M4：db_schema 认整个写成一个字符串的 JSON 数组的表名

验收时盲算、反方审稿的模型把 table 写成 '["orders", "stores"]'（一个字符串），工具只按逗号拆，回「中没有 ["orders"」
「中没有 "stores"]」，两张表明明都在，模型重试几轮才改成不传 table。表名参数的解析抽成 _schema_tables：元素都是
字符串的 JSON 数组（整个作为一个字符串传进来，或者数组里的一项）按数组认，其余照旧按逗号拆；不是 JSON 的
（[orders]、单引号）、元素不是字符串的照旧。只动 db_schema 这一段，查询工具不动。

backend/app/tools/datasource.py

@@ -644,6 +644,27 @@ async def _store_schema(source: Any, ctx: ToolContext, *,
         return None
 
 
+def _schema_tables(table: str | list[str] | None) -> list[str]:
+    """db_schema 的 table 参数 → 要看的表名。"user, roles"、["user","roles"]、整个写成一个字符串的 JSON 数组
+    '["user", "roles"]' 是同一个意思，模型三种都会写，为此拒绝一次调用纯属浪费一步。最后那种只按逗号拆就成了
+    「没有 ["user"」「没有 "roles"]」，两张表明明都在（2026-10 验收：盲算、反方审稿的核查 agent 因此白白多调了几轮
+    模型）。只认元素都是字符串的 JSON 数组，别的写法照旧按逗号拆。"""
+    wanted = [table] if isinstance(table, str) else list(table or [])
+    names: list[str] = []
+    for one in wanted:
+        text = str(one).strip()
+        if text.startswith("[") and text.endswith("]"):
+            try:
+                parsed = json.loads(text)
+            except ValueError:
+                parsed = None
+            if isinstance(parsed, list) and all(isinstance(t, str) for t in parsed):
+                names += parsed
+                continue
+        names.append(str(one))
+    return [t.strip() for one in names for t in one.split(",") if t.strip()]
+
+
 def _make_schema_tool(source: Any, *, fixed: bool = False,
                       catalog: Mapping[str, data_catalog.CatalogEntry] | None = None) -> StructuredTool:
     """source 同 _make_query_tool：上传源的表结构取绑定快照里冻结的那份。
@@ -655,10 +676,7 @@ def _make_schema_tool(source: Any, *, fixed: bool = False,
     async def _run(table: str | list[str] | None = None) -> str:
         if gone := _vanished(source, fixed):
             return f"{DATA_UNAVAILABLE}{gone}"
-        wanted = [table] if isinstance(table, str) else list(table or [])
-        # 逗号分隔也认："user, roles" 和 ["user","roles"] 是同一个意思，
-        # 而模型两种都会写。为此拒绝一次调用，纯属浪费一步
-        wanted = [t.strip() for one in wanted for t in str(one).split(",") if t.strip()]
+        wanted = _schema_tables(table)
         if wanted:
             return "\n\n".join(introspect.describe_table(source, t, catalog=notes) for t in wanted)
         tables = introspect.table_names(source)

backend/tests/test_tool_args.py

@@ -334,6 +334,52 @@ async def test_schema_tool_still_lists_everything_when_asked_for_nothing(source)
     assert "张表" in out and "orders" in out
 
 
+@pytest.fixture
+async def depots(tmp_path):
+    """两张表的库（虚构的仓储）：parcels、depots，结构已探查。"""
+    from app.data import introspect
+
+    path = tmp_path / "depots.db"
+    db = sqlite3.connect(path)
+    db.executescript("CREATE TABLE parcels (id INTEGER PRIMARY KEY, depot_id INTEGER, fee REAL);"
+                     "CREATE TABLE depots (id INTEGER PRIMARY KEY, zone TEXT);")
+    db.commit()
+    db.close()
+    src = SimpleNamespace(id="tool-args-depots-src", name="depots", kind="sqlite", database=str(path), host=None,
+                          port=None, username=None, password=None, options={}, readonly=True, description="测试库",
+                          schema_cache={})
+    src.schema_cache = await introspect.introspect(src)
+    return src
+
+
+@pytest.mark.parametrize("written", ['["parcels", "depots"]', '["parcels","depots"]', ' [ "parcels" , "depots" ] ',
+                                     ['["parcels", "depots"]'], ['["parcels"]', "depots"]])
+async def test_schema_tool_reads_a_json_array_written_as_one_string(depots, written) -> None:
+    """2026-10 验收：模型把 table 写成一整个字符串 '["orders", "stores"]'（参数说明里的 ["user","roles"] 引得它这么写），
+    只按逗号拆就成了「没有 ["orders"」「没有 "stores"]」——两张都在，模型却白白多调了几轮。写成字符串的 JSON 数组
+    和逗号分隔一样认。"""
+    out = await _make_schema_tool(depots).coroutine(table=written)
+    assert out.count("表 parcels") == 1 and out.count("表 depots") == 1, out
+    assert "没有" not in out
+
+
+async def test_schema_tool_reads_a_one_table_json_array(depots) -> None:
+    out = await _make_schema_tool(depots).coroutine(table='["parcels"]')
+    assert out.startswith("表 parcels") and "depots" not in out and "没有" not in out
+    # 空数组和不传一样：列出所有表
+    assert "共 2 张表" in await _make_schema_tool(depots).coroutine(table="[]")
+
+
+@pytest.mark.parametrize("written,missing", [
+    ("[parcels]", "[parcels]"),         # 不是 JSON（名字没加引号）：照旧整个当表名
+    ("['parcels']", "['parcels']"),     # 单引号也不是 JSON
+    ("[1, 2]", "[1"),                   # 数组里不是表名：照旧按逗号拆
+])
+async def test_schema_tool_leaves_other_bracketed_text_as_before(depots, written, missing) -> None:
+    out = await _make_schema_tool(depots).coroutine(table=written)
+    assert f"中没有 {missing}。" in out
+
+
 def test_copilot_is_told_not_to_pin_max_steps() -> None:
     """Copilot 把 max_steps: 8 写进了生成的图，后端默认值根本轮不上。"""
     from app.api.copilot import NODE_REFERENCE

backend/tests/test_verify_schema_table_arg.py

@@ -0,0 +1,87 @@
+"""核查 agent（engine/verify_agent.py）的 db_schema 认「整个写成一个字符串的 JSON 数组」的表名。
+
+2026-10 验收：盲算、反方审稿的模型把 table 写成 '["orders", "stores"]'（一个字符串），工具只按逗号拆，回「中没有
+["orders"」「中没有 "stores"]」，两张表明明都在；模型照原样重试、再失败，到第三轮才改成不传 table，白白多调了几轮
+模型。现在第一次调用就拿到两张表的结构（tools/datasource.py 的 _schema_tables），模型看到的工具结果、时间线上的
+tool.end 都是表结构。
+
+模型是 MockChatModel 剧本，数据源是现造的 SQLite 文件，表名、数字都是虚构的。
+"""
+from __future__ import annotations
+
+import json
+import sqlite3
+import uuid
+
+import pytest
+from langchain_core.messages import AIMessage, ToolMessage
+
+from app.core.events import EventType
+from app.data import introspect
+from app.data.engine import engines
+from app.db.base import SessionLocal
+from app.db.models import DataSource
+from app.engine import verify_agent
+from app.engine.verify_agent import run_verify_agent
+from app.providers.factory import ModelSpec
+from app.providers.mock_model import MockChatModel
+from app.tools.registry import Versions
+
+SCHEMA = {"type": "object", "properties": {"value": {"type": "number", "description": "包裹运费合计"}},
+          "required": ["value"]}
+
+
+@pytest.fixture
+async def depots(tmp_path):
+    """两张表：包裹（运费 12、8、5）和仓库。返回源名。"""
+    path = tmp_path / f"{uuid.uuid4().hex[:8]}.db"
+    db = sqlite3.connect(path)
+    db.executescript("CREATE TABLE parcels (id INTEGER PRIMARY KEY, depot_id INTEGER, fee INTEGER);"
+                     "CREATE TABLE depots (id INTEGER PRIMARY KEY, zone TEXT);"
+                     "INSERT INTO parcels VALUES (1, 1, 12), (2, 1, 8), (3, 2, 5);"
+                     "INSERT INTO depots VALUES (1, '东区'), (2, '西区');")
+    db.commit()
+    db.close()
+    name = f"depots_{uuid.uuid4().hex[:8]}"
+    async with SessionLocal() as session:
+        row = DataSource(name=name, kind="sqlite", database=str(path), readonly=True, enabled=True)
+        session.add(row)
+        await session.flush()
+        row.schema_cache = await introspect.introspect(row)
+        await session.commit()
+        source_id = row.id
+    yield name
+    await engines.invalidate(source_id)
+
+
+async def test_a_json_array_written_as_one_string_gets_both_tables_at_once(monkeypatch, depots):
+    seen: list[list] = []
+    steps = [[(f"db_schema__{depots}", {"table": '["parcels", "depots"]'}, "s1")],
+             [(f"db_query__{depots}", {"sql": "SELECT SUM(fee) AS total FROM parcels"}, "c1")]]
+
+    def decide(model, messages):
+        seen.append(list(messages))
+        if model.response_format:
+            return AIMessage(content=json.dumps({"value": {"alias": "Q1", "row": 0, "column": "total"},
+                                                 "rationale": "运费合计见 [[see:Q1]]。"}, ensure_ascii=False))
+        if not steps:
+            return AIMessage(content="查完了。")
+        return AIMessage(content="查一下。", tool_calls=[{"name": n, "args": a, "id": i} for n, a, i in steps.pop(0)])
+
+    async def fake_model(session, spec):
+        return MockChatModel(model_name="mock-fast"), "mock-fast"
+
+    monkeypatch.setattr(MockChatModel, "_decide", decide)
+    monkeypatch.setattr(verify_agent, "get_chat_model", fake_model)
+    events: list[tuple] = []
+    out = await run_verify_agent(spec=ModelSpec(model="mock-fast"), system="核对包裹运费。", task="算一遍包裹运费合计。",
+                                 sources=[depots], output_schema=SCHEMA, data_versions=Versions.CURRENT,
+                                 emit=lambda kind, **data: events.append((kind, data)), run_id="run-m4", node_id="dual")
+    assert out.error is None and out.result["value"] == 25
+    # 模型在循环里看到的那次工具结果：两张表的结构，没有「中没有」
+    [answer] = [str(m.content) for messages in seen for m in messages
+                if isinstance(m, ToolMessage) and m.tool_call_id == "s1"][:1]
+    assert "表 parcels" in answer and "表 depots" in answer and "没有" not in answer
+    # 时间线上那一条 tool.end 也是表结构
+    [end] = [d for kind, d in events if kind == EventType.TOOL_END and d.get("call_id") == "s1"]
+    assert "表 parcels" in end["preview"] and "没有" not in end["preview"]