# Chats (/chats)


A chat is the unit of work. It owns its own model, permission mode, network
policy, attached files, connected folders, and the outputs it produced.
Nothing is global except your credentials and defaults.

Start one with **New chat** in the sidebar, or ⌘N.

## The composer

The message box is the main surface, and most controls hang off it.

- **Tools** (the `+` button) — **Attach files**, **Attach folder**,
  **Reasoning** effort, **Network** policy, and **Plugins**.
- **Model pill** — the model this chat runs on. See
  [Models and providers](/models-and-providers).
- **Permission selector** — Plan, Ask, Auto, or Allow all. See
  [Permission modes](/permission-modes).
- **Microphone** — dictate a message. Transcription runs on a local model,
  which is not bundled: the first time you use it, Tidebreak sends you to
  **Settings → Voice input** to download one. Cloud transcription is available
  there too, once the provider it belongs to is enabled.

Two typed triggers:

- **`@`** searches your files and folders and inserts them as attachments. It
  also offers **Browse files…** and **Connect a folder…**.
- **`/`** opens the plugin library — skills, plugin bundles, and saved
  prompts. Picking a skill attaches it to the message as a chip. There are no
  slash commands beyond this picker.

## While a turn is running

The composer stays usable. Typing and sending mid-turn **redirects** the
active response rather than queueing a new one — the send button becomes
**Redirect**. A separate stop button cancels the turn.

The chat header carries a context meter and an activity chip that opens the
per-chat panels.

## Panels

The right side of a chat holds panels, opened from the activity chip:

| Panel | Contents |
| --- | --- |
| **Document** | An attached source, opened for reading. |
| **Outputs** | Files the agent produced. See [Outputs](/outputs). |
| **Folders** | Connected folders and their access. |
| **Agents** | Background agents this chat started. |

Which panels are open is part of the chat's URL, so a layout is stable across
reloads.

## Background agents

For work that splits — reading twenty files, trying three approaches — the
agent can start background agents that run their own sandboxed turns and
report back. They appear in the **Agents** panel with their activity and task
plans.

They cannot start further agents of their own; delegation is one level deep.
**Settings → Agents** caps how many may run at once for a single chat.

## The inbox

Anything waiting on you across every chat collects in the **Inbox** in the
sidebar: approvals, questions the agent asked, plan reviews, folder access
requests, and requests to save a file into a folder. The sidebar badges the
count.

## Durability

Turns are journaled as they run. Closing the app mid-turn, or losing the
window, does not lose the work — reopening reconnects to the running turn and
replays what you missed. An approval card left unanswered is still there
tomorrow.

## Managing chats

The **Chats** section of the sidebar lists them, with a filter box behind the
section menu. Each row's menu has **Rename** and **Delete**; the chat header
menu adds **Copy debug info** and **Save debug bundle…**, which are useful
when reporting a problem.

## Projects

The data model has an optional project that can own several chats and hold
approvals that apply across them. It is not surfaced in the desktop app yet.
Everything you do today lives in a chat.