diff --git a/scripts/kernel-stress/README.md b/scripts/kernel-stress/README.md
new file mode 100644
index 0000000..55b782c
--- /dev/null
+++ b/scripts/kernel-stress/README.md
@@ -0,0 +1,63 @@
+# 内核二进制压力测试（kernel-stress）
+
+对本仓库**编译出来的真实内核**做压力测试 —— 与 `go test` 的区别是：它跑真二进制、
+真插件加载、真 unix socket 协议，因此能抓到只在集成面上出现的问题
+（已有战绩：根 agent `DataDir` 漏接线、插件通道没登记为 inputch、create 后子不开工）。
+
+## 为什么必须放在私有 netns 里
+
+生产实例占着 `*:8080 / *:9890 / *:9876`，而插件的监听都设了 `SO_REUSEADDR`：
+同机再起一个实例会**在 127.0.0.1 上与之并存绑定**（实测抢到过 `127.0.0.1:9890` 约 1 分钟）。
+`unshare -n` 后实例只有 `lo`，结构上不可能碰到生产端口。
+unix socket 是文件系统对象，跨 netns 仍可驱动，所以驱动脚本在 netns 外也能用。
+
+## 前置
+
+```bash
+go build -o /tmp/homed-stress ./cmd/homed      # 被压的内核
+export GOCACHE=/tmp/gocache GOPATH=/tmp/gopath TMPDIR=/var/tmp/gotmp
+```
+
+## 用法
+
+```bash
+# 1) 准备数据目录 + 把 LLM 指向本地 mock（无外网也能跑，且快、可控）
+DATA=/var/tmp/kstress
+mkdir -p $DATA
+#  先跑一次实例建出 config.db，再写入下面这些键（也可直接复用现成目录）：
+#    core.llm.provider=mock
+#    core.llm.sources.mock.base_url=http://127.0.0.1:9099/v1
+#    core.llm.sources.mock.model=mock   api_key=mock   adapter=openai
+#    core.llm.sources.mock.adapter_path=adapters/openai.lua   priority=100
+#    core.defaults.llm_endpoints=http://127.0.0.1:9099/v1/models   # 探活端点（探活用 HEAD！）
+#    core.defaults.rollback.max_retries=100000   auto_rollback=false
+#  并把 core.llm.sources.deepseek* 删掉（netns 里它不可达，会让 agent 被判 degraded → rollback 循环）
+
+# 2) 起 mock LLM + 内核（都在同一个私有 netns 里）
+MOCK_DELAY_MS=300 MOCK_CHUNKS=8 ./launch.sh /tmp/homed-stress
+
+# 3) 取认证密钥（cli 插件回落到 webui.api_key）
+export KCLI_KEY=$(sqlite3 $DATA/config.db "select value from config_webui where key='api_key';")
+
+# 4) 压
+./kcli.py $DATA/cli.sock stats full            # 看内核状态（含 scheduler 计数）
+./stress.py $DATA/cli.sock "$KCLI_KEY" 16 12 4 20 dense 0.05   # 16 连接×12 输入 + 4 线程×20 中断
+./stress.py $DATA/cli.sock "$KCLI_KEY" 12 12 1 25 0.0 3.0      # 稀疏中断 ⇒ 压抢占/挂起/恢复
+./stress.py $DATA/cli.sock "$KCLI_KEY" 1 1 0 0 resident 0.0    # 驻留子全链路（mock 见 !resident 标记）
+```
+
+## 读结果
+
+| 指标 | 含义 |
+|---|---|
+| `executed` / `rejected` | 任务执行数 / 被拒数（压力下应为 0） |
+| 峰值 `峰值_队列` / `峰值_待处理中断` | 采样到的最大排队深度 / 待处理中断数 |
+| `suspended` / `resumed` / `preempted` | 抢占三件套。**中断要放稀**才会打在高优先级任务上：密集中断会互相同级（L4 vs L4）不抢占，数值会很低 |
+| `max_suspend_depth` | 中断栈结构上界（= 中断级数 4） |
+
+## 已知坑
+
+- 探活用 **HEAD**（`internal/network/monitor.go`）：mock 必须实现 `do_HEAD`，否则 501 → 判不可达 → agent degraded → rollback 循环（会 reset 合并目录）。
+- CLI 协议**每条连接串行**处理（`handleChat` 阻塞到本次响应产出）：单连接狂发只会排成一条线，造不出队列压力 —— 必须多连接。
+- 认证：连接后先发 `/auth <key>`，否则一切命令返回 `unauthorized`。
+- `mockllm.py` 的 `!resident` / `!notify` 标记用来让 mock 回工具调用，从而在真内核里驱动 `resident_agents` / `notify_parent`。
diff --git a/scripts/kernel-stress/kcli.py b/scripts/kernel-stress/kcli.py
new file mode 100644
index 0000000..39d7287
--- /dev/null
+++ b/scripts/kernel-stress/kcli.py
@@ -0,0 +1,81 @@
+import json, os, socket, sys, threading, time
+
+SOCK = sys.argv[1]
+
+KEY = os.environ.get("KCLI_KEY", "")
+
+def connect(auth=True):
+    s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    s.settimeout(20)
+    s.connect(SOCK)
+    if auth and KEY:
+        s.sendall(("/auth %s\n" % KEY).encode())
+        buf = b""
+        while b"\n" not in buf:
+            c = s.recv(65536)
+            if not c:
+                break
+            buf += c
+        if b"authenticated" not in buf:
+            raise RuntimeError("auth failed: %r" % buf[:120])
+    return s
+
+def kernel_stats(sock=None, timeout=15):
+    """连一条新连接查 /kernel，返回解析后的 dict。"""
+    s = sock or connect()
+    try:
+        s.sendall(b"/kernel\n")
+        buf = b""
+        deadline = time.time() + timeout
+        while time.time() < deadline:
+            chunk = s.recv(65536)
+            if not chunk:
+                break
+            buf += chunk
+            for line in buf.split(b"\n"):
+                if not line.strip():
+                    continue
+                try:
+                    obj = json.loads(line.decode("utf-8", "replace"))
+                except Exception:
+                    continue
+                if obj.get("type") == "response" and obj.get("content", "").lstrip().startswith("{"):
+                    return json.loads(obj["content"])
+        raise RuntimeError("no /kernel response")
+    finally:
+        if sock is None:
+            s.close()
+
+def blast(n_queued, n_interrupt, tag):
+    """一条连接狂发：排队输入与中断交错。响应在后台线程里丢弃，避免缓冲阻塞。"""
+    s = connect()
+    threading.Thread(target=lambda: drain(s), daemon=True).start()
+    sent = 0
+    for i in range(max(n_queued, n_interrupt)):
+        if i < n_queued:
+            s.sendall(("queued-%s-%d\n" % (tag, i)).encode())
+            sent += 1
+        if i < n_interrupt:
+            s.sendall(("/interrupt intr-%s-%d\n" % (tag, i)).encode())
+            sent += 1
+    return sent
+
+def drain(s):
+    try:
+        while s.recv(65536):
+            pass
+    except Exception:
+        pass
+
+if __name__ == "__main__":
+    cmd = sys.argv[2] if len(sys.argv) > 2 else "stats"
+    if cmd == "stats":
+        st = kernel_stats()
+        if len(sys.argv) > 3 and sys.argv[3] == "full":
+            print(json.dumps(st, ensure_ascii=False))
+        else:
+            sched = st.get("scheduler") or st
+            print(json.dumps(sched, ensure_ascii=False))
+    elif cmd == "blast":
+        n = blast(int(sys.argv[3]), int(sys.argv[4]), sys.argv[5] if len(sys.argv) > 5 else "b")
+        print("sent=%d" % n)
diff --git a/scripts/kernel-stress/launch.sh b/scripts/kernel-stress/launch.sh
new file mode 100755
index 0000000..4c4c9bf
--- /dev/null
+++ b/scripts/kernel-stress/launch.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+# 在**私有 netns** 里启动：mock LLM + 完整内核实例。
+# 私有 netns 的意义：生产实例占着 *:8080 / *:9890 / *:9876，插件设了 SO_REUSEADDR，
+# 同机再起一个实例会在 127.0.0.1 上与之并存绑定（第一次实测就这么抢到了 9890 约 1 分钟）。
+# netns 里只有 lo，结构上不可能碰到生产端口。
+set -e
+DATA=/var/tmp/gotmp/kstress
+DRIVE=/var/tmp/gotmp/kstress-drive
+DELAY=${MOCK_DELAY_MS:-300}
+CHUNKS=${MOCK_CHUNKS:-8}
+unshare -n bash -c "
+ip link set lo up
+nohup env MOCK_DELAY_MS=$DELAY MOCK_CHUNKS=$CHUNKS python3 $DRIVE/mockllm.py > $DATA/mock.log 2>&1 &
+echo \$! > $DATA/mock.pid
+nohup $1 -data $DATA -webui 127.0.0.1:18080 > $DATA/boot.log 2>&1 &
+echo \$! > $DATA/pid
+"
diff --git a/scripts/kernel-stress/mockllm.py b/scripts/kernel-stress/mockllm.py
new file mode 100644
index 0000000..ce52eb1
--- /dev/null
+++ b/scripts/kernel-stress/mockllm.py
@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+"""最小 OpenAI 兼容 mock：可控延迟 + SSE 分块 + 可选工具调用。
+
+用途：给压力测试一个**快且可控**的 LLM —— 没有它，无外网的 netns 里每条输入
+都要走 provider 重试（≈2 分钟/条），既慢又压不出调度器行为。
+"""
+import json, os, time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+DELAY_MS = int(os.environ.get("MOCK_DELAY_MS", "300"))
+CHUNKS = int(os.environ.get("MOCK_CHUNKS", "8"))   # SSE 分块数：越多，流式段越长（可被中断的窗口越大）
+
+
+class H(BaseHTTPRequestHandler):
+    protocol_version = "HTTP/1.1"
+
+    def log_message(self, *a):
+        pass
+
+    def _json(self, obj, code=200):
+        b = json.dumps(obj).encode()
+        self.send_response(code)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(b)))
+        self.end_headers()
+        self.wfile.write(b)
+
+    def do_HEAD(self):
+        # 内核探活用 HEAD（见 internal/network/monitor.go CheckOnce）：
+        # 不实现它 → BaseHTTPRequestHandler 回 501 → 判为不可达 → agent degraded/rollback。
+        self.send_response(200)
+        self.send_header("Content-Length", "0")
+        self.end_headers()
+
+    def do_GET(self):
+        self._json({"object": "list", "data": [{"id": "mock", "object": "model"}]})
+
+    def do_POST(self):
+        n = int(self.headers.get("Content-Length", "0"))
+        body = json.loads(self.rfile.read(n) or b"{}")
+        msgs = body.get("messages") or []
+        text = ""
+        for m in reversed(msgs):
+            if m.get("role") == "user":
+                c = m.get("content")
+                text = c if isinstance(c, str) else json.dumps(c, ensure_ascii=False)
+                break
+
+        # 标记 !resident ⇒ 回一个工具调用，用于在**真实内核**里驱动驻留子工具链。
+        if "!resident" in text and not any(m.get("role") == "tool" for m in msgs):
+            tc = {"id": "call_mock_1", "type": "function",
+                  "function": {"name": "resident_agents",
+                               "arguments": json.dumps({"action": "create", "id": "r1",
+                                                        "task_prompt": "驻留子任务：统计一下 !notify",
+                                                        "input_chs": "cli"}, ensure_ascii=False)}}
+            return self._respond(body, content=None, tool_calls=[tc])
+        if "!notify" in text and not any(m.get("role") == "tool" for m in msgs):
+            tc = {"id": "call_mock_2", "type": "function",
+                  "function": {"name": "notify_parent",
+                               "arguments": json.dumps({"text": "mock 汇报：子已完成统计"}, ensure_ascii=False)}}
+            return self._respond(body, content=None, tool_calls=[tc])
+        return self._respond(body, content="mock-ok:" + text[:40])
+
+    def _respond(self, body, content=None, tool_calls=None):
+        if body.get("stream"):
+            self.send_response(200)
+            self.send_header("Content-Type", "text/event-stream")
+            self.send_header("Cache-Control", "no-cache")
+            self.send_header("Transfer-Encoding", "chunked")
+            self.end_headers()
+            def emit(delta):
+                data = json.dumps({"id": "mock", "object": "chat.completion.chunk",
+                                   "model": "mock", "choices": [{"index": 0, "delta": delta}]})
+                self._chunk(("data: " + data + "\n\n").encode())
+            if tool_calls:
+                emit({"role": "assistant", "tool_calls": tool_calls})
+                time.sleep(DELAY_MS / 1000.0)
+            if content:
+                emit({"role": "assistant", "content": content[: max(1, len(content) // CHUNKS)]})
+                per = max(1, len(content) // CHUNKS)
+                for i in range(per, len(content), per):
+                    time.sleep(DELAY_MS / 1000.0 / CHUNKS)
+                    emit({"content": content[i:i + per]})
+            self._chunk(b"data: [DONE]\n\n")
+            self._chunk(b"")
+            return
+        msg = {"role": "assistant", "content": content}
+        if tool_calls:
+            msg["tool_calls"] = tool_calls
+            msg["content"] = None
+        time.sleep(DELAY_MS / 1000.0)
+        self._json({"id": "mock", "object": "chat.completion", "model": "mock",
+                    "choices": [{"index": 0, "message": msg, "finish_reason": "stop"}],
+                    "usage": {"prompt_tokens": 10, "completion_tokens": 10, "total_tokens": 20}})
+
+    def _chunk(self, b):
+        self.wfile.write(("%x\r\n" % len(b)).encode() + b + b"\r\n")
+
+
+if __name__ == "__main__":
+    port = int(os.environ.get("MOCK_PORT", "9099"))
+    ThreadingHTTPServer(("127.0.0.1", port), H).serve_forever()
diff --git a/scripts/kernel-stress/stress.py b/scripts/kernel-stress/stress.py
new file mode 100644
index 0000000..35b2d07
--- /dev/null
+++ b/scripts/kernel-stress/stress.py
@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+"""二进制级调度器压力：多并发连接轰炸排队输入 + L4 中断，并采样峰值。
+
+用法: stress.py <sock> <key> <conns> <inputs_per_conn> <interrupt_threads> <interrupts_each> [resident]
+"""
+import json, os, socket, sys, threading, time
+
+SOCK, KEY = sys.argv[1], sys.argv[2]
+NC, NI, IT, IE = (int(x) for x in sys.argv[3:7])
+MODE = sys.argv[7] if len(sys.argv) > 7 else ""
+GAP = float(sys.argv[8]) if len(sys.argv) > 8 else 0.05
+
+lock = threading.Lock()
+sent = 0
+done = 0
+errs = []
+lat = []
+peak = {"queue": 0, "pending": 0, "stack": 0}
+
+
+def connect():
+    s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    s.settimeout(60)
+    s.connect(SOCK)
+    s.sendall(("/auth %s\n" % KEY).encode())
+    buf = b""
+    while b"\n" not in buf:
+        buf += s.recv(65536)
+    return s
+
+
+def chat(s, text):
+    """发一条输入并等到终止帧（response/error）。"""
+    s.sendall((text + "\n").encode())
+    buf = b""
+    while True:
+        c = s.recv(65536)
+        if not c:
+            raise RuntimeError("closed")
+        buf += c
+        while b"\n" in buf:
+            ln, buf = buf.split(b"\n", 1)
+            if not ln.strip():
+                continue
+            try:
+                obj = json.loads(ln.decode("utf-8", "replace"))
+            except Exception:
+                continue
+            if obj.get("type") in ("response", "error"):
+                return obj
+
+
+def worker(idx, n):
+    global sent, done
+    try:
+        s = connect()
+    except Exception as e:
+        with lock:
+            errs.append("connect: %r" % e)
+        return
+    for i in range(n):
+        t = time.time()
+        try:
+            obj = chat(s, "w%d-%d%s" % (idx, i, " !resident" if MODE == "resident" else ""))
+        except Exception as e:
+            with lock:
+                errs.append("chat: %r" % e)
+            return
+        dt = time.time() - t
+        with lock:
+            sent += 1
+            if obj.get("type") == "response":
+                done += 1
+                lat.append(dt)
+    s.close()
+
+
+def interrupter(idx, n, gap):
+    for i in range(n):
+        try:
+            s = connect()
+            s.sendall(("/interrupt stress-%d-%d\n" % (idx, i)).encode())
+            time.sleep(gap)
+            s.close()
+        except Exception as e:
+            with lock:
+                errs.append("intr: %r" % e)
+            return
+
+
+def sampler(stop, dur):
+    while not stop.is_set():
+        try:
+            s = connect()
+            s.sendall(b"/kernel\n")
+            buf = b""
+            while b"\n" not in buf:
+                buf += s.recv(65536)
+            obj = json.loads(buf.split(b"\n")[0].decode())
+            sc = json.loads(obj["content"]).get("scheduler", {})
+            with lock:
+                peak["queue"] = max(peak["queue"], sc.get("ready_queue_depth", 0))
+                peak["pending"] = max(peak["pending"], sc.get("pending_interrupts", 0))
+                peak["stack"] = max(peak["stack"], sc.get("suspend_stack", 0))
+            s.close()
+        except Exception:
+            time.sleep(0.05)
+        time.sleep(0.1)
+
+
+def kstat():
+    s = connect()
+    s.sendall(b"/kernel\n")
+    buf = b""
+    while b"\n" not in buf:
+        buf += s.recv(65536)
+    return json.loads(json.loads(buf.split(b"\n")[0].decode())["content"])
+
+
+t0 = time.time()
+stop = threading.Event()
+threads = [threading.Thread(target=worker, args=(i, NI)) for i in range(NC)]
+threads += [threading.Thread(target=interrupter, args=(i, IE, GAP)) for i in range(IT)]
+smp = threading.Thread(target=sampler, args=(stop, 0), daemon=True)
+smp.start()
+for t in threads:
+    t.start()
+for t in threads:
+    t.join()
+stop.set()
+time.sleep(0.4)
+st = kstat()
+with lock:
+    lat.sort()
+    p = lambda q: (lat[int(len(lat) * q)] if lat else 0)
+    print(json.dumps({
+        "耗时s": round(time.time() - t0, 2),
+        "并发连接": NC, "每连接输入": NI, "中断线程": IT, "每线程中断": IE,
+        "完成": done, "错误": len(errs),
+        "延迟_p50": round(p(0.5), 2), "延迟_p95": round(p(0.95), 2), "延迟_max": round(lat[-1] if lat else 0, 2),
+        "峰值_队列": peak["queue"], "峰值_待处理中断": peak["pending"], "峰值_中断栈": peak["stack"],
+        "scheduler": st.get("scheduler"), "goroutines": st.get("runtime", {}).get("goroutines"),
+        "错误样本": errs[:3],
+    }, ensure_ascii=False))