@vyntral/grove-runtime

The Grove runtime: supervised agent processes, time-travel recorder, deterministic-tool cache, hot-reloadable workers, graceful shutdown.

bun add @vyntral/grove-runtime

import { start, AISDKBackend } from '@vyntral/grove-runtime'
import { tree } from './agent.ts'

const { handle, sessionId } = await start(tree, { backend: new AISDKBackend() })
const out = await handle.run('hello')
console.log(out)
await handle.stop()

What this package owns:

• Supervisor process model — OTP-style one_for_one / one_for_all / rest_for_one strategies, restart-intensity guard, hot reload.
• Recorder — every event lands in .grove/recordings.db (SQLite). Read it with getRecorder(); visualise it with @vyntral/grove-bench.
• Deterministic tool cache — content-hashed by (tool, input), persistent across processes, LRU eviction beyond maxEntries.
• Memory primitive — memory.{ephemeral,session,persistent}(key) surfaces three implicit tools (memory_get/memory_set/memory_list) to the agent.
• AISDKBackend — Vercel AI SDK v6 with retry, timeout, prompt caching.
• Graceful shutdown — SIGINT/SIGTERM run a registered cleanup chain.
• Streaming — opt in via agent({ stream: true }); tokens emit text_chunk events to the recorder as they arrive.

Site: https://vyntral.github.io/Grove · Source: https://github.com/Vyntral/Grove

MIT.