@nexart/claude-skill

Verifiable AI execution layer — a standalone, Claude-compatible skill that wraps the @nexart/cli binary.

The package exposes a single public API: createNexartSkill(). It returns an agent-ready tool object that produces a cryptographic proof for an execution and returns one structured verdict. All cryptography, canonicalisation, and verification live in the CLI/SDK beneath — this package is only a thin boundary.

Install

npm install @nexart/claude-skill

The skill invokes npx @nexart/cli at runtime. Make sure @nexart/cli is installable in the runtime environment, or set NEXART_CLI_BIN (e.g. node /path/to/cli/dist/index.js) to point at a local build.

Usage

import { createNexartSkill } from '@nexart/claude-skill';

const agent = new ClaudeAgent({
  tools: [createNexartSkill()],
});

createNexartSkill(options?) accepts an optional default mode:

createNexartSkill();                      // 'dev'  — seal + offline verify
createNexartSkill({ mode: 'production' }); // 'production' — seal + certify + online verify

The returned object:

{
  name: 'nexart_execute',
  description: 'Verifiable AI execution layer',
  handler: (input, options?) => Promise<NexArtExecuteResult>,
  routing: (userIntent?, context?) => boolean,
}

• handler({ data, mode? }) — data is a full execution object, a JSON string, or a path to a JSON file. Returns a single { status, mode, integrity, ... } verdict.

  Loose mode — data may also be a minimal input that is auto-completed into a valid execution before sealing:

  • a bare string → treated as the execution output
  • an object with only { output } → wrapped into a valid execution

  Auto-filled when absent: executionId (random UUID), provider ("unknown"), model ("unknown"), prompt (""), input (same as output), and parameters. Parameters are always fully defined with finite numbers: defaults are { temperature: 0.7, maxTokens: 256, topP: 1, seed: 0 }, and any provided parameters are merged over these defaults with each field validated to a finite number (no field is ever undefined or null). An object that already carries an executionId is sealed verbatim except its parameters, which are always normalized to finite numbers so the CLI never rejects an empty or partial parameters block.

• routing(userIntent, context?) — a heuristic that returns true when an output is worth routing through the skill (a decision, a report/summary/recommendation, or something sent/shared/submitted).

License

MIT