The assistant

The pane on the right is a conversation with your library. It has two modes, and the difference matters.

The switch is the mode picker under the message box, labelled with whichever mode you are in: Read only or Edit.

Read only

The default. Ask anything — “what did we decide about the ingest schema?”, “what's still open on the trip plan?” — and the assistant answers from your pages, citing the pages and sources it used. Read-only mode never changes your library, so it's always safe to poke around — and it never uses an agent run, so you can ask as much as you like on any plan. See Plans & limits.

By default it finds the relevant pages itself. To control exactly what it sees, tick specific pages in the library or pin a highlighted passage to the conversation.

Edit

Switch the picker to Edit and the assistant can act: update pages, restructure a section, merge duplicates. Small, safe changes are applied directly; bigger or riskier ones are queued for your review in your Inbox first. Each Edit-mode turn is one agent run against your daily allowance.

The split is not a judgement call you have to guess at. Changes that only add go straight in: filing something from the inbox, adding links, updating metadata, creating a new page from a source, appending a claim that cites a source, or recording an inference that is marked as one. Everything that could lose or distort what you already trust waits for you:

Changing or deleting a confirmed fact.
Resolving a conflict between sources, or merging duplicate pages.
Editing a page you marked protected.
Moving content from one space to another.

And some things the agent never does at all, in either mode: send or publish anything, expose private information, or take an action outside your library.

Reviewing what the agent did

Changes waiting on your approval land in your Inbox (bottom-left), alongside anything else awaiting your attention. Click anywhere on one — or its View changes link — to read the exact edit as a diff, added and removed lines per page, so you can see precisely what would change before you Approve or Reject it. The diff opens full-width in the middle of the screen and your Inbox stays open beside it, so you can work through a stack of them without losing your place.

Activity (bottom-left) is the ledger: every run and what it changed. Undo anything that landed. Each entry is tagged with how it ended:

Review (gold) — it needs you. Approve or reject it in the Inbox.
Committed (ink) — it landed. There's an Undo on it.
No action (ink) — the run finished and deliberately changed nothing, because there was nothing to change. Not a failure, and not a silent one either: a run that decides your library is already correct says so rather than inventing an edit.
Failed (terracotta) — the run did not complete. Nothing was written.

A failed entry shows the reason underneath it, and that reason is whatever the agent or its model provider actually said — sometimes a plain sentence (“your credential has no access to that model”), sometimes a line of tooling output. It is deliberately not smoothed into “a run failed”: the specific string is usually the difference between fixing it in a minute and guessing. Nothing is lost when a run fails — re-run it once the cause is dealt with.

Conversations that keep up

Chats are saved and sync across devices — start on the desktop, continue on the phone.
Chats title themselves after the first exchange.
A question already running is not tied to the screen you asked it on. Switch to another chat, reload the page, or close the tab — the assistant keeps working and its answer lands in the conversation you asked in. Come back to a chat that is still working and you see it working, trace and all.
The assistant takes one turn at a time, because it answers with your own connected subscription. While one chat is working, the others say which one it is and offer you the way back to it.
Past chats live behind the History button at the top of the assistant, newest first and grouped by age. Rename or archive one from its row there; restore an archived chat from Settings → Archive.
Enter sends; Shift+Enter makes a newline.

The assistant can also work on a clock: schedule a prompt — a morning brief, a weekly task sweep — and it runs unattended, landing results in its own conversation. See Routines.

Answers are only as trustworthy as their sources — see Sources & trust for how citations and knowledge statuses keep the assistant honest.