Merge pull request #109 from shy3130/fix/ext-concepts-cast-and-meta-normalize

fix: 扩展数据 List cast 报错 + AI 策略 META 规范化失败

backend/app/api/strategy.py

@@ -313,14 +313,24 @@ def _py_string(value: str) -> str:
 
 
 def _find_meta_dict(code: str) -> ast.Dict:
+    # 兼容两种 LLM 常见写法:
+    #   META = {...}        → ast.Assign
+    #   META: dict = {...}  → ast.AnnAssign (类型注解, 合法但旧逻辑漏匹配)
     tree = ast.parse(code)
     for node in ast.walk(tree):
+        value = None
         if isinstance(node, ast.Assign):
             for target in node.targets:
                 if isinstance(target, ast.Name) and target.id == "META":
-                    if not isinstance(node.value, ast.Dict):
-                        raise ValueError("META 必须是字面量字典")
-                    return node.value
+                    value = node.value
+                    break
+        elif isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name) \
+                and node.target.id == "META":
+            value = node.value
+        if value is not None:
+            if not isinstance(value, ast.Dict):
+                raise ValueError("META 必须是字面量字典")
+            return value
     raise ValueError("找不到 META 字典")
 
 

backend/app/services/ext_data.py

@@ -451,10 +451,23 @@ def parse_upload_file(file_path: Path, symbol_col: str = "symbol", data_dir: Pat
 
 
 def cast_df_to_schema(df: pl.DataFrame, fields: list[ExtField]) -> pl.DataFrame:
-    """按配置的字段类型转换 DataFrame 列类型。"""
+    """按配置的字段类型转换 DataFrame 列类型。
+
+    List → string 的处理: 上游接口常返回数组字段 (如 concepts: ["AI", "芯片"]),
+    若声明为 string, 直接 cast 会抛 `cannot cast List type`。
+    这里把列表元素先转字符串再以分号拼接, 与 _flatten_concept_rows 保持一致。
+    """
+    schema = df.schema
     for f in fields:
-        if f.name in df.columns:
-            target = _POLARS_DTYPE_MAP.get(f.dtype, pl.Utf8)
+        if f.name not in df.columns:
+            continue
+        target = _POLARS_DTYPE_MAP.get(f.dtype, pl.Utf8)
+        src = schema[f.name]
+        if isinstance(src, pl.List) and target == pl.Utf8:
+            df = df.with_columns(
+                pl.col(f.name).cast(pl.List(pl.Utf8)).list.join(";").cast(target)
+            )
+        else:
             df = df.with_columns(pl.col(f.name).cast(target))
     return df
 

backend/app/strategy/ai_generator.py

@@ -144,14 +144,25 @@ class AIStrategyGenerator:
 
     @staticmethod
     def _extract_meta(code: str) -> dict:
-        """从代码字符串中提取 META 字典（不执行代码, 仅接受字面量）"""
+        """从代码字符串中提取 META 字典（不执行代码, 仅接受字面量）
+
+        兼容两种声明: META = {...} (Assign) 和 META: dict = {...} (AnnAssign)。
+        与 api.strategy._find_meta_dict 保持同一套匹配逻辑。
+        """
         tree = ast.parse(code)
         for node in ast.walk(tree):
+            value = None
             if isinstance(node, ast.Assign):
                 for target in node.targets:
                     if isinstance(target, ast.Name) and target.id == "META":
-                        try:
-                            return ast.literal_eval(node.value)
-                        except (ValueError, SyntaxError) as e:
-                            raise ValueError(f"META 必须是纯字面量字典: {e}") from e
+                        value = node.value
+                        break
+            elif isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name) \
+                    and node.target.id == "META":
+                value = node.value
+            if value is not None:
+                try:
+                    return ast.literal_eval(value)
+                except (ValueError, SyntaxError) as e:
+                    raise ValueError(f"META 必须是纯字面量字典: {e}") from e
         return {}

backend/tests/test_ai_strategy_meta_normalize.py

@@ -76,3 +76,65 @@ def filter(df: pl.DataFrame, params: dict) -> pl.Expr:
     assert '"id": "ai_inserted"' in code
     assert '"name": "中文名"' in code
     assert '"description": "描述"' in code
+
+
+# --- 回归: LLM 偏移写法 -------------------------------------------------
+# 模型常给 META 加类型注解 (META: dict = {...}, ast.AnnAssign 节点)。
+# 旧版匹配器只遍历 ast.Assign, 漏掉注解形式 → 报「找不到 META 字典」。
+
+ANNOTATED_CODE = '''"""模型返回的策略 (带类型注解的 META — LLM 常见偏移)"""
+import polars as pl
+
+META: dict = {
+    "id": "annotated_wrong_id",
+    "name": "Placeholder",
+    "description": "model desc",
+    "tags": ["AI"],
+    "params": [],
+    "scoring": {},
+}
+
+def filter(df: pl.DataFrame, params: dict) -> pl.Expr:
+    return pl.lit(True)
+'''
+
+
+def test_find_meta_dict_accepts_type_annotated_form():
+    """META: dict = {...} 必须能被识别 (旧版会抛「找不到 META 字典」)。"""
+    from app.api.strategy import _find_meta_dict
+
+    node = _find_meta_dict(ANNOTATED_CODE)
+    assert node is not None  # 能找到就说明没抛异常
+
+
+def test_extract_meta_accepts_type_annotated_form():
+    from app.strategy.ai_generator import AIStrategyGenerator
+
+    meta = AIStrategyGenerator._extract_meta(ANNOTATED_CODE)
+    assert meta["id"] == "annotated_wrong_id"
+    assert meta["name"] == "Placeholder"
+
+
+def test_normalize_strategy_meta_works_on_annotated_form():
+    """端到端: AI 生成注解形式 META 时, 规范化不再报「规范化 META 失败」。"""
+    code = _normalize_strategy_meta(
+        ANNOTATED_CODE,
+        "ai_annotated_ok",
+        name="断板反包",
+        description="中文描述",
+    )
+    assert '"id": "ai_annotated_ok"' in code
+    assert '"name": "断板反包"' in code
+    assert '"description": "中文描述"' in code
+    assert "annotated_wrong_id" not in code
+
+
+def test_normalize_build_result_succeeds_on_annotated_form():
+    """模拟前端 /build/stream 的完整结果路径 (之前报错的入口)。"""
+    result = {"code": ANNOTATED_CODE, "meta": {}, "valid": True, "error": None}
+
+    normalized = _normalize_build_result(result, "ai_build_ok")
+
+    assert normalized["valid"] is True
+    assert normalized["error"] is None
+    assert normalized["meta"]["id"] == "ai_build_ok"