Slash commands

Slash commands are shortcuts you type into the chat composer to steer the agent’s next turn — start a goal, force a compression, run a coding agent, kick off research, and more.

Typing a command

In any chat with an agent, type / as the first character of a message, then the command name and any arguments:

/goal Fix every failing test in tests/ and report the command that passes

Press Enter to send it like a normal message. A few rules:

• The / must be the very first character — a slash mid-sentence is treated as plain text.
• Everything after the command name is passed as arguments; arguments can span multiple lines (useful for a Rubric: block).
• If a command name isn’t one of the built-ins below, the agent tries to run it as a skill of the same name.

All commands

Command	What it does	Capability toggle	Details
/clear	Wipes the conversation history for this session and destroys the session sandbox, giving you a clean slate.	Always on (no toggle)	—
/compress	Forces context compression right now instead of waiting for the automatic threshold. Refuses when the conversation has 5 messages or fewer.	/compress	—
/code	Runs a coding agent (Claude Code or Codex) in the session sandbox using your own plan. Sub-verbs for login, status, and runs.	/code	Coding agents
/deep-research <topic>	Delegates once to a research sub-agent that produces a cited report. Single-shot.	/deep-research	Deep research
/auto-research …	Starts an autonomous optimization run (Arbor) that iterates over a hypothesis tree against a measured baseline.	/auto-research	see below
/mission …	Rubric-judged orchestration — a strict coordinator that decomposes and drives an objective to completion.	/mission	Missions
/goal …	A session outcome the agent keeps iterating toward until an evaluator says it’s met.	/goal	Goals
/loop …	Schedules a prompt to run on a recurring interval, or self-paced when you give no interval.	/loop	Scheduled work
/stop (or /cancel)	Interrupts the turn that’s running right now. Messaging channels only.	Always on (no toggle)	—
/<skill-name> [args]	Runs an attached skill (or consults an active expert) by name.	Per skill attachment	below

/clear and /stop are always available — neither has a capability toggle. Every other built-in is gated by a per-agent toggle on the CONFIG page’s Capabilities tab (see Configure an agent). A master switch there turns all built-in commands on or off at once, with individual toggles below it. When a command is turned off, typing it has no special effect.

/clear in detail

/clear does two things at once:

1. Replaces the entire conversation history with an empty context — the next turn starts fresh.
2. Destroys the session’s sandbox if one exists, so any files, processes, or checkouts from this session are gone.

It confirms with Conversation cleared. Use it when the current context is no longer useful and you want to keep working in the same session with a clean start. The durable event log is untouched; only the model-visible context is reset.

/stop in detail

/stop (or /cancel) is how you interrupt an agent from Slack, Telegram or WhatsApp. Send it on its own, with nothing else in the message. It travels out of band — it doesn’t queue behind the turn it’s cancelling, which is why a long run can still be stopped mid-flight. The bot acknowledges with ⏹ Stopping the current run…; nothing is added to the conversation history.

On Web there is no /stop. While a turn is running the send button becomes a stop control that does the same thing (see Chatting with an agent).

On Slack, typing /stop opens Slack’s own command menu instead of sending a message — put a single space before the slash so Slack passes it through.

/compress in detail

/compress runs the configured context compressor immediately rather than waiting for the automatic threshold, then replies with a short summary of what happened. If the conversation has 5 or fewer messages it refuses with “Context is too small to compress” — there is nothing meaningful to compact yet. Use it before continuing a long session that’s approaching its context limit.

/auto-research in detail

/auto-research creates a research-kind mission: an autonomous, long-horizon optimization run that grows a hypothesis (Idea) tree, dispatches experiments into isolated worktrees, and only merges an improvement after an independently re-run held-out evaluation. It is a specialized variant of /mission — for the single-shot literature-style variant, use /deep-research instead.

Don’t hand-write the tokens on your first run. If the arbor-research skill is attached to the agent, send /arbor-research <goal> first: that intake skill finds the repo, works out the dev and held-out eval commands, measures both baselines, and hands you a complete, ready-to-send /auto-research … line with a rubric anchored to the measured numbers. It never starts the run itself — you review the command and send it.

Leading key=value tokens configure the run (all optional except repo= on create):

Token	Meaning
repo=	Workspace repo path the run operates on (required to create; must be under /workspace/)
max_iterations=	Iteration budget for the run
baseline=	The measured dev-split baseline score
baseline_test=	The measured held-out baseline — the reference the merge gate compares against

Everything after the tokens is the objective, which must be followed by a Rubric: block defining when the run counts as satisfied; a command with no rubric is rejected outright (“a research run needs a repo and a Rubric”). Bound the work with max_iterations=; the Budget: line and the budget sub-verb are /mission-only and have no effect here. Control an active run with:

Command	Behavior
/auto-research status	Show the run’s status and last verdict
/auto-research pause	Pause automatic continuation
/auto-research resume	Resume a paused run
/auto-research cancel [--cascade]	Cancel the run (and, with --cascade, its running experiments)

/auto-research requires an authenticated user or service-account session — anonymous channel sessions cannot own a research run. It also needs the arbor-executor sub-agent to be available to the agent: that is what every dispatched experiment runs as, implementing and evaluating one hypothesis in its own isolated copy of the repo. Without it the coordinator has nothing to dispatch experiments to — see Sub-agents. Missions started this way show up on the agent’s Missions page alongside /goal and /mission runs.

Running a skill

Any slash command that isn’t a built-in is resolved as a skill attached to the agent:

/research vector databases
/babysit-prs

When the command name matches an attached skill, the agent inlines that skill’s instructions (its SKILL.md) into your message before the model acts on it, and stages any supporting files the skill ships. Anything you type after the name is passed to the skill as its request. If the matched skill is an active expert, the command instead runs an expert consultation and hands the expert’s deliverable back into the conversation.

If no attached skill matches the name, your message reaches the agent unchanged (it just looks like text that starts with /). Built-in command names are reserved and are never treated as skill names. Which skills are attached — and therefore which /<skill-name> commands exist — is configured per agent on the CONFIG → Skills tab.

Related pages

• Goals (/goal) — iterate toward a session outcome.
• Missions — the dashboard for /goal and /mission runs.
• Scheduled work — the /loop schedules an agent runs.
• Coding agents (/code)
• Deep research (/deep-research)