Spaces:
Running
Running
sync: 176 file da Baida98/AI@eeb3eb47 (2026-08-25 09:19 UTC) [deploy-all]
#32
by Baida07 - opened
- agents/planner.py +3 -0
- api/agent.py +42 -6
- api/state.py +32 -1
- models/ai_client.py +42 -4
- tests/test_ai_client_byok.py +56 -0
agents/planner.py
CHANGED
|
@@ -194,6 +194,7 @@ def _parse_plan(raw: str) -> dict | None:
|
|
| 194 |
|
| 195 |
class Planner:
|
| 196 |
def __init__(self, llm_client: AIClient | None = None):
|
|
|
|
| 197 |
if llm_client is not None:
|
| 198 |
self.llm = llm_client
|
| 199 |
else:
|
|
@@ -211,6 +212,8 @@ class Planner:
|
|
| 211 |
def _get_fast_llm(self) -> AIClient:
|
| 212 |
"""Gap-5: Cerebras gpt-oss-120b (2000+ tok/s) per quick-start draft.
|
| 213 |
Fallback: Groq openai/gpt-oss-20b se CEREBRAS_API_KEY assente."""
|
|
|
|
|
|
|
| 214 |
try:
|
| 215 |
from models.role_router import RoleRouter, Role
|
| 216 |
return RoleRouter.get_client(Role.REASONER) # Cerebras 120B
|
|
|
|
| 194 |
|
| 195 |
class Planner:
|
| 196 |
def __init__(self, llm_client: AIClient | None = None):
|
| 197 |
+
self._explicit_llm = llm_client is not None
|
| 198 |
if llm_client is not None:
|
| 199 |
self.llm = llm_client
|
| 200 |
else:
|
|
|
|
| 212 |
def _get_fast_llm(self) -> AIClient:
|
| 213 |
"""Gap-5: Cerebras gpt-oss-120b (2000+ tok/s) per quick-start draft.
|
| 214 |
Fallback: Groq openai/gpt-oss-20b se CEREBRAS_API_KEY assente."""
|
| 215 |
+
if self._explicit_llm:
|
| 216 |
+
return self.llm
|
| 217 |
try:
|
| 218 |
from models.role_router import RoleRouter, Role
|
| 219 |
return RoleRouter.get_client(Role.REASONER) # Cerebras 120B
|
api/agent.py
CHANGED
|
@@ -45,7 +45,7 @@ from .auth_guard import require_role, AuthRole
|
|
| 45 |
from pydantic import BaseModel, field_validator
|
| 46 |
from typing import Literal
|
| 47 |
from .state import (
|
| 48 |
-
_agent_tasks, _task_checkpoints, _loop_registry, _run_stream_tasks,
|
| 49 |
_prune_agent_tasks, _prune_checkpoints, _prune_loop_registry,
|
| 50 |
_get_mem_manager, _get_mem_manager_async, _get_executor, _get_planner, _get_ai_client,
|
| 51 |
ReasonLoopIn, AgentTaskIn,
|
|
@@ -87,6 +87,21 @@ except Exception:
|
|
| 87 |
router = APIRouter()
|
| 88 |
|
| 89 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
# ── Deprecated run_loop ────────────────────────────────────────────────────────
|
| 91 |
|
| 92 |
@router.post('/run_loop', deprecated=True)
|
|
@@ -601,6 +616,8 @@ async def create_agent_task(body: AgentTaskIn, role: AuthRole = Depends(require_
|
|
| 601 |
|
| 602 |
# Already in memory → return immediately (normal path, includes S358 reconnect)
|
| 603 |
if task_id in _agent_tasks:
|
|
|
|
|
|
|
| 604 |
return {'taskId': task_id, 'status': _agent_tasks[task_id]['status']}
|
| 605 |
|
| 606 |
# S359: try Supabase lazy restore (only hit network after backend restart)
|
|
@@ -610,6 +627,7 @@ async def create_agent_task(body: AgentTaskIn, role: AuthRole = Depends(require_
|
|
| 610 |
# Use context from the incoming request (not persisted to save space).
|
| 611 |
restored['context'] = body.context
|
| 612 |
_agent_tasks[task_id] = restored
|
|
|
|
| 613 |
return {'taskId': task_id, 'status': restored['status'], 'restored': True}
|
| 614 |
|
| 615 |
# Brand new task
|
|
@@ -627,6 +645,10 @@ async def create_agent_task(body: AgentTaskIn, role: AuthRole = Depends(require_
|
|
| 627 |
'persona': body.persona, # P17-F5: expertise persona hint
|
| 628 |
'session_id': body.session_id or '', # P17-F2: BB session key (normalize None→'')
|
| 629 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
| 630 |
# BG-4: restore cross-session handoff context (async, non-blocking)
|
| 631 |
if body.session_id:
|
| 632 |
_hctx = await sb_restore_handoff_context(body.session_id)
|
|
@@ -735,6 +757,7 @@ async def list_agent_tasks(limit: int = 50, status: str = '', role: AuthRole = D
|
|
| 735 |
async def cancel_agent_task(task_id: str, role: AuthRole = Depends(require_role(AuthRole.MACHINE))): # GAP-1-fix
|
| 736 |
if task_id in _agent_tasks:
|
| 737 |
_agent_tasks[task_id]['status'] = 'CANCELLED'
|
|
|
|
| 738 |
reg = _loop_registry.get(task_id)
|
| 739 |
if reg and not reg.get('done'):
|
| 740 |
at = reg.get('asyncio_task')
|
|
@@ -947,8 +970,9 @@ async def stream_agent_task(task_id: str, request: Request, resume: int = 0, rol
|
|
| 947 |
async def run_loop() -> None:
|
| 948 |
try:
|
| 949 |
from agents.unified_loop import UnifiedAgentLoop
|
| 950 |
-
#
|
| 951 |
-
|
|
|
|
| 952 |
try:
|
| 953 |
from agents.critic import Critic
|
| 954 |
from agents.response_verifier import ResponseVerifier
|
|
@@ -1051,11 +1075,20 @@ async def stream_agent_task(task_id: str, request: Request, resume: int = 0, rol
|
|
| 1051 |
_hctx = task.get("_handoff_context", "")
|
| 1052 |
if _hctx:
|
| 1053 |
context_str = f"{_hctx}\n\n{context_str}".strip()
|
| 1054 |
-
#
|
| 1055 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1056 |
loop = UnifiedAgentLoop(
|
| 1057 |
llm_client=_persona_client, critic=_critic, verifier=_verifier,
|
| 1058 |
-
memory=await _get_mem_manager_async(), executor=_get_executor(), planner=
|
| 1059 |
)
|
| 1060 |
step_idx = [0]
|
| 1061 |
_backend_steps: list[dict] = [] # GAP-SYNC-FIX: log step per resume preciso
|
|
@@ -1347,6 +1380,9 @@ async def stream_agent_task(task_id: str, request: Request, resume: int = 0, rol
|
|
| 1347 |
asyncio.create_task(_tg_error(task_id, task.get('goal', ''), str(err))).add_done_callback(_log_task_exc)
|
| 1348 |
|
| 1349 |
finally:
|
|
|
|
|
|
|
|
|
|
| 1350 |
reg_entry['done'] = True
|
| 1351 |
reg_entry['finished_at'] = time.time()
|
| 1352 |
for q in list(reg_entry['subscriber_queues']):
|
|
|
|
| 45 |
from pydantic import BaseModel, field_validator
|
| 46 |
from typing import Literal
|
| 47 |
from .state import (
|
| 48 |
+
_agent_tasks, _task_ai_clients, _task_checkpoints, _loop_registry, _run_stream_tasks,
|
| 49 |
_prune_agent_tasks, _prune_checkpoints, _prune_loop_registry,
|
| 50 |
_get_mem_manager, _get_mem_manager_async, _get_executor, _get_planner, _get_ai_client,
|
| 51 |
ReasonLoopIn, AgentTaskIn,
|
|
|
|
| 87 |
router = APIRouter()
|
| 88 |
|
| 89 |
|
| 90 |
+
def _attach_byok_client(task_id: str, credentials: object) -> None:
|
| 91 |
+
"""Create a task-scoped LLM client without persisting or logging credentials."""
|
| 92 |
+
if credentials is None:
|
| 93 |
+
return
|
| 94 |
+
try:
|
| 95 |
+
runtime_config = credentials.as_runtime_config()
|
| 96 |
+
if not runtime_config:
|
| 97 |
+
return
|
| 98 |
+
from models.ai_client import AIClient
|
| 99 |
+
_task_ai_clients[task_id] = AIClient(byok_credentials=runtime_config)
|
| 100 |
+
except Exception as exc:
|
| 101 |
+
# Never include the payload or credential values in diagnostics.
|
| 102 |
+
_logger.warning("[agent] unable to initialise BYOK task client: %s", type(exc).__name__)
|
| 103 |
+
|
| 104 |
+
|
| 105 |
# ── Deprecated run_loop ────────────────────────────────────────────────────────
|
| 106 |
|
| 107 |
@router.post('/run_loop', deprecated=True)
|
|
|
|
| 616 |
|
| 617 |
# Already in memory → return immediately (normal path, includes S358 reconnect)
|
| 618 |
if task_id in _agent_tasks:
|
| 619 |
+
if task_id not in _task_ai_clients:
|
| 620 |
+
_attach_byok_client(task_id, body.provider_credentials)
|
| 621 |
return {'taskId': task_id, 'status': _agent_tasks[task_id]['status']}
|
| 622 |
|
| 623 |
# S359: try Supabase lazy restore (only hit network after backend restart)
|
|
|
|
| 627 |
# Use context from the incoming request (not persisted to save space).
|
| 628 |
restored['context'] = body.context
|
| 629 |
_agent_tasks[task_id] = restored
|
| 630 |
+
_attach_byok_client(task_id, body.provider_credentials)
|
| 631 |
return {'taskId': task_id, 'status': restored['status'], 'restored': True}
|
| 632 |
|
| 633 |
# Brand new task
|
|
|
|
| 645 |
'persona': body.persona, # P17-F5: expertise persona hint
|
| 646 |
'session_id': body.session_id or '', # P17-F2: BB session key (normalize None→'')
|
| 647 |
}
|
| 648 |
+
# Le credenziali BYOK restano in una mappa runtime separata dai metadata task
|
| 649 |
+
# e non raggiungono Supabase, checkpoint o buffer SSE.
|
| 650 |
+
_attach_byok_client(task_id, body.provider_credentials)
|
| 651 |
+
|
| 652 |
# BG-4: restore cross-session handoff context (async, non-blocking)
|
| 653 |
if body.session_id:
|
| 654 |
_hctx = await sb_restore_handoff_context(body.session_id)
|
|
|
|
| 757 |
async def cancel_agent_task(task_id: str, role: AuthRole = Depends(require_role(AuthRole.MACHINE))): # GAP-1-fix
|
| 758 |
if task_id in _agent_tasks:
|
| 759 |
_agent_tasks[task_id]['status'] = 'CANCELLED'
|
| 760 |
+
_task_ai_clients.pop(task_id, None)
|
| 761 |
reg = _loop_registry.get(task_id)
|
| 762 |
if reg and not reg.get('done'):
|
| 763 |
at = reg.get('asyncio_task')
|
|
|
|
| 970 |
async def run_loop() -> None:
|
| 971 |
try:
|
| 972 |
from agents.unified_loop import UnifiedAgentLoop
|
| 973 |
+
# Ogni task BYOK usa il suo client effimero; gli altri mantengono
|
| 974 |
+
# il singleton runtime. Le credenziali non entrano nel task dict.
|
| 975 |
+
client = _task_ai_clients.get(task_id) or _get_ai_client()
|
| 976 |
try:
|
| 977 |
from agents.critic import Critic
|
| 978 |
from agents.response_verifier import ResponseVerifier
|
|
|
|
| 1075 |
_hctx = task.get("_handoff_context", "")
|
| 1076 |
if _hctx:
|
| 1077 |
context_str = f"{_hctx}\n\n{context_str}".strip()
|
| 1078 |
+
# I router persona dipendono dalle env del backend: con BYOK il
|
| 1079 |
+
# client per task resta autorevole in ogni fase, incluso il planner.
|
| 1080 |
+
_is_byok_task = task_id in _task_ai_clients
|
| 1081 |
+
_persona_client = client if _is_byok_task else _get_persona_llm_client(_persona, client)
|
| 1082 |
+
_planner = _get_planner()
|
| 1083 |
+
if _is_byok_task:
|
| 1084 |
+
try:
|
| 1085 |
+
from agents.planner import Planner
|
| 1086 |
+
_planner = Planner(llm_client=client)
|
| 1087 |
+
except Exception as exc:
|
| 1088 |
+
_logger.warning("[agent] BYOK planner fallback: %s", type(exc).__name__)
|
| 1089 |
loop = UnifiedAgentLoop(
|
| 1090 |
llm_client=_persona_client, critic=_critic, verifier=_verifier,
|
| 1091 |
+
memory=await _get_mem_manager_async(), executor=_get_executor(), planner=_planner,
|
| 1092 |
)
|
| 1093 |
step_idx = [0]
|
| 1094 |
_backend_steps: list[dict] = [] # GAP-SYNC-FIX: log step per resume preciso
|
|
|
|
| 1380 |
asyncio.create_task(_tg_error(task_id, task.get('goal', ''), str(err))).add_done_callback(_log_task_exc)
|
| 1381 |
|
| 1382 |
finally:
|
| 1383 |
+
# Il task è terminale: rimuove l’unico riferimento alle credenziali
|
| 1384 |
+
# BYOK, lasciando replay SSE e metadati senza segreti.
|
| 1385 |
+
_task_ai_clients.pop(task_id, None)
|
| 1386 |
reg_entry['done'] = True
|
| 1387 |
reg_entry['finished_at'] = time.time()
|
| 1388 |
for q in list(reg_entry['subscriber_queues']):
|
api/state.py
CHANGED
|
@@ -7,7 +7,7 @@ import os, time, asyncio as _asyncio_mod, json as _json, re as _re
|
|
| 7 |
import logging
|
| 8 |
from typing import Optional, Any, AsyncIterator, List, Tuple
|
| 9 |
from fastapi import HTTPException, APIRouter, Request, Body
|
| 10 |
-
from pydantic import BaseModel, field_validator
|
| 11 |
from .version import RUNTIME_VERSION
|
| 12 |
|
| 13 |
_logger = logging.getLogger("api.state")
|
|
@@ -128,6 +128,9 @@ SENSITIVE = {
|
|
| 128 |
# ── In-memory stores ──────────────────────────────────────────────────────────
|
| 129 |
_mem_fallback: dict[str, dict] = {}
|
| 130 |
_agent_tasks: dict[str, dict] = {}
|
|
|
|
|
|
|
|
|
|
| 131 |
_run_stream_tasks: dict[str, dict] = {}
|
| 132 |
_loop_registry: dict[str, dict] = {}
|
| 133 |
_LOOP_REGISTRY_TTL_S: float = 10 * 60
|
|
@@ -245,10 +248,13 @@ def _prune_agent_tasks() -> None:
|
|
| 245 |
and now - v.get('created_at', 0) > _AGENT_TASK_TTL_MS]
|
| 246 |
for k in expired:
|
| 247 |
_agent_tasks.pop(k, None)
|
|
|
|
|
|
|
| 248 |
if len(_agent_tasks) > _AGENT_TASK_MAX:
|
| 249 |
oldest = sorted(_agent_tasks.items(), key=lambda x: x[1].get('created_at', 0))
|
| 250 |
for k, _ in oldest[:len(_agent_tasks) - _AGENT_TASK_MAX]:
|
| 251 |
_agent_tasks.pop(k, None)
|
|
|
|
| 252 |
|
| 253 |
def _prune_loop_registry() -> None:
|
| 254 |
now = time.time()
|
|
@@ -284,6 +290,30 @@ class ReasonLoopIn(BaseModel):
|
|
| 284 |
def coerce_str(cls, v: object) -> str:
|
| 285 |
return str(v).strip()[:2000] if v else ""
|
| 286 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 287 |
class AgentTaskIn(BaseModel):
|
| 288 |
goal: str
|
| 289 |
context: list[dict] = []
|
|
@@ -295,6 +325,7 @@ class AgentTaskIn(BaseModel):
|
|
| 295 |
resume_from_step: Optional[int] = None
|
| 296 |
persona: Optional[str] = None
|
| 297 |
negative_constraints: Optional[str] = ""
|
|
|
|
| 298 |
|
| 299 |
@field_validator('goal', mode='before')
|
| 300 |
@classmethod
|
|
|
|
| 7 |
import logging
|
| 8 |
from typing import Optional, Any, AsyncIterator, List, Tuple
|
| 9 |
from fastapi import HTTPException, APIRouter, Request, Body
|
| 10 |
+
from pydantic import BaseModel, Field, SecretStr, field_validator
|
| 11 |
from .version import RUNTIME_VERSION
|
| 12 |
|
| 13 |
_logger = logging.getLogger("api.state")
|
|
|
|
| 128 |
# ── In-memory stores ──────────────────────────────────────────────────────────
|
| 129 |
_mem_fallback: dict[str, dict] = {}
|
| 130 |
_agent_tasks: dict[str, dict] = {}
|
| 131 |
+
# Client LLM BYOK associati al task: memoria effimera, mai serializzata nei task,
|
| 132 |
+
# checkpoint, eventi SSE o persistenza Supabase.
|
| 133 |
+
_task_ai_clients: dict[str, Any] = {}
|
| 134 |
_run_stream_tasks: dict[str, dict] = {}
|
| 135 |
_loop_registry: dict[str, dict] = {}
|
| 136 |
_LOOP_REGISTRY_TTL_S: float = 10 * 60
|
|
|
|
| 248 |
and now - v.get('created_at', 0) > _AGENT_TASK_TTL_MS]
|
| 249 |
for k in expired:
|
| 250 |
_agent_tasks.pop(k, None)
|
| 251 |
+
_task_ai_clients.pop(k, None)
|
| 252 |
+
|
| 253 |
if len(_agent_tasks) > _AGENT_TASK_MAX:
|
| 254 |
oldest = sorted(_agent_tasks.items(), key=lambda x: x[1].get('created_at', 0))
|
| 255 |
for k, _ in oldest[:len(_agent_tasks) - _AGENT_TASK_MAX]:
|
| 256 |
_agent_tasks.pop(k, None)
|
| 257 |
+
_task_ai_clients.pop(k, None)
|
| 258 |
|
| 259 |
def _prune_loop_registry() -> None:
|
| 260 |
now = time.time()
|
|
|
|
| 290 |
def coerce_str(cls, v: object) -> str:
|
| 291 |
return str(v).strip()[:2000] if v else ""
|
| 292 |
|
| 293 |
+
class ProviderCredentialsIn(BaseModel):
|
| 294 |
+
"""Credenziali LLM BYOK limitate ai provider OpenAI-compatible supportati.
|
| 295 |
+
|
| 296 |
+
SecretStr evita l’esposizione accidentale del valore in repr o log Pydantic;
|
| 297 |
+
as_runtime_config() è l’unico punto che restituisce stringhe al client effimero.
|
| 298 |
+
"""
|
| 299 |
+
groq: list[SecretStr] = Field(default_factory=list)
|
| 300 |
+
openrouter: list[SecretStr] = Field(default_factory=list)
|
| 301 |
+
|
| 302 |
+
@field_validator('groq', 'openrouter', mode='before')
|
| 303 |
+
@classmethod
|
| 304 |
+
def coerce_tokens(cls, value: object) -> list[object]:
|
| 305 |
+
return value if isinstance(value, list) else []
|
| 306 |
+
|
| 307 |
+
def as_runtime_config(self) -> dict[str, list[str]]:
|
| 308 |
+
credentials: dict[str, list[str]] = {}
|
| 309 |
+
for provider in ('groq', 'openrouter'):
|
| 310 |
+
values = [item.get_secret_value().strip() for item in getattr(self, provider)]
|
| 311 |
+
values = list(dict.fromkeys(value for value in values if value))
|
| 312 |
+
if values:
|
| 313 |
+
credentials[provider] = values
|
| 314 |
+
return credentials
|
| 315 |
+
|
| 316 |
+
|
| 317 |
class AgentTaskIn(BaseModel):
|
| 318 |
goal: str
|
| 319 |
context: list[dict] = []
|
|
|
|
| 325 |
resume_from_step: Optional[int] = None
|
| 326 |
persona: Optional[str] = None
|
| 327 |
negative_constraints: Optional[str] = ""
|
| 328 |
+
provider_credentials: Optional[ProviderCredentialsIn] = None
|
| 329 |
|
| 330 |
@field_validator('goal', mode='before')
|
| 331 |
@classmethod
|
models/ai_client.py
CHANGED
|
@@ -74,8 +74,13 @@ _PROVIDER_DEFS = [
|
|
| 74 |
|
| 75 |
|
| 76 |
class AIClient:
|
| 77 |
-
def __init__(self) -> None:
|
| 78 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
self._client_cache: dict[tuple[str, str, str], OpenAI] = {}
|
| 80 |
# Round-robin e circuit breaker sono indicizzati per purpose e profilo.
|
| 81 |
self._rr_indices: dict[str, int] = {}
|
|
@@ -83,6 +88,39 @@ class AIClient:
|
|
| 83 |
self._breaker_threshold = 2
|
| 84 |
self._breaker_cooldown_s = 60.0
|
| 85 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
def _load_providers(self) -> list[ProviderConfig]:
|
| 87 |
"""Carica la flotta: prova Supabase (tabella `ai_providers`, source of
|
| 88 |
truth dichiarata in supabase/migrations/20260711_ai_providers_fleet.sql),
|
|
@@ -375,7 +413,7 @@ class AIClient:
|
|
| 375 |
return provider
|
| 376 |
|
| 377 |
async def chat(self, messages: list, model: Optional[str] = None, temperature: float = 0.7, max_tokens: int = 4096) -> str:
|
| 378 |
-
if not self.providers: self.providers = self._load_providers()
|
| 379 |
|
| 380 |
# S-CACHE-1: Lookup semantico preventivo
|
| 381 |
cached = await get_cached_response(messages)
|
|
@@ -464,7 +502,7 @@ class AIClient:
|
|
| 464 |
|
| 465 |
async def stream_chat(self, messages: list, model: Optional[str] = None, temperature: float = 0.7, max_tokens: int = 4096) -> AsyncIterator[str]:
|
| 466 |
"""Streaming con failover immediato tra i provider configurati."""
|
| 467 |
-
if not self.providers: self.providers = self._load_providers()
|
| 468 |
# Nessun provider configurato: feedback immediato all'utente invece di
|
| 469 |
# cadere silenziosamente nel loop vuoto e dare un messaggio generico.
|
| 470 |
if not self.providers:
|
|
|
|
| 74 |
|
| 75 |
|
| 76 |
class AIClient:
|
| 77 |
+
def __init__(self, byok_credentials: dict[str, list[str]] | None = None) -> None:
|
| 78 |
+
# Le chiavi BYOK appartengono a un singolo task e vivono solo in questa
|
| 79 |
+
# istanza: non vengono scritte in env, Supabase, cache semantica o log.
|
| 80 |
+
self._byok_providers = self._providers_from_byok(byok_credentials or {})
|
| 81 |
+
# I profili BYOK precedono i provider runtime: la stessa API conserva
|
| 82 |
+
# comunque il fallback server-side in caso di quota o errore upstream.
|
| 83 |
+
self.providers = self._byok_providers + self._load_providers()
|
| 84 |
self._client_cache: dict[tuple[str, str, str], OpenAI] = {}
|
| 85 |
# Round-robin e circuit breaker sono indicizzati per purpose e profilo.
|
| 86 |
self._rr_indices: dict[str, int] = {}
|
|
|
|
| 88 |
self._breaker_threshold = 2
|
| 89 |
self._breaker_cooldown_s = 60.0
|
| 90 |
|
| 91 |
+
@staticmethod
|
| 92 |
+
def _providers_from_byok(credentials: dict[str, list[str]]) -> list[ProviderConfig]:
|
| 93 |
+
"""Build task-scoped provider profiles from browser BYOK credentials.
|
| 94 |
+
|
| 95 |
+
Only known OpenAI-compatible providers are accepted. Values are copied
|
| 96 |
+
into the transient client instance and intentionally never logged or
|
| 97 |
+
persisted; empty, malformed and unsupported entries are ignored.
|
| 98 |
+
"""
|
| 99 |
+
providers: list[ProviderConfig] = []
|
| 100 |
+
for definition in _PROVIDER_DEFS:
|
| 101 |
+
raw_tokens = credentials.get(definition["name"], [])
|
| 102 |
+
if not isinstance(raw_tokens, list):
|
| 103 |
+
continue
|
| 104 |
+
seen_tokens: set[str] = set()
|
| 105 |
+
for raw_token in raw_tokens:
|
| 106 |
+
if not isinstance(raw_token, str):
|
| 107 |
+
continue
|
| 108 |
+
token = raw_token.strip()
|
| 109 |
+
if not token or token in seen_tokens:
|
| 110 |
+
continue
|
| 111 |
+
seen_tokens.add(token)
|
| 112 |
+
providers.append(ProviderConfig(
|
| 113 |
+
id=-(10_000 + len(providers)),
|
| 114 |
+
name=definition["name"],
|
| 115 |
+
api_key=token,
|
| 116 |
+
base_url=definition["base_url"],
|
| 117 |
+
default_model=os.getenv(definition["model_env"], definition["default_model"]),
|
| 118 |
+
tier=definition["tier"],
|
| 119 |
+
purpose=definition["purpose"],
|
| 120 |
+
profile=f"byok-{len(seen_tokens)}",
|
| 121 |
+
))
|
| 122 |
+
return providers
|
| 123 |
+
|
| 124 |
def _load_providers(self) -> list[ProviderConfig]:
|
| 125 |
"""Carica la flotta: prova Supabase (tabella `ai_providers`, source of
|
| 126 |
truth dichiarata in supabase/migrations/20260711_ai_providers_fleet.sql),
|
|
|
|
| 413 |
return provider
|
| 414 |
|
| 415 |
async def chat(self, messages: list, model: Optional[str] = None, temperature: float = 0.7, max_tokens: int = 4096) -> str:
|
| 416 |
+
if not self.providers: self.providers = self._byok_providers + self._load_providers()
|
| 417 |
|
| 418 |
# S-CACHE-1: Lookup semantico preventivo
|
| 419 |
cached = await get_cached_response(messages)
|
|
|
|
| 502 |
|
| 503 |
async def stream_chat(self, messages: list, model: Optional[str] = None, temperature: float = 0.7, max_tokens: int = 4096) -> AsyncIterator[str]:
|
| 504 |
"""Streaming con failover immediato tra i provider configurati."""
|
| 505 |
+
if not self.providers: self.providers = self._byok_providers + self._load_providers()
|
| 506 |
# Nessun provider configurato: feedback immediato all'utente invece di
|
| 507 |
# cadere silenziosamente nel loop vuoto e dare un messaggio generico.
|
| 508 |
if not self.providers:
|
tests/test_ai_client_byok.py
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import sys
|
| 2 |
+
import types
|
| 3 |
+
import unittest
|
| 4 |
+
from unittest.mock import patch
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
_semantic_cache = types.ModuleType("api.semantic_cache")
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
async def _cache_miss(*_args, **_kwargs):
|
| 11 |
+
return None
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
async def _cache_noop(*_args, **_kwargs):
|
| 15 |
+
return None
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
_semantic_cache.get_cached_response = _cache_miss
|
| 19 |
+
_semantic_cache.set_cached_response = _cache_noop
|
| 20 |
+
sys.modules.setdefault("api.semantic_cache", _semantic_cache)
|
| 21 |
+
|
| 22 |
+
from models.ai_client import AIClient
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class AIClientByokTests(unittest.TestCase):
|
| 26 |
+
def test_byok_profiles_are_preferred_and_do_not_require_runtime_environment(self):
|
| 27 |
+
with patch.object(AIClient, "_load_providers", return_value=[]):
|
| 28 |
+
client = AIClient(byok_credentials={
|
| 29 |
+
"groq": [" gsk-primary ", "gsk-primary", "gsk-fallback"],
|
| 30 |
+
"openrouter": ["sk-or-primary"],
|
| 31 |
+
})
|
| 32 |
+
|
| 33 |
+
self.assertEqual([provider.name for provider in client.providers], [
|
| 34 |
+
"groq", "groq", "openrouter",
|
| 35 |
+
])
|
| 36 |
+
self.assertEqual([provider.profile for provider in client.providers], [
|
| 37 |
+
"byok-1", "byok-2", "byok-1",
|
| 38 |
+
])
|
| 39 |
+
self.assertEqual([provider.api_key for provider in client.providers], [
|
| 40 |
+
"gsk-primary", "gsk-fallback", "sk-or-primary",
|
| 41 |
+
])
|
| 42 |
+
self.assertTrue(all(provider.id < 0 for provider in client.providers))
|
| 43 |
+
|
| 44 |
+
def test_unknown_or_malformed_credentials_are_ignored(self):
|
| 45 |
+
with patch.object(AIClient, "_load_providers", return_value=[]):
|
| 46 |
+
client = AIClient(byok_credentials={
|
| 47 |
+
"unsupported": ["secret"],
|
| 48 |
+
"groq": "not-a-list", # type: ignore[dict-item]
|
| 49 |
+
"openrouter": ["", " "],
|
| 50 |
+
})
|
| 51 |
+
|
| 52 |
+
self.assertEqual(client.providers, [])
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
if __name__ == "__main__":
|
| 56 |
+
unittest.main()
|