diff --git a/.dockerignore b/.dockerignore
index fa56d2d..b4c6f04 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -1,10 +1,18 @@
 .git
+.claude
 .idea
 __pycache__
 *.pyc
 .venv
+.benchmarks
+.pytest_cache
+.ruff_cache
+.coverage
 .env
-data/
+# K8s 音频冒烟镜像只需要仓库自带的少量种子音频；其他本地数据不能进入构建上下文。
+data/*
+!data/audio/
+!data/audio/**
 *.db
 *.sqlite
 *.sqlite3
diff --git a/deploy/k8s/Dockerfile.audio b/deploy/k8s/Dockerfile.audio
new file mode 100644
index 0000000..1f62a14
--- /dev/null
+++ b/deploy/k8s/Dockerfile.audio
@@ -0,0 +1,31 @@
+# 本地 K8s 音频链路的完整 CPU runtime。
+# 此镜像与项目 uv.lock 保持一致，包含真实 SenseVoice 所需依赖和 system codec；
+# Kubernetes 资源声明不申请 GPU，CUDA 相关依赖即使存在也不会被使用。
+FROM python:3.12-slim
+
+COPY --from=ghcr.io/astral-sh/uv:0.9.17 /uv /uvx /bin/
+
+WORKDIR /app
+ENV UV_COMPILE_BYTECODE=1 \
+    UV_PROJECT_ENVIRONMENT=/opt/venv \
+    # PyPI 上的 Linux ARM64 Torch 依赖包含数百 MiB 的 CUDA wheel；本地网络较慢时，
+    # 解压阶段可能超过默认超时。保留有限下载并发，并把单个请求/解压的超时提高到 15 分钟。
+    UV_HTTP_TIMEOUT=900 \
+    UV_CONCURRENT_DOWNLOADS=2 \
+    PATH=/opt/venv/bin:$PATH \
+    CUDA_VISIBLE_DEVICES=""
+
+RUN apt-get update \
+    && apt-get install --yes --no-install-recommends \
+        build-essential ffmpeg libsndfile1 wget curl git \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY pyproject.toml uv.lock README.md ./
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv sync --frozen --no-install-project --no-dev
+COPY . .
+RUN --mount=type=cache,target=/root/.cache/uv uv sync --frozen --no-dev
+# 项目锁定 Ray 本体；最终 `uv sync` 会移除不在 lock 中的额外包，故必须在其之后
+# 安装 KubeRay health check 和 Ray Dashboard 需要的 default extra。
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --python /opt/venv/bin/python "ray[default]==2.55.1"
diff --git a/deploy/k8s/minio.yaml b/deploy/k8s/minio.yaml
new file mode 100644
index 0000000..e9bc287
--- /dev/null
+++ b/deploy/k8s/minio.yaml
@@ -0,0 +1,103 @@
+# 本地专用 MinIO，不读取宿主机或现有对象存储的任何数据。
+apiVersion: v1
+kind: Secret
+metadata:
+  name: minio-credentials
+  namespace: daft-ray
+type: Opaque
+stringData:
+  MINIO_ROOT_USER: minioadmin
+  MINIO_ROOT_PASSWORD: minioadmin
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: minio-data
+  namespace: daft-ray
+spec:
+  accessModes: [ReadWriteOnce]
+  resources:
+    requests:
+      storage: 4Gi
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: minio
+  namespace: daft-ray
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: minio
+  template:
+    metadata:
+      labels:
+        app: minio
+    spec:
+      containers:
+        - name: minio
+          image: minio/minio:RELEASE.2025-07-23T15-54-02Z
+          args: [server, /data, --console-address, ":9001"]
+          envFrom:
+            - secretRef:
+                name: minio-credentials
+          ports:
+            - containerPort: 9000
+              name: api
+            - containerPort: 9001
+              name: console
+          resources:
+            requests:
+              cpu: 250m
+              memory: 512Mi
+            limits:
+              cpu: "1"
+              memory: 1Gi
+          volumeMounts:
+            - name: data
+              mountPath: /data
+      volumes:
+        - name: data
+          persistentVolumeClaim:
+            claimName: minio-data
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: minio
+  namespace: daft-ray
+spec:
+  selector:
+    app: minio
+  ports:
+    - name: api
+      port: 9000
+      targetPort: api
+    - name: console
+      port: 9001
+      targetPort: console
+---
+# 创建 benchmark bucket；Job 可安全重复执行，现有 bucket 不会被删除或清空。
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: minio-init
+  namespace: daft-ray
+spec:
+  backoffLimit: 20
+  template:
+    spec:
+      restartPolicy: OnFailure
+      containers:
+        - name: mc
+          image: minio/mc:RELEASE.2025-07-21T05-28-08Z
+          envFrom:
+            - secretRef:
+                name: minio-credentials
+          command: [/bin/sh, -ec]
+          args:
+            - >-
+              until mc alias set local http://minio:9000 "$MINIO_ROOT_USER" "$MINIO_ROOT_PASSWORD";
+              do sleep 2; done;
+              mc mb --ignore-existing local/benchmark
diff --git a/deploy/k8s/mock-llm.yaml b/deploy/k8s/mock-llm.yaml
new file mode 100644
index 0000000..8df7217
--- /dev/null
+++ b/deploy/k8s/mock-llm.yaml
@@ -0,0 +1,44 @@
+# 使用项目内 OpenAI-compatible Mock LLM，避免本地 E2E 依赖外部模型服务或密钥。
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: mock-llm
+  namespace: daft-ray
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: mock-llm
+  template:
+    metadata:
+      labels:
+        app: mock-llm
+    spec:
+      containers:
+        - name: mock-llm
+          image: multimodal-toolkit:audio-local
+          imagePullPolicy: IfNotPresent
+          command: [python, -m, benchmark.audio, serve-mock, --host, 0.0.0.0, --port, "8010", --profile, fast]
+          ports:
+            - containerPort: 8010
+              name: http
+          resources:
+            requests:
+              cpu: 100m
+              memory: 256Mi
+            limits:
+              cpu: 500m
+              memory: 512Mi
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: mock-llm
+  namespace: daft-ray
+spec:
+  selector:
+    app: mock-llm
+  ports:
+    - name: http
+      port: 8010
+      targetPort: http
diff --git a/deploy/k8s/namespace.yaml b/deploy/k8s/namespace.yaml
new file mode 100644
index 0000000..4e69405
--- /dev/null
+++ b/deploy/k8s/namespace.yaml
@@ -0,0 +1,5 @@
+# 本地 Daft/Ray 演示环境的所有资源都隔离在此命名空间，便于一键清理。
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: daft-ray
diff --git a/deploy/k8s/prometheus.yaml b/deploy/k8s/prometheus.yaml
new file mode 100644
index 0000000..d2996a2
--- /dev/null
+++ b/deploy/k8s/prometheus.yaml
@@ -0,0 +1,113 @@
+# 轻量、单副本 Prometheus：只保存本地 Ray 集群的指标，避免本地开发机负担过重。
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: prometheus-config
+  namespace: daft-ray
+data:
+  prometheus.yml: |
+    global:
+      scrape_interval: 5s
+      evaluation_interval: 5s
+    scrape_configs:
+      - job_name: ray
+        kubernetes_sd_configs:
+          # Pod discovery 配合 Ray 节点标签，避免 Endpoints 元数据在不同
+          # Prometheus/Kubernetes 版本间的端口标签差异。
+          - role: pod
+            namespaces:
+              names: [daft-ray]
+        relabel_configs:
+          - source_labels: [__meta_kubernetes_pod_label_ray_io_is_ray_node]
+            action: keep
+            # KubeRay 将 ray.io/is-ray-node 标签的值设为 "yes"。
+            regex: "yes"
+          - source_labels: [__meta_kubernetes_pod_container_port_name]
+            action: keep
+            regex: metrics
+---
+ # Prometheus 通过 Kubernetes 服务发现读取 Service、Endpoints 和 Pod；默认账号没有
+ # 这些读取权限，会导致 UI 中没有任何 active target。
+apiVersion: v1
+kind: ServiceAccount
+metadata:
+  name: prometheus
+  namespace: daft-ray
+---
+apiVersion: rbac.authorization.k8s.io/v1
+kind: Role
+metadata:
+  name: prometheus-discovery
+  namespace: daft-ray
+rules:
+  - apiGroups: [""]
+    resources: ["services", "endpoints", "pods"]
+    verbs: ["get", "list", "watch"]
+---
+apiVersion: rbac.authorization.k8s.io/v1
+kind: RoleBinding
+metadata:
+  name: prometheus-discovery
+  namespace: daft-ray
+subjects:
+  - kind: ServiceAccount
+    name: prometheus
+    namespace: daft-ray
+roleRef:
+  apiGroup: rbac.authorization.k8s.io
+  kind: Role
+  name: prometheus-discovery
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: prometheus
+  namespace: daft-ray
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: prometheus
+  template:
+    metadata:
+      labels:
+        app: prometheus
+    spec:
+      serviceAccountName: prometheus
+      containers:
+        - name: prometheus
+          image: prom/prometheus:v3.5.0
+          args:
+            - --config.file=/etc/prometheus/prometheus.yml
+            - --storage.tsdb.path=/prometheus
+            - --storage.tsdb.retention.time=2d
+          ports:
+            - containerPort: 9090
+              name: web
+          resources:
+            requests:
+              cpu: 100m
+              memory: 256Mi
+            limits:
+              cpu: 500m
+              memory: 512Mi
+          volumeMounts:
+            - name: config
+              mountPath: /etc/prometheus
+      volumes:
+        - name: config
+          configMap:
+            name: prometheus-config
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: prometheus
+  namespace: daft-ray
+spec:
+  selector:
+    app: prometheus
+  ports:
+    - name: web
+      port: 9090
+      targetPort: web
diff --git a/deploy/k8s/ray-metrics-service.yaml b/deploy/k8s/ray-metrics-service.yaml
new file mode 100644
index 0000000..422ec15
--- /dev/null
+++ b/deploy/k8s/ray-metrics-service.yaml
@@ -0,0 +1,15 @@
+# Prometheus 集群内抓取所有 Ray Pod 的 OpenMetrics endpoint（head 与 worker）。
+apiVersion: v1
+kind: Service
+metadata:
+  name: daft-ray-metrics
+  namespace: daft-ray
+  labels:
+    app.kubernetes.io/name: daft-ray
+spec:
+  selector:
+    ray.io/cluster: daft-ray
+  ports:
+    - name: metrics
+      port: 8080
+      targetPort: 8080
diff --git a/deploy/k8s/raycluster.yaml b/deploy/k8s/raycluster.yaml
new file mode 100644
index 0000000..c77f8b1
--- /dev/null
+++ b/deploy/k8s/raycluster.yaml
@@ -0,0 +1,96 @@
+# KubeRay 管理的本地 Ray 集群。
+# 镜像由 scripts/k8s/build-image.sh 构建并载入 Minikube，包含完整音频 pipeline。
+apiVersion: ray.io/v1
+kind: RayCluster
+metadata:
+  name: daft-ray
+  namespace: daft-ray
+spec:
+  rayVersion: "2.55.1"
+  enableInTreeAutoscaling: true
+  autoscalerOptions:
+    upscalingMode: Default
+  headGroupSpec:
+    rayStartParams:
+      dashboard-host: "0.0.0.0"
+      metrics-export-port: "8080"
+      num-cpus: "0"
+    template:
+      spec:
+        containers:
+          - name: ray-head
+            image: multimodal-toolkit:audio-local
+            imagePullPolicy: IfNotPresent
+            env:
+              - name: MINIO_ENDPOINT
+                value: http://minio:9000
+              - name: MINIO_REGION
+                value: us-east-1
+              - name: NO_PROXY
+                value: minio,mock-llm,127.0.0.1,localhost
+              - name: no_proxy
+                value: minio,mock-llm,127.0.0.1,localhost
+            envFrom:
+              - secretRef:
+                  name: minio-credentials
+            resources:
+              requests:
+                cpu: 500m
+                memory: 4Gi
+              limits:
+                cpu: "1"
+                memory: 4Gi
+            ports:
+              - containerPort: 6379
+                name: gcs
+              - containerPort: 8265
+                name: dashboard
+              - containerPort: 8080
+                name: metrics
+  workerGroupSpecs:
+    - groupName: cpu-workers
+      replicas: 1
+      minReplicas: 1
+      maxReplicas: 2
+      rayStartParams:
+        metrics-export-port: "8080"
+      template:
+        spec:
+          containers:
+            - name: ray-worker
+              image: multimodal-toolkit:audio-local
+              imagePullPolicy: IfNotPresent
+              resources:
+                requests:
+                  cpu: "2"
+                  memory: 5Gi
+                limits:
+                  cpu: "2"
+                  memory: 5Gi
+              env:
+                - name: MINIO_ENDPOINT
+                  value: http://minio:9000
+                - name: MINIO_REGION
+                  value: us-east-1
+                - name: NO_PROXY
+                  value: minio,mock-llm,127.0.0.1,localhost
+                - name: no_proxy
+                  value: minio,mock-llm,127.0.0.1,localhost
+                # 同一个 worker 内的 ASR actor 复用模型；Pod 重建后会重新下载，避免依赖共享 RWX 卷。
+                - name: MODELSCOPE_CACHE
+                  value: /models
+                - name: CUDA_VISIBLE_DEVICES
+                  value: ""
+              ports:
+                # 为 Prometheus 的 Pod 服务发现显式声明 Ray metrics endpoint。
+                - containerPort: 8080
+                  name: metrics
+              envFrom:
+                - secretRef:
+                    name: minio-credentials
+              volumeMounts:
+                - name: models
+                  mountPath: /models
+          volumes:
+            - name: models
+              emptyDir: {}
diff --git a/docs/k8s-ray-local.md b/docs/k8s-ray-local.md
new file mode 100644
index 0000000..1fba7f0
--- /dev/null
+++ b/docs/k8s-ray-local.md
@@ -0,0 +1,82 @@
+# 本地 Kubernetes + Ray + Daft 完整音频环境
+
+本环境在专用的 Minikube profile `daft-ray` 中运行真实音频链路，不读取宿主机已有的
+MinIO、模型服务或 S3 数据：
+
+```text
+MinIO (cluster-local) → Daft download → SenseVoice ASR → Mock LLM → Lance Blob v2
+                              │
+                         KubeRay worker
+                              │ :8080
+                         Prometheus
+```
+
+Ray Dashboard 由 KubeRay head 提供；Prometheus 仅抓取 Ray head 和 worker 的内置
+OpenMetrics 指标。本环境不安装 Grafana，也不暴露业务自定义指标。
+
+## 前置条件
+
+- Docker Desktop 或 OrbStack 已运行，并至少可分配 **4 CPU / 12 GiB 内存**；首次模型下载
+  和完整 Python 镜像还需要若干 GiB 磁盘空间。
+- 已安装 `minikube`、`kubectl`、`helm`，且能访问镜像仓库、KubeRay Helm 仓库和 PyPI。
+- 首次 audio smoke 会下载 SenseVoice/VAD 模型。模型缓存位于 worker Pod 的本地卷；Pod
+  重建后会重新下载，属于开发环境的有意取舍。
+
+## 启动与访问
+
+```bash
+scripts/k8s/up.sh
+scripts/k8s/status.sh
+```
+
+`up.sh` 会创建专用 Minikube profile、构建 `multimodal-toolkit:audio-local`、安装固定
+版本 KubeRay operator，并部署 MinIO、Mock LLM、RayCluster 和 Prometheus。镜像完整安装
+当前 `uv.lock` 的项目依赖，但声明为 CPU-only，不请求 GPU。
+
+在独立终端执行端口转发：
+
+```bash
+kubectl --context daft-ray -n daft-ray port-forward service/daft-ray-head-svc 8265:8265
+kubectl --context daft-ray -n daft-ray port-forward service/prometheus 9090:9090
+```
+
+- Ray Dashboard：<http://127.0.0.1:8265>
+- Prometheus targets：<http://127.0.0.1:9090/targets>
+
+所有 Ray target 应为 `UP`。可从以下 PromQL 起步；实际可用指标名以 target 的 `/metrics`
+为准，因为会随 Ray 版本和组件变化：
+
+```promql
+ray_tasks
+ray_object_store_memory
+ray_node_cpu_utilization
+```
+
+## 运行真实音频 smoke
+
+```bash
+scripts/k8s/run-audio-smoke.sh
+```
+
+脚本会使用 UTC run-id，在集群内 MinIO 的 `benchmark/audio/<run-id>/` 下写入 4 条种子
+音频、Parquet manifest 与独立 Lance 表，随后通过 Ray Jobs API 提交：
+
+```text
+Daft download → @daft.cls SenseVoice ASR → HTTP Mock LLM → Lance Blob v2
+```
+
+结束时脚本会验证结果表行数、所有 `status=ok` 与 `audio_blob` 的 Blob v2 格式。执行期间可在
+Ray Dashboard 的 Jobs 页面查看作业和 actor；首次模型冷启动会明显比后续运行更久。
+
+## 清理
+
+```bash
+scripts/k8s/down.sh
+```
+
+它只删除专用 profile 中的 `daft-ray` 命名空间及其 MinIO PVC，不会删除 Minikube 或
+KubeRay operator。若确认不再需要整个专用集群，再手动执行：
+
+```bash
+minikube delete --profile daft-ray
+```
diff --git a/examples/k8s/verify_audio_smoke.py b/examples/k8s/verify_audio_smoke.py
new file mode 100644
index 0000000..112ebc3
--- /dev/null
+++ b/examples/k8s/verify_audio_smoke.py
@@ -0,0 +1,26 @@
+"""验证 Kubernetes 音频 smoke 的行保留、成功状态及 Lance Blob v2 写出。"""
+
+from __future__ import annotations
+
+import sys
+
+import daft
+
+from multimodal_toolkit.storage.blob import validate_blob_v2
+from multimodal_toolkit.storage.io import daft_io_config
+
+
+def main(uri: str, expected_rows: int) -> None:
+    validate_blob_v2(uri, "audio_blob")
+    rows = daft.read_lance(uri, io_config=daft_io_config()).select("status").to_pydict()["status"]
+    if len(rows) != expected_rows:
+        raise RuntimeError(f"expected {expected_rows} rows, got {len(rows)}")
+    if set(rows) != {"ok"}:
+        raise RuntimeError(f"audio smoke contains non-success statuses: {rows}")
+    print(f"validated {len(rows)} rows with status=ok and Lance Blob v2")
+
+
+if __name__ == "__main__":
+    if len(sys.argv) != 3:
+        raise SystemExit("usage: verify_audio_smoke.py <lance-uri> <expected-rows>")
+    main(sys.argv[1], int(sys.argv[2]))
diff --git a/scripts/k8s/build-image.sh b/scripts/k8s/build-image.sh
new file mode 100755
index 0000000..5071cd9
--- /dev/null
+++ b/scripts/k8s/build-image.sh
@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+# 构建与本地源码一致的 Ray runtime 镜像，并载入 Minikube 的节点镜像缓存。
+set -euo pipefail
+
+project_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+image="multimodal-toolkit:audio-local"
+
+docker build --quiet --file "$project_root/deploy/k8s/Dockerfile.audio" --tag "$image" "$project_root"
+minikube image load "$image" --profile daft-ray
diff --git a/scripts/k8s/down.sh b/scripts/k8s/down.sh
new file mode 100755
index 0000000..90f1562
--- /dev/null
+++ b/scripts/k8s/down.sh
@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+# 仅删除专用 daft-ray profile 中本项目命名空间的资源；不会删除 Minikube 或 operator。
+set -euo pipefail
+
+project_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+kubectl --context daft-ray delete namespace daft-ray --ignore-not-found
diff --git a/scripts/k8s/run-audio-smoke.sh b/scripts/k8s/run-audio-smoke.sh
new file mode 100755
index 0000000..e651e20
--- /dev/null
+++ b/scripts/k8s/run-audio-smoke.sh
@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# 生成隔离的测试音频，并以 Ray Jobs API 提交真实 Daft/SenseVoice/Mock-LMM/Lance 链路。
+set -euo pipefail
+
+kubectl_context="daft-ray"
+namespace="daft-ray"
+run_id="k8s-smoke-$(date -u +%Y%m%dT%H%M%SZ)"
+bucket="benchmark"
+prefix="audio"
+head_pod="$(kubectl --context "$kubectl_context" -n "$namespace" get pod \
+  -l ray.io/node-type=head -o jsonpath='{.items[0].metadata.name}')"
+
+if [[ -z "$head_pod" ]]; then
+  echo "Ray head is unavailable; run scripts/k8s/up.sh first." >&2
+  exit 1
+fi
+
+common_env=(
+  "MINIO_ENDPOINT=http://minio:9000"
+  "MINIO_ROOT_USER=minioadmin"
+  "MINIO_ROOT_PASSWORD=minioadmin"
+  "MINIO_REGION=us-east-1"
+  "NO_PROXY=minio,mock-llm,127.0.0.1,localhost"
+  "no_proxy=minio,mock-llm,127.0.0.1,localhost"
+)
+
+echo "[generate] run_id=$run_id"
+kubectl --context "$kubectl_context" -n "$namespace" exec "$head_pod" -c ray-head -- \
+  env "${common_env[@]}" python -m benchmark.audio generate \
+    --run-id "$run_id" --bucket "$bucket" --prefix "$prefix" --profile smoke --count 4 \
+    --source-dir /app/data/audio
+
+manifest="s3://$bucket/$prefix/$run_id/manifest.parquet"
+output="s3://$bucket/$prefix/$run_id/output.lance"
+echo "[submit] $manifest -> $output"
+kubectl --context "$kubectl_context" -n "$namespace" exec "$head_pod" -c ray-head -- \
+  env "${common_env[@]}" ray job submit --address http://127.0.0.1:8265 \
+    --submission-id "$run_id" --working-dir /app -- \
+    python -m benchmark.audio run --run-id "$run_id" --manifest "$manifest" \
+      --lance-uri "$output" --mock-url http://mock-llm:8010 --ray-address auto \
+      --asr-actor-cpus 2 --num-partitions 4 --max-minutes 20
+
+kubectl --context "$kubectl_context" -n "$namespace" exec "$head_pod" -c ray-head -- \
+  env "${common_env[@]}" python examples/k8s/verify_audio_smoke.py "$output" 4
+echo "Audio smoke completed: $output"
diff --git a/scripts/k8s/status.sh b/scripts/k8s/status.sh
new file mode 100755
index 0000000..38c9e9c
--- /dev/null
+++ b/scripts/k8s/status.sh
@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+# 输出连接方式；port-forward 在前台运行，使用独立终端启动它们。
+set -euo pipefail
+
+kubectl --context daft-ray --namespace daft-ray get raycluster,pods,svc,jobs
+cat <<'EOF'
+
+Ray Dashboard:
+  kubectl --context daft-ray -n daft-ray port-forward service/daft-ray-head-svc 8265:8265
+  open http://127.0.0.1:8265
+
+Prometheus:
+  kubectl --context daft-ray -n daft-ray port-forward service/prometheus 9090:9090
+  open http://127.0.0.1:9090/targets
+
+MinIO Console:
+  kubectl --context daft-ray -n daft-ray port-forward service/minio 9001:9001
+  open http://127.0.0.1:9001
+
+Real audio smoke:
+  scripts/k8s/run-audio-smoke.sh
+EOF
diff --git a/scripts/k8s/up.sh b/scripts/k8s/up.sh
new file mode 100755
index 0000000..a611c46
--- /dev/null
+++ b/scripts/k8s/up.sh
@@ -0,0 +1,38 @@
+#!/usr/bin/env bash
+# 启动专用 Minikube、完整音频 Ray 集群、MinIO、Mock LLM 与 Prometheus。
+set -euo pipefail
+
+project_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+kubectl_context="daft-ray"
+
+minikube start --profile "$kubectl_context" --driver=docker --cpus=4 --memory=12288
+"$project_root/scripts/k8s/build-image.sh"
+
+helm repo add kuberay https://ray-project.github.io/kuberay-helm/ --force-update
+helm repo update kuberay
+helm upgrade --install kuberay-operator kuberay/kuberay-operator \
+  --kube-context "$kubectl_context" \
+  --namespace kuberay-system --create-namespace --version 1.4.1
+
+kubectl --context "$kubectl_context" apply -f "$project_root/deploy/k8s/namespace.yaml"
+kubectl --context "$kubectl_context" apply -f "$project_root/deploy/k8s/minio.yaml"
+kubectl --context "$kubectl_context" apply -f "$project_root/deploy/k8s/prometheus.yaml"
+kubectl --context "$kubectl_context" apply -f "$project_root/deploy/k8s/mock-llm.yaml"
+
+kubectl --context "$kubectl_context" wait --namespace kuberay-system \
+  --for=condition=available deployment/kuberay-operator --timeout=5m
+kubectl --context "$kubectl_context" rollout status --namespace daft-ray deployment/minio --timeout=5m
+kubectl --context "$kubectl_context" wait --namespace daft-ray \
+  --for=condition=complete job/minio-init --timeout=5m
+
+kubectl --context "$kubectl_context" apply -f "$project_root/deploy/k8s/raycluster.yaml"
+kubectl --context "$kubectl_context" apply -f "$project_root/deploy/k8s/ray-metrics-service.yaml"
+
+kubectl wait --context "$kubectl_context" --namespace daft-ray \
+  --for=condition=ready pod -l ray.io/node-type=head --timeout=5m
+kubectl wait --context "$kubectl_context" --namespace daft-ray \
+  --for=condition=ready pod -l ray.io/node-type=worker --timeout=5m
+kubectl rollout status --context "$kubectl_context" --namespace daft-ray deployment/prometheus --timeout=5m
+kubectl rollout status --context "$kubectl_context" --namespace daft-ray deployment/mock-llm --timeout=5m
+
+echo "Environment ready. Run scripts/k8s/run-audio-smoke.sh for the real audio pipeline."