fix(billing): unify quota counters, fix Stripe webhooks, tier/plan sync

routes/auth_routes.py

@@ -104,13 +104,14 @@ async def require_user(credentials: HTTPAuthorizationCredentials = Depends(secur
 def user_to_response(user) -> UserResponse:
     """Convert User to UserResponse with plan limits"""
     plan_limits = PLANS[user.plan]
+    tier = getattr(user, "tier", None) or user.plan
     return UserResponse(
         id=user.id,
         email=user.email,
         name=user.name,
         avatar_url=user.avatar_url,
         plan=user.plan,
-        tier=user.plan,
+        tier=tier,
         subscription_status=user.subscription_status,
         subscription_ends_at=getattr(user, 'subscription_ends_at', None),
         cancel_at_period_end=getattr(user, 'cancel_at_period_end', False),
@@ -540,7 +541,7 @@ async def login_v1(request: Request):
         )
 
     access_token = create_access_token(
-        user.id, tier=user.plan.value, expires_delta=timedelta(minutes=15)
+        user.id, tier=getattr(user, "tier", user.plan.value), expires_delta=timedelta(minutes=15)
     )
     refresh_token = create_refresh_token(user.id, expires_delta=timedelta(days=7))
 
@@ -626,7 +627,7 @@ async def google_auth_v1(body: GoogleAuthRequest):
             content={"error": "USER_CREATE_FAILED", "message": str(exc)},
         )
 
-    access_token = create_access_token(user.id)
+    access_token = create_access_token(user.id, tier=getattr(user, "tier", "free"))
     refresh_token = create_refresh_token(user.id)
 
     return JSONResponse(
@@ -724,7 +725,7 @@ async def refresh_v1(request: Request):
         )
 
     access_token = create_access_token(
-        user.id, tier=user.plan.value, expires_delta=timedelta(minutes=15)
+        user.id, tier=getattr(user, "tier", user.plan.value), expires_delta=timedelta(minutes=15)
     )
     new_refresh_token = create_refresh_token(user.id, expires_delta=timedelta(days=7))
 

routes/translate_routes.py

@@ -44,8 +44,13 @@ from typing_extensions import Annotated
 from config import config
 from translators import ExcelTranslator, WordTranslator, PowerPointTranslator
 from models.subscription import PlanType
-from middleware.tier_quota import tier_quota_service
-from services.auth_service import record_usage, check_usage_limits
+from services.auth_service import (
+    record_usage,
+    check_usage_limits,
+    reserve_translation_quota,
+    release_translation_quota,
+)
+from middleware.tier_quota import _seconds_until_next_month, _next_month_utc
 from middleware.validation import FileValidator, ValidationError, LanguageValidator, webhook_validator
 from middleware.api_key_auth import get_authenticated_user, get_user_from_api_key
 from utils import file_handler
@@ -548,6 +553,7 @@ async def translate_document_v1(
     """
     request_id = getattr(request.state, "request_id", str(uuid.uuid4())[:8])
 
+    quota_reserved = False
     try:
         if not file and not file_url:
             raise TranslateEndpointError(
@@ -615,40 +621,42 @@ async def translate_document_v1(
                 )
 
         if current_user:
-            quota = await tier_quota_service.check_quota(user_id, tier)
-            if not quota.allowed:
-                retry_after = tier_quota_service.seconds_until_reset()
-                raise HTTPException(
-                    status_code=429,
-                    detail={
-                        "error": "QUOTA_EXCEEDED",
-                        "message": f"Monthly limit reached ({quota.current_usage}/{quota.limit} documents). Upgrade your plan for more.",
-                        "details": {
-                            "current_usage": quota.current_usage,
-                            "limit": quota.limit,
-                            "tier": tier,
-                            "reset_at": quota.reset_at_utc.isoformat(),
-                        },
-                    },
-                    headers={"Retry-After": str(retry_after)},
-                )
-            
-            # Strict database plan limit check (Starter, Pro, Business, Enterprise)
             usage = check_usage_limits(current_user)
             if not usage["can_translate"]:
+                retry_after = _seconds_until_next_month()
                 raise HTTPException(
                     status_code=429,
                     detail={
                         "error": "QUOTA_EXCEEDED",
                         "message": f"Monthly limit reached ({usage['docs_used']}/{usage['docs_limit']} documents). Upgrade your plan for more.",
                         "details": {
-                            "current_usage": usage["docs_used"],
-                            "limit": usage["docs_limit"],
+                            "current_usage": usage['docs_used'],
+                            "limit": usage['docs_limit'],
                             "tier": tier,
+                            "reset_at": _next_month_utc().isoformat(),
                         },
                     },
+                    headers={"Retry-After": str(retry_after)},
                 )
-            rate_limit_remaining = quota.remaining
+            # Atomically reserve one document slot now so concurrent requests cannot
+            # overshoot the monthly quota while background jobs are still running.
+            reserved = await asyncio.to_thread(reserve_translation_quota, user_id)
+            if not reserved:
+                retry_after = _seconds_until_next_month()
+                raise HTTPException(
+                    status_code=429,
+                    detail={
+                        "error": "QUOTA_EXCEEDED",
+                        "message": "Monthly limit reached. Upgrade your plan for more.",
+                        "details": {
+                            "tier": tier,
+                            "reset_at": _next_month_utc().isoformat(),
+                        },
+                    },
+                    headers={"Retry-After": str(retry_after)},
+                )
+            quota_reserved = True
+            rate_limit_remaining = usage["docs_remaining"]
         else:
             rate_limit_remaining = -1
 
@@ -839,6 +847,8 @@ async def translate_document_v1(
         )
 
     except TranslateEndpointError as e:
+        if quota_reserved and user_id:
+            await asyncio.to_thread(release_translation_quota, user_id)
         status_code = 400
         if e.code == TranslateEndpointError.FILE_TOO_LARGE:
             status_code = 413
@@ -852,8 +862,12 @@ async def translate_document_v1(
             content=e.to_dict(),
         )
     except HTTPException:
+        if quota_reserved and user_id:
+            await asyncio.to_thread(release_translation_quota, user_id)
         raise
     except Exception as e:
+        if quota_reserved and user_id:
+            await asyncio.to_thread(release_translation_quota, user_id)
         logger.error(f"[{request_id}] Unexpected error: {e}")
         return JSONResponse(
             status_code=400,
@@ -973,6 +987,7 @@ async def _run_translation_job(
         return
 
     tracker = ProgressTracker(job_id, _translation_jobs)
+    usage_recorded = False
 
     try:
         job["status"] = "processing"
@@ -1337,14 +1352,14 @@ async def _run_translation_job(
                 else:
                     cost_factor = 5
 
-            for _ in range(cost_factor):
-                await tier_quota_service.increment_on_success(user_id)
-
             # Persist monthly usage counters in PostgreSQL (docs + pages)
             pages = await asyncio.to_thread(
                 _estimate_pages, input_path, file_extension
             )
-            await asyncio.to_thread(record_usage, user_id, pages, False, cost_factor)
+            await asyncio.to_thread(
+                record_usage, user_id, pages, cost_factor, reserved_docs=1
+            )
+            usage_recorded = True
             logger.info(f"Job {job_id}: usage recorded — {pages} page(s) with cost factor {cost_factor}")
 
             # Apply watermark for Free-tier users
@@ -1364,6 +1379,12 @@ async def _run_translation_job(
         record_translation(provider=provider, file_type=file_extension or "unknown", duration=duration, status="success")
         logger.info(f"Job {job_id}: Completed successfully")
 
+    except asyncio.CancelledError:
+        # Background task cancelled (e.g. TestClient teardown or server shutdown).
+        # The document slot was already reserved at request time; keep it consumed
+        # so quota enforcement remains deterministic.
+        logger.warning(f"Job {job_id}: translation task cancelled, keeping reserved quota")
+        raise
     except Exception as e:
         # Check if this is our structured TranslationProviderError
         if type(e).__name__ == "TranslationProviderError":
@@ -1376,6 +1397,13 @@ async def _run_translation_job(
         # Record translation failure metric
         record_translation(provider=provider, file_type=file_extension or "unknown", duration=0, status="error")
 
+        if user_id and not usage_recorded:
+            try:
+                await asyncio.to_thread(release_translation_quota, user_id)
+                logger.info(f"Job {job_id}: released reserved quota after failure")
+            except Exception as release_err:
+                logger.exception(f"Job {job_id}: failed to release reserved quota: {release_err}")
+
     finally:
         if webhook_url:
             try: