commands.md

Slash commands

A command is answered by Antares itself rather than by the model. /status costs nothing and returns instantly; asking the model the same question would cost a turn and might be wrong.

Every command is defined once, in internal/commands, and the terminal interface, the web chat, and the messaging gateways all dispatch through it. A command therefore means the same thing wherever it is typed, and adding one is a single registration rather than three.

Using them

In the terminal. Type /. The palette filters as you type; arrow keys move, Tab completes, Enter runs a command that is already whole (one that needs an argument is completed instead). Commands run even while a turn is going, which is what /stop, /steer and /answer are for. Output appears as its own block.

In the web chat. The same, in the composer. Command output appears as its own block in the transcript, marked so it is never mistaken for the model talking.

In Telegram or Discord. Send the command as a message. A leading slash is intercepted before the message reaches the model.

Everything there is

Getting your bearings

Command	What it does
/help	Every command available on this surface
/status	Model, provider, toolset, database, counts, turns in flight
/version	Which build this is
/web	The dashboard address

Model and provider

Command	What it does
/model	Show the active model
/model <id>	Switch to another, saved and applied immediately
/models [provider]	What the provider offers
/provider	List configured providers and which have keys
/provider <id>	Switch provider

Tools and skills

Command	What it does
/tools	The tools active this turn, with the toolset they came from
/toolset [name]	Show or change which tools the model gets
/skills [query]	Installed skills
/skills search [words]	Browse the hub — also takes owner/repo or a URL
/skills install <id>	Install one
/mcp	Configured MCP servers and their connection state
/mcp search [words]	Browse the MCP catalogue
/mcp install <id>	Add one to the configuration

Memory

Command	What it does
/memory	Recent long-term memories
/memory <query>	Search them
/remember <text>	Save something. key: value sets an explicit key
/forget <key>	Delete one

The session

Command	What it does
/new	Start fresh
/clear	Clear the transcript
/sessions	Recent conversations
/resume <id>	Reopen one — a prefix is enough
/compact	Start the next turn from a summary of this one
/retry	Resend the last message
/stop	Interrupt the turn in flight
/copy	Copy the last reply
/title [name]	Show or rename this conversation
/fork [name]	Copy it, to try another direction without losing this one
/undo	Remove the last exchange
/export [markdown|json]	Write it to a file

Long work

Command	What it does
/goal <text>	Set a goal to keep working towards across turns
/goal status	Where the goal stands, and how many iterations it has taken
/goal pause / /goal resume	Hold it, or pick it back up
/goal clear	Drop it
/steer <instruction>	Redirect a run that is already going
/learn [focus]	Turn this session into a reusable skill
/rollback	List the files this session changed
/rollback all | /rollback <path>	Put them back
/panel <question>	Ask several models and synthesise one answer
/team	How the specialist roles have performed

See the harness for what these actually do.

Roles and specialists

Command	What it does
/roles	List the specialist roles
/role [name]	Run this conversation as a role
/panel <question>	Ask several models and synthesise one answer
/team	How the specialist roles have performed

Authorized security testing

Command	What it does
/findings [remove|clear] [id]	The current engagement's findings
/report [title]	Compile the findings into a report
/engagement [intel]	The assessment's phase progress

See roles for what these do.

Settings and accounting

Command	What it does
/config <path>	Read a setting
/config <path> <value>	Change one
/reasoning [on|off]	Show or hide the model's reasoning
/usage [days]	Tokens and cost, by model
/cost [days]	The same thing
/setup	Open the setup wizard
/quit	Leave the terminal interface

Terminal only

The terminal interface lists everything above plus these, which need a screen and a keyboard, or change options only the terminal holds. Where a name also exists in the shared registry, the terminal's version is the one that runs.

Command	What it does
/model / /provider / /theme	Open a picker (/model <id> still sets it directly)
/undo	Undo the last turn and restore the files it changed, after a y to confirm
/revert [message-id]	The same, back to any earlier turn, picked from a list
/reasoning	Show or hide thinking blocks on screen (Ctrl+R)
/effort [level]	Reasoning effort for the next turns — the levels the active model accepts, or auto
/project [dir|clear]	Bind a folder to a new session, as the web's project picker does
/attach [path|clear]	Attach a file to the next message: images go inline, other files are copied where read_document can read them
/search <text>	Search past messages and resume the conversation you pick
/delete	Delete this conversation, after a y to confirm

Ctrl+P opens a settings menu with the same controls in one place: model, provider, reasoning effort, theme, the reasoning display, the side column and the project folder, each with its current value. | /answer [text] | Answer a question the agent put aside with Esc |

A role picked with /role before the first message is held and applied when the session starts. The active role, effort, bound project, and attachments show in the status line.

When the agent stops to ask. An ask_user question appears in the transcript with its choices numbered; the next thing you send is the answer (a number picks a choice). Esc puts the question aside — the turn stays paused until /answer <text> or /stop. With tools.approval_mode: prompt, a tool that changes something shows what it wants to do: y allows it, n or Esc refuses. Several requests are decided in the order they arrived.

Actions. /clear wipes the screen and keeps the session (like Ctrl+L); /new starts over. /retry resends the last message. /copy sends the last reply to the clipboard through OSC 52 and, when installed, pbcopy, wl-copy, xclip, xsel or clip. /compact summarises the session now, streaming like a turn. /setup tells you to quit and run antares setup. A confident goal (/goal auto …) starts working at once, and the terminal keeps starting its next turn until the goal is met, paused, or out of iterations.

Which work where

Most commands work on all three surfaces. The exceptions are the ones a surface cannot carry out: /quit only means something in a terminal, and /resume and /copy need a screen. /help lists what is available where you are, so it is always accurate.

From the shell

The commands that do not need an open conversation also run straight from a shell, without the slash: antares usage 30, antares roles, antares skills, antares memory search editor, antares mcp search github. They print plain text when piped and rendered markdown on a terminal. Commands tied to a conversation (/title, /undo, /goal, /role, …) and the ones a screen carries out (/new, /copy, …) say so instead of running. See cli.md for the full list and for the shell-only subcommands such as antares ask and antares sessions.

Adding one

register(Spec{
    Name:     "weather",
    Args:     "<city>",
    Summary:  "Current conditions",
    Surfaces: anywhere,
}, func(ctx context.Context, d Deps, in Input) (Result, error) {
    return Result{Output: "..."}, nil
})

Deps carries the config, the agent, the store, the skill library, and the MCP manager. Any of them may be nil — a command that needs one that is missing should say so rather than panic.

Return Action when only the calling surface can finish the job:

return Result{Action: Action{Kind: "resume", Value: sessionID}}, nil

The terminal, the browser, and the gateways each translate the actions they understand and ignore the rest.