Start A Chat Conversation

This guide runs from a blank composer to a saved conversation: opening a composer, choosing what you are talking to, setting the mode, sending, and what you can do with the conversation afterwards. Chat Overview covers the module as a whole.

Open A Composer

Chat is a module in the left rail, and its home view is a blank composer. You can also open the floating panel from anywhere else in the app — the trigger sits in the bottom right, and it is hidden while you are already on a Chat page because the full page is the same thing, larger.

The conversation is created when you send, not when you open. An abandoned composer leaves no row.

Choose What You Are Talking To

Two tabs sit at the top of the composer.

LLMs are models Dots calls directly and runs the tools for. Choose this when the work is about your knowledge — answering from your Dotabases, drafting, summarizing, reasoning over what you have captured.

Local Agents are coding-agent platforms that run on your own machine with their own file and shell access. Choose this when the work is about your code, in a directory on this computer.

The model pill next to the tabs opens a searchable list grouped by maker. Models that can call tools are marked. For local-agent models that support it, a reasoning-effort row appears with low, medium, and high.

A provider that is not usable names the reason: a binary that is not on your PATH, an API key that is not set, a gateway that cannot be reached. Local agents need their CLI installed and authenticated; the direct providers need a key in the environment. If a machine has no coding agents on it at all, the composer falls through to whichever direct provider is available.

Set The Posture

Open the Context Picker pill and choose a Mode. Default has full agency; Ask, Plan, Research, and HUMAN.md Builder are read-only. The mode is saved onto the conversation, and the transcript shows a divider when it changes, so a thread that started as a question and became work says so.

Local-agent conversations also carry a tool-posture control of their own — a shield toggling between full access and read-only for the workspace files. Read-only guarantees the agent makes no modifications.

Point A Local Agent At A Directory

If you picked a local agent, a workspace pill appears. This is the directory the agent runs in, and it only exists for this family — a direct model has no filesystem to point at.

Pick from your recent paths, your registered workspaces, a native folder chooser, or a typed absolute path. Registered workspaces are managed under Settings, in the Workspaces tab, where you can rename them, set a default, and remove them; local agents can only run inside folders you have registered, and the registry always keeps at least one.

Changing the directory mid-conversation is called out rather than done quietly. Before you send, a banner warns that sending will start a fresh session and offers to restore the previous one; after you send, a notice records that earlier turns ran somewhere else.

Send

Type and press send. The controls around the input:

• Attachments go through the + menu — up to ten per message, fifty megabytes each.
• @ opens the mention picker for rows, connectors, and Superpowers.
• / opens the command menu, which carries /capture for saving the draft itself into a Dotabase.
• The send button is one slot with three states. It is a microphone while the draft is empty, the send button once you type, and a Stop button while a reply is streaming.

You do not have to wait for a reply to keep going. Messages you send while one is running are queued, up to five, and run in order. The queue indicator opens a list where you can remove individual messages or clear it, and each queued message shows the mode, model, and Superpower that were active when you queued it — those are frozen at that moment, so changing the pickers afterwards does not rewrite what is already waiting. Pressing stop with a non-empty queue asks first, because stopping clears it.

The Conversation Names Itself

The row is created immediately with a placeholder title taken from the first characters of your prompt, then retitled once the first reply completes. Generated titles are kept to a few words. If titling fails, the placeholder stands rather than a blank.

You can rename it yourself at any time from the header, and a title you set is never overwritten.

Keep It

The sidebar holds your conversations in three sections — pinned, folders, and recents — with its own filter box for narrowing the loaded list by title. A conversation belongs to at most one folder, folders nest up to three levels, and they carry a color drawn from the DIIICE categories.

Once a conversation exists you can:

• Fork from here on any message, which starts a new conversation from that point and records a link back to the original.
• Edit and run, retry, or regenerate a message, which creates a branch rather than overwriting. Re-running a full-access turn confirms first, because file changes and commands may run again.
• Export as Markdown, JSON, plain text, or print.
• Save as Plan, which writes the plan into Artifacts with provenance back to the conversation.
• Select text in any message to save it as a DOT, connect it to an existing one, quote it in your reply, or ask a follow-up about it.

All of it persists because the conversation is a row. Conversations covers what that row carries: search across it, edges out to the work it produced, and metrics derived from the body rather than entered by hand.

Next

Give Chat the Right Context covers the other half: opening the right Dotabases and memories to a turn, choosing which HUMAN.md applies, and scoping a turn down to a curated set of rows.