1. 保存代理脚本

cat > ~/omlx_compress_proxy.py << 'EOF'
#!/usr/bin/env python3
import json, time, copy
from http.server import HTTPServer, BaseHTTPRequestHandler
from urllib.request import Request, urlopen
from urllib.error import HTTPError

OMLX_BASE = "http://127.0.0.1:8201"   # OMLX 新地址
PROXY_PORT = 8200                     # 接管 Hermes 原来的端口
COMPRESS_EVERY = 15
KEEP_RECENT = 1
SUMMARY_MAX_TOKENS = 400
SUMMARY_TEMP = 0.1

SUMMARY_SYSTEM = (
    "你是后台上下文压缩器。请用中文极度精简地总结以下对话，"
    "保留所有关键决策、代码、数据、未完成任务。只输出总结内容，"
    "不要任何解释、问候、格式标记。"
)

class Store:
    def __init__(self):
        self.sessions = {}
    def get(self, body):
        key = body.get("conversation_id") or body.get("session_id") or body.get("user") or "default"
        return self.sessions.get(key, []), key
    def set(self, key, messages):
        self.sessions[key] = messages
    def count(self, messages):
        return sum(1 for m in messages if m.get("role") in ("user", "assistant"))

store = Store()

def call_omlx(messages, max_tokens=None, temperature=None):
    payload = {"model": "Qwen3.5-35b-a3b-4bit-mlx", "messages": messages, "stream": False}
    if max_tokens: payload["max_tokens"] = max_tokens
    if temperature is not None: payload["temperature"] = temperature
    req = Request(f"{OMLX_BASE}/v1/chat/completions",
                  data=json.dumps(payload).encode(), headers={"Content-Type": "application/json"}, method="POST")
    return json.loads(urlopen(req, timeout=300).read().decode())

def compress(messages):
    sys_msgs = [m for m in messages if m.get("role") == "system"]
    non_sys = [m for m in messages if m.get("role") != "system"]
    if len(non_sys) <= KEEP_RECENT: return messages
    to_compress = non_sys[:-KEEP_RECENT]
    keep = non_sys[-KEEP_RECENT:]
    prompt = []
    if not sys_msgs: prompt.append({"role": "system", "content": SUMMARY_SYSTEM})
    prompt.append({"role": "user", "content": "\n\n".join(f"[{m['role']}] {m['content'][:2000]}" for m in to_compress)})
    summary = call_omlx(prompt, max_tokens=SUMMARY_MAX_TOKENS, temperature=SUMMARY_TEMP)["choices"][0]["message"]["content"].strip()
    new_msgs = list(sys_msgs)
    new_msgs.append({"role": "system", "content": f"[历史压缩] {summary}"})
    new_msgs.extend(keep)
    return new_msgs

class H(BaseHTTPRequestHandler):
    def log_message(self, *args): pass
    def do_POST(self):
        if self.path != "/v1/chat/completions": return self._proxy()
        body = json.loads(self.rfile.read(int(self.headers.get("Content-Length", 0))).decode())
        history, key = store.get(body)
        history = body.get("messages", [])
        if store.count(history) >= COMPRESS_EVERY:
            t0 = time.time()
            history = compress(history)
            print(f"[Compress] {key}: done in {time.time()-t0:.1f}s")
        store.set(key, history)
        omlx_body = copy.deepcopy(body)
        omlx_body["messages"] = history
        for k in ["conversation_id", "session_id", "user"]: omlx_body.pop(k, None)
        req = Request(f"{OMLX_BASE}/v1/chat/completions",
                      data=json.dumps(omlx_body).encode(), headers={"Content-Type": "application/json"}, method="POST")
        try:
            with urlopen(req, timeout=600) as r:
                data = r.read()
                self.send_response(r.status)
                for k, v in r.headers.items():
                    if k.lower() != "transfer-encoding": self.send_header(k, v)
                self.end_headers(); self.wfile.write(data)
                try:
                    msg = json.loads(data.decode())["choices"][0]["message"]
                    history.append(msg); store.set(key, history)
                except: pass
        except HTTPError as e:
            self.send_response(e.code); self.send_header("Content-Type", "application/json"); self.end_headers(); self.wfile.write(e.read())
    def _proxy(self):
        body = self.rfile.read(int(self.headers.get("Content-Length", 0))) if int(self.headers.get("Content-Length", 0)) else b""
        req = Request(f"{OMLX_BASE}{self.path}", data=body, headers={k:v for k,v in self.headers.items()}, method="POST" if body else "GET")
        try:
            with urlopen(req) as r: self.send_response(r.status); [self.send_header(k,v) for k,v in r.headers.items()]; self.end_headers(); self.wfile.write(r.read())
        except HTTPError as e: self.send_response(e.code); self.end_headers(); self.wfile.write(e.read())

if __name__ == "__main__":
    print(f"Proxy http://127.0.0.1:{PROXY_PORT}/v1 -> {OMLX_BASE}")
    print(f"Compress every {COMPRESS_EVERY} turns")
    HTTPServer(("127.0.0.1", PROXY_PORT), H).serve_forever()
EOF
chmod +x ~/omlx_compress_proxy.py

2. 调整端口（核心）

假设你之前 Hermes 连的是 127.0.0.1:8200：

# 先停掉 OMLX
pkill -f "omlx serve"

# OMLX 改绑到 8201（让出 8200 给代理）
omlx serve --port 8201 --model Qwen3.5-35b-a3b-4bit-mlx

3. 启动代理（占原端口 8200）

另开一个终端：

python3 ~/omlx_compress_proxy.py

日志应显示：

Proxy http://127.0.0.1:8200/v1 -> http://127.0.0.1:8201/v1
Compress every 15 turns

4. Hermes 配置（保持不变）

你在 Hermes SSH 终端界面里不要改任何端口，保持原来的 127.0.0.1:8200。因为代理已经占了 8200，Hermes 的请求会先经过代理，代理再转发给 OMLX 的 8201。

如果你之前在 Hermes 里配的是其他端口（比如 11434 或 5001），把上面脚本里的 PROXY_PORT 和 OMLX_BASE 对应改一下即可。

5. 验证压缩触发

正常聊天到第 15 轮，代理终端会打印：

[Compress] default: done in 1.8s

此时 KV Cache 从 15 轮瞬间降到 2 轮，内存压力解除。Hermes 前端看不到任何异常。

关键提醒

按这个顺序部署：先改 OMLX 端口 → 启动代理 → Hermes 保持原配置继续用。
16和24 需要改小总结轮数测试。测试机型32G M4。
请勿直接使用。