Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript
2.8 KB
Raw Blame
import http from 'http'
import type { Socket } from 'net'
import { createApp } from './app.js'
import { config } from './config.js'
import { handleDesktopUpgrade } from './routes/desktopProxy.js'
import { hostWakeCoordinator } from './services/wakeAndHold.js'

/**
 * Bounded deadline for graceful shutdown. After the wake-and-hold coordinator
 * drains, only ordinary in-flight requests remain; if `server.close()` has not
 * completed by then we force exit rather than hang the pod termination.
 */
const SHUTDOWN_DEADLINE_MS = 10_000

async function bootstrap(): Promise<void> {
  if (process.env.NODE_ENV === 'production' && !config.desktopApiToken) {
    console.warn(
      '[RPC_PROXY] WARNING: RPC_PROXY_DESKTOP_API_TOKEN is empty in production mode — ' +
        'rpc-proxy→HCC desktop API calls have no service-token auth. Set this to match CONTEXT_MAPPER_DESKTOP_API_TOKEN.'
    )
  }

  const app = createApp()
  const server = http.createServer(app)

  server.once('error', error => {
    console.error('[RPC_PROXY] server failed during startup:', error)
    process.exit(1)
  })

  // Handle WebSocket upgrade for desktop VNC proxy
  server.on('upgrade', (req, socket, head) => {
    const handled = handleDesktopUpgrade(req as any, socket as Socket, head)
    if (!handled) {
      socket.write('HTTP/1.1 404 Not Found\r\n\r\n')
      socket.destroy()
    }
  })

  // First graceful-stop path for rpc-proxy (Issue #791 §12.2): on SIGTERM/
  // SIGINT stop accepting new connections, deterministically settle every
  // parked wake-and-hold waiter, then close the server under a bounded
  // deadline. Fail loud on close error or deadline expiry.
  let shuttingDown = false
  const shutdown = (signal: string): void => {
    if (shuttingDown) return
    shuttingDown = true
    console.log(`[RPC_PROXY] ${signal} received — draining wake holds and shutting down`)
    // Settle held requests first so `server.close()` is not blocked waiting on
    // them, then stop accepting new connections.
    hostWakeCoordinator.drain('shutdown')
    const deadline = setTimeout(() => {
      console.error(
        `[RPC_PROXY] graceful shutdown deadline (${SHUTDOWN_DEADLINE_MS}ms) exceeded — forcing exit`
      )
      process.exit(1)
    }, SHUTDOWN_DEADLINE_MS)
    deadline.unref?.()
    server.close(error => {
      clearTimeout(deadline)
      if (error) {
        console.error('[RPC_PROXY] error during server.close:', error)
        process.exit(1)
      }
      console.log('[RPC_PROXY] shutdown complete')
      process.exit(0)
    })
  }
  process.on('SIGTERM', () => shutdown('SIGTERM'))
  process.on('SIGINT', () => shutdown('SIGINT'))

  server.listen(config.port, () => {
    console.log(`[RPC_PROXY] listening on :${config.port}`)
  })
}

bootstrap().catch(error => {
  console.error('[RPC_PROXY] failed to start:', error)
  process.exit(1)
})