diff --git a/k8s/README.md b/k8s/README.md
new file mode 100644
index 000000000..fb5d8e460
--- /dev/null
+++ b/k8s/README.md
@@ -0,0 +1,62 @@
+# Kubernetes manifests — gittensory-miner (AMS) fleet mode
+
+Example manifests for running **N isolated miner workers** on a small Kubernetes cluster, as an alternative to
+`docker run` / docker-compose on a single host. Built on the existing
+[`packages/gittensory-miner/Dockerfile`](../packages/gittensory-miner/Dockerfile) image (see
+[`DEPLOYMENT.md`](../packages/gittensory-miner/DEPLOYMENT.md) for the fleet-mode overview). These are starting
+points — review resource sizing, storage class, and your registry before applying to a real cluster.
+
+## Why a StatefulSet (not a Deployment)
+
+The miner keeps all state in **local SQLite ledgers** (`claim-ledger.sqlite3`, `plan-store.sqlite3`, …) under
+`GITTENSORY_MINER_CONFIG_DIR` (`/data/miner`). Those stores are **not safe for concurrent multi-pod access**, so
+each worker needs its **own** volume. A Deployment can only mount a single shared PVC across every replica; a
+**StatefulSet's `volumeClaimTemplates`** give each replica its own PersistentVolumeClaim — so scaling to N
+replicas yields N workers with fully isolated state. That per-pod isolation is the whole reason these manifests
+use a StatefulSet.
+
+## Deploy
+
+1. **Build and push the image** from the monorepo root, then set `image:` in `miner-deployment.yaml`:
+   ```sh
+   docker build -f packages/gittensory-miner/Dockerfile -t <registry>/gittensory-miner:latest .
+   docker push <registry>/gittensory-miner:latest
+   ```
+2. **Create the Secret** (fill in real values first — never commit the filled-in copy):
+   ```sh
+   cp k8s/miner-secret.example.yaml k8s/miner-secret.yaml   # edit in your real GITHUB_TOKEN + provider keys
+   kubectl apply -f k8s/miner-secret.yaml
+   ```
+3. **Deploy the workers:**
+   ```sh
+   kubectl apply -f k8s/miner-deployment.yaml
+   ```
+
+## Scale
+
+Each replica is one isolated worker with its own volume. Scale the fleet with:
+
+```sh
+kubectl scale statefulset/gittensory-miner --replicas=<N>
+```
+
+or by editing `replicas:` in `miner-deployment.yaml` and re-applying. New replicas each get a fresh
+per-pod PVC from the `volumeClaimTemplate`; scaling down retains the PVCs (Kubernetes does not delete them
+automatically), so a scaled-back worker resumes its own state when scaled up again.
+
+## Notes
+
+- **Secrets** are injected at runtime via the Secret — the image contains no credentials. `ANTHROPIC_API_KEY` /
+  `OPENAI_API_KEY` are marked `optional`, so a worker running only the providers you configure starts cleanly
+  without the others.
+- **Resources** default to a modest CLI-worker baseline (`250m`/`512Mi` request, `1`/`1Gi` limit) with headroom
+  for a coding-agent subprocess. Tune for your providers and cluster.
+- **Storage** uses the cluster's default StorageClass at `2Gi` per pod; uncomment `storageClassName` in the
+  `volumeClaimTemplate` if you need a specific class.
+- **Image tag** — the example uses `:latest` (which defaults to `imagePullPolicy: Always`, re-pulling on every
+  restart). For production, push and pin an immutable tag (e.g. a version or digest).
+- **Probes** — no `livenessProbe`/`readinessProbe` is defined: the worker is a CLI loop, not a served endpoint,
+  so there's no health port to probe. Add a process-based `livenessProbe` (e.g. an `exec` check) if your
+  platform expects one.
+- **Filesystem ownership** — `fsGroup`/`runAsGroup` are set so the non-root user owns its PVC and can write the
+  SQLite files; without them the worker cannot create state on a root-owned `ReadWriteOnce` volume.
diff --git a/k8s/miner-deployment.yaml b/k8s/miner-deployment.yaml
new file mode 100644
index 000000000..e1b33b607
--- /dev/null
+++ b/k8s/miner-deployment.yaml
@@ -0,0 +1,88 @@
+# Kubernetes workload for gittensory-miner (AMS) fleet mode (#5181).
+#
+# This is a StatefulSet, NOT a Deployment: the miner keeps all of its state in local SQLite ledgers
+# (claim-ledger.sqlite3, plan-store.sqlite3, …) under GITTENSORY_MINER_CONFIG_DIR, and those stores are NOT
+# safe for concurrent multi-pod access. A Deployment can only mount one shared PVC across every replica; a
+# StatefulSet's volumeClaimTemplates give each replica its OWN PersistentVolumeClaim, so N workers run with
+# fully isolated state — the safety property this manifest exists to guarantee. See k8s/README.md.
+#
+# Image: build from packages/gittensory-miner/Dockerfile at the monorepo root and push to your registry:
+#   docker build -f packages/gittensory-miner/Dockerfile -t <registry>/gittensory-miner:latest .
+# Then set `image:` below. Secrets come from k8s/miner-secret.example.yaml (apply that first).
+apiVersion: apps/v1
+kind: StatefulSet
+metadata:
+  name: gittensory-miner
+  labels:
+    app.kubernetes.io/name: gittensory-miner
+    app.kubernetes.io/component: fleet-worker
+spec:
+  serviceName: gittensory-miner
+  # Scale the fleet by editing `replicas` (each replica is one isolated worker) or:
+  #   kubectl scale statefulset/gittensory-miner --replicas=<N>
+  replicas: 2
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: gittensory-miner
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: gittensory-miner
+        app.kubernetes.io/component: fleet-worker
+    spec:
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 1000
+        # runAsGroup + fsGroup so the non-root user owns the mounted PVC and can create its SQLite files on a
+        # root-owned ReadWriteOnce volume; OnRootMismatch avoids re-chowning a large volume on every restart.
+        runAsGroup: 1000
+        fsGroup: 1000
+        fsGroupChangePolicy: OnRootMismatch
+      containers:
+        - name: miner
+          # Replace with the image you built + pushed from packages/gittensory-miner/Dockerfile.
+          image: gittensory-miner:latest
+          # ENTRYPOINT is `gittensory-miner`; `run` is the continuous fleet-worker loop.
+          args: ["run"]
+          env:
+            - name: GITTENSORY_MINER_CONFIG_DIR
+              value: /data/miner
+            - name: GITHUB_TOKEN
+              valueFrom:
+                secretKeyRef:
+                  name: gittensory-miner-secrets
+                  key: GITHUB_TOKEN
+            # Coding-agent provider credentials — optional; present only for the providers you enable.
+            - name: ANTHROPIC_API_KEY
+              valueFrom:
+                secretKeyRef:
+                  name: gittensory-miner-secrets
+                  key: ANTHROPIC_API_KEY
+                  optional: true
+            - name: OPENAI_API_KEY
+              valueFrom:
+                secretKeyRef:
+                  name: gittensory-miner-secrets
+                  key: OPENAI_API_KEY
+                  optional: true
+          volumeMounts:
+            - name: miner-data
+              mountPath: /data/miner
+          # A CLI worker, not a server: modest baseline with headroom for a coding-agent subprocess.
+          resources:
+            requests:
+              cpu: 250m
+              memory: 512Mi
+            limits:
+              cpu: "1"
+              memory: 1Gi
+  # Per-pod PersistentVolumeClaim — each replica gets its OWN volume, never one shared across replicas.
+  volumeClaimTemplates:
+    - metadata:
+        name: miner-data
+      spec:
+        accessModes: ["ReadWriteOnce"]
+        # storageClassName: fast-ssd   # set this if your cluster's default StorageClass isn't what you want
+        resources:
+          requests:
+            storage: 2Gi
diff --git a/k8s/miner-secret.example.yaml b/k8s/miner-secret.example.yaml
new file mode 100644
index 000000000..2971cc6ca
--- /dev/null
+++ b/k8s/miner-secret.example.yaml
@@ -0,0 +1,27 @@
+# Secret template for gittensory-miner fleet workers (#5181).
+#
+# Values are intentionally EMPTY — this is a committed example, and a real-looking placeholder string in a
+# secret-named field trips secret scanners. Fill them in out-of-band; do NOT commit the filled-in copy.
+#
+# Recommended (keeps secret values out of any file entirely):
+#   kubectl create secret generic gittensory-miner-secrets \
+#     --from-literal=GITHUB_TOKEN=<your-token> \
+#     --from-literal=ANTHROPIC_API_KEY=<optional> --from-literal=OPENAI_API_KEY=<optional>
+#
+# Or copy + edit this file, then apply it (stringData takes raw values; Kubernetes base64-encodes them):
+#   cp k8s/miner-secret.example.yaml k8s/miner-secret.yaml   # then fill in the empty values
+#   kubectl apply -f k8s/miner-secret.yaml
+apiVersion: v1
+kind: Secret
+metadata:
+  name: gittensory-miner-secrets
+  labels:
+    app.kubernetes.io/name: gittensory-miner
+type: Opaque
+stringData:
+  # Required: a GitHub token the miner uses to read issues and open PRs.
+  GITHUB_TOKEN: ""
+  # Optional: only for the coding-agent providers you enable (claude-cli / codex-cli / agent-sdk).
+  # Delete a line entirely for any provider you do not use.
+  ANTHROPIC_API_KEY: ""
+  OPENAI_API_KEY: ""
diff --git a/test/unit/miner-k8s-manifests.test.ts b/test/unit/miner-k8s-manifests.test.ts
new file mode 100644
index 000000000..301170d6b
--- /dev/null
+++ b/test/unit/miner-k8s-manifests.test.ts
@@ -0,0 +1,116 @@
+import { readFileSync } from "node:fs";
+import { join } from "node:path";
+import { describe, expect, it } from "vitest";
+import { parse } from "yaml";
+
+// Validation for the example K8s manifests (#5181). These are static infra artifacts (not src/** logic), so
+// Codecov's patch gate doesn't apply to the YAML — but the safety property the manifests exist to guarantee
+// (per-pod isolated SQLite storage, never a shared PVC across replicas) is asserted here as a real test, and the
+// structural validator is exercised against both a well-formed manifest (passes) and a malformed one (fails).
+
+const K8S_DIR = join(process.cwd(), "k8s");
+const readManifest = (name: string): Record<string, unknown> =>
+  parse(readFileSync(join(K8S_DIR, name), "utf8")) as Record<string, unknown>;
+
+/** Minimal well-formedness check every Kubernetes resource must satisfy; throws on the first violation. */
+function validateK8sResource(doc: unknown): { kind: string } {
+  if (typeof doc !== "object" || doc === null)
+    throw new Error("manifest is not a mapping");
+  const m = doc as Record<string, unknown>;
+  for (const field of ["apiVersion", "kind", "metadata"] as const) {
+    if (m[field] == null)
+      throw new Error(`manifest missing required field: ${field}`);
+  }
+  const metadata = m.metadata as Record<string, unknown>;
+  if (typeof metadata.name !== "string" || metadata.name.length === 0) {
+    throw new Error("manifest missing metadata.name");
+  }
+  return { kind: String(m.kind) };
+}
+
+/** The safety property this issue introduces: each replica gets its OWN volume, never one shared claim. */
+function assertPerPodStorage(statefulSet: Record<string, unknown>): void {
+  const spec = statefulSet.spec as Record<string, unknown> | undefined;
+  const vcts = spec?.volumeClaimTemplates;
+  if (!Array.isArray(vcts) || vcts.length === 0) {
+    throw new Error(
+      "no volumeClaimTemplates: replicas would not get per-pod storage",
+    );
+  }
+  const template = spec?.template as Record<string, unknown> | undefined;
+  const podSpec = template?.spec as Record<string, unknown> | undefined;
+  const podVolumes =
+    (podSpec?.volumes as Array<Record<string, unknown>> | undefined) ?? [];
+  for (const volume of podVolumes) {
+    if (volume.persistentVolumeClaim)
+      throw new Error("a shared PVC across replicas is not allowed");
+  }
+}
+
+describe("k8s miner manifests (#5181)", () => {
+  const deployment = readManifest("miner-deployment.yaml");
+
+  it("miner-deployment.yaml is a well-formed StatefulSet", () => {
+    expect(validateK8sResource(deployment).kind).toBe("StatefulSet");
+  });
+
+  it("gives each replica its own SQLite volume (per-pod PVC, never a shared claim)", () => {
+    expect(() => assertPerPodStorage(deployment)).not.toThrow();
+    const spec = deployment.spec as Record<string, unknown>;
+    expect((spec.volumeClaimTemplates as unknown[]).length).toBeGreaterThan(0);
+    expect(typeof spec.replicas).toBe("number");
+  });
+
+  it("runs the continuous worker with config dir, a secret-sourced token, and resource bounds", () => {
+    const containers = (deployment.spec as any).template.spec
+      .containers as Array<Record<string, any>>;
+    const container = containers[0];
+    if (!container)
+      throw new Error("no container in the StatefulSet pod template");
+    expect(container.args).toContain("run");
+    const env = container.env as Array<Record<string, any>>;
+    const configDir = env.find((e) => e.name === "GITTENSORY_MINER_CONFIG_DIR");
+    expect(configDir?.value).toBe("/data/miner");
+    const token = env.find((e) => e.name === "GITHUB_TOKEN");
+    expect(token?.valueFrom?.secretKeyRef?.key).toBe("GITHUB_TOKEN");
+    expect(container.resources.requests).toBeTruthy();
+    expect(container.resources.limits).toBeTruthy();
+  });
+
+  it("secret template is a well-formed Secret exposing GITHUB_TOKEN", () => {
+    const secret = readManifest("miner-secret.example.yaml");
+    expect(validateK8sResource(secret).kind).toBe("Secret");
+    expect(
+      (secret.stringData as Record<string, unknown>).GITHUB_TOKEN,
+    ).toBeDefined();
+  });
+
+  it("the structural validator rejects a malformed manifest", () => {
+    const malformed = parse("apiVersion: apps/v1\nmetadata:\n  name: broken\n"); // no `kind`
+    expect(() => validateK8sResource(malformed)).toThrow(
+      /missing required field: kind/,
+    );
+  });
+
+  it("the per-pod-storage check rejects a shared-PVC configuration", () => {
+    const shared = {
+      spec: {
+        volumeClaimTemplates: [{ metadata: { name: "data" } }],
+        template: {
+          spec: {
+            volumes: [
+              { name: "data", persistentVolumeClaim: { claimName: "shared" } },
+            ],
+          },
+        },
+      },
+    };
+    expect(() => assertPerPodStorage(shared)).toThrow(/shared PVC/);
+  });
+
+  it("the per-pod-storage check rejects a config with no volumeClaimTemplates", () => {
+    expect(() =>
+      assertPerPodStorage({ spec: { template: { spec: {} } } }),
+    ).toThrow(/no volumeClaimTemplates/);
+  });
+});