diff --git a/.github/workflows/validate.yaml b/.github/workflows/validate.yaml
new file mode 100644
index 0000000..435bd5d
--- /dev/null
+++ b/.github/workflows/validate.yaml
@@ -0,0 +1,65 @@
+name: Validate
+
+# Validates the Kubernetes manifests against the real agent-sandbox CRD
+# schema (server-side dry-run on a throwaway kind cluster) and runs the
+# gateway unit tests. Runs on every PR and on pushes to main.
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+permissions:
+  contents: read
+
+jobs:
+  validate-manifests:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Create kind cluster
+        uses: helm/kind-action@v1.10.0
+        with:
+          version: v0.23.0
+          cluster_name: dsh-validate
+
+      - name: Install agent-sandbox (extensions)
+        run: |
+          VERSION=v1.0.7
+          kubectl apply -f https://github.com/kubernetes-sigs/agent-sandbox/releases/download/${VERSION}/sandbox-with-extensions.yaml
+          kubectl wait --for=condition=Established crd/sandboxtemplates.extensions.agents.x-k8s.io --timeout=180s
+          kubectl wait --for=condition=Established crd/sandboxwarmpools.extensions.agents.x-k8s.io --timeout=180s
+          kubectl wait --for=condition=Established crd/sandboxclaims.extensions.agents.x-k8s.io --timeout=180s
+          kubectl wait --for=condition=Established crd/sandboxes.agents.x-k8s.io --timeout=180s
+
+      - name: Validate manifests (server dry-run)
+        run: |
+          set -euo pipefail
+          kubectl create namespace dsh-demo
+          for f in examples/dsh-agents-as-a-service/*.yaml; do
+            echo "== $f =="
+            kubectl apply --dry-run=server -f "$f"
+          done
+          kubectl apply --dry-run=server -f deploy/k8s/sandbox-template.yaml
+          echo "all manifests validated"
+
+  gateway-tests:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: examples/dsh-agents-as-a-service/gateway
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install dependencies
+        run: pip install -r requirements.txt pytest
+
+      - name: Run gateway unit tests
+        run: python -m pytest test_gateway.py -v
diff --git a/.gitignore b/.gitignore
index bb6f512..a80a91a 100644
--- a/.gitignore
+++ b/.gitignore
@@ -149,3 +149,7 @@ vite.config.ts.timestamp-*
 .turbo
 
 !.env.sample
+
+# Python
+__pycache__/
+*.py[cod]
diff --git a/README.md b/README.md
index 970b37d..ce098d7 100644
--- a/README.md
+++ b/README.md
@@ -1,178 +1,78 @@
-# deepseek-kubernetes-sandbox based on hermes-agents-as-a-service
+# deepseek-kubernetes-sandbox
 
-Original: https://github.com/kubernetes-sigs/agent-sandbox/tree/main/examples/hermes-agents-as-a-service
+Packages [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) Web (`dsh web`) into
+a multi-arch, Kubernetes-ready container image, and ships the
+[agents-as-a-service](examples/dsh-agents-as-a-service/) platform pattern — running a personal
+DSH Web instance per user on [kubernetes-sigs/agent-sandbox](https://github.com/kubernetes-sigs/agent-sandbox).
 
-After analyzing the `hermes-agents-as-a-service` example, its core pattern maps directly onto your `dsh-web` container. The example's key mechanisms are: **`SandboxTemplate` defines a reusable Pod blueprint, `SandboxWarmPool` pre-warms instances, `SandboxClaim` lets each user claim a persistent Sandbox, and `service: true` gives each Sandbox a stable network identity for gateway routing.** Your DSH container is a strong fit for this model — it needs a persistent `DSH_HOME` (like Hermes needs a persistent data directory) and exposes a web port. The main differences are how secrets are injected and how the port is surfaced to end users.
+## What's in this repo
 
-### 🔑 Mechanism Mapping: Hermes → DSH
+| Path | Purpose |
+| --- | --- |
+| [`Dockerfile`](Dockerfile) | Multi-stage build that clones DSH, `pnpm pack`s the CLI, and produces a slim runtime image |
+| [`build.sh`](build.sh) | Local build + run (builds `dsh-web:<version>` and runs it) |
+| [`.github/workflows/multi-build.yaml`](.github/workflows/multi-build.yaml) | CI build (amd64 + arm64) and push to GHCR |
+| [`deploy/k8s/sandbox-template.yaml`](deploy/k8s/sandbox-template.yaml) | Minimal single-tenant `SandboxTemplate` |
+| [`examples/dsh-agents-as-a-service/`](examples/dsh-agents-as-a-service/) | The full multi-tenant platform pattern (warm pool, per-user claims, gateway, smoke test) |
 
-| Hermes example mechanism | Purpose | Adaptation for DSH |
-| :--- | :--- | :--- |
-| **`SandboxTemplate`** | Reusable Pod blueprint | Change image to `dsh-web:0.2.1-rc.2`, expose port `3080` |
-| **`volumeClaimTemplates`** | Auto-create a PVC per user, persisting state | Mount point becomes `/home/node/.dsh` (DSH's state directory) |
-| **`SandboxWarmPool`** | Pre-warmed instances for sub-second provisioning | Keep as-is; size the pool for expected concurrent users |
-| **`SandboxClaim` + `warmPoolRef`** | User "registration" claims a lightweight object, adopts a pooled instance | Keep exactly as-is; one Claim per DSH user |
-| **`service: true`** | Creates a headless Service per Sandbox with a stable DNS name | **Enable it** so a front-end gateway or Ingress can reliably route to each user's DSH Web |
-| **`envVarsInjectionPolicy`** | Controls whether a Claim may inject/override env vars | DSH needs `DEEPSEEK_API_KEY` injected, so either allow it or reference a Secret from the template |
-| **`operatingMode`** | Cost switch: `Running` / `Suspended` (deletes Pod, keeps PVC + Service) | Reuse directly — suspend idle DSH instances to save cost |
+## Quick start (image)
 
-### 📦 Equivalent Kubernetes Manifests
-
-The manifests below mirror the Hermes example's organization. Apply them in order, all in a dedicated namespace (e.g. `dsh-demo`).
-
-#### 1. Secret: DSH API key
-
-```yaml
-apiVersion: v1
-kind: Secret
-metadata:
-  name: dsh-api-secret
-  namespace: dsh-demo
-type: Opaque
-stringData:
-  DEEPSEEK_API_KEY: "sk-..." # replace with a real key
-```
-
-#### 2. SandboxTemplate: DSH Web blueprint
-
-This is the core resource, corresponding to `10-sandboxtemplate.yaml` in the example. Key settings: **`service: true`** for stable DNS, **`volumeClaimTemplates`** to persist `/home/node/.dsh`, and a security context that runs as the non-root `node` user (UID 1000).
-
-```yaml
-apiVersion: extensions.agents.x-k8s.io/v1beta1
-kind: SandboxTemplate
-metadata:
-  name: dsh-web-template
-  namespace: dsh-demo
-spec:
-  service: true # headless Service per Sandbox → stable DNS
-  podTemplate:
-    spec:
-      securityContext:
-        runAsNonRoot: true
-        runAsUser: 1000
-        runAsGroup: 1000
-        fsGroup: 1000
-      containers:
-        - name: dsh-web
-          image: dsh-web:0.2.1-rc.2 # replace with your registry path
-          ports:
-            - containerPort: 3080
-              name: web
-          env:
-            - name: DEEPSEEK_API_KEY
-              valueFrom:
-                secretKeyRef:
-                  name: dsh-api-secret
-                  key: DEEPSEEK_API_KEY
-            - name: DSH_PUBLIC_URL
-              value: "http://localhost:3080" # change to the external URL when behind a gateway/Ingress
-          volumeMounts:
-            - name: dsh-home
-              mountPath: /home/node/.dsh
-          readinessProbe:
-            httpGet:
-              path: /
-              port: 3080
-            initialDelaySeconds: 5
-            periodSeconds: 10
-          livenessProbe:
-            httpGet:
-              path: /
-              port: 3080
-            initialDelaySeconds: 15
-            periodSeconds: 20
-      # If your cluster uses gVisor / Kata, set the runtime class here
-      # runtimeClassName: gvisor
-  volumeClaimTemplates:
-    - metadata:
-        name: dsh-home
-      spec:
-        accessModes: ["ReadWriteOnce"]
-        resources:
-          requests:
-            storage: 5Gi # adjust to your DSH state size
-```
-
-#### 3. SandboxWarmPool: pre-warmed pool
-
-Corresponds to `20-sandboxwarmpool.yaml`. Pooled Sandboxes pull the image and boot ahead of time; a user Claim adopts one instantly.
-
-```yaml
-apiVersion: extensions.agents.x-k8s.io/v1beta1
-kind: SandboxWarmPool
-metadata:
-  name: dsh-web-pool
-  namespace: dsh-demo
-spec:
-  replicas: 2 # number of pre-warmed instances
-  sandboxTemplateRef:
-    name: dsh-web-template
-```
-
-#### 4. SandboxClaim: per-user claim
-
-Corresponds to `40-claim.yaml`. Each user creates one lightweight Claim and receives a pooled instance with its own PVC and stable DNS.
-
-```yaml
-apiVersion: extensions.agents.x-k8s.io/v1beta1
-kind: SandboxClaim
-metadata:
-  name: user-alice-dsh
-  namespace: dsh-demo
-spec:
-  sandboxTemplateRef:
-    name: dsh-web-template
-  warmPoolRef:
-    name: dsh-web-pool
-  # If the template allows env injection (envVarsInjectionPolicy: Allowed),
-  # you can override/add variables here:
-  # env:
-  #   - name: DSH_PUBLIC_URL
-  #     value: "https://alice.dsh.example.com"
-```
-
-### 🌐 Networking: How Users Reach DSH Web
-
-Unlike Hermes agents, which are mainly consumed by an internal gateway, DSH Web is a **user-facing web UI** — each Sandbox's `3080` port must be reachable by end users. Two patterns:
-
-#### Pattern A: Direct via headless Service (simple)
-
-With `service: true`, each Sandbox gets a headless Service whose DNS name is exposed in the Sandbox's `status.serviceFQDN`. You can reach it from inside the cluster via that FQDN, or add an Ingress/upper-layer gateway that routes a path to the corresponding Service.
-
-#### Pattern B: Gateway-proxied (recommended, mirrors Hermes's service layer)
-
-The Hermes example ships a ~250-line Python gateway (`gateway/gateway.py`) that handles **per-user routing, auth, and auto-waking suspended Sandboxes**. You can build a simpler equivalent for DSH:
-
-1. Accept external requests (e.g. `/u/{user}/**`).
-2. Look up the `status.serviceFQDN` of that user's Sandbox.
-3. Reverse-proxy to `http://{serviceFQDN}:3080`.
-4. If the Sandbox is `Suspended`, flip its `operatingMode` back to `Running`, wait for `Ready`, then proxy.
-
-If you don't want a gateway yet, you can manually create one Ingress per user pointing at their Service — less automated, but functional.
-
-### ⚙️ Key Differences and Caveats
-
-1. **Secret-injection policy.** The Hermes example sets `envVarsInjectionPolicy: Disallowed` to strictly limit what Claims can inject. DSH needs `DEEPSEEK_API_KEY` injected, so you have two choices:
-    - **Set it to `Allowed`**, then reference the Secret from `SandboxClaim.env` (more flexible, but requires trusting Claim creators).
-    - **Keep it `Disallowed`** and reference the Secret directly inside the `SandboxTemplate`'s `podTemplate` (stricter, but all users share the same key — fine for a single-tenant internal setup).
-    Pick based on your multi-tenancy model.
-
-2. **`DSH_PUBLIC_URL` configuration.** Your entrypoint defaults to `http://localhost:3080`. When accessed via a headless Service or gateway, set `DSH_PUBLIC_URL` to the real external URL (e.g. `https://alice.dsh.example.com`), otherwise links DSH generates may be unreachable. You can override it in `SandboxClaim.env` (if injection is allowed) or rewrite URLs in the gateway.
-
-3. **Image pull and platform.** As with the Hermes example, push `dsh-web` to a registry reachable by the cluster. Your image is built for `amd64`; node architecture must match.
-
-4. **PVC size.** DSH state size (sessions, configs, etc.) depends on usage; Hermes allocated a PVC for `/opt/data`. Adjust the `storage` request in `volumeClaimTemplates` accordingly.
-
-5. **Auto-suspend via `operatingMode`.** Hermes's gateway includes an idle sweeper that suspends idle instances automatically. If your DSH gateway implements a similar sweeper, you get the same cost model: `Running` while active, `Suspended` when idle (PVC + Service retained), auto-woken when the user returns.
-
-### ✅ Deployment Order
+The image is published to GitHub Container Registry as
+`ghcr.io/sredevopsorg/deepseek-kubernetes-sandbox:<tag>`, where `<tag>` is `main`,
+`latest`, or a DSH version tag such as `dsh-v0.2.1-rc.2`.
 
 ```bash
-kubectl create namespace dsh-demo
-kubectl apply -f 00-secret.yaml
-kubectl apply -f 10-sandboxtemplate.yaml
-kubectl apply -f 20-sandboxwarmpool.yaml
-kubectl apply -f 40-claim.yaml   # one per user
+docker run --rm -it \
+  -p 3080:3080 \
+  -v dsh-home:/home/node/.dsh \
+  -e DEEPSEEK_API_KEY \
+  ghcr.io/sredevopsorg/deepseek-kubernetes-sandbox:dsh-v0.2.1-rc.2
 ```
 
-After applying, watch the warm pool spin up with `kubectl -n dsh-demo get sandboxes -w`, and inspect adoption and `status.serviceFQDN` with `kubectl -n dsh-demo get sandboxclaim user-alice-dsh -o yaml`. At that point you have a platform structurally identical to the Hermes example, but running DSH Web for multiple users.
+> **Local port note:** `dsh web` listens on `3080` inside the container. If `3080` is
+> already in use on your workstation, map a different host port (e.g. `-p 18081:3080`).
+
+## The multi-user platform pattern
+
+`examples/dsh-agents-as-a-service/` is a port of the upstream
+[hermes-agents-as-a-service](https://github.com/kubernetes-sigs/agent-sandbox/tree/main/examples/hermes-agents-as-a-service)
+example to DSH Web. It demonstrates the platform shape — one `SandboxTemplate`, one warm pool,
+one tiny `SandboxClaim` per user, and `operatingMode` as the cost dial — plus a gateway that
+turns it into a usable service (signup, per-user tokens, `/u/{user}/**` proxy with
+wake-on-connect, idle suspend). See its [README](examples/dsh-agents-as-a-service/README.md) for
+the full walkthrough.
+
+### Mechanism mapping: Hermes → DSH
+
+| Hermes example mechanism | Role | Adaptation for DSH |
+| :--- | :--- | :--- |
+| `SandboxTemplate` | Reusable Pod blueprint | Image `dsh-web`, port `3080` |
+| `volumeClaimTemplates` | Auto-create a PVC per user | Mount `/home/node/.dsh` (DSH state) |
+| `SandboxWarmPool` | Pre-warmed instances | Keep as-is; size for expected concurrency |
+| `SandboxClaim` + `warmPoolRef` | "Registration" adopts a pooled instance | Keep exactly as-is; one claim per user |
+| `service: true` | Stable DNS per Sandbox | Enabled, so the gateway routes to each user's DSH Web |
+| `envVarsInjectionPolicy: Disallowed` | Reject claims that would cold-start | `DEEPSEEK_API_KEY` injected via `envFrom` on the template (shared platform key) |
+| `operatingMode` | Cost switch: Running / Suspended | Reused directly — suspend idle instances to save cost |
+
+### Key differences from Hermes
+
+- **Multi-arch** — the DSH image is built for amd64 and arm64 (no amd64 nodeSelector).
+- **Single port** — DSH Web serves everything on `3080`; no API/dashboard port split.
+- **One secret** — only `DEEPSEEK_API_KEY`; no dashboard/session/api-server credentials.
+- **Host forwarding** — the gateway forwards `Host` + `X-Forwarded-*` (and keeps
+  `DSH_PUBLIC_URL` empty) so DSH generates reachable links.
+- **WebSocket caveat** — the requests-based proxy handles SSE but not WebSocket upgrades; use a
+  path-based Ingress per user for full web-GUI interactivity.
+
+## Build it yourself
+
+```bash
+./build.sh                              # docker build + run (dsh-web:0.2.1-rc.2)
+# or
+docker build \
+  --build-arg DSH_VERSION=dsh-v0.2.1-rc.2 \
+  -t dsh-web:0.2.1-rc.2 .
+```
+
+See [`Dockerfile`](Dockerfile) for build args (`DSH_REPO`, `DSH_VERSION`, `PNPM_VERSION`) and
+the runtime image contract (non-root `node` user, `DSH_HOME=/home/node/.dsh`, `EXPOSE 3080`).
diff --git a/deploy/k8s/sandbox-template.yaml b/deploy/k8s/sandbox-template.yaml
index 230f304..b45a1c4 100644
--- a/deploy/k8s/sandbox-template.yaml
+++ b/deploy/k8s/sandbox-template.yaml
@@ -1,30 +1,53 @@
-apiVersion: extensions.agents.x-k8s.io/v1alpha1
+# Minimal single-tenant SandboxTemplate for DeepSeek Harness Web.
+#
+# This is the "just one DSH Web" quick-start template. For the full
+# multi-tenant platform (warm pool, per-user claims, gateway, idle-suspend),
+# see examples/dsh-agents-as-a-service/.
+apiVersion: extensions.agents.x-k8s.io/v1beta1
 kind: SandboxTemplate
 metadata:
   name: dsh-web-template
-  namespace: default
+  namespace: dsh-demo
 spec:
+  service: true # headless Service per Sandbox -> stable DNS (status.serviceFQDN)
   podTemplate:
     spec:
       securityContext:
         runAsNonRoot: true
         runAsUser: 1000 # Matches the UID of the 'node' user
         runAsGroup: 1000
+        fsGroup: 1000
       containers:
         - name: dsh-web
-          image: ghcr.io/sredevopsorg/deepseek-kubernetes-sandbox/deepseek-kubernetes-sandbox:0.2.1-rc.2
+          image: ghcr.io/sredevopsorg/deepseek-kubernetes-sandbox:dsh-v0.2.1-rc.2
+          imagePullPolicy: IfNotPresent
           ports:
-            - containerPort: 3080
-          env:
-            - name: DEEPSEEK_API_KEY
-              valueFrom:
-                secretKeyRef:
-                  name: deepseek-secret
-                  key: api-key
+            - name: web
+              containerPort: 3080
+          envFrom:
+            - secretRef:
+                name: dsh-provider-keys # Secret with key DEEPSEEK_API_KEY
           volumeMounts:
             - name: dsh-home
-              mountPath: /home/node/.dsh
-      volumes:
-        - name: dsh-home
-          persistentVolumeClaim:
-            claimName: dsh-home-pvc # You will need to create this PVC
\ No newline at end of file
+              mountPath: /home/node/.dsh # DSH_HOME
+          readinessProbe:
+            httpGet:
+              path: /
+              port: 3080
+            initialDelaySeconds: 5
+            periodSeconds: 5
+          livenessProbe:
+            httpGet:
+              path: /
+              port: 3080
+            initialDelaySeconds: 15
+            periodSeconds: 15
+  volumeClaimTemplates:
+    - metadata:
+        name: dsh-home
+      spec:
+        accessModes: ["ReadWriteOnce"]
+        resources:
+          requests:
+            storage: 5Gi
+        # storageClassName omitted => cluster default
diff --git a/examples/dsh-agents-as-a-service/00-prereqs.yaml b/examples/dsh-agents-as-a-service/00-prereqs.yaml
new file mode 100644
index 0000000..7d82f77
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/00-prereqs.yaml
@@ -0,0 +1,27 @@
+# DSH "Agents as a Service" — prerequisites: namespace + shared provider key.
+#
+# The dsh-provider-keys Secret is deliberately NOT committed with a usable
+# value. DEEPSEEK_API_KEY reaches every tenant sandbox (injected via envFrom
+# on the SandboxTemplate, stamped before any user exists — a warm-pool
+# constraint), so a real key committed here would be shared by every install
+# of this example. Generate a fresh value per install (README step 1;
+# run-test-kind.sh uses a throwaway):
+#
+#   kubectl -n dsh-demo create secret generic dsh-provider-keys \
+#     --from-literal=DEEPSEEK_API_KEY="sk-..."
+#
+# All the infrastructure in this example works without a real key; only
+# actual model calls need one.
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: dsh-demo
+---
+apiVersion: v1
+kind: Secret
+metadata:
+  name: dsh-provider-keys
+  namespace: dsh-demo
+type: Opaque
+# stringData:
+#   DEEPSEEK_API_KEY: "sk-your-key-here"
diff --git a/examples/dsh-agents-as-a-service/10-sandboxtemplate.yaml b/examples/dsh-agents-as-a-service/10-sandboxtemplate.yaml
new file mode 100644
index 0000000..dcd3fe0
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/10-sandboxtemplate.yaml
@@ -0,0 +1,81 @@
+# The reusable DSH Web blueprint. Warm pools stamp pod env BEFORE any user
+# exists, so per-user secrets cannot live in pod env; the shared provider key
+# is injected via envFrom, and claims that would bypass the warm pool are
+# rejected outright (see 40-claim-rejected.yaml).
+apiVersion: extensions.agents.x-k8s.io/v1beta1
+kind: SandboxTemplate
+metadata:
+  name: dsh-web
+  namespace: dsh-demo
+spec:
+  # Claims must not inject env or volumeClaimTemplates: with Disallowed, such
+  # claims are REJECTED instead of silently taking the cold (non-warm-pool)
+  # provisioning path. Demonstrated in step 6.
+  envVarsInjectionPolicy: Disallowed
+  volumeClaimTemplatesPolicy: Disallowed
+
+  # Per-sandbox headless Service -> stable DNS (sandbox.status.serviceFQDN).
+  # The gateway routes each user to their DSH Web instance by this name.
+  service: true
+
+  # No gateway pod scopes ingress in this standalone demo, so the
+  # NetworkPolicy is left unmanaged. A production platform would use
+  # networkPolicyManagement: Managed and restrict ingress to its gateway on
+  # :3080 — with the shared provider key, that NetworkPolicy is the per-user
+  # isolation boundary.
+  networkPolicyManagement: Unmanaged
+
+  podTemplate:
+    spec:
+      # DSH Web is multi-arch (amd64 + arm64); no arch nodeSelector needed.
+      restartPolicy: "OnFailure"
+      securityContext:
+        runAsUser: 1000   # the 'node' user in the DSH image
+        runAsGroup: 1000
+        fsGroup: 1000
+        runAsNonRoot: true
+      containers:
+        - name: dsh-web
+          image: ghcr.io/sredevopsorg/deepseek-kubernetes-sandbox:dsh-v0.2.1-rc.2
+          imagePullPolicy: IfNotPresent
+          ports:
+            - name: web
+              containerPort: 3080
+          envFrom:
+            - secretRef:
+                name: dsh-provider-keys
+          env:
+            # Empty => DSH derives links from the request Host, which the
+            # gateway forwards (plus X-Forwarded-*). Point this at the
+            # gateway base URL if you want hard-coded absolute links instead.
+            - name: DSH_PUBLIC_URL
+              value: ""
+          readinessProbe:
+            httpGet:
+              path: /
+              port: 3080
+            initialDelaySeconds: 5
+            periodSeconds: 5
+            failureThreshold: 30
+          livenessProbe:
+            httpGet:
+              path: /
+              port: 3080
+            initialDelaySeconds: 15
+            periodSeconds: 15
+            failureThreshold: 3
+          resources:
+            requests: {cpu: 250m, memory: 512Mi}
+            limits: {cpu: "1", memory: 1Gi}
+          volumeMounts:
+            - name: dsh-home
+              mountPath: /home/node/.dsh # DSH_HOME: sessions, config, state
+  volumeClaimTemplates:
+    - metadata:
+        name: dsh-home
+      spec:
+        accessModes: ["ReadWriteOnce"]
+        resources:
+          requests:
+            storage: 5Gi
+        # storageClassName omitted => cluster default (kind & GKE both work)
diff --git a/examples/dsh-agents-as-a-service/20-sandboxwarmpool.yaml b/examples/dsh-agents-as-a-service/20-sandboxwarmpool.yaml
new file mode 100644
index 0000000..c3bc768
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/20-sandboxwarmpool.yaml
@@ -0,0 +1,12 @@
+# Pre-warmed spares: image pulled, PVC provisioned, DSH Web booted, probe
+# green — everything a signup would otherwise wait for is pre-paid, so
+# claiming an instance takes ~2s instead of a full cold provision.
+apiVersion: extensions.agents.x-k8s.io/v1beta1
+kind: SandboxWarmPool
+metadata:
+  name: dsh-pool
+  namespace: dsh-demo
+spec:
+  replicas: 2
+  sandboxTemplateRef:
+    name: dsh-web
diff --git a/examples/dsh-agents-as-a-service/30-claim-alice.yaml b/examples/dsh-agents-as-a-service/30-claim-alice.yaml
new file mode 100644
index 0000000..beaad28
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/30-claim-alice.yaml
@@ -0,0 +1,19 @@
+# "User signup": one claim per user. This is the entire per-user resource —
+# the platform pattern keeps it deliberately minimal:
+#
+#   - warmPoolRef ONLY. No spec.env and no spec.volumeClaimTemplates: either
+#     would force a cold start past the warm pool (and this template's
+#     Disallowed policies reject them outright — see 40-claim-rejected.yaml).
+#   - additionalPodMetadata is the one warm-compatible per-claim
+#     customization. Label keys must use the sandbox.users.io domain.
+apiVersion: extensions.agents.x-k8s.io/v1beta1
+kind: SandboxClaim
+metadata:
+  name: dsh-alice
+  namespace: dsh-demo
+spec:
+  warmPoolRef:
+    name: dsh-pool
+  additionalPodMetadata:
+    labels:
+      sandbox.users.io/dsh-user: alice
diff --git a/examples/dsh-agents-as-a-service/40-claim-rejected.yaml b/examples/dsh-agents-as-a-service/40-claim-rejected.yaml
new file mode 100644
index 0000000..2849c5d
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/40-claim-rejected.yaml
@@ -0,0 +1,18 @@
+# NEGATIVE demo: this claim carries spec.env, which the template's
+# envVarsInjectionPolicy: Disallowed rejects. The claim never binds a
+# sandbox — inspect it to see the rejection (step 6 of the README).
+#
+# Why a platform wants rejection over acceptance: env injection forces a
+# cold start that bypasses the warm pool, silently turning a ~2s signup into
+# a cold provision. Disallowed turns that mistake into a loud error.
+apiVersion: extensions.agents.x-k8s.io/v1beta1
+kind: SandboxClaim
+metadata:
+  name: dsh-mallory
+  namespace: dsh-demo
+spec:
+  warmPoolRef:
+    name: dsh-pool
+  env:
+    - name: SNEAKY_OVERRIDE
+      value: "should-not-work"
diff --git a/examples/dsh-agents-as-a-service/50-gateway.yaml b/examples/dsh-agents-as-a-service/50-gateway.yaml
new file mode 100644
index 0000000..8ca6bb5
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/50-gateway.yaml
@@ -0,0 +1,94 @@
+# The gateway: control+data plane that turns the resources into a usable
+# service. Build the image first (see README step 8):
+#   docker build -t dsh-aaas-gateway:demo gateway/
+#   kind load docker-image dsh-aaas-gateway:demo   # (kind clusters)
+apiVersion: v1
+kind: ServiceAccount
+metadata:
+  name: dsh-gateway
+  namespace: dsh-demo
+---
+# Least privilege: claims are created/deleted per user; sandboxes are only
+# read and patched (operatingMode flips for suspend/resume).
+apiVersion: rbac.authorization.k8s.io/v1
+kind: Role
+metadata:
+  name: dsh-gateway
+  namespace: dsh-demo
+rules:
+  - apiGroups: ["extensions.agents.x-k8s.io"]
+    resources: ["sandboxclaims"]
+    verbs: ["get", "list", "create", "delete"]
+  - apiGroups: ["agents.x-k8s.io"]
+    resources: ["sandboxes"]
+    verbs: ["get", "list", "patch"]
+---
+apiVersion: rbac.authorization.k8s.io/v1
+kind: RoleBinding
+metadata:
+  name: dsh-gateway
+  namespace: dsh-demo
+roleRef:
+  apiGroup: rbac.authorization.k8s.io
+  kind: Role
+  name: dsh-gateway
+subjects:
+  - kind: ServiceAccount
+    name: dsh-gateway
+    namespace: dsh-demo
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: dsh-gateway
+  namespace: dsh-demo
+spec:
+  replicas: 1 # in-memory idle clock => single replica by design
+  selector:
+    matchLabels:
+      app: dsh-gateway
+  template:
+    metadata:
+      labels:
+        app: dsh-gateway
+    spec:
+      serviceAccountName: dsh-gateway
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 10000 # matches USER in gateway/Dockerfile
+        runAsGroup: 10000
+        seccompProfile:
+          type: RuntimeDefault
+      containers:
+        - name: gateway
+          image: dsh-aaas-gateway:demo
+          imagePullPolicy: IfNotPresent
+          securityContext:
+            allowPrivilegeEscalation: false
+            readOnlyRootFilesystem: true
+            capabilities:
+              drop: ["ALL"]
+          ports:
+            - containerPort: 8080
+          env:
+            - name: NAMESPACE
+              value: dsh-demo
+            - name: POOL
+              value: dsh-pool
+            - name: IDLE_TIMEOUT
+              value: "60"
+          resources:
+            requests: {cpu: 50m, memory: 128Mi}
+            limits: {cpu: 500m, memory: 256Mi}
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: dsh-gateway
+  namespace: dsh-demo
+spec:
+  selector:
+    app: dsh-gateway
+  ports:
+    - port: 8080
+      targetPort: 8080
diff --git a/examples/dsh-agents-as-a-service/README.md b/examples/dsh-agents-as-a-service/README.md
new file mode 100644
index 0000000..21c0072
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/README.md
@@ -0,0 +1,198 @@
+# DeepSeek Harness Web as a Service: the multi-user platform pattern
+
+This example distills the upstream
+[hermes-agents-as-a-service](https://github.com/kubernetes-sigs/agent-sandbox/tree/main/examples/hermes-agents-as-a-service)
+platform pattern into pure agent-sandbox resources + kubectl, running **DeepSeek Harness Web
+(DSH)** instead of Hermes. Every user of such a platform gets a personal, long-lived DSH Web
+instance that is **provisioned in ~2 seconds from a warm pool, suspended when idle to save
+cost, and woken transparently when the user returns** — with all state (conversations,
+sessions, config) surviving on a per-user PVC at `/home/node/.dsh`.
+
+It ships in two layers:
+
+1. **The resource layer** (steps 1–7): pure manifests + kubectl, so every mechanism is visible.
+2. **The service layer** (step 8): a small included **gateway** (`gateway/`) that makes it
+   directly usable — signup API, per-user bearer tokens, a `/u/{user}/**` proxy that
+   transparently **wakes suspended instances on connect**, and an idle sweeper that suspends
+   them again.
+
+What it demonstrates, feature by feature:
+
+| agent-sandbox feature | Role in the platform |
+| --- | --- |
+| `SandboxTemplate` with `volumeClaimTemplates` | Identical DSH pods; per-user state on a PVC at `/home/node/.dsh` |
+| `envVarsInjectionPolicy` / `volumeClaimTemplatesPolicy: Disallowed` | Claims that would bypass the warm pool are **rejected loudly**, not cold-started silently |
+| `service: true` | Stable per-sandbox DNS (`status.serviceFQDN`) for the gateway to route on |
+| `SandboxWarmPool` | Signup latency is pre-paid: image pull, PVC provision, DSH boot all happen before any user exists |
+| `SandboxClaim` + `warmPoolRef` | "Signup" = create a ~10-line claim; adoption is a label flip, Ready in ~2s |
+| `additionalPodMetadata` | The one warm-compatible per-claim customization (user label on the pod) |
+| `Sandbox.spec.operatingMode` | Suspend = delete only the pod (PVC + Service survive); resume = recreate pod, reattach the same PVC |
+| Cascade delete | Deleting the claim garbage-collects sandbox **and** PVC — account deletion in one call |
+
+## Differences from the Hermes example
+
+- **Multi-arch**: the DSH image is built for amd64 and arm64, so there is no
+  `kubernetes.io/arch: amd64` nodeSelector.
+- **Single port**: DSH Web serves everything on `3080`; the gateway proxies to one upstream
+  instead of splitting `/v1/*` (API) from the dashboard.
+- **One shared secret**: only `DEEPSEEK_API_KEY` is injected (via `envFrom`); there is no
+  dashboard basic-auth, session-cookie secret, or API-server key.
+- **Host forwarding**: the gateway forwards the original `Host` + `X-Forwarded-*` (and keeps
+  `DSH_PUBLIC_URL` empty) so DSH generates reachable links from the request.
+- **WebSocket caveat**: the requests-based streaming proxy handles SSE but not WebSocket
+  upgrades. For full DSH web-GUI interactivity, front the gateway with a path-based Ingress
+  per user (or a WS-capable proxy).
+
+## Prerequisites
+
+- A Kubernetes cluster with a default StorageClass (kind and GKE both work as-is).
+- agent-sandbox with extensions installed:
+
+  ```bash
+  VERSION=v1.0.7
+  kubectl apply -f https://github.com/kubernetes-sigs/agent-sandbox/releases/download/${VERSION}/sandbox-with-extensions.yaml
+  kubectl -n agent-sandbox-system rollout status deployment --timeout=180s
+  ```
+
+Or run everything below in one shot with `./run-test-kind.sh` (expects a kind cluster + the
+manifest above already applied).
+
+## 1. Install the platform resources
+
+First the namespace, then the shared provider key. The key reaches **every** tenant sandbox
+(a warm-pool constraint: pod env is stamped before any user exists), so it is **never
+committed** — generate a fresh value per install:
+
+```bash
+kubectl apply -f 00-prereqs.yaml
+kubectl -n dsh-demo create secret generic dsh-provider-keys \
+  --from-literal=DEEPSEEK_API_KEY="sk-..."
+kubectl apply -f 10-sandboxtemplate.yaml -f 20-sandboxwarmpool.yaml
+```
+
+Watch the two warm spares boot (first time pulls the image — subsequent operations never pay
+that again; that is the point of the pool):
+
+```bash
+kubectl -n dsh-demo get sandboxes -l agents.x-k8s.io/warm-pool-sandbox -w
+```
+
+Wait until both show `READY: True`. Each spare already has its PVC bound and DSH Web serving
+on port `3080`.
+
+## 2. "Signup": claim an instance (~2s)
+
+A user signing up is one small claim — `warmPoolRef` plus a pod label:
+
+```bash
+kubectl apply -f 30-claim-alice.yaml
+kubectl -n dsh-demo wait sandboxclaim dsh-alice --for=condition=Ready --timeout=60s
+```
+
+Inspect what happened:
+
+```bash
+SB=$(kubectl -n dsh-demo get sandboxclaim dsh-alice -o jsonpath='{.status.sandbox.name}')
+echo "alice's sandbox: $SB"
+kubectl -n dsh-demo get sandboxes --show-labels
+```
+
+Three things worth noticing:
+
+- **The sandbox keeps its pool-generated name** (`dsh-pool-…`), not the claim's name —
+  adoption re-labels an existing warm sandbox rather than creating one. A platform must always
+  resolve user → sandbox through `claim.status.sandbox.name`.
+- The `agents.x-k8s.io/warm-pool-sandbox` label is **gone** from the adopted sandbox (it left
+  the pool's inventory), and the pool has already started replenishing back to 2 spares.
+- The pod carries the claim's `additionalPodMetadata` label
+  (`sandbox.users.io/dsh-user=alice`), and
+  `kubectl -n dsh-demo get sandbox $SB -o jsonpath='{.status.serviceFQDN}'` shows the stable
+  DNS name the gateway proxies to.
+
+## 3. Talk to the instance
+
+```bash
+kubectl -n dsh-demo port-forward "$SB" 18081:3080
+# open http://localhost:18081 — DSH Web for alice
+```
+
+> **Local port note:** DSH Web listens on `3080` **in-cluster**. If `3080` is already in use
+> on your workstation (e.g. by a local DSH), forward to a different local port (here `18081`),
+> never `3080`.
+
+## 4. Write state on the PVC
+
+```bash
+kubectl -n dsh-demo exec "$SB" -- sh -c 'echo "remember me" > /home/node/.dsh/marker.txt'
+```
+
+## 5. Suspend: delete the pod, keep the PVC + Service
+
+```bash
+kubectl -n dsh-demo patch sandbox "$SB" --type merge -p '{"spec":{"operatingMode":"Suspended"}}'
+kubectl -n dsh-demo wait --for=delete "pod/$SB" --timeout=120s
+kubectl -n dsh-demo get pvc "dsh-home-$SB"   # still here
+kubectl -n dsh-demo get svc "$SB"            # still here
+```
+
+## 6. Resume and verify state survived
+
+```bash
+kubectl -n dsh-demo patch sandbox "$SB" --type merge -p '{"spec":{"operatingMode":"Running"}}'
+kubectl -n dsh-demo wait sandbox "$SB" --for=condition=Ready --timeout=180s
+kubectl -n dsh-demo exec "$SB" -- cat /home/node/.dsh/marker.txt   # "remember me"
+```
+
+## 7. Injection policy rejects env-carrying claims
+
+```bash
+kubectl apply -f 40-claim-rejected.yaml
+kubectl -n dsh-demo get sandboxclaim dsh-mallory \
+  -o jsonpath='{.status.conditions[?(@.type=="Ready")].reason}'
+# EnvVarsInjectionRejected — the claim never bound a sandbox
+kubectl delete -f 40-claim-rejected.yaml
+```
+
+## 8. The gateway: signup, proxy, idle-suspend, wake-on-connect
+
+Build the gateway image and deploy it:
+
+```bash
+docker build -t dsh-aaas-gateway:demo gateway/
+kind load docker-image dsh-aaas-gateway:demo   # (kind clusters)
+kubectl apply -f 50-gateway.yaml
+kubectl -n dsh-demo rollout status deploy/dsh-gateway --timeout=120s
+kubectl -n dsh-demo port-forward svc/dsh-gateway 18080:8080
+```
+
+Then:
+
+```bash
+# signup → claim over the warm pool + one-time bearer token
+SIGNUP=$(curl -s -X POST http://localhost:18080/users -H 'Content-Type: application/json' -d '{"user":"bob"}')
+TOKEN=$(echo "$SIGNUP" | sed -n 's/.*"token":"\([^"]*\)".*/\1/p')
+
+# proxy to bob's DSH Web (the gateway wakes it if suspended)
+curl -s -H "Authorization: Bearer $TOKEN" http://localhost:18080/u/bob/
+
+# tokenless is rejected
+curl -s -o /dev/null -w '%{http_code}\n' http://localhost:18080/u/bob/   # 401
+
+# after IDLE_TIMEOUT (60s) the sweeper suspends bob
+curl -s -H "Authorization: Bearer $TOKEN" http://localhost:18080/users/bob   # state: Suspended
+
+# delete bob (cascades sandbox + PVC)
+curl -s -X DELETE -H "Authorization: Bearer $TOKEN" http://localhost:18080/users/bob
+```
+
+## Cost model
+
+The same economics as the upstream pattern: suspend-when-idle plus warm pools turn an
+always-on instance into a sub-second cold-start pause per return-after-absence, at a fraction
+of the cost.
+
+## Cleanup
+
+```bash
+kubectl delete namespace dsh-demo
+```
diff --git a/examples/dsh-agents-as-a-service/gateway/Dockerfile b/examples/dsh-agents-as-a-service/gateway/Dockerfile
new file mode 100644
index 0000000..f83f0bf
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/gateway/Dockerfile
@@ -0,0 +1,7 @@
+FROM python:3.12-slim
+WORKDIR /app
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+COPY gateway.py .
+USER 10000
+CMD ["python", "gateway.py"]
diff --git a/examples/dsh-agents-as-a-service/gateway/gateway.py b/examples/dsh-agents-as-a-service/gateway/gateway.py
new file mode 100644
index 0000000..b52a14f
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/gateway/gateway.py
@@ -0,0 +1,348 @@
+# Minimal agents-as-a-service gateway for DeepSeek Harness Web (DSH).
+#
+# A compact, single-file control+data plane adapted from the upstream
+# hermes-agents-as-a-service gateway (kubernetes-sigs/agent-sandbox). It turns
+# the DSH sandbox resources into a usable multi-user service:
+#
+#   POST   /users {"user": "alice"}  signup: create a SandboxClaim over the
+#                                    warm pool, wait Ready, mint a bearer
+#                                    token (only its SHA-256 is stored, as a
+#                                    claim annotation)
+#   GET    /users/<user>             derived state: Ready/Waking/Suspended/...
+#                                    (requires the user's bearer token)
+#   DELETE /users/<user>             delete the claim (cascades sandbox + PVC;
+#                                    requires the user's bearer token)
+#   ANY    /u/<user>/<path>          per-user proxy. Requests to a suspended
+#                                    sandbox are HELD while it resumes
+#                                    (wake-on-connect); everything is proxied
+#                                    to the DSH Web UI on :3080.
+#
+# A background sweeper suspends sandboxes idle for IDLE_TIMEOUT seconds —
+# suspend deletes only the pod; the PVC (and the user's state) survives.
+#
+# Differences from the Hermes original:
+#   - Single upstream on :3080 (no OpenAI-compatible :8642 / dashboard :9119
+#     split, and no platform API key to inject).
+#   - The original Host header and X-Forwarded-* are forwarded (the gateway's
+#     own bearer token and any ?token= are still stripped) so DSH generates
+#     reachable links from the request Host.
+#   - Requests streaming proxies SSE but does NOT handle WebSocket upgrades;
+#     a per-user path-based Ingress is the alternative for full WS support.
+#
+# Deliberately NOT production code: single replica, in-memory idle clock,
+# no TLS.
+
+import hashlib
+import os
+import re
+import secrets
+import threading
+import time
+
+import requests
+from flask import Flask, Response, jsonify, request
+from kubernetes import client, config
+
+GROUP, VERSION = "extensions.agents.x-k8s.io", "v1beta1"
+CORE_GROUP = "agents.x-k8s.io"
+NAMESPACE = os.environ.get("NAMESPACE", "dsh-demo")
+POOL = os.environ.get("POOL", "dsh-pool")
+IDLE_TIMEOUT = int(os.environ.get("IDLE_TIMEOUT", "60"))
+WAKE_TIMEOUT = int(os.environ.get("WAKE_TIMEOUT", "120"))
+TOKEN_ANNOTATION = "aaas.example.com/token-sha256"
+# RFC 1123 label: lowercase alphanumerics and '-', must start/end
+# alphanumeric. The claim name is "dsh-<user>", so cap the user part at
+# 63 - len("dsh-") to keep the claim a valid Kubernetes object name.
+DNS1123_LABEL = re.compile(r"^[a-z0-9]([-a-z0-9]*[a-z0-9])?$")
+MAX_USER_LEN = 63 - len("dsh-")
+
+try:
+    config.load_incluster_config()
+except config.ConfigException:
+    config.load_kube_config()
+crd = client.CustomObjectsApi()
+
+app = Flask(__name__)
+# Cap proxied request bodies: Flask rejects anything larger with 413 before
+# request.get_data() would buffer it in memory.
+app.config["MAX_CONTENT_LENGTH"] = \
+    int(os.environ.get("MAX_BODY_MB", "32")) * 1024 * 1024
+last_activity: dict[str, float] = {}
+# Admitted proxy requests per user. While > 0 the sweeper must not suspend:
+# a request being served (or still streaming) after IDLE_TIMEOUT is
+# activity, not idleness. The sweeper checks-and-suspends under the same
+# lock that admission increments under, so the pair is atomic.
+in_flight: dict[str, int] = {}
+flight_lock = threading.Lock()
+
+
+def end_flight(user: str):
+    with flight_lock:
+        in_flight[user] = max(in_flight.get(user, 1) - 1, 0)
+    last_activity[user] = time.time()
+
+
+def claim_name(user: str) -> str:
+    return f"dsh-{user}"
+
+
+def get_claim(user: str):
+    try:
+        return crd.get_namespaced_custom_object(
+            GROUP, VERSION, NAMESPACE, "sandboxclaims", claim_name(user))
+    except client.ApiException as e:
+        if e.status == 404:
+            return None
+        raise
+
+
+def get_sandbox(name: str):
+    return crd.get_namespaced_custom_object(
+        CORE_GROUP, VERSION, NAMESPACE, "sandboxes", name)
+
+
+def sandbox_of(claim):
+    name = (claim.get("status") or {}).get("sandbox", {}).get("name")
+    if not name:
+        return None
+    try:
+        return get_sandbox(name)
+    except client.ApiException as e:
+        if e.status == 404:  # stale claim status, e.g. mid-cascade-delete
+            return None
+        raise
+
+
+def condition(obj, ctype: str) -> bool:
+    for c in (obj.get("status") or {}).get("conditions", []):
+        if c.get("type") == ctype:
+            return c.get("status") == "True"
+    return False
+
+
+def derive_state(sandbox) -> str:
+    """The state machine a platform shows users (mode x conditions)."""
+    if sandbox is None:
+        return "Provisioning"
+    mode = sandbox["spec"].get("operatingMode", "Running")
+    if mode == "Suspended":
+        return "Suspended" if condition(sandbox, "Suspended") else "Suspending"
+    return "Ready" if condition(sandbox, "Ready") else "Waking"
+
+
+def set_operating_mode(sandbox_name: str, mode: str):
+    crd.patch_namespaced_custom_object(
+        CORE_GROUP, VERSION, NAMESPACE, "sandboxes", sandbox_name,
+        {"spec": {"operatingMode": mode}})
+
+
+def wait_ready(user: str, timeout: int):
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        claim = get_claim(user)
+        sandbox = sandbox_of(claim) if claim else None
+        if sandbox is not None and derive_state(sandbox) == "Ready":
+            return sandbox
+        time.sleep(1)
+    return None
+
+
+def authorized(claim) -> bool:
+    # Header-only on purpose: a ?token= query param would leak the bearer
+    # token into access logs, browser history and Referer headers.
+    token = request.headers.get("Authorization", "").removeprefix("Bearer ")
+    want = (claim["metadata"].get("annotations") or {}).get(TOKEN_ANNOTATION, "")
+    got = hashlib.sha256(token.encode()).hexdigest()
+    return bool(token) and secrets.compare_digest(got, want)
+
+
+@app.post("/users")
+def create_user():
+    user = (request.get_json(silent=True) or {}).get("user", "")
+    if not isinstance(user, str) or not DNS1123_LABEL.fullmatch(user) \
+            or len(user) > MAX_USER_LEN:
+        return jsonify(error="body must be {'user': '<dns-1123 label>'}"), 400
+    if get_claim(user) is not None:
+        return jsonify(error="user exists"), 409
+    token = secrets.token_urlsafe(32)
+    # The claim IS the user record: warmPoolRef only (env/VCTs would be
+    # rejected by the template policies), token hash as an annotation.
+    try:
+        crd.create_namespaced_custom_object(
+            GROUP, VERSION, NAMESPACE, "sandboxclaims", {
+                "apiVersion": f"{GROUP}/{VERSION}",
+                "kind": "SandboxClaim",
+                "metadata": {
+                    "name": claim_name(user),
+                    "annotations": {
+                        TOKEN_ANNOTATION: hashlib.sha256(token.encode()).hexdigest()},
+                },
+                "spec": {
+                    "warmPoolRef": {"name": POOL},
+                    "additionalPodMetadata": {
+                        "labels": {"sandbox.users.io/dsh-user": user}},
+                },
+            })
+    except client.ApiException as e:
+        if e.status == 409:  # lost a concurrent-signup race for this user
+            return jsonify(error="user exists"), 409
+        raise
+    sandbox = wait_ready(user, WAKE_TIMEOUT)
+    last_activity[user] = time.time()
+    return jsonify(
+        user=user,
+        token=token,  # shown once; only the hash is stored
+        state=derive_state(sandbox),
+        sandbox=(sandbox or {}).get("metadata", {}).get("name"),
+    ), 201
+
+
+@app.get("/users/<user>")
+def get_user(user):
+    claim = get_claim(user)
+    if claim is None:
+        return jsonify(error="not found"), 404
+    if not authorized(claim):  # only the token holder may read their state
+        return jsonify(error="unauthorized"), 401
+    sandbox = sandbox_of(claim)
+    return jsonify(user=user, state=derive_state(sandbox),
+                   sandbox=(sandbox or {}).get("metadata", {}).get("name"))
+
+
+@app.delete("/users/<user>")
+def delete_user(user):
+    claim = get_claim(user)
+    if claim is None:
+        return jsonify(error="not found"), 404
+    if not authorized(claim):  # deletion cascades the PVC: token required
+        return jsonify(error="unauthorized"), 401
+    try:
+        crd.delete_namespaced_custom_object(
+            GROUP, VERSION, NAMESPACE, "sandboxclaims", claim_name(user))
+    except client.ApiException as e:
+        if e.status != 404:  # already gone => idempotent success
+            raise
+    last_activity.pop(user, None)
+    return jsonify(user=user, note="claim deleted; sandbox and PVC cascade")
+
+
+@app.route("/u/<user>/", defaults={"path": ""},
+           methods=["GET", "POST", "PUT", "DELETE", "PATCH"])
+@app.route("/u/<user>/<path:path>",
+           methods=["GET", "POST", "PUT", "DELETE", "PATCH"])
+def proxy(user, path):
+    claim = get_claim(user)
+    if claim is None:
+        return jsonify(error="unknown user"), 404
+    if not authorized(claim):
+        return jsonify(error="unauthorized"), 401
+
+    # Reserve activity the moment the request is admitted — BEFORE the
+    # sandbox is resolved or woken — so the sweeper can never suspend
+    # between admission and the upstream dial. Every early return releases
+    # the reservation; the streaming path hands it to relay(), which
+    # releases it when the response finishes (or the client disconnects).
+    with flight_lock:
+        in_flight[user] = in_flight.get(user, 0) + 1
+    last_activity[user] = time.time()
+    streaming = False
+    try:
+        sandbox = sandbox_of(claim)
+        if sandbox is None:
+            return jsonify(error="sandbox not provisioned yet"), 503
+        if derive_state(sandbox) != "Ready":
+            # Wake-on-connect: flip the mode (safe in ANY state, including
+            # mid-suspension) and HOLD the request until Ready.
+            set_operating_mode(sandbox["metadata"]["name"], "Running")
+            sandbox = wait_ready(user, WAKE_TIMEOUT)
+            if sandbox is None:
+                return jsonify(error="DSH is waking up, retry shortly"), \
+                    503, {"Retry-After": "10"}
+            last_activity[user] = time.time()
+
+        fqdn = (sandbox.get("status") or {}).get("serviceFQDN")
+        # Forward the original Host + X-Forwarded-* so DSH generates
+        # reachable links; still never forward the gateway's bearer token or
+        # a stray ?token= upstream where the app (or its logs) would see it.
+        headers = {k: v for k, v in request.headers
+                   if k.lower() != "authorization"}
+        headers.setdefault("X-Forwarded-For", request.remote_addr or "")
+        headers.setdefault("X-Forwarded-Proto", request.scheme)
+        headers.setdefault("X-Forwarded-Host", request.host)
+        params = [(k, v) for k, v in request.args.items(multi=True)
+                  if k != "token"]
+        upstream = f"http://{fqdn}:3080/{path}"
+
+        # The pod can be Ready a beat before the server binds; retry the
+        # dial.
+        for attempt in range(5):
+            try:
+                r = requests.request(
+                    request.method, upstream, params=params,
+                    data=request.get_data(), headers=headers,
+                    stream=True, timeout=300)
+                break
+            except requests.ConnectionError:
+                if attempt == 4:
+                    return jsonify(error="upstream unavailable"), 502
+                time.sleep(1)
+
+        def relay():
+            try:
+                yield from r.iter_content(chunk_size=None)
+            finally:  # runs on stream completion AND client disconnect
+                r.close()  # release the upstream connection either way
+                end_flight(user)
+
+        streaming = True  # relay() now owns the in-flight reservation
+        return Response(
+            relay(),
+            status=r.status_code,
+            headers=[(k, v) for k, v in r.headers.items()
+                     if k.lower() not in ("transfer-encoding", "connection")])
+    finally:
+        if not streaming:
+            end_flight(user)
+
+
+def idle_sweeper():
+    """Suspend sandboxes idle for IDLE_TIMEOUT (pod deleted, PVC survives)."""
+    while True:
+        time.sleep(15)
+        try:
+            claims = crd.list_namespaced_custom_object(
+                GROUP, VERSION, NAMESPACE, "sandboxclaims")["items"]
+            for c in claims:
+                user = c["metadata"]["name"].removeprefix("dsh-")
+                try:
+                    sandbox = sandbox_of(c)
+                except client.ApiException:
+                    continue
+                if sandbox is None or derive_state(sandbox) != "Ready":
+                    continue
+                # Check-and-suspend atomically with proxy admission (which
+                # increments in_flight under this same lock).
+                with flight_lock:
+                    if in_flight.get(user, 0) > 0:
+                        continue  # requests in flight => the user is active
+                    idle = time.time() - \
+                        last_activity.setdefault(user, time.time())
+                    if idle < IDLE_TIMEOUT:
+                        continue
+                    print(f"suspending {user} (idle {int(idle)}s)",
+                          flush=True)
+                    set_operating_mode(sandbox["metadata"]["name"],
+                                       "Suspended")
+        except Exception as e:
+            # A transient API error must not kill this daemon thread — that
+            # would silently stop idle suspension for good.
+            print(f"sweeper: transient error, retrying next cycle: {e}",
+                  flush=True)
+
+
+if __name__ == "__main__":
+    threading.Thread(target=idle_sweeper, daemon=True).start()
+    # threaded=True so one user's wake-on-connect (which can hold a request
+    # for up to WAKE_TIMEOUT) doesn't block everyone else. Still Werkzeug's
+    # dev server — a real deployment would sit behind gunicorn or similar.
+    app.run(host="0.0.0.0", port=8080, threaded=True)
diff --git a/examples/dsh-agents-as-a-service/gateway/requirements.txt b/examples/dsh-agents-as-a-service/gateway/requirements.txt
new file mode 100644
index 0000000..ad59624
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/gateway/requirements.txt
@@ -0,0 +1,3 @@
+flask==3.1.*
+requests==2.32.*
+kubernetes==31.0.*
diff --git a/examples/dsh-agents-as-a-service/gateway/test_gateway.py b/examples/dsh-agents-as-a-service/gateway/test_gateway.py
new file mode 100644
index 0000000..eba2a68
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/gateway/test_gateway.py
@@ -0,0 +1,382 @@
+"""
+Unit tests for the dsh-agents-as-a-service gateway example.
+
+gateway.py runs as the control+data plane in front of agent-sandbox; these
+tests cover its security/parsing logic (bearer-token auth, DNS-1123 user
+validation, the mode x conditions state machine) and the /users and
+idle-sweeper decision logic, with the Kubernetes CustomObjectsApi mocked.
+No kind cluster or container build is needed.
+
+gateway.py isn't a valid module name to import as a package member from a
+test file that also wants a fresh module per test (its top level calls
+config.load_incluster_config()/load_kube_config(), both of which would blow
+up or reach a real cluster), so each test loads its own copy by file path
+via the `gw` fixture, mirroring examples/hpa-swp-scaling/test_create_claim.py.
+"""
+
+import hashlib
+import importlib.util
+import json
+import os
+import time
+from types import SimpleNamespace
+from unittest import mock
+
+import pytest
+from kubernetes import client as k8s_client
+from kubernetes.config import ConfigException
+
+GATEWAY_PATH = os.path.join(os.path.dirname(__file__), "gateway.py")
+
+
+class _StopLoop(Exception):
+    """Sentinel used to break idle_sweeper()'s `while True` after one pass."""
+
+
+def _load_gateway(monkeypatch, mock_crd):
+    monkeypatch.setenv("NAMESPACE", "test-ns")
+    monkeypatch.setenv("POOL", "test-pool")
+    monkeypatch.setenv("IDLE_TIMEOUT", "60")
+    monkeypatch.setenv("WAKE_TIMEOUT", "120")
+
+    with mock.patch("kubernetes.config.load_incluster_config",
+                     side_effect=ConfigException("not running in a cluster")), \
+         mock.patch("kubernetes.config.load_kube_config"), \
+         mock.patch("kubernetes.client.CustomObjectsApi", return_value=mock_crd):
+        spec = importlib.util.spec_from_file_location("dsh_gateway", GATEWAY_PATH)
+        module = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(module)
+    return module
+
+
+@pytest.fixture
+def gw(monkeypatch):
+    """A freshly loaded gateway module (own last_activity/in_flight state)
+    with its Kubernetes CustomObjectsApi mocked, plus a Flask test client."""
+    mock_crd = mock.MagicMock()
+    module = _load_gateway(monkeypatch, mock_crd)
+    module.app.testing = True  # propagate unhandled exceptions instead of a bare 500
+    return SimpleNamespace(module=module, crd=mock_crd, client=module.app.test_client())
+
+
+def crd_get_side_effect(claims=None, sandboxes=None):
+    """Builds a get_namespaced_custom_object side_effect that serves
+    "sandboxclaims" from `claims` and "sandboxes" from `sandboxes` (both
+    {name: obj_or_None} maps), raising a 404 ApiException for anything
+    missing -- mirroring how get_claim()/get_sandbox() interpret 404."""
+    claims = claims or {}
+    sandboxes = sandboxes or {}
+
+    def _get(group, version, namespace, plural, name):
+        if plural == "sandboxclaims":
+            table = claims
+        elif plural == "sandboxes":
+            table = sandboxes
+        else:
+            raise AssertionError(f"unexpected plural {plural!r}")
+        obj = table.get(name)
+        if obj is None:
+            raise k8s_client.ApiException(status=404)
+        return obj
+
+    return _get
+
+
+# --- condition() ---------------------------------------------------------
+
+def test_condition_true_when_status_matches(gw):
+    obj = {"status": {"conditions": [{"type": "Ready", "status": "True"}]}}
+    assert gw.module.condition(obj, "Ready") is True
+
+
+def test_condition_false_when_status_is_false(gw):
+    obj = {"status": {"conditions": [{"type": "Ready", "status": "False"}]}}
+    assert gw.module.condition(obj, "Ready") is False
+
+
+def test_condition_false_when_type_absent(gw):
+    obj = {"status": {"conditions": [{"type": "Suspended", "status": "True"}]}}
+    assert gw.module.condition(obj, "Ready") is False
+
+
+def test_condition_false_when_no_status(gw):
+    assert gw.module.condition({}, "Ready") is False
+
+
+# --- derive_state() --------------------------------------------------------
+
+def test_derive_state_provisioning_when_sandbox_missing(gw):
+    assert gw.module.derive_state(None) == "Provisioning"
+
+
+def test_derive_state_ready(gw):
+    sandbox = {"spec": {}, "status": {"conditions": [{"type": "Ready", "status": "True"}]}}
+    assert gw.module.derive_state(sandbox) == "Ready"
+
+
+def test_derive_state_waking_when_not_yet_ready(gw):
+    sandbox = {"spec": {"operatingMode": "Running"}, "status": {"conditions": []}}
+    assert gw.module.derive_state(sandbox) == "Waking"
+
+
+def test_derive_state_suspended(gw):
+    sandbox = {"spec": {"operatingMode": "Suspended"},
+               "status": {"conditions": [{"type": "Suspended", "status": "True"}]}}
+    assert gw.module.derive_state(sandbox) == "Suspended"
+
+
+def test_derive_state_suspending_while_mode_flipped_but_not_settled(gw):
+    sandbox = {"spec": {"operatingMode": "Suspended"}, "status": {"conditions": []}}
+    assert gw.module.derive_state(sandbox) == "Suspending"
+
+
+# --- authorized() ----------------------------------------------------------
+
+def _claim_for_token(module, token):
+    return {"metadata": {"annotations": {
+        module.TOKEN_ANNOTATION: hashlib.sha256(token.encode()).hexdigest()}}}
+
+
+def test_authorized_accepts_correct_bearer_token(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    with gw.module.app.test_request_context(headers={"Authorization": "Bearer s3cret"}):
+        assert gw.module.authorized(claim) is True
+
+
+def test_authorized_rejects_wrong_token(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    with gw.module.app.test_request_context(headers={"Authorization": "Bearer wrong"}):
+        assert gw.module.authorized(claim) is False
+
+
+def test_authorized_rejects_missing_header(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    with gw.module.app.test_request_context():
+        assert gw.module.authorized(claim) is False
+
+
+def test_authorized_rejects_empty_bearer_token(gw):
+    claim = _claim_for_token(gw.module, "")
+    with gw.module.app.test_request_context(headers={"Authorization": "Bearer "}):
+        # bool(token) guards against an empty token matching an empty/unset
+        # annotation hash.
+        assert gw.module.authorized(claim) is False
+
+
+# --- POST /users (create_user) ---------------------------------------------
+
+@pytest.mark.parametrize("bad_user", ["Alice", "-alice", "alice-", "", "a_b", 123, None])
+def test_create_user_rejects_invalid_dns1123_label(gw, bad_user):
+    response = gw.client.post("/users", json={"user": bad_user})
+    assert response.status_code == 400
+    gw.crd.create_namespaced_custom_object.assert_not_called()
+
+
+def test_create_user_rejects_label_over_max_len(gw):
+    too_long = "a" * (gw.module.MAX_USER_LEN + 1)
+    response = gw.client.post("/users", json={"user": too_long})
+    assert response.status_code == 400
+    gw.crd.create_namespaced_custom_object.assert_not_called()
+
+
+def test_create_user_conflict_when_claim_already_exists(gw):
+    gw.crd.get_namespaced_custom_object.side_effect = crd_get_side_effect(
+        claims={"dsh-alice": {"metadata": {"name": "dsh-alice"}}})
+
+    response = gw.client.post("/users", json={"user": "alice"})
+
+    assert response.status_code == 409
+    gw.crd.create_namespaced_custom_object.assert_not_called()
+
+
+def test_create_user_conflict_on_concurrent_signup_race(gw):
+    gw.crd.get_namespaced_custom_object.side_effect = k8s_client.ApiException(status=404)
+    gw.crd.create_namespaced_custom_object.side_effect = k8s_client.ApiException(status=409)
+
+    response = gw.client.post("/users", json={"user": "alice"})
+
+    assert response.status_code == 409
+
+
+def test_create_user_success_stores_only_token_hash(gw, monkeypatch):
+    gw.crd.get_namespaced_custom_object.side_effect = k8s_client.ApiException(status=404)
+    ready_sandbox = {
+        "metadata": {"name": "dsh-alice-sbx"},
+        "spec": {},
+        "status": {"conditions": [{"type": "Ready", "status": "True"}]},
+    }
+    # wait_ready() polls with time.sleep(1); stub it out rather than racing
+    # a real poll loop in a unit test.
+    monkeypatch.setattr(gw.module, "wait_ready", lambda user, timeout: ready_sandbox)
+
+    response = gw.client.post("/users", json={"user": "alice"})
+
+    assert response.status_code == 201
+    body = response.get_json()
+    assert body["user"] == "alice"
+    assert body["state"] == "Ready"
+    assert body["sandbox"] == "dsh-alice-sbx"
+    assert body["token"]  # shown once in the response...
+
+    gw.crd.create_namespaced_custom_object.assert_called_once()
+    claim_body = gw.crd.create_namespaced_custom_object.call_args.args[4]
+    assert claim_body["metadata"]["name"] == "dsh-alice"
+    # ...but only its SHA-256 is ever sent to the API server.
+    want_hash = hashlib.sha256(body["token"].encode()).hexdigest()
+    assert claim_body["metadata"]["annotations"][gw.module.TOKEN_ANNOTATION] == want_hash
+    assert body["token"] not in json.dumps(claim_body)
+    assert claim_body["spec"]["warmPoolRef"]["name"] == "test-pool"
+    assert claim_body["spec"]["additionalPodMetadata"]["labels"]["sandbox.users.io/dsh-user"] == "alice"
+
+
+# --- GET /users/<user> and DELETE /users/<user> -----------------------------
+
+def test_get_user_not_found(gw):
+    gw.crd.get_namespaced_custom_object.side_effect = k8s_client.ApiException(status=404)
+    response = gw.client.get("/users/alice")
+    assert response.status_code == 404
+
+
+def test_get_user_requires_valid_token(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    gw.crd.get_namespaced_custom_object.side_effect = crd_get_side_effect(
+        claims={"dsh-alice": claim})
+
+    response = gw.client.get("/users/alice", headers={"Authorization": "Bearer wrong"})
+
+    assert response.status_code == 401
+
+
+def test_get_user_returns_state_when_authorized(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    sandbox = {"metadata": {"name": "dsh-alice-sbx"}, "spec": {},
+               "status": {"conditions": [{"type": "Ready", "status": "True"}]}}
+    claim["status"] = {"sandbox": {"name": "dsh-alice-sbx"}}
+    gw.crd.get_namespaced_custom_object.side_effect = crd_get_side_effect(
+        claims={"dsh-alice": claim}, sandboxes={"dsh-alice-sbx": sandbox})
+
+    response = gw.client.get("/users/alice", headers={"Authorization": "Bearer s3cret"})
+
+    assert response.status_code == 200
+    assert response.get_json()["state"] == "Ready"
+
+
+def test_delete_user_not_found(gw):
+    gw.crd.get_namespaced_custom_object.side_effect = k8s_client.ApiException(status=404)
+    response = gw.client.delete("/users/alice")
+    assert response.status_code == 404
+
+
+def test_delete_user_requires_valid_token(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    gw.crd.get_namespaced_custom_object.side_effect = crd_get_side_effect(
+        claims={"dsh-alice": claim})
+
+    response = gw.client.delete("/users/alice", headers={"Authorization": "Bearer wrong"})
+
+    assert response.status_code == 401
+    gw.crd.delete_namespaced_custom_object.assert_not_called()
+
+
+def test_delete_user_cascades_and_clears_activity(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    gw.crd.get_namespaced_custom_object.side_effect = crd_get_side_effect(
+        claims={"dsh-alice": claim})
+    gw.module.last_activity["alice"] = time.time()
+
+    response = gw.client.delete("/users/alice", headers={"Authorization": "Bearer s3cret"})
+
+    assert response.status_code == 200
+    gw.crd.delete_namespaced_custom_object.assert_called_once()
+    assert "alice" not in gw.module.last_activity
+
+
+def test_delete_user_already_gone_is_idempotent(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    gw.crd.get_namespaced_custom_object.side_effect = crd_get_side_effect(
+        claims={"dsh-alice": claim})
+    gw.crd.delete_namespaced_custom_object.side_effect = k8s_client.ApiException(status=404)
+
+    response = gw.client.delete("/users/alice", headers={"Authorization": "Bearer s3cret"})
+
+    assert response.status_code == 200
+
+
+def test_delete_user_unexpected_api_error_propagates(gw):
+    claim = _claim_for_token(gw.module, "s3cret")
+    gw.crd.get_namespaced_custom_object.side_effect = crd_get_side_effect(
+        claims={"dsh-alice": claim})
+    gw.crd.delete_namespaced_custom_object.side_effect = k8s_client.ApiException(status=500)
+
+    with pytest.raises(k8s_client.ApiException):
+        gw.client.delete("/users/alice", headers={"Authorization": "Bearer s3cret"})
+
+
+# --- idle_sweeper() ----------------------------------------------------------
+
+def _run_one_sweep(gw):
+    # time.sleep(15) runs before the try/except in idle_sweeper's loop body,
+    # so the first call must succeed (to let one pass execute) and only the
+    # second call breaks out.
+    with mock.patch("time.sleep", side_effect=[None, _StopLoop()]):
+        with pytest.raises(_StopLoop):
+            gw.module.idle_sweeper()
+
+
+def test_idle_sweeper_suspends_idle_ready_sandbox(gw):
+    sandbox = {"metadata": {"name": "dsh-alice-sbx"}, "spec": {},
+               "status": {"conditions": [{"type": "Ready", "status": "True"}]}}
+    claim_item = {"metadata": {"name": "dsh-alice"},
+                  "status": {"sandbox": {"name": "dsh-alice-sbx"}}}
+    gw.crd.list_namespaced_custom_object.return_value = {"items": [claim_item]}
+    gw.crd.get_namespaced_custom_object.return_value = sandbox
+    gw.module.last_activity["alice"] = time.time() - gw.module.IDLE_TIMEOUT - 10
+
+    _run_one_sweep(gw)
+
+    gw.crd.patch_namespaced_custom_object.assert_called_once()
+    args = gw.crd.patch_namespaced_custom_object.call_args.args
+    assert args[4] == "dsh-alice-sbx"
+    assert args[5] == {"spec": {"operatingMode": "Suspended"}}
+
+
+def test_idle_sweeper_skips_user_with_in_flight_request(gw):
+    sandbox = {"metadata": {"name": "dsh-alice-sbx"}, "spec": {},
+               "status": {"conditions": [{"type": "Ready", "status": "True"}]}}
+    claim_item = {"metadata": {"name": "dsh-alice"},
+                  "status": {"sandbox": {"name": "dsh-alice-sbx"}}}
+    gw.crd.list_namespaced_custom_object.return_value = {"items": [claim_item]}
+    gw.crd.get_namespaced_custom_object.return_value = sandbox
+    gw.module.last_activity["alice"] = time.time() - gw.module.IDLE_TIMEOUT - 10
+    gw.module.in_flight["alice"] = 1
+
+    _run_one_sweep(gw)
+
+    gw.crd.patch_namespaced_custom_object.assert_not_called()
+
+
+def test_idle_sweeper_skips_recently_active_user(gw):
+    sandbox = {"metadata": {"name": "dsh-alice-sbx"}, "spec": {},
+               "status": {"conditions": [{"type": "Ready", "status": "True"}]}}
+    claim_item = {"metadata": {"name": "dsh-alice"},
+                  "status": {"sandbox": {"name": "dsh-alice-sbx"}}}
+    gw.crd.list_namespaced_custom_object.return_value = {"items": [claim_item]}
+    gw.crd.get_namespaced_custom_object.return_value = sandbox
+    gw.module.last_activity["alice"] = time.time()
+
+    _run_one_sweep(gw)
+
+    gw.crd.patch_namespaced_custom_object.assert_not_called()
+
+
+def test_idle_sweeper_skips_sandbox_not_ready(gw):
+    sandbox = {"metadata": {"name": "dsh-alice-sbx"}, "spec": {}, "status": {"conditions": []}}
+    claim_item = {"metadata": {"name": "dsh-alice"},
+                  "status": {"sandbox": {"name": "dsh-alice-sbx"}}}
+    gw.crd.list_namespaced_custom_object.return_value = {"items": [claim_item]}
+    gw.crd.get_namespaced_custom_object.return_value = sandbox
+    gw.module.last_activity["alice"] = time.time() - gw.module.IDLE_TIMEOUT - 10
+
+    _run_one_sweep(gw)
+
+    gw.crd.patch_namespaced_custom_object.assert_not_called()
diff --git a/examples/dsh-agents-as-a-service/run-test-kind.sh b/examples/dsh-agents-as-a-service/run-test-kind.sh
new file mode 100755
index 0000000..d502f87
--- /dev/null
+++ b/examples/dsh-agents-as-a-service/run-test-kind.sh
@@ -0,0 +1,150 @@
+#!/usr/bin/env bash
+# Smoke test for the dsh-agents-as-a-service example: runs the full README
+# walkthrough non-interactively against the current kubectl context.
+# Expects agent-sandbox (with extensions) already installed.
+#
+# NOTE: the only local port bound is the gateway port-forward on 18080.
+# The DSH Web container port (3080) is in-cluster only and is never bound
+# on the local host.
+set -euo pipefail
+cd "$(dirname "$0")"
+
+NS=dsh-demo
+pass() { echo "PASS: $1"; }
+fail() { echo "FAIL: $1" >&2; exit 1; }
+cleanup() { kubectl delete namespace "$NS" --ignore-not-found --wait=false >/dev/null 2>&1 || true; }
+trap cleanup EXIT
+
+echo "=== 1. install prereqs (throwaway key) + template + warm pool"
+kubectl apply -f 00-prereqs.yaml
+# The provider key is never committed; generate a throwaway value.
+kubectl -n "$NS" delete secret dsh-provider-keys --ignore-not-found >/dev/null
+kubectl -n "$NS" create secret generic dsh-provider-keys \
+  --from-literal=DEEPSEEK_API_KEY="sk-test-$(openssl rand -hex 16)"
+kubectl apply -f 10-sandboxtemplate.yaml -f 20-sandboxwarmpool.yaml
+for _ in $(seq 1 120); do
+  READY=$(kubectl -n "$NS" get sandboxes -l agents.x-k8s.io/warm-pool-sandbox \
+    -o jsonpath='{range .items[*]}{.status.conditions[?(@.type=="Ready")].status}{"\n"}{end}' | grep -c True || true)
+  [ "$READY" -ge 2 ] && break
+  sleep 5
+done
+[ "${READY:-0}" -ge 2 ] || fail "warm pool did not reach 2 Ready spares"
+pass "warm pool: 2 spares Ready"
+
+echo "=== 2. signup via claim (warm adoption)"
+T0=$(date +%s)
+kubectl apply -f 30-claim-alice.yaml
+kubectl -n "$NS" wait sandboxclaim dsh-alice --for=condition=Ready --timeout=60s
+ELAPSED=$(( $(date +%s) - T0 ))
+SB=$(kubectl -n "$NS" get sandboxclaim dsh-alice -o jsonpath='{.status.sandbox.name}')
+case "$SB" in dsh-pool-*) ;; *) fail "expected warm adoption, got sandbox '$SB'";; esac
+pass "claim Ready in ${ELAPSED}s, adopted $SB"
+
+echo "=== 3. write state on the PVC"
+POD="$SB"
+kubectl -n "$NS" exec "$POD" -- sh -c 'echo "remember me" > /home/node/.dsh/marker.txt'
+pass "marker written in $POD"
+
+echo "=== 4. suspend (pod deleted, PVC + Service retained)"
+kubectl -n "$NS" patch sandbox "$SB" --type merge -p '{"spec":{"operatingMode":"Suspended"}}'
+kubectl -n "$NS" wait --for=delete "pod/$POD" --timeout=120s
+kubectl -n "$NS" get pvc "dsh-home-$SB" >/dev/null || fail "PVC was not retained"
+kubectl -n "$NS" get svc "$SB" >/dev/null || fail "Service was not retained"
+pass "suspended: pod gone, PVC and Service retained"
+
+echo "=== 5. resume and verify state survived"
+kubectl -n "$NS" patch sandbox "$SB" --type merge -p '{"spec":{"operatingMode":"Running"}}'
+kubectl -n "$NS" wait sandbox "$SB" --for=condition=Ready --timeout=180s
+POD="$SB"
+MARKER=$(kubectl -n "$NS" exec "$POD" -- cat /home/node/.dsh/marker.txt)
+[ "$MARKER" = "remember me" ] || fail "marker did not survive suspend/resume"
+pass "resumed: state survived on the reattached PVC"
+
+echo "=== 6. injection policy rejects env-carrying claims"
+kubectl apply -f 40-claim-rejected.yaml
+# Wait for the POSITIVE rejection signal, not merely "not bound yet".
+REASON=""
+for _ in $(seq 1 12); do
+  REASON=$(kubectl -n "$NS" get sandboxclaim dsh-mallory \
+    -o jsonpath='{.status.conditions[?(@.type=="Ready")].reason}' 2>/dev/null || true)
+  [ "$REASON" = "EnvVarsInjectionRejected" ] && break
+  sleep 5
+done
+[ "$REASON" = "EnvVarsInjectionRejected" ] \
+  || fail "claim was not rejected (Ready reason='$REASON')"
+BOUND=$(kubectl -n "$NS" get sandboxclaim dsh-mallory -o jsonpath='{.status.sandbox.name}' 2>/dev/null || true)
+[ -z "$BOUND" ] || fail "env-carrying claim unexpectedly bound sandbox '$BOUND'"
+kubectl delete -f 40-claim-rejected.yaml >/dev/null
+pass "claim rejected with EnvVarsInjectionRejected and never bound"
+
+echo "=== 7. gateway: signup, proxy, idle-suspend, wake-on-connect"
+if command -v docker >/dev/null && command -v kind >/dev/null; then
+  KIND_CLUSTER="${KIND_CLUSTER:-$(kubectl config current-context | sed 's/^kind-//')}"
+  docker build -q -t dsh-aaas-gateway:demo gateway/ >/dev/null
+  kind load docker-image dsh-aaas-gateway:demo --name "$KIND_CLUSTER" >/dev/null
+  kubectl apply -f 50-gateway.yaml
+  kubectl -n "$NS" rollout status deploy/dsh-gateway --timeout=120s
+  # Local port-forward uses 18080 (NOT 3080, which stays in-cluster).
+  kubectl -n "$NS" port-forward svc/dsh-gateway 18080:8080 >/dev/null 2>&1 &
+  PF=$!
+  sleep 3
+  GW=localhost:18080
+
+  SIGNUP=$(curl -s -X POST "http://$GW/users" -H 'Content-Type: application/json' -d '{"user":"bob"}')
+  TOKEN=$(echo "$SIGNUP" | sed -n 's/.*"token":"\([^"]*\)".*/\1/p')
+  [ -n "$TOKEN" ] || fail "signup returned no token: $SIGNUP"
+  echo "$SIGNUP" | grep -q '"state":"Ready"' || fail "signup not Ready: $SIGNUP"
+  pass "gateway signup: bob Ready with one-time token"
+
+  curl -sf -H "Authorization: Bearer $TOKEN" "http://$GW/u/bob/" >/dev/null \
+    || fail "proxied web call failed"
+  curl -s -o /dev/null -w '%{http_code}' "http://$GW/u/bob/" | grep -q 401 \
+    || fail "tokenless request was not rejected"
+  pass "gateway proxy: authed request 200, tokenless 401"
+
+  echo "waiting for idle sweeper (~75s)..."
+  for _ in $(seq 1 30); do
+    STATE=$(curl -s -H "Authorization: Bearer $TOKEN" "http://$GW/users/bob" \
+      | sed -n 's/.*"state":"\([^"]*\)".*/\1/p')
+    [ "$STATE" = "Suspended" ] && break
+    sleep 5
+  done
+  [ "$STATE" = "Suspended" ] || fail "bob was not idle-suspended (state=$STATE)"
+  pass "idle sweeper suspended bob"
+
+  T0=$(date +%s)
+  curl -sf -m 150 -H "Authorization: Bearer $TOKEN" "http://$GW/u/bob/" >/dev/null \
+    || fail "wake-on-connect request failed"
+  pass "wake-on-connect: request held and served after $(( $(date +%s) - T0 ))s"
+
+  curl -sf -X DELETE -H "Authorization: Bearer $TOKEN" "http://$GW/users/bob" >/dev/null \
+    || fail "gateway user deletion failed"
+  kill "$PF" 2>/dev/null || true
+elif [ "${SKIP_GATEWAY:-}" = "1" ]; then
+  echo "WARNING: gateway checks SKIPPED (SKIP_GATEWAY=1) — signup, auth," >&2
+  echo "WARNING: proxy, idle-suspend and wake-on-connect were NOT validated" >&2
+  GATEWAY_SKIPPED=1
+else
+  fail "docker + kind are required for the gateway checks (set SKIP_GATEWAY=1 to run only the resource-layer checks)"
+fi
+
+echo "=== 8. cascade delete"
+kubectl -n "$NS" delete sandboxclaim dsh-alice
+for _ in $(seq 1 24); do
+  kubectl -n "$NS" get sandbox "$SB" >/dev/null 2>&1 || break
+  sleep 5
+done
+kubectl -n "$NS" get sandbox "$SB" >/dev/null 2>&1 && fail "sandbox not garbage-collected"
+for _ in $(seq 1 24); do
+  kubectl -n "$NS" get pvc "dsh-home-$SB" >/dev/null 2>&1 || break
+  sleep 5
+done
+kubectl -n "$NS" get pvc "dsh-home-$SB" >/dev/null 2>&1 && fail "PVC not garbage-collected"
+pass "claim delete cascaded: sandbox + PVC gone"
+
+echo
+if [ -n "${GATEWAY_SKIPPED:-}" ]; then
+  echo "Resource-layer checks passed; gateway checks were SKIPPED."
+else
+  echo "All checks passed."
+fi