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
2026-10-13 10:01:11 +03:30
parent a2284db4f4
commit da5a2a5d1a
29 changed files with 906 additions and 995 deletions
@@ -1,7 +1,10 @@
 """Shared test fixtures.

-Tests run against an in-memory SQLite database and the ``mock`` adapters, so the
-suite needs no Postgres, no Redis and no credentials.
+Tests run against an in-memory SQLite database with the real agy-shaped adapter
+wiring and a fake model provider, so the suite needs no Postgres, no Redis, no
+Google and no accounts. agy is the only model provider the platform talks to
+(see ``app/infrastructure/adapters/registry.py``); the fake keeps completions
+deterministic and offline.
 """

 from __future__ import annotations
@@ -96,6 +99,76 @@ async def client(engine, session_factory):
    app.dependency_overrides.clear()


+# ---------------------------------------------------------------------------
+# Fake model provider — the only sanctioned stand-in for agy in tests
+# ---------------------------------------------------------------------------
+
+
+class _FakeModelProvider:
+    """Offline stand-in for the agy adapter.
+
+    The canned text must read like model output, not like a debug tag: the
+    provider/model identity is platform-internal and must never surface in
+    anything a customer can read (see the e2e smoke test).
+    """
+
+    async def complete(
+        self,
+        *,
+        model,
+        messages,
+        system="",
+        temperature=0.2,
+        max_tokens=2048,
+        credential=None,
+        credentials=None,
+    ):
+        import asyncio
+        import time
+
+        from app.application.ports import ModelResponse
+
+        started = time.perf_counter()
+        last_user = next((m.content for m in reversed(messages) if m.role == "user"), "")
+        prompt_tokens = sum(len(m.content.split()) for m in messages)
+        # A tiny sleep so latency bookkeeping runs a real code path.
+        await asyncio.sleep(0)
+        content = (
+            "این پاسخ آزمایشی است. "
+            f"درخواست دریافتی: {last_user[:280] or '—'}"
+        )
+        return ModelResponse(
+            content=content,
+            model_internal_name=model.internal_name,
+            input_tokens=prompt_tokens + len(system.split()),
+            output_tokens=len(content.split()),
+            latency_ms=int((time.perf_counter() - started) * 1000),
+            raw={},
+        )
+
+    async def health(self):
+        from app.application.ports import ModelHealth
+
+        return ModelHealth(ok=True, detail="fake agy pool")
+
+
+@pytest.fixture(autouse=True)
+def _fake_model_provider(monkeypatch):
+    """Swap the real model adapter out of every request path.
+
+    ``container`` and the health router bind ``get_model_provider`` at import
+    time, so both import sites are patched; without this a completion test
+    would open the developer's real database and reach the real Google API.
+    """
+    from app.application import container as di_container
+    from app.interfaces.http.routers import health as health_router
+
+    provider = _FakeModelProvider()
+    monkeypatch.setattr(di_container, "get_model_provider", lambda: provider)
+    monkeypatch.setattr(health_router, "get_model_provider", lambda: provider)
+    return provider
+
+
 async def signup(
    client: AsyncClient, email: str, password: str = "password123", name: str = ""
 ) -> dict:
@@ -149,10 +222,31 @@ SYSTEM_ROUTING_POLICIES: list[tuple[str, str, bool, str]] = [
    ("استدلال عمیق", "reasoning", False, "برای تحلیل پیچیده."),
 ]

-#: The mock models ``app/seed.py`` installs: (internal_name, display, capabilities).
-SYSTEM_MODELS: list[tuple[str, str, list[str]]] = [
-    ("mock-fast", "Mock Fast", ["text"]),
-    ("mock-reasoning", "Mock Reasoning", ["text", "reasoning", "code"]),
+#: The agy models ``app/seed.py`` installs: (internal_name, display,
+#: capabilities, reasoning_effort). agy is the only provider, so this is the
+#: whole catalogue.
+SYSTEM_MODELS: list[tuple[str, str, list[str], str]] = [
+    ("gemini-3-flash", "Gemini 3 Flash", ["text", "code", "long_context"], ""),
+    (
+        "gemini-3-pro-high",
+        "Gemini 3 Pro",
+        ["text", "reasoning", "code", "vision", "long_context"],
+        "high",
+    ),
+    ("claude-sonnet-4-5", "Claude Sonnet 4.5", ["text", "code", "tools"], ""),
+    (
+        "claude-sonnet-4-5-thinking",
+        "Claude Sonnet 4.5 (Thinking)",
+        ["text", "code", "reasoning"],
+        "high",
+    ),
+    (
+        "claude-opus-4-5-thinking",
+        "Claude Opus 4.5 (Thinking)",
+        ["text", "reasoning", "code", "long_context"],
+        "high",
+    ),
+    ("gpt-oss-120b-medium", "GPT-OSS 120B Medium", ["text", "code"], ""),
 ]

 #: The rate cards ``app/seed.py`` installs: (unit, customer_price, internal_cost).
@@ -174,11 +268,11 @@ SYSTEM_SANDBOX_TEMPLATES: list[tuple[str, str, str]] = [
 async def seeded(session_factory):
    """Install the platform-level seed rows a fresh install would have.

-    Agent templates, routing policies, the mock provider/models, the rate cards
-    and the sandbox templates are all *platform* data (no organization owns
-    them), so without this an employee cannot be created, a run cannot resolve a
-    model to route to, usage cannot be priced and no sandbox can be started —
-    none of which the real seeded install lacks.
+    Agent templates, routing policies, the agy provider with its model catalog,
+    the rate cards and the sandbox templates are all *platform* data (no
+    organization owns them), so without this an employee cannot be created, a
+    run cannot resolve a model to route to, usage cannot be priced and no
+    sandbox can be started — none of which the real seeded install lacks.
    """
    from app.infrastructure.db.models import (
        AgentTemplateRow,
@@ -214,9 +308,11 @@ async def seeded(session_factory):
        # surfaces as a ``ProviderUnavailableError`` and a *failed* run.
        provider_id = new_id()
        session.add(
-            AiProviderRow(id=provider_id, name="mock", type="mock", base_url="", is_active=True)
+            AiProviderRow(
+                id=provider_id, name="agy", type="agy", base_url="", is_active=True
+            )
        )
-        for internal, display, capabilities in SYSTEM_MODELS:
+        for internal, display, capabilities, effort in SYSTEM_MODELS:
            session.add(
                AiModelRow(
                    provider_id=provider_id,
@@ -225,6 +321,7 @@ async def seeded(session_factory):
                    capabilities=capabilities,
                    input_price_per_1m=0.3,
                    output_price_per_1m=1.2,
+                    reasoning_effort=effort,
                    is_active=True,
                )
            )