agy تنها ارائه‌دهندهٔ مدل؛ حذف بخش ارائه‌دهنده‌ها و افزودن زنجیرهٔ انتخاب مدل

- حذف کامل providerهای mock و openai_compatible: دامنه، آداپوترها، رجیستری، کانفیگ، seed و رابطٔ مدیریت
- بخش «ارائه‌دهنده‌ها» از کنسول حذف شد؛ ورود اکانت‌ها فقط از صفحهٔ اکانت‌های agy
- صفحهٔ مدل‌ها بازطراحی شد: ارائه‌دهنده → اکانت → مدل (کاتالوگ لحظه‌ای و زندهٔ همان اکانت از Google) → افورت استدلال؛ بدون هیچ ورود دستی نام مدل
- مدل‌ها ستون reasoning_effort گرفتند؛ آداپتور آن را به thinkingConfig (بودجهٔ ۱۰۲۴ تا ۳۲۷۶۸) ترجمه می‌کند و فقط به مدل‌های استدلالی می‌فرستد
- مهاجرت c2d3e4f5b6a7: افزودن reasoning_effort و حذف ردیف‌های providerهای قدیمی
- تست‌ها: ۲۴۶ تست بک‌اند و ۵۱ تست فرانت سبز؛ ruff و mypy و tsc و eslint تمیز

apps/api/app/infrastructure/adapters/agy.py

@@ -91,6 +91,33 @@ SAFETY_SETTINGS: list[dict[str, str]] = [
 #: manual paste must fit inside it comfortably.
 CHALLENGE_TTL_SECONDS = 600
 
+#: agy ``reasoning_effort`` → thinking budget (tokens), mirroring the Antigravity
+#: client. An unknown value falls back to ``medium`` server-side, and the UI
+#: only ever offers these five.
+EFFORT_THINKING_BUDGETS: dict[str, int] = {
+    "minimal": 1024,
+    "low": 4096,
+    "medium": 8192,
+    "high": 16384,
+    "xhigh": 32768,
+}
+
+#: Where thinking is honoured. A request must never carry ``thinkingConfig`` to
+#: a model that does not support it — that is a hard 400 from the backend.
+_THINKING_MARKERS: tuple[str, ...] = (
+    "thinking",
+    "gemini-3",
+    "gemini-2.5-pro",
+    "gemini-2.5-flash",
+)
+
+
+def _supports_thinking(model_name: str) -> bool:
+    lowered = model_name.lower()
+    if "image" in lowered:
+        return False
+    return any(marker in lowered for marker in _THINKING_MARKERS)
+
 _StatusFn = Callable[..., Awaitable[None]]
 _LoaderFn = Callable[[], Awaitable[list[ModelCredential]]]
 
@@ -603,7 +630,25 @@ class AgyModelAdapter:
         accounts = [a for a in (_account_of(c) for c in await self._load()) if a.refresh_token]
         if not accounts:
             return []
-        token = await self._access_token(accounts[0])
+        return await self._fetch_models(accounts[0])
+
+    async def available_models_for(self, credential: ModelCredential) -> list[str]:
+        """Live model catalog of exactly one account.
+
+        Called per account selection in the admin console, so the list is
+        fetched fresh from Google every time (the access token is cached, the
+        ``fetchAvailableModels`` round-trip is not) and reflects what *this*
+        subscription is entitled to.
+        """
+        account = _account_of(credential)
+        if not account.refresh_token:
+            raise ProviderUnavailableError(
+                "این اکانت اطلاعات ورود ندارد؛ یک بار دیگر لاگین کنید."
+            )
+        return await self._fetch_models(account)
+
+    async def _fetch_models(self, account: AgyAccount) -> list[str]:
+        token = await self._access_token(account)
         for base in BASE_URLS:
             try:
                 async with httpx.AsyncClient(timeout=30.0) as client:
@@ -713,7 +758,8 @@ class AgyModelAdapter:
     ) -> ModelResponse:
         token = await self._access_token(account)
         payload = _build_generate_payload(
-            account, model.internal_name, messages, system, temperature, max_tokens
+            account, model.internal_name, messages, system, temperature, max_tokens,
+            reasoning_effort=model.reasoning_effort,
         )
         headers = {
             "Authorization": f"Bearer {token}",
@@ -775,6 +821,7 @@ def _build_generate_payload(
     system: str,
     temperature: float,
     max_tokens: int,
+    reasoning_effort: str = "",
 ) -> dict:
     """Translate the port's OpenAI-shaped call into the Antigravity envelope."""
     system_parts: list[str] = [system] if system else []
@@ -786,13 +833,27 @@ def _build_generate_payload(
         role = "model" if message.role == "assistant" else "user"
         contents.append({"role": role, "parts": [{"text": message.content}]})
 
+    generation: dict = {
+        "temperature": temperature,
+        "maxOutputTokens": max_tokens,
+    }
+    # Reasoning effort is only sent to models that can think; the backend
+    # rejects ``thinkingConfig`` everywhere else with a 400.
+    if reasoning_effort and _supports_thinking(model_name):
+        budget = EFFORT_THINKING_BUDGETS.get(reasoning_effort, 8192)
+        generation["thinkingConfig"] = {
+            "thinkingBudget": budget,
+            "includeThoughts": True,
+        }
+        if max_tokens <= budget:
+            # The budget must fit inside the answer window or the thinking
+            # model has nowhere to emit its final answer.
+            generation["maxOutputTokens"] = budget + 8000
+
     request: dict = {
         "contents": contents,
         "sessionId": _random_session_id(),
-        "generationConfig": {
-            "temperature": temperature,
-            "maxOutputTokens": max_tokens,
-        },
+        "generationConfig": generation,
         "toolConfig": {"functionCallingConfig": {"mode": "VALIDATED"}},
         "safetySettings": SAFETY_SETTINGS,
     }