feat: 接入模型动态 Fallback 与调用审计

core/backend/apps/ai/free_video.py

@@ -15,6 +15,7 @@ import logging
 import re
 import uuid
 from datetime import timedelta
+from decimal import Decimal
 from io import BytesIO
 
 from django.conf import settings
@@ -272,18 +273,33 @@ def _reap_stale_free_video_tasks(*, team) -> None:
     · SUBMITTED/POLLING 超 2 小时:轮询链早已断且无人认领(正常出片 5-10 分钟)→ 标失败退费;
     · POSTPROCESSING 超 30 分钟:转存/结算中途崩溃 → 标失败退费(火山可能已出片,平台承担该笔成本)。"""
     now = timezone.now()
+    from .routing_policy import load_model_routing_policy
+
+    video_policy = load_model_routing_policy().video
     buckets = [
-        ([AITask.Status.RESERVED], now - timedelta(minutes=10), "任务未在预期时间内提交(自动回收)"),
-        ([AITask.Status.SUBMITTED, AITask.Status.POLLING], now - timedelta(hours=2), "生成超时(自动回收)"),
-        ([AITask.Status.POSTPROCESSING], now - timedelta(minutes=30), "视频结果处理超时(自动回收)"),
+        (
+            [AITask.Status.RESERVED],
+            {"updated_at__lt": now - timedelta(seconds=video_policy.submit_total_timeout)},
+            "任务未在配置的提交总时限内完成(自动回收)",
+        ),
+        (
+            [AITask.Status.SUBMITTED, AITask.Status.POLLING],
+            {"submitted_at__lt": now - timedelta(seconds=video_policy.generation_timeout)},
+            "生成超过配置的成片等待总时限(自动回收)",
+        ),
+        (
+            [AITask.Status.POSTPROCESSING],
+            {"updated_at__lt": now - timedelta(minutes=30)},
+            "视频结果处理超时(自动回收)",
+        ),
     ]
-    for statuses, cutoff, reason in buckets:
+    for statuses, stale_filter, reason in buckets:
         stale = AITask.objects.filter(
             team=team,
             project__isnull=True,
             task_type=AITask.Type.FREE_VIDEO,
             status__in=statuses,
-            updated_at__lt=cutoff,
+            **stale_filter,
         )
         for task in stale:
             try:
@@ -404,6 +420,7 @@ def submit_free_video(*, team, user, params: dict) -> AITask:
         "price_multiplier": quote.meta.get("price_multiplier", "1"),
         "points_per_yuan_snapshot": quote.meta.get("rate", ""),
         "references": built["snapshots"],
+        "model_routing_v1": True,
     }
 
     # 建任务 + 预留同一事务:余额不足/限额拦截时回滚任务行,不留半套
@@ -418,7 +435,7 @@ def submit_free_video(*, team, user, params: dict) -> AITask:
             idempotency_key=f"free_video:{team.id}:{uuid.uuid4()}",
             request_payload=request_payload,
             estimated_cost=quote.points,
-            base_cost=quote.base_cost_yuan,
+            base_cost=Decimal("0"),
         )
         try:
             reserve_credit(team=team, user=user, task=task, amount=reserve_amount)
@@ -431,26 +448,44 @@ def submit_free_video(*, team, user, params: dict) -> AITask:
 
     # 火山调用在事务外(不持锁调外网)
     try:
-        from .services import build_provider
+        from .services import execute_routed_video_submit
 
-        provider = build_provider(model_config)
-        response = provider.create_video_task(
-            model=model_config.name,
-            endpoint=model_config.endpoint,
+        routed = execute_routed_video_submit(
+            task=task,
+            primary_model=model_config,
             prompt=built["api_prompt"],
             ratio=aspect_ratio,
             duration=duration,
             resolution=resolution,
-            generate_audio=generate_audio,
+            reference_images=[],
             content_items=built["content_items"],
+            pricing_references=built["snapshots"],
+            generate_audio=generate_audio,
             seed=seed if seed != -1 else None,
             search_mode=search_mode,
+            request_summary={"feature": "free_video", "mode": mode},
         )
-        task.provider_task_id = str(response.get("id") or response.get("task_id") or "")
+        response, provider_task_id = routed.value
+        # execute_model_call 以原子 F 表达式累计实际尝试的平台成本；刷新内存对象，
+        # 保证本方法返回值与数据库中的 AITask.base_cost 完全一致。
+        task.refresh_from_db(fields=["base_cost"])
+        task.provider_task_id = provider_task_id
         task.response_payload = response
+        payload = dict(task.request_payload or {})
+        payload["actual_model_config_id"] = str(routed.actual_model.id)
+        task.request_payload = payload
         task.status = AITask.Status.SUBMITTED
         task.submitted_at = timezone.now()
-        task.save(update_fields=["provider_task_id", "response_payload", "status", "submitted_at", "updated_at"])
+        task.save(
+            update_fields=[
+                "provider_task_id",
+                "response_payload",
+                "request_payload",
+                "status",
+                "submitted_at",
+                "updated_at",
+            ]
+        )
     except Exception as exc:  # noqa: BLE001 — 创建失败:标失败退费,返回失败卡(不向上抛)
         code, raw_message = parse_provider_error(exc)
         public_error = classify_generation_error(
@@ -577,11 +612,47 @@ def finalize_free_video(*, task: AITask) -> AITask:
     if not task.provider_task_id:
         return task
 
-    from .services import build_provider
+    from .routing_policy import load_model_routing_policy
+    from .services import get_video_provider
 
-    provider = build_provider(task.model_config)
+    video_policy = load_model_routing_policy().video
+    if task.submitted_at and (
+        timezone.now() - task.submitted_at
+    ).total_seconds() >= video_policy.generation_timeout:
+        timeout_message = "视频生成超过配置的成片等待总时限"
+        public_error = classify_generation_error(
+            TimeoutError(timeout_message),
+            operation="video_generate",
+            reference_id=str(task.id),
+        )
+        with transaction.atomic():
+            locked = AITask.objects.select_for_update().get(id=task.id)
+            if locked.status not in (AITask.Status.SUBMITTED, AITask.Status.POLLING):
+                return locked
+            locked.status = AITask.Status.FAILED
+            locked.error_code = "GenerationTimeout"
+            locked.error_message = timeout_message
+            locked.completed_at = timezone.now()
+            locked.save(
+                update_fields=["status", "error_code", "error_message", "completed_at", "updated_at"]
+            )
+            release_credit(reservation=locked.credit_reservation, reason=timeout_message)
+        _notify_failure(locked, raw=timeout_message, hint=public_error.fallback_message)
+        return locked
+
+    submit_attempt = (
+        task.model_attempts.filter(status="succeeded", operation="video_generate")
+        .select_related("model_config__provider")
+        .order_by("-sequence")
+        .first()
+    )
+    actual_model = submit_attempt.model_config if submit_attempt and submit_attempt.model_config else task.model_config
+
+    provider = get_video_provider(actual_model)
     response = provider.poll_video_task(
-        endpoint=task.model_config.endpoint, provider_task_id=task.provider_task_id
+        endpoint=actual_model.endpoint,
+        provider_task_id=task.provider_task_id,
+        timeout=video_policy.poll_request_timeout,
     )
     remote_status = str(response.get("status") or "")
 
@@ -647,7 +718,7 @@ def finalize_free_video(*, task: AITask) -> AITask:
             from decimal import Decimal
 
             settle = quote_video_actual(
-                locked.model_config, tokens=total_tokens, with_video_ref=with_video_ref, resolution=resolution,
+                actual_model, tokens=total_tokens, with_video_ref=with_video_ref, resolution=resolution,
                 multiplier=Decimal(str(payload.get("price_multiplier") or "1")),
             )
             actual, base_cost = settle.points, settle.base_cost_yuan