feat: scope RuntimeDefault seccomp profile to node-24+ run pods

Only node-24+ runtimes create io_uring file descriptors (libuv >= ~1.52)
that block CRIU checkpointing; node/node-22/bun do not. Apply the
RuntimeDefault seccomp profile only to those cold pods instead of all run
pods, so existing runtimes keep their current syscall surface.

Plumbs the canonical runtime identifier through the dequeue message
(backgroundWorker.runtime) to the Kubernetes workload manager.

Author: carderne

apps/supervisor/src/index.ts

@@ -615,6 +615,7 @@ class ManagedSupervisor {
         projectId: message.project.id,
         deploymentFriendlyId: message.deployment.friendlyId,
         deploymentVersion: message.backgroundWorker.version,
+        runtime: message.backgroundWorker.runtime,
         runId: message.run.id,
         runFriendlyId: message.run.friendlyId,
         version: message.version,

apps/supervisor/src/workloadManager/kubernetes.ts

@@ -14,7 +14,10 @@ import { PlacementTagProcessor } from "@trigger.dev/core/v3/serverOnly";
 import { env } from "../env.js";
 import { type K8sApi, createK8sApi, type k8s } from "../clients/kubernetes.js";
 import { getRunnerId } from "../util.js";
-import { withRuntimeDefaultSeccompProfile } from "./kubernetesPodSpec.js";
+import {
+  runtimeRequiresSeccompProfile,
+  withRuntimeDefaultSeccompProfile,
+} from "./kubernetesPodSpec.js";
 
 type ResourceQuantities = {
   [K in "cpu" | "memory" | "ephemeral-storage"]?: string;
@@ -106,6 +109,11 @@ export class KubernetesWorkloadManager implements WorkloadManager {
     const runnerId = getRunnerId(opts.runFriendlyId, opts.nextAttemptNumber);
 
     try {
+      const basePodSpec = this.addPlacementTags(this.#defaultPodSpec, opts.placementTags);
+      const podSpec = runtimeRequiresSeccompProfile(opts.runtime)
+        ? withRuntimeDefaultSeccompProfile(basePodSpec)
+        : basePodSpec;
+
       await this.k8s.core.createNamespacedPod({
         namespace: this.namespace,
         body: {
@@ -120,7 +128,7 @@ export class KubernetesWorkloadManager implements WorkloadManager {
             },
           },
           spec: {
-            ...this.addPlacementTags(this.#defaultPodSpec, opts.placementTags),
+            ...podSpec,
             affinity: this.#getAffinity(opts),
             tolerations: this.#getScheduleTolerations(this.#isScheduledRun(opts)),
             terminationGracePeriodSeconds: 60 * 60,
@@ -310,7 +318,7 @@ export class KubernetesWorkloadManager implements WorkloadManager {
   }
 
   get #defaultPodSpec(): Omit<k8s.V1PodSpec, "containers"> {
-    return withRuntimeDefaultSeccompProfile({
+    return {
       restartPolicy: "Never",
       automountServiceAccountToken: false,
       imagePullSecrets: this.getImagePullSecrets(),
@@ -333,7 +341,7 @@ export class KubernetesWorkloadManager implements WorkloadManager {
             },
           }
         : {}),
-    });
+    };
   }
 
   get #defaultResourceRequests(): ResourceQuantities {

apps/supervisor/src/workloadManager/kubernetesPodSpec.ts

@@ -1,5 +1,21 @@
 import type { k8s } from "../clients/kubernetes.js";
 
+/**
+ * Node >= 24 (libuv >= ~1.52) unconditionally creates io_uring file descriptors,
+ * which cannot be checkpointed. Launching those pods under the RuntimeDefault
+ * seccomp profile makes io_uring_setup fail so libuv falls back to epoll, keeping
+ * the pod checkpointable. "node" (21.x), "node-22" (UV_USE_IO_URING=0 still works)
+ * and "bun" do not create these descriptors, so the profile is scoped to node-24+
+ * to avoid changing the syscall surface of existing runtimes.
+ *
+ * Tolerant of an "experimental-" prefix in case a non-normalized value reaches here.
+ */
+export function runtimeRequiresSeccompProfile(runtime: string | null | undefined): boolean {
+  if (!runtime) return false;
+  const match = /^(?:experimental-)?node-(\d+)$/.exec(runtime);
+  return match ? Number(match[1]) >= 24 : false;
+}
+
 export function withRuntimeDefaultSeccompProfile(
   podSpec: Omit<k8s.V1PodSpec, "containers">
 ): Omit<k8s.V1PodSpec, "containers"> {

apps/supervisor/src/workloadManager/types.ts

@@ -40,6 +40,9 @@ export interface WorkloadManagerCreateOptions {
   projectId: string;
   deploymentFriendlyId: string;
   deploymentVersion: string;
+  // Canonical runtime identifier (e.g. "node", "node-22", "node-24"). Used to
+  // scope the RuntimeDefault seccomp profile to runtimes that require it.
+  runtime?: string;
   runId: string;
   runFriendlyId: string;
   snapshotId: string;

internal-packages/run-engine/src/engine/systems/dequeueSystem.ts

@@ -597,6 +597,7 @@ export class DequeueSystem {
                   id: result.worker.id,
                   friendlyId: result.worker.friendlyId,
                   version: result.worker.version,
+                  runtime: result.worker.runtime ?? undefined,
                 },
                 // TODO: use a discriminated union schema to differentiate between dequeued runs in dev and in deployed environments.
                 // Would help make the typechecking stricter

packages/core/src/v3/schemas/runEngine.ts

@@ -272,6 +272,9 @@ export const DequeuedMessage = z.object({
     id: z.string(),
     friendlyId: z.string(),
     version: z.string(),
+    // Canonical runtime identifier (e.g. "node", "node-22", "node-24", "bun").
+    // Consumed by orchestrators to decide per-runtime pod settings.
+    runtime: z.string().optional(),
   }),
   deployment: z.object({
     id: z.string().optional(),