Self-host on GKE sandboxes

The GKE backend dispatches Session Work to Agent Sandbox pods on a Google Kubernetes Engine cluster. It uses the poll_unacked_then_dispatch strategy: the dispatcher reserves Work without ACK, reaps stale SandboxClaims, binds one warm pod through an atomic get-or-create claim, and posts only the Session/Work ids. The in-pod ant beta:worker run then owns the Work ACK, heartbeat, and stop.

Prerequisites

A GKE cluster with the Agent Sandbox CRDs installed, a claim template, and a warm pool.
A GkeSandboxClaimControlPlanePort implementation — either a Kubernetes SDK adapter or the OpenMA-owned in-cluster driver service.
A GkeManagedWorkControlPort so the dispatcher can force-stop Work that exceeds maxRedispatch.

Configuration

GKE is wired through the provider_dispatch mode of createManagedEnvironmentWorkerInstallation, or directly via createManagedEnvironmentWorkDispatchPort in @open-managed-agents/environment-dispatch-gke:

const dispatch = createManagedEnvironmentWorkDispatchPort({
  controlPlane,   // SandboxClaim list/delete/create/postSession
  workControl,    // Work stop
  namespace: "openma",
  template: "openma-sandbox-template",
  warmPool: "openma-warm-pool",
  readyTimeoutMs: 120_000,  // default
  maxRedispatch: 3,         // default
});

namespace, template, and warmPool are required, non-empty strings; the warmPool names the Kubernetes Agent Sandbox warm pool each claim binds to. The port reports provider gke-agent-sandbox with strategy poll_unacked_then_dispatch.

Claims carry labels openma.environment-id, openma.session-id, openma.work-id, and openma.dispatch-count. Each dispatch lists the session’s existing claims, deletes stale ones for other Work ids, and creates the next claim under the idempotency key <workId>:<dispatchCount>. When dispatchCount would exceed maxRedispatch, the port stops the Work instead of looping forever.

Capabilities


Provider	`gke-agent-sandbox`
Strategy	`poll_unacked_then_dispatch`
Work ownership	in-pod `ant beta:worker run` (ACK, heartbeat, stop)
Stale claims	reaped on every dispatch
Redispatch	`maxRedispatch` bound (default 3)