Composer
The composer is where you write your prompts. The toolbar picks the model and the servers. Quick settings set how this conversation runs, as its own copy of your defaults. While a run is in progress, the next thing you type can steer that run or wait as the next message. Layout is a separate button in the Studio header.
| Control | What it does |
|---|---|
| Model | Which model this session calls |
| MCP | Toggle external servers for this run |
| Context gauge | How full the window is |
| Attachments | Paperclip: pick files or images from the host |
| Quick settings | Mode, container (until the first message), Bypass, persona, trace speed |
| Send | Blue by default. Orange while Bypass is on. The placeholder says which |
On a narrow composer, MCP and the paperclip drop out so the row can wrap. Model, the gauge, Quick settings, and Send stay.
The composer border stays neutral. Risk colour is on the send button only.
Mentions (@) and commands (/) are typed in the box, not toolbar buttons.
To bring forward or pick an open shell, desktop, or tool tab, use Active windows in the launch menu in the Studio header. See Terminals.
Attachments
The paperclip opens the OS file picker and attaches from your host (the machine running the editor), not from the container.
- Text files become context chips the model reads at send (large files may be truncated)
- Images go through the multimodal path when the selected model supports images
- You can also paste an image into the composer on models that accept images
Chips above the composer show what will go with the next message. Content is snapshotted when you send, not when you attach.
Mentions
Type @ in the composer for a mention menu. What appears:
| Mention | What it does |
|---|---|
@path/to/file or folder | Read file content or list a folder from the container (@container:/path when several containers are active) |
@<terminal> | An open terminal’s live output |
@https://… | Fetch the page as text at send |
@brief:… | A brief page (full document). Type @brief (or a collection name) to drill into ingested sources, then pick a page |
@<agent> | Start of a turn that runs that agent instead of the main agent. Message must begin with @name, then the task |
@killchain | Ask the agent to reconstruct / save the kill chain (optional @killchain:/folder) |
Skills are not @ mentions. Use Commands (/) for those.
Also from the editor or a terminal:
- Add Selection to Chat
- Add Terminal Selection to Chat
You can keep typing while a run is in progress. What that message does is covered under While the agent is working.
Commands
Type / for a commands menu: builtins plus user-invocable skills.
| Command | What it does |
|---|---|
/clear | Start a fresh session |
/compact | Summarize and shrink the context |
/copy | Copy the last response. /copy N copies an earlier one |
/help | Short help |
/<skill> | Run a user-invocable skill by name |
Model
The model picker lists what your configured providers can actually call. The pick is this session's. The model a new session starts on is the default under Sessions. Reasoning effort and Fast mode, when the model supports them, are per model. See Providers and models.
MCP
The MCP button turns MCPs on or off for the current run without editing Settings. A red dot on the button means one or more servers failed; an amber mark means a server is waiting (for example on a container) without failing. Open the menu for per-server status, then fix or retry under Settings → MCP.
Context gauge
The ring shows how full the model window is. Hidden until a model is selected. Click to open or close the breakdown.
What fills the window (see Context for the full story):
- Exegol System: Studio's built-in layer (not editable)
- Your side: conversation, personas, harness (skills, rules, agents, hooks, briefs), MCPs, attachments and mentions
- Kept free for auto-compaction: headroom held while auto-compaction is on; past it, older turns are summarized
- Free space: room left under that reserve
Compaction loses detail: keep findings in /workspace and the kill chain, or run /compact from Commands.
Quick settings
Quick settings are the switches for the conversation you have open: mode, Bypass, persona, trace speed, and (until the first message) which container it runs in.
They start from a default, set once under Sessions → Default. A new session copies that default. After that, the copy is the session's. What you pick here changes this conversation. A second session running beside it keeps its own copy, including a different model, mode, or Bypass.
Open the menu from the composer. The button always shows the mode that is on, and wears that mode's colour on the pill (quiet grey for Auto, yellow for Plan, blue for Answer, or the colour you picked on a mode you built). Shift+Tab in the prompt cycles every mode in that menu, built-in and custom. The menu stays open while you pick, so several switches can be set in one visit. A cog beside a heading opens the Settings page for that switch.
| Setting | Effect |
|---|---|
| Mode | Auto, Plan, Answer, then any mode you built, in one list (Modes) |
| Bypass | Approvals off for this session: exec, write, and edit run without cards (Approvals). Own band under Mode, so it does not read as a fourth mode |
| Container | Where this blank session will run. Only before the first message, and only when the project has more than one container. After the first message the row is gone |
| Persona | Working style for this session (Personas). The pencil edits one you wrote |
| Trace speed | Step, Human, or Machine. How streamed text is painted. It does not slow the model or the container |
To change a session you are not looking at, use its row on Settings → Sessions. Changing Default applies to the next session you create.
Layout is a separate choice: the Layout button in the Studio header, beside Settings. See Layouts. It is the same for every session in the window.
Concurrent messages
You can type while a run is in progress. The new message can join the work already happening, or wait and become the next turn. A command that is already running is left to finish either way.
Steer hands the text to the agent after the step it is on. The agent reads it and carries on with the same task. Use it for a fact the run needs now: the right target, a password, "skip that host".
Queue holds the message until the agent finishes this task, then sends it as the next turn. Use it for a follow-up that should leave the current task alone: "when this scan is done, write the findings to notes.md".
Enter does the one you chose under Settings → Interface → Concurrent messages. Steered is the starting choice. The button beside Send, and Ctrl or Cmd+Enter, do the other. The placeholder names what Enter will do.
Anything you have sent that the agent has not read yet sits in a list above the composer. Each line says what it is waiting for, for example after the current command finishes, or when the agent finishes. Until the agent reads it you can switch that message between Steer and Queue, pull it back into the box to edit, or remove it.
A message with an attachment, or one that starts with /, waits until the run finishes. The line says why. Stop and send stops the agent and sends everything waiting as a new message.
Surfaces that share this thread
The same conversation appears in the sidebar, the full-width chat tab, Room tiles, and Monitor. Closing a panel does not stop the run. For several sessions at once, see Rooms.