form research run even if the settings say "fast mode." After a long web scrape, she now summarizes findings in her own words instead of pasting raw page content. If you grant access to a folder, she can see and read the files inside it.

Still not implemented

Nothing is code-signed yet. Ollama and the voice model weights are still not bundled into the .exe.