@keboola/agent-vocabulary

The vocabulary Keboola's agents are taught and Keboola's renderers draw.

A prompt tells the model it may write icon:table; a renderer has to know what table looks like. Both halves live in different packages — often in different repositories — so the list of legal names is published on its own, with zero dependencies, and both halves import it.

Installation

npm install @keboola/agent-vocabulary
# or
pnpm add @keboola/agent-vocabulary

Usage

Prompt side — tell the model exactly which names exist:

import { CHAT_INLINE_ICON_NAMES } from '@keboola/agent-vocabulary';

const promptFragment = `Available icons: ${CHAT_INLINE_ICON_NAMES.join(', ')}`;

Renderer side — satisfies turns a missing glyph into a compile error:

import { type ChatInlineIconName, CHAT_INLINE_ICON_NAMES } from '@keboola/agent-vocabulary';
import { Table, Zap, type LucideIcon } from 'lucide-react';

const ICONS = {
  table: Table,
  zap: Zap,
  // …one entry per name
} satisfies Record<ChatInlineIconName, LucideIcon>;

Exports

| Export | What it is | | ---------------------------- | ------------------------------------------------------------ | | CHAT_INLINE_ICON_NAMES | Names legal in a chat answer's inline icon:<name> token | | ChatInlineIconName | The union of those names | | MERMAID_DIAGRAM_ICON_NAMES | Lucide names legal in a mermaid node's lucide:<name> shape | | MermaidDiagramIconName | The union of those names |

The two lists are deliberately different vocabularies: chat names are product-facing aliases (analytics, arrow-rotate-back), mermaid names are Lucide's own (chart-line, undo-2).

Adding a name

1. Add it to the list here and release.
2. Add the matching glyph to every renderer — satisfies fails the build until you do.

Removing a name is the reverse order: teach the model to stop emitting it first, then drop the glyph.