Dots Constellation Die
Docs

Chat

Conversations held against your own workspace: two families of model, five modes, a context aperture you set per message, and a transcript that lands back in your workspace as a row.

Chat is where you hold a conversation against your own workspace. Each message goes to a model you picked, alongside the context you opened to it for that turn: the Dotabases it may read, the slice of the Memory Bank it may draw on, and the HUMAN.md that describes how you work.

None of that is a separate store built for the chat. The model reaches the same rows the CLI and the web app read, and the conversation itself lands back in your workspace as a row you can search, fork, and link to the work it produced. Whichever model you pick answers as the Dots co-worker, the persona the assembled system prompt opens with.

One distinction shapes the rest. Selecting a source is not the same as pasting one: opening a to a turn authorizes the model to go looking, rather than pushing its rows into the prompt. That is what lets a workspace of any size sit behind a single message without the message growing.

Key Features

  • Two Families — LLMs that Dots calls over an API and runs the tools for, and local coding agents on your own machine, which bring their own shell and file access.
  • Five Modes — Default, Ask, Plan, Research, and Builder. Only Default runs at full tool posture; the other four are read-only.
  • Per-Message Context — the Context Picker's whole lens is snapshotted onto each message as you send it, so changing the picker never rewrites a message already sent.
  • Workspace Retrieval — the model searches the rows of the Dotabases open to that turn, rather than answering from a copy pasted into the prompt.
  • Superpowers On The Turn — a resolves at dispatch and is prepended ahead of the system prompt; a Power-Up Playbook also confines the turn's reads to the rows inside it.
  • Automations — a saved prompt on a schedule. Each fire opens a fresh conversation with every source switched off, so what the run needs sits in the prompt itself.
  • Recorded Once — a turn persists exactly once, carrying its input and output token counts, duration, model, and cost. The typed dispatch failures throw before anything is written, so a half-finished turn never lands.

Get Started

Resources

  • How Chat Works — the anatomy of one turn: the system prompt blocks, retrieval, dispatch, and persistence.
  • Conversations — the Dotabase your transcripts land in, and what you can do with one afterwards.
  • Dotabases Overview — the knowledge a conversation reads from and writes back into.

Get Started