feat(llm): OCG key池——按/api/keys实时状态排序轮换(key6优先),空输出同key升pro再切下一key

deploy/profile-scripts/llm_client.py

@@ -49,6 +49,75 @@ def _load_ocg_key():
     return None
 
 
+def _load_key_pool():
+    """从 hermes config.yaml 解析全部 OCG key，按 AgentsMeeting /api/keys 的
+    实时状态排序（可用优先、用量低优先）。失败回退到单 key6。
+    返回 [(key_id, api_key), ...]，按优先级排序。"""
+    import re
+    fallback = [("key6", _load_ocg_key())] if _load_ocg_key() else []
+    try:
+        with open(_HERMES_CONFIG, encoding="utf-8") as f:
+            text = f.read()
+        # 解析 providers 下的 ocg 块: name + api_key
+        blocks = re.findall(r'^\s{2}(ocg[\w\-]*):\s*\n\s*api_key:\s*(\S+)', text, re.M)
+        if not blocks:
+            return fallback
+        # provider名 → key_id 映射（ocg-keyN→keyN, ocg-3→key3, ocg-hy3→key1, ocg-new→key2）
+        name2id = {"ocg-key6": "key6", "ocg-key5": "key5", "ocg-key4": "key4",
+                   "ocg-3": "key3", "ocg-hy3": "key1", "ocg-new": "key2"}
+        kv = {}
+        for name, key in blocks:
+            kid = name2id.get(name)
+            if kid and kid != "key7":
+                kv[kid] = key
+        if not kv:
+            return fallback
+        # 查询 AgentsMeeting /api/keys 实时状态
+        import json as _json
+        status = {}
+        try:
+            req = urllib.request.Request("http://127.0.0.1:5803/api/keys")
+            with urllib.request.build_opener(urllib.request.ProxyHandler({})).open(req, timeout=5) as r:
+                data = _json.loads(r.read().decode())
+            for k in data.get("keys", []):
+                m_ok = k.get("monthly", {}).get("status") == "ok"
+                r_ok = k.get("rolling", {}).get("status") == "ok"
+                usage = k.get("monthly", {}).get("usage_percent", 99)
+                status[k["key_id"]] = (m_ok and r_ok, usage)
+        except Exception:
+            # API 不可达时默认全可用、key6 优先
+            order = ["key6", "key5", "key4", "key2", "key1", "key3"]
+            return [(k, kv[k]) for k in order if k in kv]
+        # 可用且用量低优先；key6 在用量相同时优先（沿用现有配置）
+        scored = []
+        for kid, key in kv.items():
+            ok, usage = status.get(kid, (True, 50))
+            if ok:
+                scored.append((usage - (2 if kid == "key6" else 0), kid, key))
+        scored.sort()
+        pool = [(kid, key) for _, kid, key in scored]
+        return pool or fallback
+    except Exception as e:
+        print(f"  [LLM] key池加载失败，回退单key: {e}", flush=True)
+        return fallback
+
+
+_KEY_POOL = None
+_KEY_POOL_TS = 0
+
+
+def _get_key_pool():
+    """key 池缓存 5 分钟。"""
+    global _KEY_POOL, _KEY_POOL_TS
+    import time as _t
+    if not _KEY_POOL or (_t.time() - _KEY_POOL_TS) > 300:
+        _KEY_POOL = _load_key_pool()
+        _KEY_POOL_TS = _t.time()
+        if _KEY_POOL:
+            print(f"  [LLM] key池: {[k for k, _ in _KEY_POOL]}", flush=True)
+    return _KEY_POOL
+
+
 _OCG_KEY = _load_ocg_key()
 OCG_HEADERS = {
     "Content-Type": "application/json",
@@ -128,9 +197,16 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
         "max_tokens": max_tokens,
     }).encode()
 
+    # ── 构建通道列表：key池(OCG直连，按/api/keys实时状态排序) → gateway 兜底 ──
     channels = []
-    if OCG_HEADERS:
-        channels.append(("ocg", OCG_URL, OCG_HEADERS))
+    pool = _get_key_pool()
+    for kid, key in pool:
+        if key:
+            channels.append((f"ocg:{kid}", OCG_URL, {
+                "Content-Type": "application/json",
+                "Authorization": f"Bearer {key}",
+                "User-Agent": "curl/8.5.0",
+            }))
     channels.append(("gateway", GATEWAY, {
         "Content-Type": "application/json",
         "Authorization": GATEWAY_AUTH,
@@ -149,8 +225,8 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
                 elapsed = time.monotonic() - t0
                 if ok:
                     # ── 空输出升级：flash 对部分 prompt 稳定返回空（688617 实测），
-                    #    换 FALLBACK_MODEL(pro) 重试一次 ──
-                    if not result.strip() and model_name != FALLBACK_MODEL and OCG_HEADERS:
+                    #    先同 key 换 FALLBACK_MODEL(pro) 重试；仍空则让外循环切下一个 key ──
+                    if not result.strip() and model_name != FALLBACK_MODEL and ch_name.startswith("ocg"):
                         print(f"  [LLM] {ch_name} {model_name} 空输出({elapsed:.1f}s)，"
                               f"升级 {FALLBACK_MODEL} 重试...", flush=True)
                         esc_payload = json.dumps({
@@ -158,7 +234,7 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
                             "max_tokens": max_tokens,
                         }).encode()
                         try:
-                            ok2, result2 = _post(OCG_URL, OCG_HEADERS, esc_payload, timeout)
+                            ok2, result2 = _post(ch_url, ch_headers, esc_payload, timeout)
                             total_attempts += 1
                             if ok2 and result2.strip():
                                 print(f"  [LLM] 升级 {FALLBACK_MODEL} 成功, 输出{len(result2)}字", flush=True)
@@ -168,7 +244,7 @@ def call_llm(prompt, model=None, max_tokens=4096, timeout=150,
                                     "elapsed": time.monotonic() - t_start,
                                     "attempts": total_attempts, "channel": ch_name + "+esc",
                                 }
-                            print(f"  [LLM] 升级 {FALLBACK_MODEL} 仍空/失败", flush=True)
+                            print(f"  [LLM] {ch_name} 升级 {FALLBACK_MODEL} 仍空/失败", flush=True)
                         except Exception as e2:
                             print(f"  [LLM] 升级 {FALLBACK_MODEL} 异常: {str(e2)[:100]}", flush=True)
                     print(f"  [LLM] {ch_name} 尝试{attempt+1}/{retries+1} 成功, "