Core Runtime

@karkata-ai/core owns the Agent lifecycle, sessions, tool dispatch, cancellation, and state. It does not depend on a model vendor, the DOM, or Node.js APIs.

Create and run

import { Agent, type LLMAdapter } from '@karkata-ai/core'

declare const llm: LLMAdapter
const agent = new Agent({ llm, timeoutMs: 30_000, maxSteps: 12 })
const result = await agent.send('Summarize the current order')

if (result.status === 'completed') console.log(result.content)
if (result.status === 'error') console.error(result.error.code)

An instance permits only one active send(). A successful run commits its complete message sequence as history for the next turn. Errors, timeouts, and manual aborts do not commit the incomplete turn. clearHistory() removes committed history and cannot run during an active turn.

State and lifecycle

agent.subscribe() immediately receives an isolated snapshot. A listener exception cannot affect the Runtime, and mutating a snapshot does not write back to the Agent.

const unsubscribe = agent.subscribe((state) => {
  renderStatus(state.status)
  if (state.activeTool) renderTool(state.activeTool.name)
})

agent.abort()
unsubscribe()
await agent.dispose()

abort() first guarantees that the current send() settles promptly. The Runtime passes its AbortSignal to model, tool, and user callbacks, but cannot guarantee that an external side effect already in flight will stop. Continue with Tools, Streaming, and Human Input.