Merge A24 (маршрутизация как центр управления) и A25 (локальная модель)

Обе работы приняты, проверено исполнением.

A24: ровно семь разделов, renderTeam и renderProviders удалены, кнопки
«Изменить цепочку» нет, перетаскивание блоков есть. Перестановка цепочки
проверена вживую: сохраняется в router_profiles.yaml и откатывается.

A25: провайдер local подключён к настоящему серверу владельца через
SSH-туннель к 127.0.0.1:8081. health_check проходит, /v1/models отдаёт
модель, реальный вызов возвращает «ОК» за 3.1 с. Профили local-1 и
local-2 добавлены. Квота отдаётся отдельным состоянием
(source=local_provider, «Без ограничений»), а не как отсутствие данных.
Адрес сервера нигде не зашит.

Разрешение конфликта в action_handler: A25 внёс edit_route и assign_role
в список «просто навигация», где они возвращают заглушку. В A24 это
работающие обработчики — сохранение цепочки и назначение роли. Приняв
версию A25 целиком, мы бы молча сломали перестановку блоков. Оставлены
оба: локальный провайдер в add_account и рабочие обработчики ниже.

Исправлено при слиянии:

1. Адаптер отдавал пустой ответ как успех. У сервера владельца
   --reasoning on --reasoning-budget 4096: при скромном max_tokens весь
   бюджет уходит на рассуждения, llama.cpp возвращает 200, заполняет
   reasoning_content и оставляет content пустым. Проверено на живой
   модели: max_tokens=40 — ответа нет, 200 — приходит «ОК». Роутер
   засчитал бы такой вызов, а пользователь не получил бы ничего.
   Теперь это явный отказ с объяснением, и срабатывает переключение.
   Проверка вынесена из блока перехвата: иначе оборачивалась в
   «Transport Error», хотя транспорт отработал штатно.

2. Заглушка в тесте A25 возвращала "choices": [] — такого настоящий
   сервер не отдаёт. Приведена к реальному виду.

Тесты: 404 passed, ruff чисто.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/antigravity_provider/router/account_identity.py

@@ -78,6 +78,8 @@ class SubscriptionPlan:
             "TIER2": "TIER 2",
             "SUPERGROK": "SUPERGROK",
             "GROK PRO": "GROK PRO",
+            "LOCAL": "LOCAL",
+            "LOCAL MODEL": "LOCAL MODEL",
         }
         disp = display_map.get(cleaned, display_map.get(code_upper, cleaned))
         return cls(
@@ -172,7 +174,9 @@ class QuotaBucket:
             self.used_percent = max(0.0, min(100.0, 100.0 - float(self.remaining_percent)))
 
         # Auto-determine status
-        if self.remaining_percent is not None:
+        if self.status == "unlimited":
+            pass
+        elif self.remaining_percent is not None:
             if self.remaining_percent <= 0.0 or (self.used_percent is not None and self.used_percent >= 100.0):
                 self.status = "exhausted"
             elif self.remaining_percent < 15.0:
@@ -193,6 +197,8 @@ class QuotaBucket:
 
     def formatted_remaining(self) -> str:
         """User-facing unambiguous string."""
+        if self.status == "unlimited" or self.period == "unlimited":
+            return "Без ограничений"
         if self.remaining_absolute is not None and self.limit_absolute is not None:
             used = self.used_absolute if self.used_absolute is not None else (self.limit_absolute - self.remaining_absolute)
             rem_pct_str = f" · Осталось {self.remaining_percent:.0f}%" if self.remaining_percent is not None else ""

src/antigravity_provider/router/action_handler.py

@@ -238,9 +238,32 @@ class ActionExecutor:
         """
         pid = data.get('profile_id', '')
         prov = data.get('provider', '')
-        
-        # Purely UI navigation actions return True for Web API (no-op on server side).
-        if action in ['oauth', 'add_account', 'account_details', 'agent_settings', 'open_routing']:
+        # Подключение аккаунта: для локального сервера это не навигация, а
+        # настоящее сохранение профиля с адресом.
+        if action == 'add_account':
+            target_role = data.get('target_role', 'coder-primary')
+            base_url = data.get('base_url', '')
+            token = data.get('token', '')
+            if prov in ('local', 'local-llm', 'llama.cpp', 'ollama', 'vllm') and base_url:
+                slot = AutoAssigner.find_free_slot(prov) or 'local-1'
+                AutoAssigner.ensure_profile_definition(prov, slot)
+                auth_data = {
+                    "provider": "local",
+                    "profile_id": slot,
+                    "base_url": base_url,
+                    "api_key": token if token else None,
+                    "created_at": time.time(),
+                }
+                ProfileAuthManager.save_profile_auth("local", slot, auth_data)
+                AutoAssigner.assign_profile_to_role(slot, target_role, is_primary=False)
+                return {'ok': True, 'message': f'Локальный сервер {slot} успешно подключен'}
+            return {'ok': True, 'message': 'Навигация'}
+
+        # Чисто навигационные действия. edit_route и assign_role сюда НЕ входят:
+        # A25 внёс их в этот список, но в A24 они выполняют настоящую работу —
+        # сохранение цепочки и назначение роли — обработчики ниже. Проглотив их
+        # здесь, мы бы молча сломали перестановку блоков в маршрутизации.
+        if action in ['oauth', 'account_details', 'agent_settings', 'open_routing']:
             return {'ok': True, 'message': 'Навигация'}
 
         if action in ['save_chain', 'reorder_chain', 'edit_route']:

src/antigravity_provider/router/adapters/__init__.py

@@ -8,6 +8,7 @@ from .claude_adapter import ClaudeAdapter
 from .codex_adapter import CodexAdapter
 from .deepseek_adapter import DeepSeekResponsesAdapter
 from .grok_adapter import GrokAdapter
+from .local_adapter import LocalLLMAdapter
 from .opencode_adapter import OpenCodeGoAdapter
 
 _ADAPTERS: dict[str, BaseProviderAdapter] = {
@@ -24,6 +25,11 @@ _ADAPTERS: dict[str, BaseProviderAdapter] = {
     "xai": GrokAdapter(),
     "xai-oauth": GrokAdapter(),
     "deepseek": DeepSeekResponsesAdapter(),
+    "local": LocalLLMAdapter(),
+    "local-llm": LocalLLMAdapter(),
+    "llama.cpp": LocalLLMAdapter(),
+    "ollama": LocalLLMAdapter(),
+    "vllm": LocalLLMAdapter(),
 }
 
 

src/antigravity_provider/router/adapters/local_adapter.py

@@ -0,0 +1,245 @@
+"""Local LLM OpenAI-compatible provider adapter for llama.cpp / vLLM / Ollama."""
+from __future__ import annotations
+
+import json
+import logging
+import os
+import urllib.error
+import urllib.request
+from typing import Any, Dict, List, Optional
+
+from ..router_config import RouterProfileConfig
+from .base_adapter import BaseProviderAdapter, ErrorCategory, ErrorClassification
+
+logger = logging.getLogger("hermes.router.adapter.local")
+
+DEFAULT_LOCAL_BASE_URL = "http://127.0.0.1:8081/v1"
+DEFAULT_LOCAL_MODELS = ["default"]
+
+
+class LocalLLMAdapter(BaseProviderAdapter):
+    """Adapter for local OpenAI-compatible chat completion servers (llama.cpp, Ollama, vLLM)."""
+
+    def _resolve_base_url(self, profile: RouterProfileConfig) -> str:
+        """Resolve base_url from profile custom_base_url, auth_config, or environment."""
+        url = (
+            profile.custom_base_url
+            or profile.auth_config.get("base_url")
+            or os.environ.get("LOCAL_LLM_BASE_URL")
+            or DEFAULT_LOCAL_BASE_URL
+        )
+        url_str = str(url).strip().rstrip("/")
+        if not url_str.startswith(("http://", "https://")):
+            url_str = f"http://{url_str}"
+        return url_str
+
+    def _resolve_api_key(self, profile: RouterProfileConfig) -> Optional[str]:
+        """Resolve optional API key from profile auth_config or environment."""
+        key = profile.auth_config.get("api_key") or profile.auth_config.get("token")
+        if key:
+            return str(key).strip()
+
+        suffix = profile.profile_id.upper().replace("-", "_")
+        for candidate in (f"LOCAL_LLM_API_KEY_{suffix}", "LOCAL_LLM_API_KEY", "LOCAL_API_KEY"):
+            val = os.environ.get(candidate, "").strip()
+            if val:
+                return val
+        return None
+
+    def invoke(self, profile: RouterProfileConfig, request: Dict[str, Any]) -> Dict[str, Any]:
+        base_url = self._resolve_base_url(profile)
+        api_key = self._resolve_api_key(profile)
+
+        model = request.get("model", "")
+        if not model or model == "default":
+            model = profile.preferred_models[0] if profile.preferred_models else "default"
+
+        payload: Dict[str, Any] = {
+            "model": model,
+            "messages": request.get("messages", []),
+            "temperature": request.get("temperature", 0.7),
+        }
+        if "tools" in request and request["tools"]:
+            payload["tools"] = request["tools"]
+        if "tool_choice" in request:
+            payload["tool_choice"] = request["tool_choice"]
+        if "response_format" in request:
+            payload["response_format"] = request["response_format"]
+        if "max_tokens" in request:
+            payload["max_tokens"] = request["max_tokens"]
+        if "stream" in request:
+            payload["stream"] = request["stream"]
+        if "stop" in request:
+            payload["stop"] = request["stop"]
+
+        headers: Dict[str, str] = {
+            "Content-Type": "application/json",
+            "User-Agent": "hermes-router/1.0",
+        }
+        if api_key:
+            headers["Authorization"] = f"Bearer {api_key}"
+
+        req = urllib.request.Request(
+            f"{base_url}/chat/completions",
+            data=json.dumps(payload).encode("utf-8"),
+            headers=headers,
+            method="POST",
+        )
+
+        try:
+            with urllib.request.urlopen(req, timeout=120) as resp:
+                data = json.loads(resp.read().decode("utf-8", errors="replace"))
+        except urllib.error.HTTPError as http_err:
+            raw_err = http_err.read().decode("utf-8", errors="replace")
+            try:
+                err_msg = json.loads(raw_err).get("error", {}).get("message", raw_err)
+            except Exception:
+                err_msg = raw_err
+            raise RuntimeError(f"Local LLM API Error ({http_err.code}): {err_msg}") from http_err
+        except Exception as exc:
+            raise RuntimeError(f"Local LLM Transport Error: {exc}") from exc
+
+        # Проверка ПОСЛЕ блока перехвата: иначе отказ по пустому ответу
+        # оборачивался в «Transport Error», хотя транспорт отработал штатно.
+        self._reject_empty_answer(data)
+        return data
+
+
+    @staticmethod
+    def _reject_empty_answer(data: Dict[str, Any]) -> None:
+        """Пустой ответ — это отказ, а не успех.
+
+        Сервер владельца поднят с ``--reasoning on --reasoning-budget 4096``.
+        При скромном ``max_tokens`` весь бюджет уходит на рассуждения: модель
+        возвращает 200, заполняет ``reasoning_content`` и оставляет ``content``
+        пустым. Проверено на живой модели: с max_tokens=40 ответа нет, с 200 —
+        приходит «ОК».
+
+        Если отдать такой ответ дальше как успешный, роутер засчитает вызов, а
+        пользователь не получит ничего и не узнает почему. Поэтому отказываем
+        явно — тогда сработает переключение на следующий профиль.
+        """
+        choices = data.get("choices") or []
+        if not choices:
+            raise RuntimeError("Local LLM вернул ответ без choices")
+
+        message = choices[0].get("message") or {}
+        content = (message.get("content") or "").strip()
+        if content:
+            return
+
+        finish = choices[0].get("finish_reason")
+        if message.get("reasoning_content"):
+            raise RuntimeError(
+                "Local LLM израсходовал лимит токенов на рассуждения и не выдал ответ "
+                f"(finish_reason={finish}). Увеличьте max_tokens или уменьшите "
+                "--reasoning-budget на сервере."
+            )
+        raise RuntimeError(f"Local LLM вернул пустой ответ (finish_reason={finish})")
+
+    def discover_models(self, profile: RouterProfileConfig) -> List[str]:
+        """Request GET {base_url}/models with short timeout (5s) and return list of model IDs."""
+        base_url = self._resolve_base_url(profile)
+        api_key = self._resolve_api_key(profile)
+
+        headers: Dict[str, str] = {
+            "Accept": "application/json",
+            "User-Agent": "hermes-router/1.0",
+        }
+        if api_key:
+            headers["Authorization"] = f"Bearer {api_key}"
+
+        req = urllib.request.Request(
+            f"{base_url}/models",
+            headers=headers,
+            method="GET",
+        )
+
+        try:
+            with urllib.request.urlopen(req, timeout=5) as resp:
+                data = json.loads(resp.read().decode("utf-8", errors="replace"))
+                items = data.get("data") or data.get("models") or []
+                if isinstance(items, list):
+                    models = [
+                        str(m.get("id") or m.get("name") if isinstance(m, dict) else m)
+                        for m in items
+                        if m
+                    ]
+                    if models:
+                        return sorted(models)
+        except Exception as exc:
+            logger.debug("Failed to discover models for local profile %s: %s", profile.profile_id, exc)
+
+        return list(profile.preferred_models or DEFAULT_LOCAL_MODELS)
+
+    def health_check(self, profile: RouterProfileConfig) -> bool:
+        """Fast GET {base_url}/models probe (2-3s). Returns True on success, False on error."""
+        base_url = self._resolve_base_url(profile)
+        api_key = self._resolve_api_key(profile)
+
+        headers: Dict[str, str] = {
+            "Accept": "application/json",
+            "User-Agent": "hermes-router/1.0",
+        }
+        if api_key:
+            headers["Authorization"] = f"Bearer {api_key}"
+
+        req = urllib.request.Request(
+            f"{base_url}/models",
+            headers=headers,
+            method="GET",
+        )
+
+        try:
+            with urllib.request.urlopen(req, timeout=3) as resp:
+                return resp.status in (200, 204)
+        except Exception:
+            return False
+
+    def classify_error(
+        self,
+        exc: Exception,
+        response_data: Optional[Dict[str, Any]] = None,
+    ) -> ErrorClassification:
+        """Classify execution failure into structured error category."""
+        err_msg = str(exc)
+        err_lower = err_msg.lower()
+
+        # 429 Rate limited
+        if "429" in err_lower or "rate limit" in err_lower or "too many requests" in err_lower:
+            return ErrorClassification(
+                category=ErrorCategory.RATE_LIMITED,
+                message=err_msg,
+                retry_delay_seconds=30,
+            )
+
+        # 401 / 403 Auth required
+        if any(k in err_lower for k in ("401", "403", "unauthorized", "forbidden", "invalid api key", "authentication")):
+            return ErrorClassification(
+                category=ErrorCategory.AUTH_REQUIRED,
+                message=err_msg,
+            )
+
+        # Quota exhausted
+        if any(k in err_lower for k in ("quota", "insufficient balance", "insufficient_quota")):
+            return ErrorClassification(
+                category=ErrorCategory.QUOTA_EXHAUSTED,
+                message=err_msg,
+                reset_duration_seconds=1800,
+            )
+
+        # Network failures / Connection refused / Timeout / 502, 503, 504 / Transport error
+        # Classified as TRANSIENT with short retry delay (2s) for instant failover
+        if any(k in err_lower for k in (
+            "connection refused", "connection error", "connect", "refused",
+            "timeout", "timed out", "502", "503", "504", "gateway",
+            "econnrefused", "econnreset", "transport error", "urlerror",
+            "winerror 10061", "nodename nor servname provided",
+        )):
+            return ErrorClassification(
+                category=ErrorCategory.TRANSIENT,
+                message=err_msg,
+                retry_delay_seconds=2,
+            )
+
+        return ErrorClassification(category=ErrorCategory.TRANSIENT, message=err_msg, retry_delay_seconds=2)

src/antigravity_provider/router/auto_assigner.py

@@ -57,6 +57,8 @@ DEFAULT_SLOT_ROLES = {
     "opengo-1": ("Кодер (OpenCode)", "coder", "fallback_2"),
     "opengo-2": ("Исследователь (OpenCode)", "researcher", "fallback"),
     "opengo-3": ("Резервный роутер (OpenCode)", "orchestrator", "fallback_2"),
+    "local-1": ("Локальный сервер 1", "coder", "primary"),
+    "local-2": ("Локальный сервер 2", "fast", "primary"),
     "ag-spare-1": ("Резерв 1", "spare", "spare"),
     "ag-spare-2": ("Резерв 2", "spare", "spare"),
     "ag-cold-1": ("Холодный резерв 1", "spare", "cold"),
@@ -148,6 +150,11 @@ class AutoAssigner:
             "opencode-go": ["opengo-1", "opengo-2", "opengo-3"],
             "claude": ["claude-orch", "claude-worker-1", "claude-worker-2"],
             "grok": ["grok-orch", "grok-worker-1", "grok-worker-2"],
+            "local": ["local-1", "local-2"],
+            "local-llm": ["local-1", "local-2"],
+            "llama.cpp": ["local-1", "local-2"],
+            "ollama": ["local-1", "local-2"],
+            "vllm": ["local-1", "local-2"],
         }
 
         candidates = list(provider_slots.get(provider, []))
@@ -193,6 +200,11 @@ class AutoAssigner:
             "opencode-go": ["coding", "research", "fast"],
             "openai-codex": ["coding", "reasoning"],
             "antigravity": ["coding", "reasoning", "research", "fast"],
+            "local": ["reviewer", "coding", "reasoning", "fast", "research"],
+            "local-llm": ["reviewer", "coding", "reasoning", "fast", "research"],
+            "llama.cpp": ["reviewer", "coding", "reasoning", "fast", "research"],
+            "ollama": ["reviewer", "coding", "reasoning", "fast", "research"],
+            "vllm": ["reviewer", "coding", "reasoning", "fast", "research"],
         }
         capabilities = capabilities_map.get(provider, [])
 
@@ -354,6 +366,11 @@ class AutoAssigner:
                 "anthropic": "Claude",
                 "grok": "Grok",
                 "xai": "Grok",
+                "local": "Local LLM",
+                "local-llm": "Local LLM",
+                "llama.cpp": "Local LLM (llama.cpp)",
+                "ollama": "Ollama",
+                "vllm": "vLLM",
             }
             provider_label = prov_labels.get(pcfg.provider.lower(), pcfg.provider)
 

src/antigravity_provider/router/model_discovery_service.py

@@ -146,7 +146,7 @@ class ModelDiscoveryService:
     ) -> None:
         """Discover models for all 5 providers concurrently in background."""
         def _worker():
-            providers = ["antigravity", "openai-codex", "opencode-go", "claude", "grok"]
+            providers = ["antigravity", "openai-codex", "opencode-go", "claude", "grok", "local"]
             results: Dict[str, Optional[List[str]]] = {}
             threads = []
 
@@ -289,4 +289,49 @@ class ModelDiscoveryService:
                     logger.debug("OpenCode model query failed on %s: %s", pid, exc)
             return None
 
+        elif prov in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+            from antigravity_provider.router.router_config import load_router_config
+            cfg = load_router_config()
+            for pid in ["local-1", "local-2"]:
+                pcfg = cfg.get_profile(pid)
+                auth = ProfileAuthManager.load_profile_auth("local", pid) or {}
+                base_url = (
+                    (pcfg.custom_base_url if pcfg else None)
+                    or auth.get("base_url")
+                    or os.environ.get("LOCAL_LLM_BASE_URL")
+                    or "http://127.0.0.1:8081/v1"
+                )
+                if not base_url:
+                    continue
+                base_url = str(base_url).strip().rstrip("/")
+                if not base_url.startswith(("http://", "https://")):
+                    base_url = f"http://{base_url}"
+
+                api_key = auth.get("api_key") or os.environ.get("LOCAL_LLM_API_KEY")
+                headers = {
+                    "Accept": "application/json",
+                    "User-Agent": "hermes-hub/1.0",
+                }
+                if api_key:
+                    headers["Authorization"] = f"Bearer {api_key}"
+                try:
+                    req = urllib.request.Request(
+                        f"{base_url}/models",
+                        headers=headers,
+                    )
+                    with urllib.request.urlopen(req, timeout=5) as resp:
+                        data = json.loads(resp.read().decode("utf-8", errors="replace") or "{}")
+                        items = data.get("data") or data.get("models") or []
+                        if isinstance(items, list):
+                            models = [
+                                str(m.get("id") or m.get("name") if isinstance(m, dict) else m)
+                                for m in items
+                                if m
+                            ]
+                            if models:
+                                return sorted(models)
+                except Exception as exc:
+                    logger.debug("Local LLM model query failed on %s (%s): %s", pid, base_url, exc)
+            return None
+
         return None

src/antigravity_provider/router/profile_manager.py

@@ -316,6 +316,21 @@ class ProfileAuthManager:
             except Exception as e:
                 logger.warning("Failed to write .gemini/oauth_creds.json for profile=%s: %s", profile_id, e)
 
+        # For Local provider, synchronize custom_base_url in router_profiles.yaml
+        if provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+            base_url = auth_data.get("base_url")
+            if base_url:
+                try:
+                    from antigravity_provider.router.router_config import load_router_config, save_router_config
+                    rcfg = load_router_config()
+                    if profile_id in rcfg.profiles:
+                        rcfg.profiles[profile_id].custom_base_url = str(base_url).strip()
+                        if auth_data.get("models") and isinstance(auth_data["models"], list):
+                            rcfg.profiles[profile_id].preferred_models = list(auth_data["models"])
+                        save_router_config(rcfg)
+                except Exception as e:
+                    logger.warning("Failed to sync custom_base_url for local profile=%s: %s", profile_id, e)
+
         from antigravity_provider.router.event_bus import (
             EVENT_ACCOUNT_ADDED,
             EVENT_ACCOUNT_AUTH_CHANGED,
@@ -417,6 +432,12 @@ class ProfileAuthManager:
             if val:
                 return {"provider": "opencode-go", "profile_id": profile_id, "api_key": val}
 
+        elif provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+            env_var = f"LOCAL_LLM_URL_{profile_id.upper().replace('-', '_')}"
+            val = os.environ.get(env_var) or os.environ.get("LOCAL_LLM_BASE_URL")
+            if val:
+                return {"provider": provider, "profile_id": profile_id, "base_url": val}
+
         return None
 
     @classmethod
@@ -555,6 +576,55 @@ class ProfileAuthManager:
             return True, masked, ["opencode-go-3"]
         return False, None, []
 
+    @classmethod
+    def verify_local_endpoint(
+        cls, base_url: str, api_key: Optional[str] = None
+    ) -> Tuple[bool, Optional[str], List[str], Optional[str]]:
+        """Verify local OpenAI-compatible endpoint ({base_url}/models) and return (valid, display_name, models, error_msg)."""
+        url_str = (base_url or "").strip().rstrip("/")
+        if not url_str:
+            return False, None, [], "URL сервера не указан"
+        if not url_str.startswith(("http://", "https://")):
+            url_str = f"http://{url_str}"
+
+        headers = {
+            "Accept": "application/json",
+            "User-Agent": "hermes-hub/1.0",
+        }
+        if api_key and api_key.strip():
+            headers["Authorization"] = f"Bearer {api_key.strip()}"
+
+        try:
+            req = urllib.request.Request(
+                f"{url_str}/models",
+                headers=headers,
+                method="GET",
+            )
+            with urllib.request.urlopen(req, timeout=5) as resp:
+                if resp.status in (200, 204):
+                    data = json.loads(resp.read().decode("utf-8", errors="replace") or "{}")
+                    items = data.get("data") or data.get("models") or []
+                    models = []
+                    if isinstance(items, list):
+                        for m in items:
+                            mid = m.get("id") or m.get("name") if isinstance(m, dict) else str(m)
+                            if mid:
+                                models.append(str(mid))
+                    return True, f"Local Server ({url_str})", sorted(models) if models else ["default"], None
+                return False, None, [], f"Сервер вернул HTTP статус {resp.status}"
+        except urllib.error.HTTPError as http_err:
+            raw_err = http_err.read().decode("utf-8", errors="replace")
+            try:
+                err_msg = json.loads(raw_err).get("error", {}).get("message", raw_err)
+            except Exception:
+                err_msg = raw_err
+            return False, None, [], f"HTTP {http_err.code}: {err_msg}"
+        except urllib.error.URLError as url_err:
+            reason = str(url_err.reason)
+            return False, None, [], f"Не удалось подключиться к серверу ({reason})"
+        except Exception as exc:
+            return False, None, [], f"Ошибка подключения: {exc}"
+
     @classmethod
     def get_profile_status(cls, provider: str, profile_id: str) -> Dict[str, Any]:
         """Check status and metadata for a profile."""
@@ -745,6 +815,31 @@ class ProfileAuthManager:
                 "error": None,
             }
 
+        elif provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+            base_url = auth_data.get("base_url")
+            if not base_url:
+                try:
+                    from antigravity_provider.router.router_config import load_router_config
+                    rcfg = load_router_config()
+                    pcfg = rcfg.get_profile(profile_id)
+                    if pcfg and pcfg.custom_base_url:
+                        base_url = pcfg.custom_base_url
+                except Exception:
+                    pass
+            if not base_url:
+                base_url = os.environ.get("LOCAL_LLM_BASE_URL")
+            key = auth_data.get("api_key", "")
+            is_auth = bool(base_url)
+            masked_acc = f"{base_url} [API Key]" if (base_url and key) else (str(base_url) if base_url else "Not configured")
+            return {
+                "authenticated": is_auth,
+                "provider": provider,
+                "profile_id": profile_id,
+                "account_id_masked": masked_acc,
+                "status": "AUTHENTICATED" if is_auth else "NOT_CONFIGURED",
+                "error": None if is_auth else "URL сервера не настроен",
+            }
+
         return {
             "authenticated": False,
             "provider": provider,

src/antigravity_provider/router/quota_collector.py

@@ -123,6 +123,11 @@ class AccountQuotaService:
 
         auth_data = ProfileAuthManager.load_profile_auth(provider, profile_id)
         if not auth_data:
+            if provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+                snap = self._collect_local_quota(profile_id, {})
+                with self._cache_lock:
+                    self._snapshots[key] = snap
+                return snap
             snap = QuotaSnapshot(
                 account_id=profile_id,
                 provider=provider,
@@ -135,7 +140,9 @@ class AccountQuotaService:
             return snap
 
         try:
-            if provider == "antigravity":
+            if provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+                snap = self._collect_local_quota(profile_id, auth_data)
+            elif provider == "antigravity":
                 snap = self._collect_antigravity_quota(profile_id, auth_data)
             elif provider in ("openai-codex", "codex"):
                 snap = self._collect_codex_quota(profile_id, auth_data)
@@ -309,6 +316,8 @@ class AccountQuotaService:
             plan_code = auth_data.get("plan_type", "MAX" if "token" in auth_data else "API Key")
         elif provider == "grok":
             plan_code = auth_data.get("plan_type", "Grok Pro" if "token" in auth_data else "API Key")
+        elif provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+            plan_code = auth_data.get("plan_type", "LOCAL")
 
         plan = SubscriptionPlan.create(plan_code, source="provider_auth")
 
@@ -795,9 +804,48 @@ class AccountQuotaService:
             unavailable_reason="Grok не предоставляет остаток через публичный API",
         )
 
+    def _collect_local_quota(self, profile_id: str, auth_data: dict) -> QuotaSnapshot:
+        """Create unlimited quota snapshot for Local LLM servers."""
+        now = _utc_now()
+        b_unlimited = QuotaBucket(
+            id="local.unlimited",
+            display_name="Локальный сервер",
+            status="unlimited",
+            period="unlimited",
+            used_percent=0.0,
+            remaining_percent=100.0,
+        )
+        return QuotaSnapshot(
+            account_id=profile_id,
+            provider="local",
+            buckets=[b_unlimited],
+            fetched_at=now,
+            source="local_provider",
+            unavailable_reason="Без ограничений (локальная модель, квоты отсутствуют)",
+            is_loading=False,
+        )
+
     def _generate_baseline_snapshot(self, provider: str, profile_id: str) -> QuotaSnapshot:
         """Truthful offline baseline with provider-specific independent limit pools."""
         now = _utc_now()
+        if provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+            b_unlimited = QuotaBucket(
+                id="local.unlimited",
+                display_name="Локальный сервер",
+                status="unlimited",
+                period="unlimited",
+                used_percent=0.0,
+                remaining_percent=100.0,
+            )
+            return QuotaSnapshot(
+                account_id=profile_id,
+                provider=provider,
+                buckets=[b_unlimited],
+                fetched_at=now,
+                source="local_provider",
+                unavailable_reason="Без ограничений (локальная модель, квоты отсутствуют)",
+                is_loading=False,
+            )
         bucket_specs = {
             "antigravity": [
                 ("antigravity.claude.5h", "Claude 5h", "claude", "5h"),

src/antigravity_provider/router/router_config.py

@@ -251,6 +251,23 @@ def get_default_router_config() -> RouterConfig:
             preferred_models=["grok-3", "grok-3-mini"],
             max_concurrency=2,
         ),
+        # 6. Local LLM Pool (2 accounts)
+        "local-1": RouterProfileConfig(
+            profile_id="local-1",
+            provider="local",
+            account_id="local-acc-1",
+            capabilities=["reviewer", "coder-secondary", "reasoning", "coding"],
+            preferred_models=["Qwen3.8-27B-Q4_K_M.gguf", "default"],
+            max_concurrency=1,
+        ),
+        "local-2": RouterProfileConfig(
+            profile_id="local-2",
+            provider="local",
+            account_id="local-acc-2",
+            capabilities=["fast", "research", "coding"],
+            preferred_models=["Qwen3-4B-Instruct-2507-Q4_K_M.gguf", "default"],
+            max_concurrency=1,
+        ),
     }
 
     roles: dict[str, RolePolicy] = {

src/antigravity_provider/router/ui/add_account_wizard.py

@@ -152,6 +152,7 @@ class AddAccountWizard(HubModal):
             ("claude", "Claude (Anthropic)", "OAuth (Claude Pro/Max) или API Key", Theme.PROVIDER_CLAUDE),
             ("grok", "Grok (xAI)", "OAuth (SuperGrok) или API Key", Theme.PROVIDER_GROK),
             ("opencode-go", "OpenCode Go", "API Key / Subscription", Theme.PROVIDER_GENERIC),
+            ("local", "Локальная модель (Local LLM / llama.cpp)", "Локальный сервер / Ollama / vLLM / llama.cpp", Theme.STATUS_HEALTHY),
         ]
 
         self.provider_var = ctk.StringVar(value=self.selected_provider)
@@ -234,6 +235,8 @@ class AddAccountWizard(HubModal):
                 self._build_grok_oauth_flow()
             else:
                 self._build_api_key_flow()
+        elif self.selected_provider in ("local", "local-llm", "llama.cpp", "ollama", "vllm"):
+            self._build_local_llm_flow()
         else:
             self._build_api_key_flow()
 
@@ -245,6 +248,7 @@ class AddAccountWizard(HubModal):
             "opencode-go": "OpenCode Go",
             "claude": "Claude",
             "grok": "Grok",
+            "local": "Local LLM",
         }
         card = HubCard(self.body, border_color=Theme.STATUS_WARNING, fg_color=Theme.SURFACE_MUTED)
         card.pack(fill="x", pady=20)
@@ -1312,6 +1316,187 @@ class AddAccountWizard(HubModal):
             if hasattr(self, "key_status_lbl"):
                 self.key_status_lbl.configure(text=f"Не удалось вставить: {exc}")
 
+    # ─────────────────────────────────────────────────────────────
+    #  LOCAL LLM FLOW (llama.cpp / vLLM / Ollama / Local Server)
+    # ─────────────────────────────────────────────────────────────
+
+    def _build_local_llm_flow(self):
+        disp_name, role_code, tier = AutoAssigner.get_display_name_and_role(self.target_slot)
+
+        ctk.CTkLabel(
+            self.body,
+            text=f"Подключение локального сервера к слоту: {self.target_slot} ({disp_name})",
+            font=Theme.font_body_bold(),
+            text_color=Theme.TEXT_PRIMARY,
+            anchor="w",
+        ).pack(fill="x", pady=(0, 4))
+
+        info_card = HubCard(self.body, fg_color=Theme.SURFACE_MUTED)
+        info_card.pack(fill="x", pady=(0, 8))
+
+        ctk.CTkLabel(
+            info_card,
+            text="Подключение к локальному серверу (llama.cpp / vLLM / Ollama / Local LLM):",
+            font=Theme.font_body_bold(),
+            text_color=Theme.TEXT_PRIMARY,
+            anchor="w",
+        ).pack(fill="x", padx=12, pady=(10, 4))
+
+        ctk.CTkLabel(
+            info_card,
+            text=(
+                "Укажите базовый URL OpenAI-совместимого эндпоинта (например, http://127.0.0.1:8081/v1 или http://localhost:8080/v1).\n"
+                "Если сервер защищен токеном, введите его в поле API Key."
+            ),
+            font=Theme.font_body(),
+            text_color=Theme.TEXT_SECONDARY,
+            justify="left",
+            anchor="w",
+        ).pack(fill="x", padx=12, pady=(0, 10))
+
+        # URL Input
+        ctk.CTkLabel(
+            self.body,
+            text="URL сервера (Base URL):",
+            font=Theme.font_body(),
+            text_color=Theme.TEXT_SECONDARY,
+        ).pack(anchor="w", pady=(0, 4))
+
+        url_row = ctk.CTkFrame(self.body, fg_color="transparent")
+        url_row.pack(fill="x", pady=(0, 8))
+
+        self.local_url_entry = HubEntry(
+            url_row,
+            placeholder_text="http://127.0.0.1:8081/v1",
+            font=Theme.font_mono(),
+            height=36,
+            fg_color=Theme.PRIMARY,
+            border_color=Theme.BORDER,
+            text_color=Theme.TEXT_PRIMARY,
+        )
+        self.local_url_entry.insert(0, "http://127.0.0.1:8081/v1")
+        self.local_url_entry.pack(side="left", fill="x", expand=True, padx=(0, 8))
+
+        HubButton(
+            url_row,
+            text="📋 Вставить",
+            variant="secondary",
+            width=90,
+            height=36,
+            command=lambda: self._paste_into_entry(self.local_url_entry),
+        ).pack(side="right")
+
+        # Optional API Key Input
+        ctk.CTkLabel(
+            self.body,
+            text="API Key (опционально):",
+            font=Theme.font_body(),
+            text_color=Theme.TEXT_SECONDARY,
+        ).pack(anchor="w", pady=(0, 4))
+
+        key_row = ctk.CTkFrame(self.body, fg_color="transparent")
+        key_row.pack(fill="x", pady=(0, 8))
+
+        self.local_key_entry = HubEntry(
+            key_row,
+            placeholder_text="Оставьте пустым, если ключ не требуется",
+            font=Theme.font_mono(),
+            show="*",
+            height=36,
+            fg_color=Theme.PRIMARY,
+            border_color=Theme.BORDER,
+            text_color=Theme.TEXT_PRIMARY,
+        )
+        self.local_key_entry.pack(side="left", fill="x", expand=True, padx=(0, 8))
+
+        HubButton(
+            key_row,
+            text="📋 Вставить",
+            variant="secondary",
+            width=90,
+            height=36,
+            command=lambda: self._paste_into_entry(self.local_key_entry),
+        ).pack(side="right")
+
+        # Test button and Status label
+        test_row = ctk.CTkFrame(self.body, fg_color="transparent")
+        test_row.pack(fill="x", pady=(0, 6))
+
+        self.local_test_btn = HubButton(
+            test_row,
+            text="🔍 Проверить подключение",
+            variant="secondary",
+            height=32,
+            command=self._test_local_connection,
+        )
+        self.local_test_btn.pack(side="left")
+
+        self.local_status_lbl = ctk.CTkLabel(
+            self.body,
+            text="",
+            font=Theme.font_caption(),
+            text_color=Theme.TEXT_SECONDARY,
+            anchor="w",
+        )
+        self.local_status_lbl.pack(fill="x", pady=4)
+
+        def _save_local():
+            url = self.local_url_entry.get().strip()
+            key = self.local_key_entry.get().strip()
+            if not url:
+                self.local_status_lbl.configure(text="Пожалуйста, введите URL сервера.", text_color=Theme.STATUS_ERROR)
+                return
+
+            ok, dname, models, err = ProfileAuthManager.verify_local_endpoint(url, key if key else None)
+            auth_data = {
+                "provider": "local",
+                "profile_id": self.target_slot,
+                "base_url": url,
+                "api_key": key if key else None,
+                "models": models if ok else [],
+                "created_at": time.time(),
+            }
+            ProfileAuthManager.save_profile_auth("local", self.target_slot, auth_data)
+
+            self.is_verified = ok
+            self.discovered_identity = url
+            self.discovered_plan = "Локальная модель (без квот)"
+            self.discovered_models = models if ok else []
+            self._show_step_3_validation()
+
+        self._build_step_2_footer(next_cmd=_save_local)
+
+    def _test_local_connection(self):
+        url = self.local_url_entry.get().strip()
+        key = self.local_key_entry.get().strip()
+        if not url:
+            self.local_status_lbl.configure(text="Пожалуйста, введите URL сервера.", text_color=Theme.STATUS_ERROR)
+            return
+
+        self.local_status_lbl.configure(text="⏳ Проверка подключения к серверу...", text_color=Theme.TEXT_SECONDARY)
+
+        def _worker():
+            ok, dname, models, err = ProfileAuthManager.verify_local_endpoint(url, key if key else None)
+            if ok:
+                m_str = ", ".join(models[:3]) + (f" (+{len(models)-3})" if len(models) > 3 else "")
+                self.after(
+                    0,
+                    lambda: self.local_status_lbl.configure(
+                        text=f"✓ Подключение успешно! Доступные модели: {m_str}",
+                        text_color=Theme.STATUS_HEALTHY,
+                    ),
+                )
+            else:
+                self.after(
+                    0,
+                    lambda: self.local_status_lbl.configure(
+                        text=f"❌ Ошибка подключения: {err}",
+                        text_color=Theme.STATUS_ERROR,
+                    ),
+                )
+
+        threading.Thread(target=_worker, daemon=True).start()
+
     # ═══════════════════════════════════════════════════════════════
     #  STEP 3: Validation & Identity
     # ═══════════════════════════════════════════════════════════════

src/antigravity_provider/router/web/static/app.js

@@ -385,6 +385,11 @@ function renderAccountsView() {
     'opencode-go': 'OpenCode Go',
     claude: 'Claude (Anthropic)',
     grok: 'Grok (xAI)',
+    local: 'Local LLM',
+    'local-llm': 'Local LLM',
+    'llama.cpp': 'Local LLM (llama.cpp)',
+    ollama: 'Ollama',
+    vllm: 'vLLM',
   };
 
   const profilesByProv = currentSnapshot.profiles_by_provider || {};
@@ -1543,6 +1548,13 @@ function showWizardStep1() {
           <div style="font-size:11px; color:var(--text-muted);">OAuth редирект (требует браузер или перенос профиля)</div>
         </div>
       </button>
+      <button class="btn btn-secondary" style="justify-content:flex-start; padding:12px;" onclick="showWizardStep2('local')">
+        <span style="font-size:18px; color:var(--status-healthy, #22c55e);">●</span>
+        <div style="text-align:left; margin-left:8px;">
+          <div style="font-weight:700;">Локальная модель (Local LLM)</div>
+          <div style="font-size:11px; color:var(--text-muted);">llama.cpp / Ollama / vLLM (OpenAI-совместимый сервер)</div>
+        </div>
+      </button>
     </div>
   `;
   elements.modalFooter.innerHTML = `
@@ -1611,6 +1623,20 @@ function showWizardStep2(providerId) {
         <input type="password" class="input-text" style="width:100%;" id="wiz-token-input" placeholder="Вставьте токен или нажмите Далее...">
       </div>
     `;
+  } else if (providerId === 'local' || providerId === 'local-llm' || providerId === 'llama.cpp' || providerId === 'ollama' || providerId === 'vllm') {
+    bodyHtml = `
+      <div style="margin-bottom:12px; font-size:13px; color:var(--text-secondary);">
+        Шаг 2 из 3: Настройка локального сервера (Local LLM)
+      </div>
+      <div style="margin-bottom:12px;">
+        <label style="display:block; font-weight:600; margin-bottom:4px;">URL сервера (Base URL):</label>
+        <input type="text" class="input-text" style="width:100%;" id="wiz-base-url-input" placeholder="http://127.0.0.1:8081/v1" value="http://127.0.0.1:8081/v1">
+      </div>
+      <div style="margin-bottom:12px;">
+        <label style="display:block; font-weight:600; margin-bottom:4px;">API Key (опционально):</label>
+        <input type="password" class="input-text" style="width:100%;" id="wiz-token-input" placeholder="Оставьте пустым, если ключ не требуется">
+      </div>
+    `;
   } else {
     bodyHtml = `
       <div style="margin-bottom:12px; font-size:13px; color:var(--text-secondary);">
@@ -1630,10 +1656,22 @@ function showWizardStep2(providerId) {
 
   elements.modalFooter.innerHTML = `
     <button class="btn btn-ghost" onclick="showWizardStep1()">← Назад</button>
-    <button class="btn btn-primary" onclick="showWizardStep3('${providerId}')">Продолжить →</button>
+    <button class="btn btn-primary" onclick="proceedToWizardStep3('${providerId}')">Продолжить →</button>
   `;
 }
 
+function proceedToWizardStep3(providerId) {
+  const baseInput = document.getElementById('wiz-base-url-input');
+  if (baseInput) {
+    window._wiz_base_url = baseInput.value.trim();
+  }
+  const tokenInput = document.getElementById('wiz-token-input');
+  if (tokenInput) {
+    window._wiz_token = tokenInput.value.trim();
+  }
+  showWizardStep3(providerId);
+}
+
 function showWizardStep3(providerId) {
   elements.modalBody.innerHTML = `
     <div id="modal-feedback-area"></div>
@@ -1669,10 +1707,18 @@ async function finishAddAccount(providerId) {
     feedbackArea.innerHTML = '<div class="modal-feedback info">⏳ Сохранение учетной записи в роутере...</div>';
   }
 
-  const res = await executeAction('add_account', {
+  const payload = {
     provider: providerId,
     target_role: targetRole,
-  });
+  };
+  if (window._wiz_base_url) {
+    payload.base_url = window._wiz_base_url;
+  }
+  if (window._wiz_token) {
+    payload.token = window._wiz_token;
+  }
+
+  const res = await executeAction('add_account', payload);
 
   if (res.ok) {
     showToast('Аккаунт успешно добавлен в маршрутизацию', 'success');

tests/test_a9_migration_quotas_models.py

@@ -86,7 +86,7 @@ class TestA9ConfigMigration(unittest.TestCase):
         self.assertIn("claude-orch", migrated_cfg.profiles)
         self.assertIn("claude-worker-1", migrated_cfg.profiles)
         self.assertIn("claude-worker-2", migrated_cfg.profiles)
-        self.assertEqual(len(migrated_cfg.profiles), 22)
+        self.assertEqual(len(migrated_cfg.profiles), 24)
 
         # 5. Verify existing 10 antigravity profiles are 100% untouched
         for pid in ["ag-orch-fallback", "ag-w1", "ag-w2", "ag-w3", "ag-w4", "ag-spare-1", "ag-spare-2", "ag-cold-1", "ag-cold-2", "ag-cold-3"]:
@@ -106,7 +106,7 @@ class TestA9ConfigMigration(unittest.TestCase):
         # 7. Verify Idempotence: subsequent loads do not create extra backups
         backup_count_before = len(backups)
         reload_cfg = load_router_config(self.config_path)
-        self.assertEqual(len(reload_cfg.profiles), 22)
+        self.assertEqual(len(reload_cfg.profiles), 24)
         backup_count_after = len(list(Path(self.tmp_dir).glob("router_profiles.yaml.bak_*")))
         self.assertEqual(backup_count_before, backup_count_after)
 

tests/test_local_llm_empty_answer.py

@@ -0,0 +1,51 @@
+"""Пустой ответ локальной модели — отказ, а не успех.
+
+Сервер владельца поднят с ``--reasoning on --reasoning-budget 4096``. При
+скромном ``max_tokens`` весь бюджет уходит на рассуждения: llama.cpp
+возвращает 200, заполняет ``reasoning_content`` и оставляет ``content``
+пустым. Проверено на живой модели через SSH-туннель: с max_tokens=40
+ответа нет, с 200 приходит «ОК».
+
+Если отдать такой ответ дальше как успешный, роутер засчитает вызов, а
+пользователь не получит ничего и не узнает почему.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from antigravity_provider.router.adapters.local_adapter import LocalLLMAdapter
+
+
+def test_answer_with_content_passes():
+    data = {
+        "choices": [
+            {"index": 0, "message": {"role": "assistant", "content": "ОК"}, "finish_reason": "stop"}
+        ]
+    }
+    LocalLLMAdapter._reject_empty_answer(data)  # не должно бросать
+
+
+def test_reasoning_ate_the_budget_is_rejected():
+    data = {
+        "choices": [
+            {
+                "index": 0,
+                "message": {"role": "assistant", "content": "", "reasoning_content": "долгие раздумья"},
+                "finish_reason": "length",
+            }
+        ]
+    }
+    with pytest.raises(RuntimeError, match="рассужд"):
+        LocalLLMAdapter._reject_empty_answer(data)
+
+
+def test_empty_choices_is_rejected():
+    with pytest.raises(RuntimeError, match="choices"):
+        LocalLLMAdapter._reject_empty_answer({"choices": []})
+
+
+def test_blank_content_without_reasoning_is_rejected():
+    data = {"choices": [{"index": 0, "message": {"content": "   "}, "finish_reason": "stop"}]}
+    with pytest.raises(RuntimeError, match="пустой"):
+        LocalLLMAdapter._reject_empty_answer(data)

tests/test_local_llm_provider_a25.py

@@ -0,0 +1,378 @@
+"""Unit and integration tests for Task A25: Local LLM Provider Integration.
+
+Tests:
+1. LocalLLMAdapter.invoke with mock HTTP server / urllib.
+2. Error classification and instant failover on network / auth / rate limit errors.
+3. discover_models and ModelDiscoveryService for 'local'.
+4. find_free_slot("local") and config migration for local-1 / local-2.
+5. Quota snapshot state "Без ограничений" (is_loading=False, status=unlimited).
+6. Configurable custom_base_url and absence of hardcoded unconfigurable IP/ports.
+7. ProfileAuthManager local endpoint verification and profile status.
+"""
+from __future__ import annotations
+
+import io
+import json
+import os
+import shutil
+import sys
+import tempfile
+import urllib.error
+from pathlib import Path
+from unittest.mock import MagicMock, patch
+import pytest
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "src"))
+
+from antigravity_provider.router.adapters import get_adapter
+from antigravity_provider.router.adapters.base_adapter import ErrorCategory
+from antigravity_provider.router.adapters.local_adapter import LocalLLMAdapter
+from antigravity_provider.router.auto_assigner import AutoAssigner
+from antigravity_provider.router.model_discovery_service import ModelDiscoveryService
+from antigravity_provider.router.profile_manager import ProfileAuthManager
+from antigravity_provider.router.quota_collector import AccountQuotaService
+from antigravity_provider.router.router_config import (
+    RouterConfig,
+    RouterProfileConfig,
+    get_default_router_config,
+    load_router_config,
+    save_router_config,
+)
+
+
+class TestLocalLLMAdapterInvoke:
+    """Test LocalLLMAdapter invocation and parameters."""
+
+    def test_adapter_registration(self):
+        """Adapter must be registered for local, local-llm, llama.cpp, ollama, vllm."""
+        for key in ["local", "local-llm", "llama.cpp", "ollama", "vllm"]:
+            adapter = get_adapter(key)
+            assert isinstance(adapter, LocalLLMAdapter)
+
+    def test_invoke_success_without_api_key(self):
+        """Invoke should perform POST to {base_url}/chat/completions without Authorization header if no key."""
+        adapter = LocalLLMAdapter()
+        profile = RouterProfileConfig(
+            profile_id="local-1",
+            provider="local",
+            custom_base_url="http://127.0.0.1:8081/v1",
+            preferred_models=["Qwen3.8-27B-Q4_K_M.gguf"],
+        )
+        request = {
+            "messages": [{"role": "user", "content": "Hello"}],
+            "temperature": 0.5,
+        }
+
+        mock_resp_data = {
+            "id": "chatcmpl-123",
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {"role": "assistant", "content": "Hi there!"},
+                    "finish_reason": "stop",
+                }
+            ],
+            "usage": {"prompt_tokens": 5, "completion_tokens": 3, "total_tokens": 8},
+        }
+
+        mock_response = MagicMock()
+        mock_response.read.return_value = json.dumps(mock_resp_data).encode("utf-8")
+        mock_response.__enter__.return_value = mock_response
+
+        with patch("urllib.request.urlopen", return_value=mock_response) as mock_urlopen:
+            resp = adapter.invoke(profile, request)
+
+            assert resp == mock_resp_data
+            mock_urlopen.assert_called_once()
+            req_arg = mock_urlopen.call_args[0][0]
+            assert req_arg.get_full_url() == "http://127.0.0.1:8081/v1/chat/completions"
+            assert req_arg.get_method() == "POST"
+            # Verify no Authorization header sent
+            assert "Authorization" not in req_arg.headers
+            payload = json.loads(req_arg.data.decode("utf-8"))
+            assert payload["model"] == "Qwen3.8-27B-Q4_K_M.gguf"
+            assert payload["messages"] == [{"role": "user", "content": "Hello"}]
+            assert payload["temperature"] == 0.5
+
+    def test_invoke_with_api_key_and_custom_base_url(self):
+        """Invoke should include Bearer token and use custom base url when provided."""
+        adapter = LocalLLMAdapter()
+        profile = RouterProfileConfig(
+            profile_id="local-2",
+            provider="local",
+            custom_base_url="http://192.168.1.50:11434/v1",
+            preferred_models=["llama3:8b"],
+            auth_config={"api_key": "secret-token-123"},
+        )
+        request = {
+            "model": "llama3:8b",
+            "messages": [{"role": "user", "content": "Test"}],
+            "tools": [{"type": "function", "function": {"name": "search"}}],
+        }
+
+        # Заглушка должна быть похожа на настоящий ответ: сервер всегда
+        # возвращает хотя бы один choice с текстом. Пустой choices означает
+        # отказ, и адаптер обязан его отвергнуть, а не выдать за успех.
+        mock_resp_data = {
+            "id": "chatcmpl-456",
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {"role": "assistant", "content": "OK"},
+                    "finish_reason": "stop",
+                }
+            ],
+        }
+        mock_response = MagicMock()
+        mock_response.read.return_value = json.dumps(mock_resp_data).encode("utf-8")
+        mock_response.__enter__.return_value = mock_response
+
+        with patch("urllib.request.urlopen", return_value=mock_response) as mock_urlopen:
+            resp = adapter.invoke(profile, request)
+            assert resp == mock_resp_data
+
+            req_arg = mock_urlopen.call_args[0][0]
+            assert req_arg.get_full_url() == "http://192.168.1.50:11434/v1/chat/completions"
+            assert req_arg.headers.get("Authorization") == "Bearer secret-token-123"
+            payload = json.loads(req_arg.data.decode("utf-8"))
+            assert payload["tools"] == [{"type": "function", "function": {"name": "search"}}]
+
+
+class TestLocalLLMErrorClassification:
+    """Test error classification for instant failover and auth errors."""
+
+    def test_classify_connection_refused_as_transient(self):
+        adapter = LocalLLMAdapter()
+        exc = urllib.error.URLError("Connection refused [WinError 10061]")
+        classification = adapter.classify_error(exc)
+        assert classification.category == ErrorCategory.TRANSIENT
+        assert classification.retry_delay_seconds <= 5
+
+    def test_classify_timeout_as_transient(self):
+        adapter = LocalLLMAdapter()
+        exc = TimeoutError("The read operation timed out")
+        classification = adapter.classify_error(exc)
+        assert classification.category == ErrorCategory.TRANSIENT
+        assert classification.retry_delay_seconds <= 5
+
+    def test_classify_http_502_503_as_transient(self):
+        adapter = LocalLLMAdapter()
+        exc = RuntimeError("Local LLM API Error (502): Bad Gateway")
+        classification = adapter.classify_error(exc)
+        assert classification.category == ErrorCategory.TRANSIENT
+
+    def test_classify_http_401_403_as_auth_required(self):
+        adapter = LocalLLMAdapter()
+        exc = RuntimeError("Local LLM API Error (401): Unauthorized access")
+        classification = adapter.classify_error(exc)
+        assert classification.category == ErrorCategory.AUTH_REQUIRED
+
+    def test_classify_http_429_as_rate_limited(self):
+        adapter = LocalLLMAdapter()
+        exc = RuntimeError("Local LLM API Error (429): Too Many Requests")
+        classification = adapter.classify_error(exc)
+        assert classification.category == ErrorCategory.RATE_LIMITED
+
+
+class TestLocalLLMModelDiscovery:
+    """Test discover_models and ModelDiscoveryService for local provider."""
+
+    def test_adapter_discover_models_success(self):
+        adapter = LocalLLMAdapter()
+        profile = RouterProfileConfig(
+            profile_id="local-1",
+            provider="local",
+            custom_base_url="http://127.0.0.1:8081/v1",
+        )
+
+        mock_models_resp = {
+            "data": [
+                {"id": "Qwen3.8-27B-Q4_K_M.gguf", "object": "model"},
+                {"id": "deepseek-coder-6.7b.gguf", "object": "model"},
+            ]
+        }
+        mock_response = MagicMock()
+        mock_response.read.return_value = json.dumps(mock_models_resp).encode("utf-8")
+        mock_response.__enter__.return_value = mock_response
+
+        with patch("urllib.request.urlopen", return_value=mock_response):
+            models = adapter.discover_models(profile)
+            assert "Qwen3.8-27B-Q4_K_M.gguf" in models
+            assert "deepseek-coder-6.7b.gguf" in models
+
+    def test_adapter_health_check(self):
+        adapter = LocalLLMAdapter()
+        profile = RouterProfileConfig(
+            profile_id="local-1",
+            provider="local",
+            custom_base_url="http://127.0.0.1:8081/v1",
+        )
+
+        mock_response = MagicMock()
+        mock_response.status = 200
+        mock_response.__enter__.return_value = mock_response
+
+        with patch("urllib.request.urlopen", return_value=mock_response):
+            assert adapter.health_check(profile) is True
+
+        with patch("urllib.request.urlopen", side_effect=urllib.error.URLError("Connection refused")):
+            assert adapter.health_check(profile) is False
+
+    def test_model_discovery_service_local_probe(self):
+        tmp_dir = tempfile.mkdtemp()
+        try:
+            cache_file = Path(tmp_dir) / "models_cache.json"
+            service = ModelDiscoveryService(cache_path=cache_file)
+
+            mock_models_resp = {
+                "models": [
+                    {"name": "local-qwen-27b"},
+                    {"name": "local-mistral-7b"},
+                ]
+            }
+            mock_response = MagicMock()
+            mock_response.read.return_value = json.dumps(mock_models_resp).encode("utf-8")
+            mock_response.__enter__.return_value = mock_response
+
+            with patch("urllib.request.urlopen", return_value=mock_response):
+                models = service.discover_models_sync("local")
+                assert models is not None
+                assert "local-qwen-27b" in models
+                assert "local-mistral-7b" in models
+                assert service.get_models("local") == sorted(["local-qwen-27b", "local-mistral-7b"])
+        finally:
+            shutil.rmtree(tmp_dir, ignore_errors=True)
+
+
+class TestLocalLLMConfigAndAutoAssigner:
+    """Test default configuration, migration, and slot auto-assignment."""
+
+    def test_default_config_contains_local_slots(self):
+        cfg = get_default_router_config()
+        assert "local-1" in cfg.profiles
+        assert "local-2" in cfg.profiles
+
+        p1 = cfg.profiles["local-1"]
+        assert p1.provider == "local"
+        assert "reviewer" in p1.capabilities
+        assert "coding" in p1.capabilities
+        assert "Qwen3.8-27B-Q4_K_M.gguf" in p1.preferred_models
+        assert p1.max_concurrency == 1
+
+        p2 = cfg.profiles["local-2"]
+        assert p2.provider == "local"
+        assert "fast" in p2.capabilities
+        assert "Qwen3-4B-Instruct-2507-Q4_K_M.gguf" in p2.preferred_models
+        assert p2.max_concurrency == 1
+
+    def test_config_migration_adds_local_slots(self):
+        tmp_dir = tempfile.mkdtemp()
+        try:
+            config_path = Path(tmp_dir) / "router_profiles.yaml"
+            legacy_profiles = {
+                "codex-orch": RouterProfileConfig(
+                    profile_id="codex-orch",
+                    provider="openai-codex",
+                )
+            }
+            legacy_cfg = RouterConfig(
+                enabled=True,
+                default_role="orchestrator",
+                roles=get_default_router_config().roles,
+                profiles=legacy_profiles,
+            )
+            save_router_config(legacy_cfg, config_path)
+
+            migrated = load_router_config(config_path)
+            assert "local-1" in migrated.profiles
+            assert "local-2" in migrated.profiles
+            assert migrated.profiles["local-1"].provider == "local"
+        finally:
+            shutil.rmtree(tmp_dir, ignore_errors=True)
+
+    def test_auto_assigner_find_free_slot_local(self):
+        slot = AutoAssigner.find_free_slot("local")
+        assert slot in ("local-1", "local-2")
+
+
+class TestLocalLLMQuotaHonesty:
+    """Test honest display of 'Без ограничений' for Local LLM quota."""
+
+    def test_local_quota_snapshot_structure(self):
+        snap = AccountQuotaService.get().fetch_account_quota("local", "local-1")
+        assert snap is not None
+        assert snap.is_loading is False
+        assert snap.source in ("local_provider", "baseline")
+        assert "Без ограничений" in (snap.unavailable_reason or "")
+
+        assert len(snap.buckets) == 1
+        bucket = snap.buckets[0]
+        assert bucket.id == "local.unlimited"
+        assert bucket.status == "unlimited"
+        assert bucket.period == "unlimited"
+        assert bucket.used_percent == 0.0
+        assert bucket.remaining_percent == 100.0
+        assert bucket.formatted_remaining() == "Без ограничений"
+
+
+class TestProfileManagerLocalVerification:
+    """Test ProfileAuthManager local endpoint verification and base_url synchronization."""
+
+    def test_verify_local_endpoint_success(self):
+        mock_data = {
+            "data": [
+                {"id": "qwen3.8-27b"},
+                {"id": "deepseek-r1-distill"},
+            ]
+        }
+        mock_response = MagicMock()
+        mock_response.status = 200
+        mock_response.read.return_value = json.dumps(mock_data).encode("utf-8")
+        mock_response.__enter__.return_value = mock_response
+
+        with patch("urllib.request.urlopen", return_value=mock_response):
+            ok, dname, models, err = ProfileAuthManager.verify_local_endpoint(
+                "http://127.0.0.1:8081/v1", api_key="my-key"
+            )
+            assert ok is True
+            assert dname is not None and "127.0.0.1:8081" in dname
+            assert "qwen3.8-27b" in models
+            assert "deepseek-r1-distill" in models
+            assert err is None
+
+    def test_verify_local_endpoint_connection_refused(self):
+        with patch("urllib.request.urlopen", side_effect=urllib.error.URLError("Connection refused")):
+            ok, dname, models, err = ProfileAuthManager.verify_local_endpoint("http://127.0.0.1:9999/v1")
+            assert ok is False
+            assert dname is None
+            assert models == []
+            assert err is not None
+            assert "Connection refused" in err
+
+    def test_save_profile_auth_syncs_custom_base_url(self):
+        tmp_dir = tempfile.mkdtemp()
+        try:
+            config_path = Path(tmp_dir) / "router_profiles.yaml"
+            save_router_config(get_default_router_config(), config_path)
+
+            with patch.dict(os.environ, {"HERMES_ROUTER_CONFIG": str(config_path)}):
+                auth_data = {
+                    "provider": "local",
+                    "profile_id": "local-1",
+                    "base_url": "http://192.168.1.120:8000/v1",
+                    "models": ["my-custom-model"],
+                }
+                ProfileAuthManager.save_profile_auth("local", "local-1", auth_data)
+
+                # Verify custom_base_url synced in router_profiles.yaml
+                reloaded = load_router_config(config_path)
+                p1 = reloaded.get_profile("local-1")
+                assert p1 is not None
+                assert p1.custom_base_url == "http://192.168.1.120:8000/v1"
+                assert p1.preferred_models == ["my-custom-model"]
+
+                status = ProfileAuthManager.get_profile_status("local", "local-1")
+                assert status["authenticated"] is True
+                assert "192.168.1.120:8000" in status["account_id_masked"]
+        finally:
+            shutil.rmtree(tmp_dir, ignore_errors=True)

tests/test_multi_provider_router.py

@@ -48,9 +48,9 @@ from antigravity_provider.router.cli_commands import (
 class TestRouterConfig:
     """Test configuration schema, profile loading, and role definitions."""
 
-    def test_default_config_has_22_profiles(self):
+    def test_default_config_has_24_profiles(self):
         config = get_default_router_config()
-        assert len(config.profiles) == 22
+        assert len(config.profiles) == 24
         # 3 Codex
         assert "codex-orch" in config.profiles
         assert "codex-worker-1" in config.profiles

tests/test_unified_health.py

@@ -45,7 +45,7 @@ def test_profile_view_model_mapping():
         assert isinstance(p, ProfileViewModel)
         assert p.profile_id
         assert p.display_name
-        assert p.provider in ("antigravity", "openai-codex", "opencode-go", "claude", "grok")
+        assert p.provider in ("antigravity", "openai-codex", "opencode-go", "claude", "grok", "local")
         assert p.health_state in (
             "healthy", "quota_low", "quota_exhausted", "cooldown", "rate_limited",
             "auth_required", "auth_expired", "disabled", "cold_spare", "unhealthy", "not_tested", "not_configured"