TypeScript usage

The @banh libraries are at 0.1.0-alpha.1 (MIT). They need Node.js 20+ and ESM. APIs may change during alpha.

npm install @banh/dsl@alpha @banh/runtime@alpha @banh/laya@alpha

Using the example files:

import { readFile } from 'node:fs/promises';
import { parseProcess } from '@banh/dsl';
import { ProcessRuntime } from '@banh/runtime';
import { LayaBackend } from '@banh/laya';

const definition = parseProcess(
  await readFile('examples/support-triage.yaml', 'utf8'),
);
const input = JSON.parse(
  await readFile('examples/inputs/support-ticket.json', 'utf8'),
);
const backend = await LayaBackend.create();
try {
  const runtime = new ProcessRuntime(backend);
  const result = await runtime.execute(definition, input);
  console.log(result.output);
  console.log(result.decisions);
  // Reuse this runtime and backend for subsequent inputs.
} finally {
  await backend.close();
}

Backend lifetime

You own the backend: reuse it across runs and close it when done. Parsing and validation don’t load Laya. There’s no automatic model fallback.

Execution result

Field	Meaning
process	Workflow process name.
output	Value from the first matching rule.
decisions	Per decision: normalized and raw answer, policy result, metadata.
timing.totalMs	Execution duration, excluding model loading.
timing.inferenceMs	Inference duration, when available.
usage	Optional inputTokens and outputTokens.
diagnostics	Optional HTTP attempts, latency, size and status.

Pass { onEvent } to new ProcessRuntime(backend, { onEvent }) to observe inference and flow events.

Select a provider

Install @banh/providers@alpha and set BANH_INFERENCE_TOKEN. This runs the same definition and input on Jev:

import { createProvider } from '@banh/providers';

const backend = await createProvider({
  provider: 'http',
  model: 'jev',
  options: {
    token: process.env.BANH_INFERENCE_TOKEN,
    modelId: 'jev-1.13.0',
    timeoutMs: 120_000,
    maxAttempts: 2,
  },
});
try {
  const result = await new ProcessRuntime(backend).execute(definition, input, {
    signal: AbortSignal.timeout(140_000),
  });
  console.log(result.output);
} finally {
  await backend.close();
}

createProvider() with no arguments returns native Laya. Unlike the CLI, the runtime doesn’t read the YAML model; pass the backend you want. See provider presets.

HTTP only, no native dependency: install @banh/dsl, @banh/runtime and @banh/typesafe (all @alpha). Construct TypeSafeHttpBackend with baseUrl and optional modelId, token, timeoutMs, maxAttempts (1 or 2) and retryDelayMs.