Deploy the FS-free harness on OpenShift with a durable sandbox (P0′)

[pdettori]

Summary

P0′ of the two-tier FS-free harness epic (#47) — the P1 slice: make the merged FS-free harness deploy and pass its full leaf smoke on the live OpenShift 4.20.8 cluster with a durable RWO sandbox, by reframing the never-applied pre-P1 OCP overlay + setup-ocp.sh onto the agent-sandbox Sandbox-CR world.

Spec: docs/specs/2026-07-02-p0prime-ocp-fs-free-deployment-design.md
Plan: docs/superpowers/plans/2026-07-02-p0prime-ocp-fs-free-deployment.md

What changed

• sandbox.Dockerfile — bake ripgrep into the pre-baked OCP sandbox image (the agent's find/grep tools shell out to rg inside the sandbox).
• overlays/ocp/patch-sandbox.yaml — rewritten as a JSON6902 patch onto the Sandbox CR podTemplate: non-root runAsUser/fsGroup 65532, seccomp RuntimeDefault, drop ALL caps, serviceAccountName: serverless-harness-sandbox, command: [sleep infinity], pre-baked internal-registry image. Durable PVC preserved — the patch deliberately does not touch volumeClaimTemplates/volumeMounts, so /workspace stays on the RWO EBS PVC (no emptyDir).
• overlays/ocp/kustomization.yaml — drop the deleted leaf-pvc.yaml resource, retarget the patch to Sandbox/sandbox-0, remove the dead alpine image transform, refresh header comments.
• setup-ocp.sh — install the agent-sandbox v0.5.0 controller (before the overlay apply), create + nonroot-v2-grant the sandbox SA, poll the Sandbox .status.selector → pod Ready (after the apply), tolerate a pre-provisioned llm-credentials secret, and retire the post-P1 leaf-work wording.
• README-ocp.md — replace leaf-work PVC references with the durable Sandbox /workspace PVC, document the controller-install step, and add explicit oc delete pvc to teardown (StatefulSet-style PVCs are not GC'd on CR deletion).
• leaf-smoke.sh — thread $CURL_OPTS (-k) + the guarded CURL_HDR array through dispatch_item and the claim-5 curl, mirroring lib.sh's existing Route handling (these two call sites were the only ones missing it). This file was outside the original 5-file plan scope; the edit was made after live debugging and with explicit approval.

Verification (live, OpenShift 4.20.8)

• Stage-1 walking skeleton: confirmed the agent-sandbox v0.5.0 controller propagates podTemplate securityContext/fsGroup/serviceAccountName + volumeClaimTemplates on OCP; pod non-root uid=65532, PVC Bound gp3-csi RWO, /workspace writable.
• setup-ocp.sh end-to-end: SUCCESS — controller present, in-cluster sandbox build with ripgrep 14.1.0 confirmed in the running pod, overlay applied, Sandbox pod Ready (non-root, durable PVC), ksvc Ready, Route printed.
• Full leaf-smoke.sh via the Route (SH_MODEL=claude-haiku-4-5): 10 passed / 0 failed — FS-free envelope, workspace isolation on the durable PVC, parallel fan-out (i1→FLAGGED / i2→CLEAR / i3→CLEAR), per-call model routing, malformed→HTTP 400, idempotent re-invoke, crash/resume from Redis (durable-PVC survival), and scale-to-zero.

Notes / follow-ups

• The full leaf smoke requires an LLM endpoint reachable from the cluster (a direct api.anthropic.com key was used for verification; base-url/auth-token are optional secret keys).
• Two minor plan-text inaccuracies surfaced during review (code is correct): Task 4's "deleting the Sandbox CR reclaims its PVC" (it does not — hence the explicit oc delete pvc), and Task 2's emptyDir grep expecting 0 (the render legitimately has 1 — the harness's own /tmp mount).
• Post-merge follow-up (per spec §4): once this merges and CI republishes the GHCR sandbox image with ripgrep, switch the OCP overlay to pull it and drop the in-cluster BuildConfig.

Assisted-By: Claude Code