Skip to main content

Slash Commands

Slash commands are a client-facing command surface for common actions. Clients translate commands into typed requests to the gateway.

Commands are handled by the gateway (not by the model). This keeps control-plane actions deterministic, policy-enforced, and auditable.

Command classes​

  • Standalone commands: a message that is only /... runs as a command.
  • Directives: certain commands persist per-session settings and are stripped before model inference.
  • Side-effecting commands: commands that change state or send messages are subject to policy and may require approvals.

Common commands (examples)​

Session and execution​

  • /new — start a new session (fresh context and new session id).
  • /reset — reset the current session state (policy-defined).
  • /stop — cancel the active run and clear queued followups for the current session.
  • /compact — request compaction of older history into a summary.

Context and usage​

  • /status — show runtime + session status (model, lane, queue, policy mode).
  • /context list — show a context breakdown summary for the last run.
  • /context detail — show a detailed breakdown including tool schema overhead.
  • /usage — show current session usage summary (tokens/time/cost).
  • /usage provider — show provider-reported usage/quota when available.
  • /presence — show connected gateway/client/node presence entries.

Models and auth​

  • /model — show the current model and available options.
  • /model <provider/model> — set the model for the current session.
  • /model <provider/model>@<profile> — pin an auth profile for the session.

Messaging behavior​

  • /queue <collect|followup|steer|steer_backlog|interrupt> — set the inbound queue mode for the session.
  • /send <on|off|inherit> — set or clear a per-session send policy override (operator-scoped).

Policy overrides​

  • /policy overrides list — list active and historical policy overrides (filterable by agent/tool/status).
  • /policy overrides describe <policy_override_id> — show override scope, pattern, and audit linkage.
  • /policy overrides revoke <policy_override_id> — revoke an override (audited, optionally with a reason).

Design guidelines​

  • Prefer unambiguous names (/context list over /ctx).
  • Commands should have typed request/response contracts.
  • Commands that can cause side effects should require explicit confirmation when risk is non-trivial.