@robota-sdk/agent-command

The slash-command (/cmd) implementations for agents, as command modules you register with an agent-framework runtime. The configured CLI builds its command set from this package; your own agent can take the whole default set, a filtered subset, or individual modules.

Installation

npm install @robota-sdk/agent-command

Quick start

createDefaultCommandModules builds the default command set. It returns the selected modules and any enabledCommandModules/disabledCommandModules names that matched no module, so a host can report a typo instead of silently ignoring it.

import { createDefaultCommandModules } from '@robota-sdk/agent-command';
import type { IProviderCommandSettingsAdapter } from '@robota-sdk/agent-command';
import type { IAIProvider, IProviderDefinition } from '@robota-sdk/agent-core';
import { createAgentRuntime } from '@robota-sdk/agent-framework';

declare const provider: IAIProvider;
declare const providerDefinitions: readonly IProviderDefinition[]; // e.g. agent-builtin-providers
declare const providerSettingsAdapter: IProviderCommandSettingsAdapter; // reads/writes provider profiles
declare const userLocalStorageRoot: string; // host-chosen directory for user-local state

const cwd = process.cwd();
const { modules, unknownModuleNames } = createDefaultCommandModules({
  cwd,
  userLocalStorageRoot,
  providerDefinitions,
  providerSettingsAdapter,
  disabledCommandModules: ['agent-command-remote-control'],
});
void unknownModuleNames;

const runtime = createAgentRuntime({ cwd, provider, commandModules: modules });

Each module also has its own factory, so you can compose a smaller set by hand:

import {
  createExitCommandModule,
  createHelpCommandModule,
  createModeCommandModule,
} from '@robota-sdk/agent-command';

const modules = [createHelpCommandModule(), createModeCommandModule(), createExitCommandModule()];

Commands

The default set covers these commands. Modules, which enabledCommandModules and disabledCommandModules select by name, are named agent-command-<area>: for example, agent-command-session provides /clear through /validate-session, and agent-command-schedule provides /schedule, /monitor, and /loop.

| Area | Commands | | ------------------ | ---------------------------------------------------------------------------------------------------------------- | | Help and exit | /help, /exit | | Session | /clear, /rename, /cd, /resume, /cost, /validate-session, /fork, /rewind, /compact, /context | | Model and behavior | /provider, /effort, /advisor, /preset, /output-style, /language, /mode, /permissions, /sandbox | | Agents and work | /agent, /background, /goal, /plan, /schedule, /monitor, /loop | | Terminal | /shell, /editor, /keybindings, /theme, /statusline | | Project and tools | /git, /memory, /skills, /mcp, /plugin, /reload-plugins |

| Settings and state | /settings, /reset, /user-local, /doctor | | Other sessions | /peers, /handoff, /events, /remote-control, /devices |

/memory correct <user|feedback|project|reference> <topic> <text> replaces every entry in that topic. /memory forget <topic> removes the source topic file and its active index entries. Both operations are user-only and block subsequent automatic or model additions to that topic; use correct to deliberately restore it. Project memory retains only the current replacement or a topic tombstone in lifecycle.json, which remains authoritative if source cleanup is interrupted. Historical transcripts and backups remain. A configured semantic backend must support removeTopic for these commands; an index cleanup failure is reported, while active recall continues to read current durable source content. The next model turn refreshes its retained startup memory.

/doctor and /devices are registered only when the host supplies their inputs (doctorInputs, devicesPort). /shell, /editor, /keybindings and /theme belong to the terminal the user sits at: a terminal attached to a workspace daemon runs them itself (see createTerminalClientCommands).

For what each command does and its arguments, see the CLI guide.

Project access

Project-aware commands are capability-based. Skills come only from the contribution sources and skill roots the host passes in, and provider setup writes through the injected settings adapter. The package never turns cwd into permission to read or write the project; a project-local write without an authority-backed store is refused explicitly.

Dependencies

• @robota-sdk/agent-core — provider and tool types
• @robota-sdk/agent-framework — ICommandModule, command host contracts, and the shared command APIs
• @robota-sdk/agent-interface-command, @robota-sdk/agent-interface-execution, @robota-sdk/agent-interface-session — command, execution, and session contracts
• @robota-sdk/agent-preset — presets listed and switched by /preset

Documentation

• docs/SPEC.md — package contract and per-command guarantees
• CLI guide — using the commands in the configured CLI