Chat: talking to your workspace

You just saved a stack of meeting notes when a question hits you: "what did we decide about the launch date last month?" Instead of digging through folders, just ask. The chat pane is where you talk to Consilience's AI agent. Its answers come from your own workspace: it searches your notes and knowledge graph, then shows exactly where each claim came from.

Chat lives where the window tour left it: in the left sidebar, under the titlebar switcher's "Agent" segment. Each conversation opens as its own tab in the right pane, next to your documents. This lesson covers the everyday moves: sessions, the composer, handing over context, steering, and citations. Approvals, plans, and rewind wait for the lessons on working with the agent.

Sessions: start, find, rename

A chat session lives on disk and belongs to the whole workspace; every document sees the same session list. The Agent segment gives you a "New chat" button, a "Search chats" field, and one row per past conversation, newest first with relative timestamps. A fresh chat reads "New chat". Once you send your first message, the AI names it, and you'll see "Naming chat..." while it works.

• Rename: right-click a session row and choose "Rename". Once you set a title yourself, auto-naming leaves it alone.
• "Pin as case…": turns a conversation into an investigation Case (a later lesson covers this).
• "Delete chat": asks you to confirm first, with "This cannot be undone." There's no trash for conversations, so it's really gone.

The composer

The composer sits at the bottom of the chat pane. Its placeholder, "Plan, Build, / for command, @ for context", is a quick map of what it can do. On the toolbar around it, the model badge is where you pick which frontier model answers and how hard it thinks.

Session

One saved conversation with the agent. It belongs to the workspace and shows up in the Agent sidebar.

Composer

The input bar at the bottom of the chat pane: your draft, plus send, stop, attachments, mic, and the mode badges.

Mention chip

An inline @ token in your draft that carries a file, folder, entity, past chat, terminal, or URL along as context.

Attachment chip

A file that rides along with your message: an image, PDF, text, document, audio, or a voice recording.

Steering

Dropping a message into a running response so the agent picks it up at its next step, instead of cutting it off.

Point at context with @

@ pulls anything from your workspace into a message, and / runs a command. Below is the real composer, with the @ picker open.

Type @ at the start of a word and a picker floats up above the composer. It ranks your workspace files and knowledge-graph entities side by side: exact matches first, then prefix and substring matches, with files winning ties. The categories below open submenus. Whatever you pick becomes a mention chip in your draft, and the real content travels with it when you send.

Chips act like solid objects. Click a file chip to open the file, hover to reveal the ×, and one Backspace next to a chip deletes the whole thing at once. The "Terminals" category hands the agent a build error with no copy-paste. The capture is a snapshot, so re-mention the terminal when you want fresher output.

Context flows back the other way too, with nothing to type. References to your notes in the agent's answers resolve against your live workspace, so a click lands you on the exact note, PDF, or entity behind the answer. Even a bare filename in the prose, like "report.md", becomes a clickable link when it matches a file you have.

Attachments, paste, and voice

There are four ways to attach a file, and they all feed the same pipeline. Each file becomes a chip labeled by kind: "Image", "PDF", "Text", "Document", "Audio", "Recording".

• "Attach files" in the "+" menu, the plain route through a file dialog.
• Drag and drop anywhere on the pane; the overlay reads "Drop files to attach".
• A clipboard paste. Screenshots and files land as attachments, while plain text comes through as text.
• Right-click or drag straight from the file tree.

Voice is a first-class input. Click the mic, labeled "Record voice"; its tooltip reminds you that stopping attaches the transcript. A recorder card floats over the composer with a timer and a live transcript. Press "Stop" and the transcript lands as a "Voice memo" chip. A second pass cleans up the text and splits it by speaker, and if it hears two or more voices, a "Name the speakers" dialog lets you label them. The footer is worth heeding: "Get consent before recording other people."

Steering and stopping

You don't have to wait while the agent works. Just keep typing. The placeholder switches to a "Responding" prompt with two options: Enter queues the message, ⌘Enter sends it right now. Queued messages wait in a strip above the composer and arrive at the agent's next step, so your nudge lands without derailing the run.

Where answers cite your notes

When the agent searches your workspace, you get back more than text. A small, interactive 3D mini-graph appears in the thread, titled "Where this came from in your notes", with a count line like "12 items · 18 relations" and a legend that separates "Directly matched" entities from "Related" ones. Click any node to open that entity in the graph view, or hit "Open in graph" to roam the full 3D graph.

The daily rhythm

That's the daily rhythm: open a chat, point at context with @, attach or dictate whatever's missing, steer without stopping, and follow the citations back into your workspace. Every citation lands you in a note that somebody had to write, so the next lesson opens the editor where those notes get made. How the graph behind the citations gets built comes soon after.