<!-- Source: https://console.uubright.com/docs/chat-mode/ -->
# Chat mode

> **Chat workspaces are being retired.** The app no longer offers a way to create one or
> to move a workspace into chat mode, and SciMigo Studio is moving to its own app at
> scimigo.com. A chat workspace you already have keeps working as described here.

A workspace that isn't about code (lesson slides, Shorts, courseware, writing) opens in
**chat mode**: a ChatGPT-style window with **New chat**, your **Chats**, **Projects**,
the **Library** and **Apps**. There's no Manager, workers or board to look after. It's the
same engine underneath: a chat is an ordinary task, so questions, approvals, attachments,
deliverables, notifications and phone access all work as they do in work mode.

```
SciMigo Studio ⌄
✎ New chat
Search chats
Projects            2
Library
Apps                1
TODAY
  Ch. 4 Photosynthesis…  Needs you
  A short on why leaves change color
YESTERDAY
  Cell division slides, Grade 8
⇄ Use work mode here
```

## Your chats

The list groups chats by day: Today, Yesterday, Previous 7 days and Older. **Pinned** chats sit above
them. Each row's **⋯** (on a phone, also holding the row) offers:

- **Rename**: type the new title and press Enter. Escape keeps the old one.
- **Pin** or **Unpin**.
- **Archive**: the chat leaves the list for **Archived · N** at its end; **Unarchive** brings it back.
  Nothing is lost, and an archived chat still takes messages.
- **Delete…**: tap again to confirm. It removes the chat, its history and its files. Only the owner can
  delete; a paired phone sees the other actions.

**Search chats** finds chats by title and by what was said, both your messages and the answers,
in this workspace only. A match in the conversation shows the words around it under the title.

## Which workspaces use it

The mode is set per workspace. If you haven't chosen one, a named workspace with
[plugins](https://console.uubright.com/docs/workspaces/) and no repositories counts as a non-coding one and opens in chat mode.
Default and every workspace with a repository open in work mode.

- The workspace menu (the workspace's name at the top of the sidebar) switches workspaces and opens
  **Workspace settings**.
- **Use work mode here**, at the bottom of the sidebar (or under **More** on a phone),
  switches the workspace to work mode: the Manager-first view in [Your Manager](https://console.uubright.com/docs/manager/).
- **Mode** in **Workspace settings** shows the current mode and switches it either way.
- On the server: `npm run workspace -- set-mode w-... chat` (or `work`).

Switching changes only how the app shows the workspace. Chats stay; in work mode they are
listed as tasks.

## Templates

**New workspace…** (in the workspace menu, or under **More** on a phone) asks what the workspace is
for. Choose **Non-coding**, then a template to start from:

| Template | What it sets up |
|---|---|
| **Teacher** | Lessons, quizzes and slides. Instructions for lesson structure and answer keys, math written as LaTeX ($…$ and $$…$$), which answers show as formulas. Starters: *Plan a lesson*, *Make a quiz*, *Make slides*, *Explain it simply*. If this computer has a teaching app installed, it's checked for you. |
| **Writer** | Outline, rewrite and summarize, keeping your voice. Starters: *Outline a piece*, *Rewrite this*, *Summarize*. |
| **Research** | Searches the web, reads the pages and cites every claim with a link. Starters: *Research a question*, *Compare options*, *Find sources*. |
| **Blank** | Just chats. Add instructions and apps later. |

Every template opens in chat mode. Its starters are plain prompts kept in the workspace, so they
need no app: choosing one fills the message box with a prompt such as *Plan a 45-minute lesson on
[topic] for [grade]…*, ready for you to fill in the brackets. The template's name fills an empty
**Name**. The instructions and starters are copied into the workspace when it is made, so later
changes to a template only affect new workspaces. On the server, templates are JSON files in
`src/server/templates/`: add a file to add a template.

## New chat

In the built-in **SciMigo Studio** workspace on This Mac, sending from New chat opens a
native Studio thread and asks the question there. **Open Studio conversations** opens
its saved threads without sending. This uses SciMigo device sign-in and the question
service; it does not run a coding agent or create a task. See the [Studio product
boundary](https://console.uubright.com/docs/workspaces/#scimigo-studio-workspace). Existing legacy connector workspaces
keep their earlier chat behavior.

**New chat** opens a greeting (by your device's own clock) and a message box. Describe what
you need, or attach a chapter, slides or notes with **+** (or drop them on the box).

If the workspace has [projects](#projects), **+** also offers **Choose a project**. The project
you pick shows as a small chip above the box; its **×** takes it off again. A project chat starts
with the project's instructions and memory. Without projects, **+** attaches files straight away.
The workspace's apps are listed under **Apps** (on a phone, under **More**).

Below are **starters**: the workspace's own, from its [template](#templates), then the cards the workspace's apps offer: two to a row on a computer, and
on a phone a row of small cards you swipe sideways, so the message box stays the first thing
you see however many starters there are. An app can give each card a picture, or a color and
an emoji ([Starters](https://console.uubright.com/docs/plugins/#starters)). Choosing a card that asks for details opens a short form;
**Use** fills the box with the app's prompt and your answers. A card without a form fills
the box directly. Check or add to the text, then send. An app that declares no starters
gets one card per command (the plugin's `commands/`), which fills the box with the command.
If a chat is waiting for you, a **Needs you** line under the starters opens it.

**Your first chat.** In a workspace made from a template (see [Workspaces](https://console.uubright.com/docs/workspaces/)), New chat
shows **Try asking** with three example prompts until you start your first chat. Tapping one puts it
in the message box to edit or send. The examples are shown once on each device: they go when you use
one, hide them with **×**, or the workspace has a chat.

The first message starts the chat, and it gets a title from the agent as usual.

## A chat

A chat opens in the ordinary task view, drawn as a conversation. Your messages are bubbles on
the right (long-press to copy one). Answers are plain text, with a small row under each finished
answer: **Copy**, **Share** where the device has a share sheet, and
[**Save to project**](#project-files). While the agent works, one
quiet line says what it's doing, such as *Searching the web…* or *Reading 2 pages…*. Afterward
that line becomes, for example, *Searched the web · 3 sources*, which opens on the searches and
the pages read, with links. Decision, question and app form cards appear inline, the
same cards as everywhere else.

**Answer actions.** Under the latest answer, **Regenerate** (↻) asks for a fresh attempt at your
last message; the chat shows *Regenerated the answer* and the new answer below the old one.
Under your latest message, **Edit** (✎) puts it back in the message box with *Editing your message.
Send replaces it.*; the new version appears marked *Edited*, and the agent answers it instead of
the earlier one (both stay in the conversation). **Cancel** leaves it as it was. While the agent
works, a line under the conversation says what it's doing, and **Stop** (■) is in the message box.
Edit and Regenerate wait until the answer has finished.

**Approvals are in plain words.** When the agent needs your OK (even in Auto, some actions still
ask), the card says what it would do, such as *Search the web for San Jose weather*, *Read the PDF
you attached (chapter-4.pdf)*, *Run `rm -rf old-drafts`* or *Use studio to render deck · 450
credits*, with **Allow** and **Don't**. The headline is built from what the request actually does:
the command itself, the file's path (in full when it is outside the chat's folder), the app and the
amounts in its input. What the agent or app says about it is a second line, *Claude says: …*, so a
reassuring description can't hide what will run. A tool the app doesn't know shows its own name and
what it was given, on one line, such as *Use tool_x · cmd: rm -rf old-drafts*. **Details** shows the tool's name and exactly what
it was given. A web search also offers **Allow web searches in this chat**. Coding tasks keep the
full approval card.

The heading is just the chat's title and **⋯**. The status (*Ready*, *Working*) is in **⋯**, and so
is **Show details**, which brings back the task's **Details** and the runner's activity log for
when something goes wrong. Branches, changes, pull requests and the model and permission chips
are hidden. Coding tasks keep the full task view.

Under the message box, chips show the chat's project, **Project memory · N** (it opens the
project's memory) and the apps.

**Chats run in Auto.** Claude reviews each action itself instead of asking you, so a question
like "what's the weather in Paris?" searches the web without an approval card. Tasks in work
mode keep the server's default. To approve each action in a workspace's chats instead, choose
**Chats run in: Ask me** in **Workspace settings** (chat-mode workspaces). Only the owner can
change it; a paired phone shows the setting without a control. Either
way, an app's own confirmations, such as spending money or making media, still come to you as
the app's form: Auto doesn't answer them. See [Permissions](https://console.uubright.com/docs/tasks/#permissions) for what Auto does.

**Idle chats cost nothing.** A chat whose agent has stopped (it was idle, or the server
restarted) says *This chat is resting*. Just reply: your message starts it again where it
left off. There's no Resume button.

## Projects

A project keeps related chats together, like a folder with a brief. Each project has:

- **Chats**: the chats started in it.
- **Instructions**: what every chat in the project is told first, for example
  "Grade 8 science for Greenleaf Press. British spelling. Keep decks to 12 slides."
- **Files**: reference material every chat in the project can read, such as a syllabus, a style
  guide or a chapter.
- **Project memory**: short notes the agent reads at the start of every chat in the project.

**Projects** in the sidebar lists them, and **New project** adds one. A project's page has a box
for a new chat in that project. **Rename** changes its name, and **Delete project** removes its
instructions, files and memory. Its chats stay, out of the project.

### Project files

**Add files** on the project page (or drop files on the **Files** card) adds PDFs, Word files,
Markdown, text and images. Each one goes through the same checks as a file you attach to a message:
its real type is checked, programs are refused, it's scanned for viruses if the server has a
scanner, and the workspace's per-file size limit applies (25 MB unless the server sets another). A
refused file says why; **Dismiss** clears it.

Each file shows its type, size and **Used in N chats**: how many chats have been sent it.

- **Every chat in the project gets the files with its next message**, as attachments in its
  workspace, so it can read them however it runs (on this server, in a container, through the
  relay). A new chat gets them with its first message. Your message shows only the files you
  attached yourself.
- **Replace** uploads a new version of the same file. Chats get the new version with their next
  message.
- **Remove** takes two taps. The file leaves the project; chats that already have it keep their copy.

**Save to project** puts something a chat made into a project's files: the folder icon under an answer
(saved as a Markdown file named after its first heading or line), **Save to project** on a deliverable in
the chat, or in the Library's preview. Pick the project; the chat's own project is first. The file goes
through the same checks as an upload and says which chat it came from. Saving the same thing twice keeps
one copy.

A project holds up to 20 files, and together they can be no larger than one message may carry
(100 MB unless the server sets another), so a new chat gets all of them at once. If you also
attach files to that message and they don't all fit, the rest come with the next message. So do files that would take the server below its free-disk floor (2 GB unless the server sets another). The
agent is told the files' names and that they are material to read, not instructions to follow.

### Project memory

The agent asks before it remembers anything. When it learns something lasting in a project
chat, such as a preference, a decision, or a fact about the audience or the publisher, it
proposes a note. The note arrives as a decision card, **Save to project memory?**:

- **Save** adds the note to the project's memory.
- **Don't save**, typed text, or **Dismiss** saves nothing. The agent hears your answer.

Nothing is saved silently. On the project page, each note shows where it came from (the chat
that proposed it, or *added by you*). **Edit** changes a note, **×** forgets it, and **Add a
note yourself** writes one directly. A project remembers at most 50 notes of up to 300
characters each.

**In host mode, memory is a guardrail, not a boundary.** A task on a server in host mode runs as
the server's Linux user, which also owns the project's records and files
(`<data>/workspaces/<id>/chat-projects/*/`). Claude Code's Edit and Write tools are
denied there, but a shell command in any task could still change them without a card. See
[Security model](https://console.uubright.com/docs/security/#known-limitations). On This Mac, the local sandbox keeps tasks out
of the data folder.

Chats read the memory when they start. If you edit a note, a chat that is running sees the
change the next time it starts.

## Library

**Library** is everything the workspace's chats and tasks published: slides and PowerPoint decks,
documents, images and infographics, videos and PDFs. It's a grid of previews, two across on a phone
and four on a computer. Each card shows a picture where there is one (an image, a video's first frame,
a deck's or Word file's first page once this computer has [made its previews](https://console.uubright.com/docs/artifacts/)), its type,
and **from:** the chat that made it.

- **Search the Library** matches titles.
- The type chips (**Images**, **Decks**, **Documents**, **Videos**, **PDFs**, **Other**) show only the
  types there are. A PowerPoint file and an HTML deck are both Decks; a Word file and Markdown are Documents.
- **All projects** narrows to one project, or to chats in no project; **Any time** to today, the past 7
  days or the past 30 days.

A card opens a large preview (full screen on a phone) with **Download**, **Open** (in a new tab),
**Copy link** and [**Save to project**](#project-files). The copied link opens the file for anyone who has it, for an hour. Over the relay there is
no such link, so Copy link isn't offered; download the file to share it. **from:** opens the chat. A deck
the agent revised has a **Version** menu. The address of an open preview opens it again.

Only published results are in the Library, each once, as its newest version. Files a chat made but didn't
publish, and the review steps (Approve, Mark published), are in work mode's
[Deliverables](https://console.uubright.com/docs/artifacts/#deliverable-states).

## Apps

**Apps** lists the workspace's enabled plugins: each one's name, description, starters and
the services it connects to (its MCP servers, by name only). An app's keys stay on the
server. Add or remove apps on the server with `npm run workspace -- enable-plugin` (see
[Workspaces](https://console.uubright.com/docs/workspaces/)). App settings and **Connect** come with the plugin contract.

## On a phone

The bottom tabs are **Chats · Library · Projects · More**. **Chats** lists your chats by day,
with the number that need you on the tab. Open one to reply. **More** has Apps, the
workspace switch, **Use work mode here** and settings. New chat is the screen the app opens on;
its starters are a row you swipe sideways.

## Words in chat mode

These names apply in chat mode only; work mode keeps its own.

| In chat mode | In work mode and the [API](https://console.uubright.com/docs/api/#chat-mode) |
|---|---|
| Chat | A task started as a chat (`chat` on its record). A scratch task started in work mode isn't one |
| Project | A chat project inside the workspace (not a repository project) |
| Library | Deliverables (artifacts), drawn as a grid without the review steps |
| Apps | Enabled plugins |

## Limits

- **History stays on this server.** Chats, projects and memory live in the server's data
  directory. Your phone sees them through the relay; another server does not.
- **One agent per chat**, the workspace's (Claude Code with the workspace's apps). There's
  no switching agents mid-chat.
- **Files are per chat**, unless they're [project files](#project-files). A file attached to one
  chat isn't in the next one.
- **No sharing** of a chat or project yet. Share the Library item instead.
- **No summary of long chats yet.** A long chat relies on the agent's own compaction. Start a
  new chat in the same project to begin fresh with the project's memory.
- **Search** matches chat titles only.

## Starter cards in the composer

A card selects a starter above the main message box. Type the topic or outline in
that full-size box; choose the starter's other options in the compact row above it.
Press Send once to start. Command and skill syntax is assembled for the selected
agent at submission, rather than inserted into the text you are editing. Existing
plugin approval and paid-service confirmation rules are unchanged.

Removing or switching a starter keeps your prose. The selected starter and option
values are saved per execution host and workspace alongside the normal draft and
restored after reload. A removed or changed starter contract clears the selection
without deleting that prose. Project selection and attachments keep working.
Schema limits and required fields are still validated before submission; selecting
a card never truncates an existing draft to make it fit.
