What’s new? v4.82 — Voice coming soon

WebDAV improvements and, soon, voice.

v4.82.3 — The network folder stops promising what it can’t deliver

Our own WebDAV client found it. iuredav diagnoses the server before mounting it, and it reported that the server advertised things it didn’t deliver. It was right, and the exact spot was a test file that doesn’t exist yet — which is precisely what a diagnostic creates. There, “lock while editing” returned a server error.

That matters more than it seems: a server error reads as a transient failure, so the program retries indefinitely. Our client ended up dropping locking altogether so as not to leave operations half-done, which is exactly the right thing to do with a server that breaks. Now it answers with a clear refusal that says what to do: upload the file first. On an actual document or folder, locking and setting the date still work — which is what Finder, File Explorer and Office do around every upload.

A made-up tag now ceases to exist. In tag views, any name resolved to an empty folder: the tree claimed to have something it didn’t.

And desktop junk no longer gets into the case file. macOS creates a hidden file for every file you touch in a network folder. They were slipping in because they inherit the original’s extension —a ._Minuta.md is a perfectly valid .md— and were being saved without the leading dot, that is, looking like a project document. Mounting the case file from a Mac kept filling it with ghosts. All ten instances were checked before closing this out: none had been saved yet.

Voice dictation now has a switch, instance by instance, from the central console and without updating anything. It remains off: the browser path still needs to be verified.

Updating doesn’t change any data or turn anything on.

v4.82.2 — The chat microphone stays off until it’s verified

The previous release introduced voice dictation. It failed in real use, and we’d rather turn it off than leave it half-working. In some browsers the microphone kept asking for permissions that had already been granted.

Looking into it revealed something deeper. Dictation had two paths, and the browser picked one without asking: through the firm’s server path, the audio is transcribed in-house and never goes to anyone else’s cloud —that one worked—; through the browser path, Chrome and Edge send the recording to their vendor’s servers, and that was precisely the one that was failing and the one those browsers always took.

The previous release notes said the audio never went to third-party services under any circumstances, and that was false for the second path. It has been corrected there and here. In a product sold on data sovereignty, that sentence couldn’t stand.

The “Read aloud” switch is also gone: it only reads answers to dictated questions, so without a microphone it was a control that turned on, went green and could never do anything — which is why tapping it did nothing.

Typing in the chat changes absolutely nothing. The administrator has the switch to turn dictation back on once it’s resolved, with no need to update.

v4.82.1 — The “Mark all as read” button works again

On the Notifications screen, that button did nothing — and gave no warning either. It asked the server for an address that doesn’t exist; the server said no, and that error was written to the browser console, where nobody sees it. You clicked it and absolutely nothing happened.

The bell in the top bar always worked fine. There are two paths in the product to mark everything as read, and the broken one was the screen’s — the one fewer people open, because the bell is enough for day-to-day use. That’s why it stayed that way for a while without anyone reporting it.

Now it actually marks them and, if something fails, it says so on screen instead of staying silent. A button that can’t do its job and stays mute is worse than one that shows an error: whoever clicks it tries again, thinking they missed.

If you’re coming from 4.81.x, jump straight here: you also get voice in the chat and the MCP server fix from 4.82.0, updating just once.

v4.82.0 — You can talk to the assistant, and it can answer out loud

The chat gets a microphone. You dictate the question instead of typing it, and the answer to a dictated question is read aloud with whatever voice the device itself has installed. Typing remains exactly as silent as before: only what was asked out loud is read.

Where the audio is transcribed, stated precisely. When the clip travels to the firm’s server, it is transcribed there and never goes to anyone else’s cloud. But when the browser transcribes it, the browser may send it to its vendor’s servers: Chrome and Edge do. The first version of this post claimed that the audio never went to third parties under any circumstances, and that was false for that path — corrected the same day. That’s why dictation is off by default until that part is resolved (see v4.82.2).

When you dictate, the answer changes shape, and that’s on purpose. Two to four sentences, in prose, with no tables, lists or code blocks: what’s read aloud is listened to, not skimmed. Citations are said once (“according to such-and-such document”) instead of reading out the entire marker. When you ask in writing, the assistant answers as always.

Scoped so it doesn’t get in the way of long-running work. Clips of up to 30 seconds and one dictation at a time across the whole firm. The reason is concrete: transcribing a one-hour meeting is the most expensive thing the server runs, and an unlimited microphone would take away its turn. The administrator can turn off the microphone for the entire organization without updating anything.

Security fix. Deactivating a person now also cuts off their external agent. The check was written, but it looked at a field that doesn’t exist in the user model, so it never fired: an agent kept answering after its owner had been deactivated. It’s worth knowing the rule, which applies equally to network folder passwords: deactivating doesn’t delete credentials, it closes the door; reactivating the person gives them back their access without generating anything, and if the deactivation is permanent, the credentials must be revoked.

In the Android app: you can now switch servers while offline —before, you needed a network connection to reach that screen—, the session is no longer backed up to Google’s cloud, and the app targets Android 16.

Updating the instance doesn’t turn on the microphone by itself: dictation requires the central server to offer the voice service. If it doesn’t, the chat simply doesn’t show the button.

voice webdavs

← Back to the index