Dxos Skills

DXOS Skills

Skills define AI toolkits for a domain (e.g. markdown, kanban). They combine operation definitions (for tool schemas) with operation handlers (for runtime execution). See the operations skill (.cursor/skills/operations/SKILL.md) for defining operations.

Skill definition

A skill has three parts:

Example (see packages/plugins/plugin-markdown/src/skills/markdown-skill.ts):

import { type AppCapabilities } from '@dxos/app-toolkit';
import { Skill, Template } from '@dxos/compute';
import { trim } from '@dxos/util';

import { MarkdownHandlers, Create, Open, Update } from './operations';

const SKILL_KEY = 'org.dxos.skill.markdown';

const make = () =>
  Skill.make({
    key: SKILL_KEY,
    name: 'Markdown',
    tools: Skill.toolDefinitions({ operations: [Create, Open, Update] }),
    instructions: Template.make({
      source: trim`
        You can create, read and update markdown documents.
        When asked to edit or update documents return updates as a set of compact diff string pairs.
      `,
    }),
  });

const skill: AppCapabilities.SkillDefinition = {
  key: SKILL_KEY,
  operations: MarkdownHandlers,
  make,
};

export default skill;

• Definitions (Create, Open, Update) go to Skill.toolDefinitions({ operations }) — they provide schemas for the AI.
• Handlers (MarkdownHandlers) go to the skill's operations field — they are invoked at runtime.

File structure

Skills live inside plugins. Structure:

plugin-my-domain/
├── src/
│   ├── skills/
│   │   ├── index.ts              # Re-exports skill
│   │   ├── my-skill.ts       # Skill definition (key, operations, make)
│   │   └── operations/          # Operations (definitions + handlers)
│   │       ├── definitions.ts
│   │       ├── create.ts
│   │       ├── open.ts
│   │       ├── update.ts
│   │       └── index.ts
│   └── capabilities/
│       └── skill-definition/
│           ├── index.ts          # Lazy export
│           └── skill-definition.ts  # Contributes to AppCapabilities.SkillDefinition

Where skills are registered (Composer)

1. Capability module — contributes the skill to the app:

// capabilities/skill-definition/skill-definition.ts
import * as Effect from 'effect/Effect';
import { Capability } from '@dxos/app-framework';
import { AppCapabilities } from '@dxos/app-toolkit';

import { MarkdownSkill } from '../../skills';

const skillDefinition = Capability.makeModule<[], Capability.Capability<typeof AppCapabilities.SkillDefinition>[]>(() =>
  Effect.succeed([Capability.contribute(AppCapabilities.SkillDefinition, MarkdownSkill)]),
);

export default skillDefinition;

2. Plugin — registers the capability module:

// MarkdownPlugin.tsx
import { SkillDefinition } from './capabilities';

export const MarkdownPlugin = Plugin.define(meta).pipe(
  AppPlugin.addSkillDefinitionModule({ activate: SkillDefinition }),
  // ...other modules
  Plugin.make,
);

3. Composer — plugin must be in the app's plugin list (packages/apps/composer-app/src/plugin-defs.tsx). If the plugin is imported there, its skills are available.

Testing with AssistantTestLayer

Use AssistantTestLayer from @dxos/assistant/testing to test operations and AI flows that use skills. Operation definitions and OperationHandlerSet wiring follow the same patterns as production code (see .cursor/skills/operations/SKILL.md).

import { AssistantTestLayer } from '@dxos/assistant/testing';
import { Skill } from '@dxos/compute';
import MarkdownSkill from '../markdown-skill';

const TestLayer = AssistantTestLayer({
  operationHandlers: MarkdownSkill.operations,
  types: [SpaceProperties, Collection.Collection, Skill.Skill, Markdown.Document, HasSubject.HasSubject],
  tracing: 'pretty',
});

• operationHandlers — the skill's operations (OperationHandlerSet.OperationHandlerSet). Required so the runtime can resolve handlers when you call Operation.invoke / Operation.Service (same mechanism as in-app operation execution).
• types — ECHO types the test needs (e.g. Markdown.Document, Skill.Skill). Mirror any types declared on your Operation.make definitions where relevant.
• skills — optional; use when the test binds skills via AiContextService.bindContext({ skills: [...] }) and you need the registry to know about them.

To invoke an operation directly, use Operation.invoke from @dxos/operation (not FunctionInvocationService from @dxos/functions, which is deprecated):

import { Operation } from '@dxos/operation';
import { Create } from './definitions';

const result =
  yield *
  Operation.invoke(Create, {
    name: 'My Doc',
    content: 'Hello world.',
  });

Inside another operation's handler, the same API applies (yield* Operation.invoke(...), yield* Operation.schedule(...)); see the operations skill section Invoking Operations.