Build/kvm ci runner

.github/workflows/rust-ci.yml

@@ -2,7 +2,7 @@ name: Rust CI
 
 on:
   push:
-    branches: [main]
+    branches: [build/kvm-ci-runner]
     # Documentation-only edits do not affect the Rust artifact graph; skipping
     # avoids spinning ~30min nextest lanes and cancelling in-flight runs via
     # extra pushes (see paths-ignore semantics for bundled PR diffs).
@@ -41,9 +41,20 @@ env:
   OTEL_EXPORTER_OTLP_ENDPOINT: ""
 
 jobs:
+  # Wakes the self-hosted GCP runner (auto-stops after idle). All kvm-gcp
+  # jobs gate on this transitively via `changes`.
+  # id-token: write is required for the reusable workflow's WIF auth; it
+  # has to be granted at the caller side too (intersection semantics).
+  start-runner:
+    permissions:
+      id-token: write
+      contents: read
+    uses: ./.github/workflows/start-kvm-runner.yml
+
   changes:
     name: Detect sandbox lane changes
-    runs-on: arc-runner-set
+    runs-on: kvm-gcp
+    needs: start-runner
     outputs:
       sandbox_changed: ${{ steps.filter.outputs.sandbox }}
     steps:
@@ -66,7 +77,7 @@ jobs:
 
   doctor:
     name: Workspace integrity (doctor)
-    runs-on: arc-runner-set
+    runs-on: kvm-gcp
     needs: changes
     steps:
       - uses: actions/checkout@v4
@@ -104,7 +115,7 @@ jobs:
 
   nextest:
     name: Nextest (workspace)
-    runs-on: arc-runner-set
+    runs-on: kvm-gcp
     needs: changes
     steps:
       - uses: actions/checkout@v4
@@ -312,7 +323,7 @@ jobs:
 
   sandbox-adapter:
     name: Sandbox adapter (stdio E2E + protocol)
-    runs-on: arc-runner-set
+    runs-on: kvm-gcp
     needs: [changes, nextest]
     if: needs.changes.outputs.sandbox_changed == 'true'
     steps:
@@ -386,7 +397,7 @@ jobs:
 
   sandbox-adapter-image:
     name: Sandbox adapter image (no-publish)
-    runs-on: arc-runner-set
+    runs-on: kvm-gcp
     needs: [changes, nextest]
     if: needs.changes.outputs.sandbox_changed == 'true'
     env:
@@ -575,7 +586,7 @@ jobs:
 
   sandbox-adapter-image-handoff:
     name: Sandbox image metadata handoff validation
-    runs-on: arc-runner-set
+    runs-on: kvm-gcp
     needs: [changes, nextest, sandbox-adapter-image]
     if: needs.changes.outputs.sandbox_changed == 'true'
     steps:

.github/workflows/sandbox-e2e-kvm.yml

@@ -18,9 +18,16 @@ env:
   ROOTFS_TAR: sandbox-echo-rootfs.tar
 
 jobs:
+  start-runner:
+    permissions:
+      id-token: write
+      contents: read
+    uses: ./.github/workflows/start-kvm-runner.yml
+
   prepare-fixture:
     name: Prepare local sandbox fixture bind rootfs artifact
-    runs-on: kvm-enabled
+    runs-on: kvm-gcp
+    needs: start-runner
 
     steps:
       - uses: actions/checkout@v4
@@ -127,7 +134,7 @@ jobs:
   kvm-e2e:
     name: KVM E2E (baml-rt-tools)
     needs: prepare-fixture
-    runs-on: kvm-enabled
+    runs-on: kvm-gcp
 
     steps:
       - uses: actions/checkout@v4

.github/workflows/start-kvm-runner.yml

@@ -0,0 +1,85 @@
+name: Start KVM Runner
+
+# Reusable workflow that wakes the self-hosted GCP runner (auto-stopped after
+# 20m idle) before jobs that need `runs-on: kvm-gcp`. Authenticates to GCP via
+# Workload Identity Federation — no long-lived keys.
+#
+# Repo-level vars consumed (set via `gh variable set`):
+#   GCP_WIF_PROVIDER  e.g. projects/123456789/locations/global/workloadIdentityPools/github/providers/github-oidc
+#   GCP_RUNNER_SA     e.g. ci-agentium-os-ci-1@agentium-os.iam.gserviceaccount.com
+
+on:
+  workflow_call:
+    inputs:
+      instance:
+        description: GCE instance name
+        type: string
+        default: agentium-os-ci-1
+      zone:
+        description: GCE zone
+        type: string
+        default: us-central1-a
+      project:
+        description: GCP project ID
+        type: string
+        default: agentium-os
+      wait_seconds:
+        description: Seconds to wait after `start` returns, so the actions-runner service has time to register with GitHub.
+        type: number
+        default: 60
+  workflow_dispatch:
+    inputs:
+      instance:
+        type: string
+        default: agentium-os-ci-1
+      zone:
+        type: string
+        default: us-central1-a
+      project:
+        type: string
+        default: agentium-os
+
+permissions:
+  id-token: write
+  contents: read
+
+# Two concurrent CI runs both call this — the second blocks on the first
+# rather than racing the gcloud start API.
+concurrency:
+  group: start-kvm-runner-${{ inputs.instance }}
+  cancel-in-progress: false
+
+jobs:
+  start:
+    name: Start ${{ inputs.instance }}
+    runs-on: ubuntu-latest
+    steps:
+      - name: Authenticate to GCP
+        uses: google-github-actions/auth@v2
+        with:
+          workload_identity_provider: ${{ vars.GCP_WIF_PROVIDER }}
+          service_account: ${{ vars.GCP_RUNNER_SA }}
+
+      - name: Set up gcloud
+        uses: google-github-actions/setup-gcloud@v2
+
+      - name: Start instance if not running
+        run: |
+          set -euo pipefail
+          status=$(gcloud compute instances describe "${{ inputs.instance }}" \
+            --zone="${{ inputs.zone }}" --project="${{ inputs.project }}" \
+            --format='value(status)')
+          echo "current status: $status"
+          if [ "$status" = "RUNNING" ]; then
+            echo "instance already running, nothing to do"
+            exit 0
+          fi
+          gcloud compute instances start "${{ inputs.instance }}" \
+            --zone="${{ inputs.zone }}" --project="${{ inputs.project }}"
+
+      - name: Wait for actions-runner to register
+        # Cold-boot + idempotent startup-script + runner service ≈ 60s on
+        # resume. Listing self-hosted runners would need admin perms on the
+        # repo; instead we sleep and let GitHub queue dependent jobs until
+        # the runner re-registers itself.
+        run: sleep ${{ inputs.wait_seconds }}