Slash commands and @ on every engine
Switch on Slash commands and @ on every engine in Settings → Lab. The composer's / commands then work in every session, whichever engine runs it, and typing @ lets you attach a file, a folder, the session's changes or a web page to your message. A command an engine cannot carry is never sent to the model as text — you get a plain line saying it is not available there.
On this page
What it is
Two things that make the message box behave the same on every engine.
Slash commands. Commands such as /clear, /model or /cost used to work only in Claude Code
sessions; on other engines the text simply went to the model, which then guessed. With this switch on,
the app's own commands work everywhere. Each engine carries a command in the best way it can: through
the engine's own control, a verb the engine understands, the app's own figures, or a plain note saying
the engine has no equivalent.
@ attachments. Pointing an agent at "this file", "that folder", "what you changed" or "this page"
is something people type by hand all the time. Typing @ opens a small picker instead. Omniscio — not
the window — reads what you picked, keeps it inside the session's folder or passes it through the web
safety checks, and gives every engine exactly the same text.
Where to find it
- Switch it on: Settings → Lab → Slash commands and @ on every engine. It is off by default, and it needs the native slash commands setting on as well.
- Slash commands: type
/at the start of a message in any session. The menu lists the app's commands for that engine; a command the engine cannot carry is greyed with the reason, and pressing Enter or Tab on a greyed command does nothing — it is neither filled in nor sent. On engines that run Claude skills, your skills are listed too; an engine that advertises its own commands lists those as well. - @ attachments: type
@in a session's message box. The Attach with @ picker offers matching files and folders as you type, Changes in this session (with how many files changed and how big the block is), and A web page… (enter an address starting withhttp://orhttps://). Each pick becomes a chip above the box; remove one with its close button.
How it behaves
Which engine carries which command:
| Command | Claude Code sessions (including Claude Code on other providers' models) | Codex, Gemini, Hermes, Kimi Code, DSH, Pi | Cursor, Grok, Antigravity, OpenCode, Devin, OpenClaw, Grokbot | Terminal and shared sessions |
|---|---|---|---|---|
/new, /clear |
the engine's own | a fresh context the engine's own way | not available | sent as typed |
/compact |
the engine's own | not available | not available | sent as typed |
/model, /cost, /context, /status |
the engine's own | the app's own | the app's own | sent as typed |
/stop, /help |
the app's own | the app's own | the app's own | sent as typed |
/rollback |
the app's own (needs Turn checkpoints on) | same | same | sent as typed |
"Not available" means you get a plain line saying so — for /compact, a suggestion to use /new
instead where /new works. On an engine that can do neither (Cursor, Grok and the others in that
column), /compact says the engine can't compact its conversation or start a fresh one inside the
session, and suggests starting a new session. Nothing summarizes the conversation behind your back. In a terminal or shared session, every
/name is sent to the program exactly as you typed it.
While a turn is running. /stop goes straight through and ends the running turn. Any other command
typed mid-turn is answered with a note that it can run once this turn finishes.
Commands only you can run. /clear, /new, /model, /rollback and /rewind take effect only when
you type them yourself in the app's message box. The same text arriving from a quick reply, a recipe,
a chat bot, the control server, the agents API, a queued message or another agent changes nothing: the
session answers with a short note that the command only runs when you type it yourself, and the text
is not passed on to the engine either, so the engine cannot run it as its own command.
/rollback. /rollback undoes the latest turn that changed files, and /rollback 4 undoes turn 4.
/rewind does the same while Turn checkpoints are on. Both only open the confirm step that lists every
file; nothing changes until you click there.
/model on other engines. On engines other than Claude Code, a new model takes effect the next time
the engine starts for that session.
How @ attachments are sent. The attachments are added under an "Attached with @:" heading in the message the engine receives. A held message — one queued because the session was busy — keeps its attachments and reads them when it is finally delivered.
When an attachment cannot be read. It becomes one short note in the message, such as
[Left out @notes.md: it is too large.], and the rest of the message is still sent. Reasons include a
missing file, a file that is not text, a size limit, an address the app does not open, or no changes
to show.
The limits worth knowing:
- Up to 20 attachments per message, and about 256 KB of attached text in total.
- A file is read up to 64 KB; a folder lists up to 200 entries; the session's changes are cut at 128 KB, and the cut is said plainly.
- A web page must be a text page. It may redirect at most three times, must answer within ten seconds, is downloaded up to 512 KB and kept up to 64 KB of text; a page cut at that limit ends with "[The page was cut at the size limit.]". Private and local addresses are refused at every step, and the page's text is marked as outside content so the engine treats it as data, not instructions.
- Paths stay inside the session's folder. The picker never offers anything outside it, and a path or link that leads out is refused.
- Sessions on another computer over SSH: the picker offers no files, the changes show as "no changes", and web pages still work.
- With the switch off, typing
@opens nothing and every send is exactly as it was before.
Related
- Native slash commands — what the Claude Code commands do in a Claude Code session.
- Turn checkpoints — the undo history
/rollbackopens. - Pause or stop a session — the other ways to stop an agent.
- Context details — the context figures
/contextshows.
Last verified 2026-11-01