Audrail Agent Runtime

@audrail/agent-runtime is the local-first runtime layer that sits between an AI host and the repo.

Its job is to make prompt capture durable before upstream model execution starts, then convert run evidence into a checkpoint and handoff brief.

Responsibilities

• open or continue an agent session
• create or extend run records
• append prompt, command, file, verification, and sync events
• generate checkpoint and handoff files after each run
• keep the repo ready for resumable work across agents and teammates

Install

npm install @audrail/agent-runtime

Usage

import { AudrailAgentRuntime } from "@audrail/agent-runtime";

const runtime = new AudrailAgentRuntime({
  projectRoot: process.cwd(),
  adapter: "codex",
  provider: "openai",
  agent: "codex",
});

await runtime.capturePrompt({
  prompt: "Fix the failing landing-page build",
  title: "landing-page build fix",
  threadId: "thread_123",
});

await runtime.recordCommand({
  summary: "Ran landing build",
  command: "npm run build --prefix apps/landing",
});

await runtime.recordVerification({
  summary: "Landing build passed",
  command: "npm run build --prefix apps/landing",
  status: "passed",
});

const result = await runtime.finishRun({
  status: "completed",
  owner: "[email protected]",
  nextAction: "Sync the summary to the dashboard",
});

console.log(result.checkpoint);
console.log(result.handoff);

The runtime is intentionally host-agnostic. Codex, Claude, Cursor, or an MCP gateway should call this package before sending the prompt upstream.

On finish, the runtime records checkpoint and handoff artifacts that now include:

• owner
• safeToResume
• blocker and next-step state
• local GitHub bridge metadata in buildRuntimeSnapshot() when the repo has a GitHub remote