When the gateway restarts (e.g., after config.patch), any in-flight agent run that triggers a new command during the drain window gets GatewayDrainingError. This falls through to the generic error handler in agent-runner.runtime and surfaces to the user as: This is a transient error — the gateway comes back up seconds later. But the user sees an error and thinks something is broken. GatewayDrainingError should be treated like isTransientHttp errors — auto-retry after a short delay (e.g., wait for the restart to complete, then retry). The error should never surface to the user since it always resolves on its own. In agent-runner.runtime, the error handling chain checks for billing, context overflow, role ordering, session corruption, and transient HTTP — but GatewayDrainingError is not checked and falls to the generic Agent failed before reply message. Add a check before the generic error handler: if (message.includes('Gateway is draining') || error?.name === 'GatewayDrainingError') {

Fix Plan

To resolve the GatewayDrainingError issue, we need to modify the error handling chain in agent-runner.runtime to auto-retry after a short delay when this error occurs. Here are the steps:

• Modify the error handling chain to check for GatewayDrainingError:

if (message.includes('Gateway is draining') || error?.name === 'GatewayDrainingError') {
  // Wait for restart to complete (poll gateway health or fixed delay)
  await new Promise(r => setTimeout(r, 15000)); // 15-second delay
  continue; // retry the run
}

• Alternatively, poll the gateway health instead of using a fixed delay:

if (message.includes('Gateway is draining') || error?.name === 'GatewayDrainingError') {
  while (true) {
    const gatewayHealth = await getGatewayHealth(); // implement getGatewayHealth function
    if (gatewayHealth === 'healthy') {
      break;
    }
    await new Promise(r => setTimeout(r, 1000)); // 1-second poll interval
  }
  continue; // retry the run
}

Verification

To verify that the fix worked, restart the gateway while an agent run is in progress and check that the GatewayDrainingError does not surface to the user. The agent run should auto-retry after the gateway restart is complete.

Extra Tips

• Make sure to implement the getGatewayHealth function to poll the gateway health.
• Adjust the delay or poll interval as needed to ensure that the agent run retries after the gateway restart is complete.
• Consider adding logging to track the number of retries and the time it takes for the gateway to become healthy again.