feat: 接入模型动态 Fallback 与调用审计

core/backend/apps/ai/script_agent.py

@@ -20,6 +20,7 @@ from __future__ import annotations
 
 import json
 import re
+from decimal import Decimal
 from functools import lru_cache
 from pathlib import Path
 
@@ -574,7 +575,7 @@ def stream_script_agent(
 ):
     """生成 SSE 帧字符串的同步生成器,供 StreamingHttpResponse 包裹。
     target_index 非空 = 精准只改第 N 镜(读全脚本上下文,后端强制保留其余镜原样)。"""
-    from apps.ai.services import build_provider, create_ai_task
+    from apps.ai.services import create_ai_task, stream_routed_text_request
 
     yield _sse({"type": "tool", "id": "skill", "label": "加载电商脚本技能", "status": "running"})
     skill_loaded = bool(load_ecommerce_skill())
@@ -619,6 +620,9 @@ def stream_script_agent(
                 "mode": mode,
                 "aspect_ratio": aspect_ratio,
                 "total_duration": total_duration,
+                "base_version_id": str(base_version_id or ""),
+                "target_index": target_index,
+                "model_routing_v1": True,
             },
         )
     except Exception as exc:  # noqa: BLE001 — 多为额度不足
@@ -639,13 +643,45 @@ def stream_script_agent(
             task.status = AITask.Status.SUBMITTED
             task.submitted_at = timezone.now()
             task.save(update_fields=["status", "submitted_at", "updated_at"])
-            provider = build_provider(model_config)
-            for ev in provider.chat_completion_stream(
-                model=model_config.name,
-                endpoint=model_config.endpoint,
+            def validate_script_text(raw_text: str) -> dict:
+                candidate = normalize_draft(
+                    raw_text,
+                    aspect_ratio=aspect_ratio,
+                    total_duration=effective_duration,
+                )
+                if target_index is not None and base_draft:
+                    return _merge_single_segment(
+                        base_draft,
+                        candidate,
+                        target_index,
+                        aspect_ratio,
+                        effective_duration,
+                    )
+                return candidate
+
+            routed_stream = stream_routed_text_request(
+                task=task,
+                primary_model=model_config,
                 messages=messages,
+                streaming=True,
+                structured_output=True,
+                business_operation="script_generate",
                 temperature=0.85,
-            ):
+                validate_text=validate_script_text,
+                request_summary={
+                    "mode": mode,
+                    "target_index": target_index,
+                    "base_version_id": str(base_version_id or ""),
+                    "aspect_ratio": aspect_ratio,
+                    "total_duration": effective_duration,
+                },
+            )
+            while True:
+                try:
+                    ev = next(routed_stream)
+                except StopIteration as completed:
+                    routed = completed.value
+                    break
                 et = ev.get("type")
                 if et == "reasoning":
                     # 思考流:推理模型在出 JSON 前会先想很久,把思考逐字下发给前端(像对话一样可见),
@@ -668,12 +704,8 @@ def stream_script_agent(
                             if piece.strip():
                                 yield _sse({"type": "delta", "text": piece})
                 elif et == "done":
-                    break
-            raw = "".join(full)
-            draft = normalize_draft(raw, aspect_ratio=aspect_ratio, total_duration=effective_duration)
-            if target_index is not None and base_draft:
-                # 精准改一镜:只采用新稿的第 target_index 镜,其余镜强制保持基准稿原样
-                draft = _merge_single_segment(base_draft, draft, target_index, aspect_ratio, effective_duration)
+                    continue
+            raw, _provider_response, draft = routed.value
         except Exception as exc:  # noqa: BLE001
             _fail_task(task, reservation, str(exc))
             settled = True
@@ -819,7 +851,7 @@ def regenerate_segment_via_agent(*, project, user, model_config: ModelConfig, se
     与 stream_script_agent 的 target_index 分支同源,但同步返回(不走 SSE)。计费 reserve→charge/release 闭环。"""
     from django.db import transaction
 
-    from apps.ai.services import build_provider, create_ai_task
+    from apps.ai.services import create_ai_task, execute_routed_text_request
     from apps.billing.services.ledger import charge_reserved_credit
 
     base_draft = _draft_from_version(segment.script_version)  # 用 DB 行重建基准,别用可能 stale 的 content
@@ -850,18 +882,40 @@ def regenerate_segment_via_agent(*, project, user, model_config: ModelConfig, se
             "endpoint": model_config.endpoint,
             "mode": "revise",
             "target_index": target_index,
+            "model_routing_v1": True,
         },
     )
     reservation = task.credit_reservation
+    # 每条真实尝试的平台成本由统一执行器累计；用户积分仍只结算这一条脚本任务。
+    task.base_cost = Decimal("0")
+    task.save(update_fields=["base_cost", "updated_at"])
+    # 实际平台成本由每条 AIModelAttempt 累加；用户积分仍只结算这一条逻辑任务。
+    task.base_cost = Decimal("0")
+    task.save(update_fields=["base_cost", "updated_at"])
     try:
         task.status = AITask.Status.SUBMITTED
         task.submitted_at = timezone.now()
         task.save(update_fields=["status", "submitted_at", "updated_at"])
-        provider = build_provider(model_config)
-        response = provider.chat_completion(model=model_config.name, endpoint=model_config.endpoint, messages=messages)
-        raw = provider.extract_text(response)
-        draft = normalize_draft(raw, aspect_ratio=aspect_ratio, total_duration=total_duration)
-        draft = _merge_single_segment(base_draft, draft, target_index, aspect_ratio, total_duration)
+        def validate_segment_text(raw_text: str) -> dict:
+            candidate = normalize_draft(raw_text, aspect_ratio=aspect_ratio, total_duration=total_duration)
+            return _merge_single_segment(base_draft, candidate, target_index, aspect_ratio, total_duration)
+
+        routed = execute_routed_text_request(
+            task=task,
+            primary_model=model_config,
+            messages=messages,
+            streaming=False,
+            structured_output=True,
+            business_operation="script_generate",
+            temperature=0.3,
+            validate_text=validate_segment_text,
+            request_summary={
+                "mode": "revise",
+                "target_index": target_index,
+                "base_version_id": str(segment.script_version_id),
+            },
+        )
+        raw, _response, draft = routed.value
         with transaction.atomic():
             task.status = AITask.Status.SUCCEEDED
             task.response_payload = {"raw": raw[:8000]}
@@ -871,6 +925,6 @@ def regenerate_segment_via_agent(*, project, user, model_config: ModelConfig, se
             charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost)
             script = persist_script_draft(project=project, user=user, task=task, draft=draft, source="revise")
         return script
-    except Exception:
-        _fail_task(task, reservation, "单镜重跑失败")
+    except Exception as exc:
+        _fail_task(task, reservation, str(exc) or "单镜重跑失败")
         raise