Checking a large model file in pieces covers what happens when someone in your group shares a local AI model: NAOMS fetches it straight from them over a peer-to-peer connection and checks every byte against its signature before keeping anything. The receiving side used to read the whole file into memory to do that, capping models at 256 MiB. It now streams — a measured import of a 5 GiB file needed about 20 MiB of extra memory, not 5 GiB of it.

Weather should not be written down is a bug-fix article with a clean rule behind it: some state only matters for a moment — who is online right now, where a shared cursor currently sits — and a tool meant for exactly that kind of passing state had ended up mirroring an entire chain's graph and rewriting a multi-megabyte database record on every change. The fix, and the rule it leaves behind: passing state lives in memory and is never written to disk.

Four tabs for things that need you tours the Approvals window: a new device asking to join your identity, an app asking to install, an AI plan waiting to run, a group vote where it's your turn, an invitation, a question from an agent — all of it lands in one window, split into four tabs (Approvals, Governance, Memberships, Agents). Two of the tabs say, in opposite ways, what happens if you say nothing at all.

A whole office suite inside NAOMS covers Docs, Spreadsheet, Presentations and a PDF viewer running on your own machine, with live co-editing between two machines built in — one participant's own computer keeps the changes in order, so there is no document server in the picture. It shipped inside the main NAOMS codebase on 18 September and moved into its own repository shortly after, as part of the broader "1797" package-extraction rewrite; NAOMS now loads it externally. Its own commits are not part of this week's area counts for that reason.