fix(core): 脚本 agent 用通用字段解析,根治模型 schema 漂移致旁白/画面空

模型每次生成都换字段名(scene/screenDescription/画面…;dialogue 字符串 /
lines:[{role,content}] / caption…),逐一追变体是治标。改为通用解析:
- _pick_field:优先键命中 → 否则按关键词模糊匹配,跳过 bgMusic/note/shotNo 等非内容键
- 画面来源:visual/scene/screenDescription/画面… 任意命中
- 旁白来源:结构化 dialogue/lines 优先,其次整句 dialogue,再退 narration/voiceover/caption/字幕
实测覆盖 4 种真实/契约变体,全部正确填充;新增通用解析回归用例,23 测试通过。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

core/backend/apps/ai/script_agent.py

@@ -211,6 +211,35 @@ def _nearest_duration(value) -> int:
     return min(DURATION_TIERS, key=lambda t: abs(t - value))
 
 
+# 模型每次生成都可能换字段名(scene/screenDescription/visual…、dialogue/lines/caption…),
+# 与其逐一追变体,不如「优先键命中 → 否则按关键词模糊匹配」通用解析。SKIP 掉明显的非内容键,
+# 避免误抓(shotNo/duration/bgMusic/note 等)。
+_PICK_SKIP_KEYS = {
+    "shotno", "shotsize", "shotsizetype", "duration", "starttime", "endtime", "timerange",
+    "bgmusic", "bg_music", "music", "sound", "sfx", "note", "notes", "tips", "index",
+    "role", "speaker", "entity_refs", "product_exposure", "id", "no", "transition",
+}
+_VISUAL_EXACT = ("visual", "visual_prompt", "visual_description", "scene", "screen_description", "screendescription", "screen", "picture", "shot_description")
+_VISUAL_FUZZY = ("scene", "screen", "visual", "picture", "画面", "镜头描述", "分镜画面", "描述")
+_NARRATION_EXACT = ("narration", "voiceover", "voice_over", "line", "caption", "subtitle", "speech", "口播", "旁白")
+_NARRATION_FUZZY = ("narrat", "voice", "旁白", "口播", "台词", "dialog", "caption", "subtitle", "字幕", "speech", "monolog")
+
+
+def _pick_field(seg: dict, exact: tuple[str, ...], fuzzy: tuple[str, ...]) -> str:
+    """从一镜里取某类文本字段:先按优先键精确命中,再按关键词在剩余键里模糊匹配(跳过非内容键)。"""
+    for key in exact:
+        val = seg.get(key)
+        if isinstance(val, str) and val.strip():
+            return val.strip()
+    for key, val in seg.items():
+        kl = str(key).lower()
+        if kl in _PICK_SKIP_KEYS or not (isinstance(val, str) and val.strip()):
+            continue
+        if any(f in kl for f in fuzzy):
+            return val.strip()
+    return ""
+
+
 def normalize_draft(raw_text: str, *, aspect_ratio: str, total_duration: int) -> dict:
     """把模型输出抽成 JSON 并按铁律1契约规范化。宽容:小问题就地修,不轻易抛错。"""
     blob = _extract_json(raw_text)
@@ -301,16 +330,16 @@ def normalize_draft(raw_text: str, *, aspect_ratio: str, total_duration: int) ->
                 continue
             sp = d.get("speaker")
             dialogue.append({"speaker": sp if sp in valid_ids else None, "line": line})
-        # 旁白:narration > 字符串 dialogue(整句口播)> 结构化对白/lines 拼接 > subtitle/caption(字幕兜底)
+        # 旁白:结构化对白/lines 优先,其次整句字符串 dialogue,再退到通用字段解析(narration/voiceover/caption/字幕…)
-        narration = (seg.get("narration") or seg.get("voiceover") or "").strip()
+        narration = ""
-        if not narration and isinstance(raw_dialogue, str):
+        if isinstance(raw_dialogue, str) and raw_dialogue.strip():
             narration = raw_dialogue.strip()
-        if not narration and dialogue:
+        elif dialogue:
             narration = " ".join(d["line"] for d in dialogue)  # 扁平拼接,兼容下游字幕/配音
         if not narration:
-            narration = (seg.get("subtitle") or seg.get("caption") or "").strip()
+            narration = _pick_field(seg, _NARRATION_EXACT, _NARRATION_FUZZY)
-        # 画面:visual > visual_prompt > scene > shot/visual_description(模型变体常用 scene 描述画面)
+        # 画面:通用解析(visual/scene/screenDescription/画面… 都能命中,跳过 shotNo/bgMusic 等)
-        visual = (seg.get("visual") or seg.get("visual_prompt") or seg.get("scene") or seg.get("visual_description") or "").strip()
+        visual = _pick_field(seg, _VISUAL_EXACT, _VISUAL_FUZZY)
         norm_segments.append(
             {
                 "index": i,

core/backend/apps/ai/tests.py

@@ -51,6 +51,25 @@ class NormalizeDraftTests(SimpleTestCase):
         self.assertIn("防晒泛白太尴尬", draft["segments"][0]["narration"])
         self.assertEqual(len(draft["segments"][0]["dialogue"]), 2)  # lines→结构化对白
 
+    def test_generic_resolver_covers_unseen_field_names(self):
+        """模型每次换字段名(scene/screenDescription/画面…、dialogue/lines/caption…)。
+        通用解析应「优先键 + 关键词模糊匹配」都能填,且跳过 bgMusic/note 等非内容键。"""
+        variants = {
+            "canonical": {"total_duration": 15, "segments": [{"role": "钩子", "narration": "口播A", "visual": "画面A"}]},
+            "scene+dialogue_str+subtitle": {"total_duration": 15, "shots": [{"scene": "画面B", "dialogue": "口播B", "subtitle": "字幕B"}]},
+            "scene+lines+caption": {"total_duration": 15, "shots": [{"scene": "画面C", "lines": [{"role": "女主", "content": "口播C"}], "caption": "字幕C"}]},
+            "screenDescription+dialogue+bgMusic+note": {"total_duration": 15, "shots": [{"screenDescription": "画面D", "dialogue": "口播D", "bgMusic": "音乐D", "note": "备注D"}]},
+        }
+        for name, raw in variants.items():
+            draft = normalize_draft(json.dumps(raw, ensure_ascii=False), aspect_ratio="9:16", total_duration=15)
+            seg = draft["segments"][0]
+            self.assertTrue(seg["narration"], f"{name}: 旁白为空")
+            self.assertTrue(seg["visual"], f"{name}: 画面为空")
+        # bgMusic/note 不得被误当画面/旁白
+        d = normalize_draft(json.dumps(variants["screenDescription+dialogue+bgMusic+note"], ensure_ascii=False), aspect_ratio="9:16", total_duration=15)
+        self.assertEqual(d["segments"][0]["visual"], "画面D")
+        self.assertEqual(d["segments"][0]["narration"], "口播D")
+
     def test_string_dialogue_not_iterated_as_chars(self):
         """dialogue 为整句字符串时,要当作旁白而不是逐字符遍历。"""
         raw = json.dumps({