sync: 158 file da Baida98/AI@c4f6430d (2026-08-15 16:49 UTC) [deploy-all] (#41)

- sync: 158 file da Baida98/AI@c4f6430d (2026-08-15 16:49 UTC) [deploy-all] (28207b9d22a8dca431e1e96056469e901c1b8f33)

.env.example

@@ -61,5 +61,5 @@ UPSTASH_REDIS_REST_TOKEN=
 # ── 9. Feature Flags ─────────────────────────────────────────
 VITE_ENABLE_BROWSER_SANDBOX=false
 UNIFIED_LOOP_MAX_STEPS=8
-LLM_MODEL=google/gemini-2.0-flash-exp:free
+LLM_MODEL=openai/gpt-oss-20b:free
 

api/exec.py

@@ -574,8 +574,8 @@ async def llm_fix_code(
     or_key = os.getenv('OPENROUTER_API_KEY', '')
     if or_key:
         for m in [
+            'openai/gpt-oss-20b:free',
             'meta-llama/llama-3.1-8b-instruct:free',
-            'google/gemini-2.0-flash-exp:free',
             'qwen/qwen-2.5-coder-7b-instruct:free',
         ]:
             _FIX_CHAIN.append(('https://openrouter.ai/api/v1', or_key, m))

api/providers.py

@@ -883,37 +883,40 @@ async def update_provider_models(role: AuthRole = Depends(require_role(AuthRole.
     if _sb is None:
         return {"ok": False, "error": "Supabase non configurato", "updated": 0}
 
-    # Mappa: modello_vecchio -> modello_nuovo
+    # Mappa provider-specifica: (provider, modello_vecchio, modello_nuovo).
+    # Lo stesso ID modello può essere valido su un provider e non su un altro:
+    # filtrare per `name` evita di applicare un formato incompatibile alla riga sbagliata.
+    # GPT-OSS 120B non compare come vecchio valore perché è già un modello supportato.
     MODEL_FIXES = [
-        ("llama-3.1-70b-versatile",    "openai/gpt-oss-120b"),
-        ("llama3.1-70b",               "llama-4-scout"),
-        ("llama-3.1-405b-instruct",    "meta/llama-3.3-70b-instruct"),
-        ("llama-3.1-405b",             "meta-llama/llama-4-scout:free"),
-        ("llama3-70b",                 "DeepSeek-V3.2"),
-        ("gemini-1.5-flash",           "gemini-2.5-flash-lite"),
-        ("gemini-1.5-pro",             "gemini-2.5-flash-lite"),
-        ("gpt-oss-120b",               "llama-4-scout"),
-        ("claude-3.5-sonnet",          "meta-llama/llama-4-scout:free"),
+        ("groq",       "llama-3.1-70b-versatile", "openai/gpt-oss-120b"),
+        ("cerebras",   "llama3.1-70b",            "gpt-oss-120b"),
+        ("nvidia",     "llama-3.1-405b-instruct", "meta/llama-3.3-70b-instruct"),
+        ("openrouter", "llama-3.1-405b",          "openai/gpt-oss-20b:free"),
+        ("sambanova",  "llama3-70b",              "DeepSeek-V3.2"),
+        ("gemini",     "gemini-1.5-flash",        "gemini-3.5-flash-lite"),
+        ("gemini",     "gemini-1.5-pro",          "gemini-3.6-flash"),
+        ("openrouter", "claude-3.5-sonnet",       "openai/gpt-oss-20b:free"),
     ]
 
     import asyncio as _aio
     total_updated = 0
     results = []
 
-    for old_model, new_model in MODEL_FIXES:
+    for provider_name, old_model, new_model in MODEL_FIXES:
         try:
             r = await _aio.to_thread(
-                lambda om=old_model, nm=new_model: _sb.table("ai_providers")
+                lambda pn=provider_name, om=old_model, nm=new_model: _sb.table("ai_providers")
                     .update({"default_model": nm})
+                    .eq("name", pn)
                     .eq("default_model", om)
                     .execute()
             )
             n = len(r.data) if r.data else 0
             total_updated += n
             if n > 0:
-                results.append({"old": old_model, "new": new_model, "rows": n})
+                results.append({"provider": provider_name, "old": old_model, "new": new_model, "rows": n})
         except Exception as exc:
-            results.append({"old": old_model, "new": new_model, "error": str(exc)[:100]})
+            results.append({"provider": provider_name, "old": old_model, "new": new_model, "error": str(exc)[:100]})
 
     # Disattiva provider E2B (non sono LLM provider)
     try:

models/role_router.py

@@ -37,13 +37,13 @@ _logger = logging.getLogger("models.role_router")
 
 class Role(str, Enum):
     FAST       = "fast"        # greetings, math semplice, identity — openai/gpt-oss-20b
-    ARCHITECT  = "architect"   # planning, ragionamento complesso — llama-4-scout (10M ctx)
+    ARCHITECT  = "architect"   # planning, ragionamento complesso — GPT-OSS 120B
     CODER      = "coder"       # coding, debug — openai/gpt-oss-120b
     TESTER     = "tester"      # test gen, debug hints — openai/gpt-oss-120b
     CONTEXT    = "context"     # summarization, context compression — openai/gpt-oss-120b
     DEFAULT    = "default"     # AIClient() primary
-    RESEARCHER = "researcher"  # web research + document synthesis — gemini-2.0-flash-exp
-    REASONER   = "reasoner"    # throughput massimo — Cerebras llama-4-scout (2000+ tok/s)
+    RESEARCHER = "researcher"  # web research + document synthesis — GPT-OSS 120B
+    REASONER   = "reasoner"    # throughput massimo — Cerebras GPT-OSS 120B
     SAMBANOVA  = "sambanova"
     NVIDIA     = "nvidia"    # NVIDIA NIM — nemotron-3-ultra-550b (1M ctx)   # DeepSeek-V3.2 via SambaNova (404ms, 100% qualità benchmark)
 
@@ -107,8 +107,8 @@ class RoleRouter:
 
     @staticmethod
     def _architect_client() -> Any:
-        """NVIDIA NIM deepseek-v4-flash (1M ctx) come primario — massima potenza per architettura.
-        Fallback 1: Groq llama-4-scout (10M ctx, 480ms). Fallback 2: OpenRouter llama-4-scout:free."""
+        """NVIDIA NIM come primario per architettura.
+        Fallback 1: Groq GPT-OSS 120B. Fallback 2: OpenRouter GPT-OSS 20B gratuito."""
         from models.ai_client import AIClient, ProviderConfig
         nvidia_key = os.getenv("NVIDIA_API_KEY")
         if nvidia_key:
@@ -125,7 +125,7 @@ class RoleRouter:
             client.default_model = nvidia.default_model
             client.client = client._client_for(nvidia)
             return client
-        # Fallback 1: Groq llama-4-scout (10M ctx, 480ms)
+        # Fallback 1: Groq GPT-OSS 120B, modello production supportato.
         groq_key = os.getenv("GROQ_API_KEY")
         if groq_key:
             client = AIClient()
@@ -133,7 +133,7 @@ class RoleRouter:
                 name="groq-architect",
                 api_key=groq_key,
                 base_url="https://api.groq.com/openai/v1",
-                default_model=os.getenv("ARCHITECT_MODEL", "llama-4-scout"),
+                default_model=os.getenv("ARCHITECT_MODEL", "openai/gpt-oss-120b"),
             )
             rest = [p for p in client.providers if p.name not in ("groq", "groq-architect")]
             client.providers = [architect, *rest]
@@ -141,7 +141,7 @@ class RoleRouter:
             client.default_model = architect.default_model
             client.client = client._client_for(architect)
             return client
-        # Fallback: OpenRouter meta-llama/llama-4-scout:free (1645ms ma 100% qualità)
+        # Fallback: OpenRouter GPT-OSS 20B gratuito, con ID verificato nel catalogo.
         openrouter_key = os.getenv("OPENROUTER_API_KEY")
         if openrouter_key:
             client = AIClient()
@@ -149,7 +149,7 @@ class RoleRouter:
                 name="openrouter-architect",
                 api_key=openrouter_key,
                 base_url="https://openrouter.ai/api/v1",
-                default_model="meta-llama/llama-4-scout:free",
+                default_model=os.getenv("OPENROUTER_MODEL", "openai/gpt-oss-20b:free"),
             )
             rest = [p for p in client.providers if not p.name.startswith("openrouter")]
             client.providers = [fallback, *rest]
@@ -205,7 +205,7 @@ class RoleRouter:
                 name="openrouter-coder",
                 api_key=openrouter_key,
                 base_url="https://openrouter.ai/api/v1",
-                default_model="meta-llama/llama-4-scout:free",
+                default_model=os.getenv("OPENROUTER_MODEL", "openai/gpt-oss-20b:free"),
                 purpose="coding",
             )
             rest = [p for p in client.providers if not p.name.startswith("openrouter")]
@@ -286,7 +286,7 @@ class RoleRouter:
             name="cerebras-reasoner",
             api_key=cerebras_key,
             base_url="https://api.cerebras.ai/v1",
-            default_model=os.getenv("CEREBRAS_MODEL", "llama-4-scout"),
+            default_model=os.getenv("CEREBRAS_MODEL", "gpt-oss-120b"),
         )
         rest = [provider for provider in client.providers if not provider.name.startswith("cerebras")]
         client.providers = [reasoner, *rest]

tests/test_provider_model_migrations.py

@@ -0,0 +1,116 @@
+import os
+import unittest
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from api import providers
+from models.ai_client import AIClient
+from models.role_router import Role, RoleRouter
+
+
+class _RecordingTable:
+    def __init__(self):
+        self.calls: list[dict] = []
+        self._pending: dict = {}
+
+    def update(self, values: dict):
+        self._pending = {"values": values, "filters": []}
+        return self
+
+    def eq(self, field: str, value: str):
+        self._pending["filters"].append((field, value))
+        return self
+
+    def execute(self):
+        self.calls.append(self._pending)
+        return SimpleNamespace(data=[{"id": len(self.calls)}])
+
+
+class _RecordingSupabase:
+    def __init__(self):
+        self.table_ref = _RecordingTable()
+
+    def table(self, name: str):
+        if name != "ai_providers":
+            raise AssertionError(f"Tabella inattesa: {name}")
+        return self.table_ref
+
+
+class RoleRouterModelMigrationTests(unittest.TestCase):
+    @patch.object(AIClient, "_load_providers", return_value=[])
+    @patch.dict(
+        os.environ,
+        {"GROQ_API_KEY": "test-groq-key"},
+        clear=True,
+    )
+    def test_architect_uses_supported_groq_default(self, _load_providers):
+        client = RoleRouter.get_client(Role.ARCHITECT)
+
+        self.assertEqual(client.provider_name, "groq-architect")
+        self.assertEqual(client.default_model, "openai/gpt-oss-120b")
+
+    @patch.object(AIClient, "_load_providers", return_value=[])
+    @patch.dict(
+        os.environ,
+        {"OPENROUTER_API_KEY": "test-openrouter-key"},
+        clear=True,
+    )
+    def test_openrouter_role_fallbacks_use_available_free_model(self, _load_providers):
+        architect = RoleRouter.get_client(Role.ARCHITECT)
+        coder = RoleRouter.get_client(Role.CODER)
+
+        self.assertEqual(architect.default_model, "openai/gpt-oss-20b:free")
+        self.assertEqual(coder.default_model, "openai/gpt-oss-20b:free")
+
+    @patch.object(AIClient, "_load_providers", return_value=[])
+    @patch.dict(
+        os.environ,
+        {"CEREBRAS_API_KEY": "test-cerebras-key"},
+        clear=True,
+    )
+    def test_reasoner_uses_cerebras_gpt_oss_default(self, _load_providers):
+        client = RoleRouter.get_client(Role.REASONER)
+
+        self.assertEqual(client.provider_name, "cerebras-reasoner")
+        self.assertEqual(client.default_model, "gpt-oss-120b")
+
+
+class ProviderTableMigrationTests(unittest.IsolatedAsyncioTestCase):
+    async def test_update_models_filters_by_provider_and_never_downgrades_gpt_oss(self):
+        original_supabase = providers._sb
+        fake_supabase = _RecordingSupabase()
+        providers._sb = fake_supabase
+        try:
+            payload = await providers.update_provider_models(role=None)
+        finally:
+            providers._sb = original_supabase
+
+        self.assertTrue(payload["ok"])
+        self.assertEqual(payload["total_updated"], 8)
+        self.assertEqual(len(fake_supabase.table_ref.calls), 8)
+
+        expected = {
+            ("groq", "llama-3.1-70b-versatile", "openai/gpt-oss-120b"),
+            ("cerebras", "llama3.1-70b", "gpt-oss-120b"),
+            ("nvidia", "llama-3.1-405b-instruct", "meta/llama-3.3-70b-instruct"),
+            ("openrouter", "llama-3.1-405b", "openai/gpt-oss-20b:free"),
+            ("sambanova", "llama3-70b", "DeepSeek-V3.2"),
+            ("gemini", "gemini-1.5-flash", "gemini-3.5-flash-lite"),
+            ("gemini", "gemini-1.5-pro", "gemini-3.6-flash"),
+            ("openrouter", "claude-3.5-sonnet", "openai/gpt-oss-20b:free"),
+        }
+        actual = {
+            (
+                dict(call["filters"])["name"],
+                dict(call["filters"])["default_model"],
+                call["values"]["default_model"],
+            )
+            for call in fake_supabase.table_ref.calls
+        }
+        self.assertEqual(actual, expected)
+        self.assertNotIn("llama-4-scout", {new for _, _, new in actual})
+        self.assertNotIn(("cerebras", "gpt-oss-120b", "llama-4-scout"), actual)
+
+
+if __name__ == "__main__":
+    unittest.main()