SWE-bench dedicated sandbox pool using the baked OCP image (Plan B Task 4)

[pdettori]

Summary

Plan B Task 4: a dedicated, OCP-only swebench Sandbox-CR pool that runs the baked x86_64 image produced by Task 3 (#124), plus a hermetic structural test. Purely additive / off-by-default — nothing selects this pool until Plan C sets KAGENTI_SANDBOX_POOL_SELECTOR=sh.kagenti.io/sandbox-pool=swebench. No changes to the default pool, service.yaml, the OCP overlay, setup-ocp.sh, or any Plan A/B code.

What's in it

• deploy/knative/swebench-sandbox-pool.yaml — 3 Sandbox CRs (agents.x-k8s.io/v1beta1, swebench-sandbox-0/1/2, ns default):
  • image = the Task-3 baked ref image-registry.openshift-image-registry.svc:5000/default/swebench-sandbox:ff962cb83fe5c624-15of15 (15 relocated conda envs + 8 bare repo mirrors + git/ripgrep, USER 65532)
  • pool discovery label sh.kagenti.io/sandbox-pool: swebench (distinct from the default pool's default); CR label app: sandbox (queryable — issue setup-ocp.sh: AB2 sandbox-pool readiness poll spins full timeout (queries a podTemplate label on the Sandbox CR) #103 selector gotcha)
  • command: ["sleep","infinity"] (tools baked, no apk), workingDir: /workspace, imagePullPolicy: IfNotPresent
  • 50Gi RWO /workspace PVC (a shared pod time-shares many concurrent large-repo worktrees + per-leaf venvs; baked /repos+/opt/miniconda3 live in image layers, not the PVC)
  • OCP nonroot securityContext + serviceAccountName: serverless-harness-sandbox baked in (canonical shape from overlays/ocp/patch-sandbox.yaml: pod 65532/fsGroup 65532/seccomp RuntimeDefault, container allowPrivilegeEscalation:false/drop ALL)
• packages/knative-server/test/swebench-sandbox-pool.test.ts — hermetic YAML-parse structural test (28 assertions), mirrors authbridge-manifests.test.ts; no cluster/kustomize, runs in pnpm -r test CI.

Design notes (deviations from the original Kind-centric plan text, recorded)

The baked image is x86_64 and lives only in the OCP internal registry — there is no Kind/dev.local variant (this host's Kind is arm64). So the manifest is OCP-native directly: the securityContext + SA are baked into the manifest (not a Kind base + overlays/ocp JSON6902 patch as the default/-ab2 pools use), and the image tag is a hardcoded content pin.

Live gate (OCP 4.20) — GREEN

Applied to a live cluster: 3 PVCs Bound 50Gi, all 3 pods Ready; in-pod verify confirmed uid 65532, git safe.directory '*', git+ripgrep, 15 env dirs (all relocated pythons run; 11 import numpy, 4 numpy-less as expected), 8 bare repo mirrors, /workspace writable. Cluster restored to baseline afterward.

Assisted-By: Claude Code