The Assistant page

The Assistant is offered to a person who may use the assistant, once an assistant is set up for the desk, even before the rest of setup is done. The desk forwards the assistant’s doors only for such a person, and the assistant checks again. Its pages in the side are New conversation, All conversations, Shared, Memory, Plans waiting and Standing grants, then your recent conversations.

Talk

1. Press New conversation.
2. Start from one of the ways to begin, or write your own. A way to begin puts its words in the box, to change or to send. Nothing is sent by itself.
3. Press Enter. There is one input and one assistant: it asks about the registry, builds a query, plans work or tunes the rules, whichever your words need. You never choose a station.

While a turn works, one live line says what the assistant is doing. ? opens the steps behind it. A plan of several steps shows while they run. What the assistant did in a turn folds into one line of steps.

• A change waits on one approval card. The card says what kind of change it is, that it waits for you, and its steps. Nothing runs until you press Approve (the card’s own word). Change it hands the words back to you to say what to change. Not now declines.
• A question comes as one card with its choices as buttons. A click sends the choice as your next message.

A conversation is named by its first words, then, once its first answer settles, by a few words the assistant chooses from your first message alone. A name you give always stays.

How full a conversation is

A model reads a conversation within its context window. The meter below the message box shows how much of the window the conversation holds, such as 18% of 64k, and turns amber from 70%. Hover it for the tokens.

Before the window fills, the assistant summarizes the earlier turns and keeps the latest ones word for word, and the thread notes that it did. The whole conversation still shows on the page; only what the model reads is shorter. /summarize summarizes the earlier turns now. A conversation too short to summarize says so, and no model is asked.

The thread

Answers come as formatted text: headings, lists, quotes, notes, tables, footnotes, code coloured by its language with a button that copies it, and mathematics written between dollars. A link opens only to a web page or a mail address, and an image is named, not shown.

What a model reasoned before answering shows folded above the answer. It is never copied with the answer, exported or shared.

Hover over a message you sent to copy or edit it. An answer can be copied, asked for again, or marked Good answer or Missed, with a reason if you give one. Editing a message, or asking for an answer again, continues the conversation as another version from that message; the arrows beside the message, such as 2 / 3, move between the versions.

The message box

Key	What it does
Enter	sends the message
Shift+Enter	starts a new line
Esc	stops the answer being written
Up, in an empty box	edits your last message
Tab	completes a command, or takes the first thing @ offers

A message that starts with a slash is a command:

Command	What it does
/new	starts a new conversation
/fork	continues in a copy of this conversation
/share	opens the share dialog
/remember what	keeps something for your later conversations
/plan question	plans a pipeline run; it opens filled in on the Pipelines page, for you to start
/export	saves this conversation as a markdown file
/summarize	summarizes the earlier conversation, so it takes less of the context
/rename name	renames this conversation
/status	shows the model, the station and how full the context is
/help	lists the commands

Type @ to name one of your query cards, a cohort, a result kept for you, or a saved selection, and pick it from the list that opens. It shows in your message as a chip, and a card’s chip opens the card on the Query page. The assistant reads what you name with its own tools, so naming something lets it read nothing it could not read before.

Find a conversation

1. Open All conversations.
2. Search by title, or read the groups: Pinned, Today, Yesterday, This week and Earlier, by when each was last used. The archived are on a tab of their own.
3. Rename, Pin, Archive, Export and Delete are on each row there and at the head of the conversation. Export saves the conversation as a markdown file: what was said, the steps and the query versions proposed, never what a tool returned. Delete asks first.

When you refer to something you asked before, the assistant can find your own earlier conversations by the words of their titles and of the answers they ended with. It never reads their messages, and never anyone else’s.

Work a query beside the chat

A query answer is one quiet card in the thread: “Query · v2”, its title, the answer’s funnel as subjects, visits and scans, and the top three bars of its first chart.

1. Press Open on the card. The version opens on a panel beside the chat, which narrows on the left. The panel follows the version the conversation is on.
2. On the panel: the question’s conditions as chips to take away or add, the answer’s funnel as steps to choose, its charts (a bar narrows the question when clicked), and the versions.
3. Every change on the panel is a move the engine offered and makes the next version. Saying it in the chat makes one too.
4. Press Save as a selection to keep the version under a name.
5. On a version the assistant proposed, press Keep (it names the version). It is a draft until you do; kept, it is a card on the Query page’s Cards list and the panel says kept. Unkeep puts it back among the drafts and deletes nothing. Keeping needs work on Query.
6. The icon at the panel’s head opens the card on the Query page. On the card in the thread, Download CSV saves the answer’s rows at your own detail level.

A conversation started from a card’s discussion opens on that card here too.

Plans waiting

Plans waiting is the inbox for what the assistant planned for you.

Section	What it holds	You
To confirm	a plan the assistant restated in a conversation	press Confirm
Held with a reason	a step it may not take yet, for want of a standing grant or a part that is down	read why
To decide	a change it may only propose	decide it in its conversation
Going	plans that run	follow them

A plan lists its steps and what each one will queue. Nothing runs until you confirm it.

Standing grants

Standing grants is what the assistant may do for you without asking each time. Each is revocable.

1. Read what stands now, in plain words, under the page it belongs to (Data, Review, Pipelines and so on).
2. Press Give on one action, Give all for a page, or Give everything. Only actions your own grants open are offered.
3. Press Revoke to take one back, or Revoke all for a page.

An admin is told the assistant may do everything they may, with nothing to give. A person who holds the assistant’s settings can also read Everyone’s grants and revoke any of them.

Share a conversation

1. Press Share, in a conversation’s head, or write /share.
2. Choose people from the desk, or everyone on the desk who uses the assistant.
3. The dialog shows who has opened the share and gives its link. Update the share brings it up to the latest turn, and Stop sharing closes it at once.

A share shows what was said and the query versions proposed, as the conversation stood when you shared it or last updated it. It never shows what a tool returned. A card in a share opens on the Query page with what the reader may see and do. Depending on what you see in records, a conversation may have read identifying details or sensitive values; a reader who does not see them is refused, and told which.

Shared lists what others share with you and what you share. A share opens read-only, and Continue as mine starts a conversation of your own from it: the assistant reads what it had read up to the share, and works with what you may see and do from there. The people to choose from are the desk’s users when the desk keeps its own, and everyone who has signed in at least once through a provider.

Memory

Memory is what the assistant reads at the start of each new conversation:

• The install’s instructions, written by a person with work on the assistant’s settings: the cohorts and what they mean, the conventions, a short glossary. Every version is kept, and everyone reads the current one.
• What you asked it to keep, at most 300 characters each: a preference, a fact about your work, or a way you work. Add, edit or delete them on the page, say “remember that …” in a conversation, or use /remember. When the assistant offers to keep something, it waits for Save or Not now.
• Notes from your work: a query you settled or a proposal you disregarded leaves a short note. You can delete them, and they go with the retention.

A conversation reads memory once, at its first message. It reads all you asked it to keep while that fits in 3,000 characters; beyond that, it reads what is closest to your first message. Pause memory stops both reading and keeping, and Delete everything removes what the assistant keeps about you. What you keep is yours alone: a conversation someone continues from your share reads their memory, not yours.

The assistant refuses a memory that looks like a person’s data: an identity number, a UID, a long identifier, a full date, an e-mail address, a subject code or a series description.