The composer

The composer is the box at the bottom of every chat. It starts three lines tall and grows with what you type, up to nine lines, then scrolls. Under the box is a bar with the attach, editor and dictation buttons on the left, the model, effort and mode pills, any switches the agent publishes, the Follow live edits toggle, and the send button on the right.

What you type is kept per chat tab. Switch to another tab, or close the chat and reopen it, and the unsent draft is still there.

Send a message

Type and press Enter. Shift+Enter starts a new line. The send button does the same as Enter; while a turn is running it turns into a stop button.

A message can be text, attachments, or both. A message with only an attachment is fine: “look at this” can be the image alone.

Code blocks

When the caret is inside an open code fence (you typed three backticks and have not closed them), Enter adds a line instead of sending half a block. A hint under the box says Enter adds a line inside the code block. Cmd+Enter sends.

Cmd+Enter always sends: inside a fence, with a completion menu open, from anywhere in the box.

Long drafts

Once a draft passes about 2,000 characters, the bar shows a rough size, for example about 1.2k tokens. It is an estimate, at four characters a token, so you can tell before sending that a paste is bigger than you thought.

Keys in the composer

Recall what you sent

Press ArrowUp with the caret at the very start of the box to bring back the last message you sent in this chat. Keep pressing to go further back, up to your last 50 messages. ArrowDown walks forward again, and past the newest one gives you an empty box. Once you start walking, the caret can be anywhere; ArrowUp keeps going back.

A recalled message is a fresh draft. Change it and send it as a new message; the original stays where it was in the transcript.

Clear the box

Ctrl+C empties the text in one keystroke, so a prompt abandoned halfway through does not have to be selected and deleted. Attachments stay, each with its own remove button. With an empty box, Ctrl+C does nothing.

Stash a draft for later

Halfway through a prompt, something more urgent comes up. Press Cmd+S and the draft, with its attachments, goes into the stash and the box empties. Write and send the urgent thing, then press Cmd+S in the empty box to get the draft back.

• With one stashed draft, Cmd+S in an empty box brings it straight back.
• With several, it opens a menu above the box, newest first. Each row shows the first line of the draft, how many attachments it has, and how long ago it was stashed. Move with ArrowUp and ArrowDown and press Enter, or click a row, to bring it back. Backspace, or the x on a row, discards that draft. Escape closes the menu.
• Bringing a draft back takes it out of the stash.

The stash is one list for all of Tori, not per chat or per project, so a draft stashed in one chat can come back in another. Its attachment labels are numbered again for the chat it lands in: [Image 1] may come back as [Image 2], and the text changes to match. An attachment the new chat’s agent cannot open stays in the stash on its own, its label is taken out of the text, and a message says what that agent does open.

The stash holds 20 drafts. Stashing a 21st drops the oldest. It is kept on disk, so it survives a reload and a relaunch.

If you start typing before a draft has come back, your typing wins and the draft stays in the stash.

Cmd+S does nothing in the composer while you are editing a queued message, while the draft is open in the editor, or while a new chat’s first message is waiting for its session to start.

Discarding a draft, or one dropping off the end, deletes the files you pasted or dropped into it, unless a chat or another stashed draft still uses them.

Mention a file with @

Type @ and start typing a file name. A menu above the box lists matching files in the project, best match first. Move with ArrowUp and ArrowDown and press Enter or Tab to pick one, or click it. Escape closes the menu without picking.

The file becomes an attachment chip above the box, and its label, like [file 1], takes the place of what you typed, so your sentence still reads “look at [file 1]”. The agent is handed the file’s path and reads it off disk. Nothing is uploaded.

The first @ in a chat lists the project’s files once; later mentions use that list.

You can also mention a file by dragging it into the composer from the Files tree or from an editor tab. A dragged file is a mention, the same as @.

Reference a session with @

The @ menu also lists the repository’s other sessions, above the files: up to five that match what you typed, matched on the session’s name or title. They come from the whole repository, every worktree included, not just the folder this chat runs in. The chat you are typing in is never listed.

To see only one kind, start with a scope:

• @session/ lists sessions only, all of them that match.
• @file/ lists files only. Use it for a folder that is literally named file or session, as @file/session/....

Picking a session puts [Session: title] in your sentence and a chip above the box. Brackets in the title are dropped, and a second session with the same title in one message becomes [Session: title (2)]. The agent gets the session’s id, title, agent and project, and reads the conversation itself with the session_history tool, so nothing from that session is copied into your message. In the transcript the token is a chip: hover it for the title and agent, click it to open or focus that session.

Session references need the tori tool server, which Claude, Codex and opencode chats get. Pi does not, so a Pi chat’s @ menu lists files only. Pi sessions can still be referenced from the other agents.

Reference a project or a space with @

@projects/ and @spaces/ walk your projects the way the sidebar lays them out, one level per /:

• @spaces/ lists your spaces, and @spaces/work/ lists the projects in one of them.
• @projects/ lists the projects in the space this chat belongs to.
• Under a project you get its checkouts (worktrees, or the checked out branch of a plain repository) and its sessions from every checkout.
• Under a checkout you get its files.

Each level is named by the last part of its folder, so a space called “Client Work” in a folder named client-work is @spaces/client-work/. Two projects whose folders share a name are told apart as tori and tori-2, in the order the sidebar shows them. The menu shows that name next to each row. Type part of it to filter. With a row highlighted, / opens it and fills in its name. On a session / does nothing, since there is nothing under it. Under a checkout / is just part of the file path you are typing.

Enter, Tab or a click picks the highlighted row:

• A space puts [Space: name] in your sentence. The agent gets the space’s name and each of its projects’ names and folders.
• A project puts [Project: name] in your sentence. The agent gets the project’s name, its folder and its space. For more, it can use the projects_list and sessions_list tools.
• A session is referenced the same way as above.
• A checkout opens, the same as /.
• A file is attached by its full path, as [File 1].

Brackets in a name are dropped, and two different projects or spaces with the same name in one message become name (2). In the transcript, click a project chip to show that project in the sidebar, or a space chip to switch to that space. Neither click changes which branch you have selected.

Like sessions, projects and spaces need the tori tool server, so Pi chats do not offer them.

Reference a pull request with

Type # at the start of a word to list the project’s open pull requests, newest first, each with its title. Type a number to filter by number, or a word to match titles. Enter, Tab or a click picks the highlighted one, and [PR 123] takes the place of what you typed. A # inside a fenced code block never opens the menu.

When you type a number the list does not hold, such as an old or merged pull request, the first row offers to look it up. Tab fetches it and makes the reference. Enter sends your message exactly as typed, so “fixes #42” still works as plain text. If the number is not a pull request in this repository, Tori says so and leaves your text alone.

The reference shows as a chip above the box. It carries the pull request’s number, title, link, state, draft flag and branches as they were when you picked it, and tells the agent to read the rest with the pr_get tool. In the transcript, [PR 123] is drawn as a chip: hover it for the title and state, click it to open the pull request in a tab. Removing the chip removes the token from your sentence.

A reference rides along in the queue, the stash and a relaunch the same way an attachment does.

Run a slash command with /

Type / at the start of a word to list the agent’s own slash commands, the same ones it offers in its terminal interface: /compact, /usage, your own custom commands and so on. The list comes from the agent itself, so it matches whatever agent and version the chat is on, and a new chat shows the list the agent gave last time.

Each row shows the command, the argument it expects if any, and a description. Picking one puts /name in the box with the caret after it. The argument hint is shown, never inserted.

What a command prints, like the output of /usage, /context or /cost, is shown in the transcript and is still there when you reopen the chat.

Attach files and images

There are three ways to attach a file:

• Click the + on the bar and pick one or more files.
• Paste. A copied image or file attaches itself.
• Drop files from Finder onto the composer. While you drag, the composer says what this agent can take, for example Drop to attach: image, pdf, file.

Every attachment is written under ~/.config/tori/attachments/ and the agent is handed its path, the same way a mention works. Nothing is embedded in the conversation as bytes.

Each attachment gets a label by kind, numbered per chat and never reused: [image 1], [pdf 1], [file 2]. The label goes into your message where the caret was, or where you dropped the file, so you can write “compare [image 1] with [image 2]” and the agent knows which is which.

What each agent can take

An agent that cannot open what you give it refuses it with a message saying what it does open. Where nothing at all can be attached, the + button is greyed out. Those agents cannot be given a file from outside their project, so drag the file in from the Files tree, or mention it with @, instead.

Limits

• Up to 10 attachments per message. An eleventh is refused with Send some first.
• An image or a file can be up to 5 MB, a PDF up to 32 MB.
• Images are PNG, JPEG, GIF and WebP.
• Video, audio, archives, fonts, binaries and Office documents are refused, because no agent can read them.

Attachment tiles

Attachments sit above the box as tiles: a small preview for an image, the file’s icon otherwise, with the file name and its label under it.

• Click a tile to open the file in the editor.
• Cmd+click a tile, or drag it into the text, to put its label in your message again.
• Click the x in the corner of a tile to remove it. With the tile focused, Delete or Backspace removes it too. Removing an attachment also takes its label out of your message.

A mentioned file is checked when the chat opens and every time you click back into the box. If it has been deleted or moved, its tile turns red and says where the file was. The tile stays, because your sentence still names it; remove it yourself if the agent no longer needs it.

When you reopen a chat later, the attachments in its past messages come back with it.

Long pastes become files

A paste over 30 lines or 3,000 characters becomes an attachment named pasted.txt instead of a wall of text in the box. The agent reads it the same way, and your draft stays readable. Turn Attach long pastes as files off in Settings > Chat to keep every paste inline. Only Claude takes a pasted file, so with any other agent a long paste always stays in the box.

Write a long prompt in the editor

For a prompt that needs real editing, multiple cursors, find and replace, or vim keys, open it in the editor.

Click edit here under the composer to take the draft back without closing the editor’s text in, or close the editor tab: the composer keeps whatever you last saved. The scratch file is deleted when the link ends.

While a draft is open in the editor you cannot start editing a queued message.

Dictate a message

Click the waveform button on the composer’s bar, Dictate a message, and talk. Click it again to stop. What you said is turned into text on this Mac and put into the box at the caret, after a space unless the caret is at the start of the text or of a line. Nothing is sent: the words wait in the box like anything you typed.

Dictation needs an Apple Silicon Mac. On an Intel Mac the button is not there, and Settings > Chat > Dictation says Needs an Apple Silicon Mac.

While you talk, a waveform covers the text box and moves with your voice, with Listening. Press Escape to drop it under it, and the button turns red. Click it again, Stop and put the words in the message, to finish. The bars then stop following your voice and roll across the box, Turning what you said into text, while the model works. A click does nothing here; the text lands in a moment and the box comes back. The box takes no typing while the waveform covers it.

Escape while recording drops the recording and inserts nothing. While a turn is running, that first Escape goes to the recording and the next one stops the turn.

Downloading a model

The first click with no model downloaded opens Download a dictation model. Like the voice for Read aloud, it runs on your Mac and your voice never leaves it. Pick one:

Download fetches it and the dialog shows the progress, Downloading, 40 of 67 MB, with Cancel download. Not now closes the dialog, and a download already running carries on. A download that fails says why, with Retry. Once it lands, click the waveform button again to record. One model is installed at a time; Change in Settings swaps it for the other, and Delete removes it. See Settings.

The first time

macOS asks whether Tori may use the microphone the first time you record. If you refuse, or turned it off later, or macOS gives Tori silence instead of sound, the recording stops within a couple of seconds and a notice says Tori isn’t getting sound from the microphone. Allow it under System Settings, Privacy & Security, Microphone.

What happens to the clip

• A clip that never rises above the room’s noise is dropped without a word: nothing is inserted and nothing is said.
• A long clip is cut at its quietest gaps into pieces under fifty seconds before the model sees it, and the pieces are joined with a space. A recording left running is cut at fifteen minutes.
• The model loads on the first use, about half a second for Moonshine and a second or two for Parakeet, and is dropped after five idle minutes.
• A transcription that fails says so in a notice, Couldn’t transcribe, with the reason.

The dictation button is greyed out while the draft is open in the editor.

Quote from the transcript

Select some text in the transcript and a small toolbar appears beside the selection, with Quote and Read. Click Quote to put those lines into the composer as a Markdown quote, each line starting with >, with a blank line under it and the caret ready for your reply. Read reads the selection aloud instead, see Read a reply aloud. Selections inside an input, like the free text box of a question, do not offer the toolbar.

The same toolbar appears when you select text in a PDF. There Quote sends the quote, with the page number, to the selected session.

Things that land in the composer from elsewhere

Several parts of Tori put text into a chat for you to read and send. None of them ever presses Enter for you: the text waits in the composer, or is typed at the prompt of a terminal agent, until you send it.

If the target session is waiting on a permission prompt, the send is refused with Session is waiting for permission, answer it first. Answer the prompt and try again.

Things that stop the composer

The composer is disabled, with the reason on the bar, when:

• The agent is turned off in Settings > Agents.
• The project only allows other agents or accounts (see Limiting agents per project).
• The agent is not installed, is signed out, or needs a program it cannot find, for example Chat needs npx.