agent-runtime-kit

A provider-agnostic SDK for orchestrating AI agent sessions, streaming events, and structured outputs across multiple coding agent backends.

Installation

npm install agent-runtime-kit

Usage

import { createCodexRuntime } from 'agent-runtime-kit';

const runtime = createCodexRuntime();

const result = await runtime.run(
  {
    cwd: process.cwd(),
    instructions: 'Summarize the current repository structure.',
  },
  {
    includeLogs: true,
  }
);

console.log(result.outputText);
console.log(result.logs);

const eventFilter = (event: { type: string }) => event.type !== 'turn.started';

for await (const event of runtime.runStream(
  {
    cwd: process.cwd(),
    instructions: 'Summarize the current repository structure.',
  },
  {
    eventFilter,
  }
)) {
  console.log(event);
}

GitHub E2E

The repository includes a fixed end-to-end test for a user program built on top of this wrapper. The user program targets exKAZUu/agent-benchmark issue #1, asks either the Codex SDK or the Claude Agent SDK to implement the fix and open a PR, then the test waits until the PR's required GitHub checks pass.

The e2e test is skipped only when CI is set. Outside CI, bun test runs the fixed e2e for every provider listed in SUPPORTED_RUNTIME_PROVIDERS, so adding a new SDK requires updating that provider list and the tests will cover it automatically.

License

Apache License 2.0