diff --git a/backend/tests/test_agent_tool_loop.py b/backend/tests/test_agent_tool_loop.py
deleted file mode 100644
index 39823e7a..00000000
--- a/backend/tests/test_agent_tool_loop.py
+++ /dev/null
@@ -1,180 +0,0 @@
-"""Agent 的工具循环，跑真图验收。
-
-单测能保证 split_tool_calls 分得对、prepare_args 纠得对，但保证不了它们在
-run_agent 里被接对了。这里从 run_manager.start 进去、从落库的 RunEvent 出来，
-验的是用户真正会看到的那条事件流。
-
-两件必须守住的事：
-
-- 每个 tool_use 都要有配对的 tool_result。少一条，下一次模型调用直接 400。
-- 没执行的调用绝不能发 tool.start。前端靠 call_id 配对 start/end
-  （frontend/src/run/decode.ts），一条没有结尾的 start 就是界面上一个永远转圈的卡片。
-"""
-
-from __future__ import annotations
-
-import asyncio
-import sqlite3
-
-import pytest
-from langchain_core.messages import AIMessage, ToolMessage
-from sqlalchemy import select
-
-from app.db.base import SessionLocal
-from app.db.models import DataSource, Run, RunEvent
-from app.engine.runner import run_manager
-
-
-@pytest.fixture
-async def runner():
-    """真图要跑起来就得有 checkpointer —— 平时是 FastAPI 的 lifespan 建的。"""
-    await run_manager.setup()
-    yield run_manager
-    await run_manager.shutdown()
-
-
-@pytest.fixture
-async def shop(tmp_path):
-    """一个真实的小库 + 一条 DataSource 记录，工具要能真的查到数。"""
-    path = tmp_path / "shop.db"
-    db = sqlite3.connect(path)
-    db.executescript("CREATE TABLE orders (id INTEGER PRIMARY KEY, amount REAL NOT NULL);")
-    db.executemany("INSERT INTO orders(id,amount) VALUES(?,?)", [(i, i * 2.0) for i in range(1, 8)])
-    db.commit()
-    db.close()
-
-    name = f"shop_{tmp_path.name[-6:].replace('-', '_')}"
-    async with SessionLocal() as session:
-        session.add(DataSource(
-            name=name, kind="sqlite", database=str(path), readonly=True,
-            description="测试库", options={}, schema_cache={}, enabled=True,
-        ))
-        await session.commit()
-    return name, str(path)
-
-
-def _graph(tool_name: str) -> dict:
-    return {
-        "nodes": [
-            {"id": "in", "type": "input", "data": {"config": {}}},
-            {"id": "bot", "type": "agent", "data": {"config": {
-                "prompt": "查一下订单总额",
-                "tools": [tool_name],
-                "approval": "never",      # 审批会挂起，这里要一路跑到底
-                "max_steps": 3,
-            }}},
-            {"id": "out", "type": "output", "data": {"config": {}}},
-        ],
-        "edges": [
-            {"source": "in", "target": "bot"},
-            {"source": "bot", "target": "out"},
-        ],
-    }
-
-
-async def _run_to_completion(graph: dict) -> list[RunEvent]:
-    run = await run_manager.start(graph=graph, input_payload={"question": "订单总额"})
-    for _ in range(300):
-        await asyncio.sleep(0.05)
-        async with SessionLocal() as session:
-            row = (await session.execute(select(Run).where(Run.id == run.id))).scalar_one()
-            if row.status in ("succeeded", "failed", "cancelled"):
-                break
-    else:
-        pytest.fail("运行没有在 15 秒内结束")
-
-    async with SessionLocal() as session:
-        events = list((await session.execute(
-            select(RunEvent).where(RunEvent.run_id == run.id).order_by(RunEvent.seq)
-        )).scalars())
-    return events, row
-
-
-async def test_serial_loop_runs_one_tool_and_answers_every_call(runner, shop, monkeypatch) -> None:
-    """模型一轮发两个工具调用：只跑第一个，但两个都要有回音。
-
-    顺带把第一个的参数名写错成 query —— 这正是那条
-    "_run() got an unexpected keyword argument 'query'" 的来源。
-    """
-    name, _ = shop
-    tool_name = f"db_query__{name}"
-
-    from app.providers import mock_model
-
-    turns = {"n": 0}
-
-    def _decide(self, messages):
-        turns["n"] += 1
-        if turns["n"] == 1:
-            return AIMessage(
-                content="先查一下。",
-                tool_calls=[
-                    # 参数名故意写错：能被纠正成 sql 才算过
-                    {"name": tool_name, "args": {"query": "SELECT SUM(amount) AS total FROM orders"},
-                     "id": "call_a"},
-                    {"name": tool_name, "args": {"sql": "SELECT COUNT(*) FROM orders"},
-                     "id": "call_b"},
-                ],
-            )
-        return AIMessage(content="订单总额是 56。")
-
-    monkeypatch.setattr(mock_model.MockChatModel, "_decide", _decide)
-
-    events, run = await _run_to_completion(_graph(tool_name))
-    kinds = [(e.type, e.data) for e in events]
-
-    assert run.status == "succeeded", [k for k, _ in kinds]
-
-    # 一轮只跑了一个工具
-    starts = [d for t, d in kinds if t == "tool.start"]
-    assert len(starts) == 1, f"应该只有一次 tool.start，实际 {len(starts)} 次"
-    assert starts[0]["call_id"] == "call_a"
-    # 发出去的是纠正后的参数，不是模型原样那份
-    assert "sql" in starts[0]["args"] and "query" not in starts[0]["args"]
-
-    # 每个 tool.start 都有配对的结尾，界面上不会留下转圈的卡片
-    ends = [d for t, d in kinds if t in ("tool.end", "tool.error")]
-    assert {d["call_id"] for d in ends} == {"call_a"}
-
-    # 纠正这件事说出来了，没有安静地把事办了
-    warns = [d["message"] for t, d in kinds if t == "log" and d.get("level") == "warn"]
-    assert any("query" in w and "sql" in w for w in warns), warns
-
-    # 工具真的查到了数
-    assert "56" in str([d.get("preview") for t, d in kinds if t == "tool.end"])
-
-
-async def test_bad_args_feed_the_schema_back_instead_of_dying(runner, shop, monkeypatch) -> None:
-    """参数全不认识时不执行，把"它接受什么"喂回去——模型下一步才改得对。"""
-    name, _ = shop
-    tool_name = f"db_query__{name}"
-
-    from app.providers import mock_model
-
-    seen: list[str] = []
-    turns = {"n": 0}
-
-    def _decide(self, messages):
-        turns["n"] += 1
-        if turns["n"] == 1:
-            return AIMessage(
-                content="查一下。",
-                tool_calls=[{"name": tool_name, "args": {"foo": 1, "bar": 2}, "id": "call_x"}],
-            )
-        # 第二轮：把上一轮喂回来的内容记下来，这就是模型能看到的全部依据
-        seen.extend(m.content for m in messages if isinstance(m, ToolMessage))
-        return AIMessage(content="好的。")
-
-    monkeypatch.setattr(mock_model.MockChatModel, "_decide", _decide)
-
-    events, run = await _run_to_completion(_graph(tool_name))
-    assert run.status == "succeeded"
-
-    assert seen, "参数错了也必须给模型一条 tool_result，否则下一轮直接 400"
-    text = seen[0]
-    assert "sql" in text and "必填" in text        # 说清楚该填什么
-    assert "unexpected keyword argument" not in text  # 不再是那句天书
-
-    # 没执行的调用不发 tool.start，界面上不会留下转不完的卡片
-    assert not [e for e in events if e.type == "tool.start"]
-    assert [e for e in events if e.type == "tool.error"]
diff --git a/backend/tests/test_conversations.py b/backend/tests/test_conversations.py
index 84d1209a..cbbd2849 100644
--- a/backend/tests/test_conversations.py
+++ b/backend/tests/test_conversations.py
@@ -51,41 +51,6 @@ async def _turn(client, conv_id: str, question: str, **patch) -> dict:
 # --------------------------------------------------------------------------
 
 
-async def test_a_conversation_survives_and_lists(client) -> None:
-    conv = await _conversation(client)
-    await _turn(client, conv["id"], "shop 里有几家店", answer="12 家", status="done")
-
-    # 列表里认得出是哪次聊天
-    rows = (await client.get("/api/conversations")).json()
-    mine = next(r for r in rows if r["id"] == conv["id"])
-    assert mine["turn_count"] == 1
-    assert mine["last_question"] == "shop 里有几家店"
-
-    # 完整取回：这是刷新页面后重建界面的唯一依据
-    detail = (await client.get(f"/api/conversations/{conv['id']}")).json()
-    assert [t["question"] for t in detail["turns"]] == ["shop 里有几家店"]
-    assert detail["turns"][0]["answer"] == "12 家"
-
-
-async def test_title_comes_from_the_first_question(client) -> None:
-    """一排"新对话"等于没有列表。"""
-    conv = await _conversation(client)
-    assert conv["title"] == "新对话"
-    await _turn(client, conv["id"], "shop 里有几家店")
-    assert (await client.get(f"/api/conversations/{conv['id']}")).json()["title"] == "shop 里有几家店"
-
-    # 第二轮不该再改标题
-    await _turn(client, conv["id"], "再按月份拆一下")
-    assert (await client.get(f"/api/conversations/{conv['id']}")).json()["title"] == "shop 里有几家店"
-
-
-async def test_a_renamed_conversation_keeps_its_name(client) -> None:
-    conv = await _conversation(client)
-    await client.patch(f"/api/conversations/{conv['id']}", json={"title": "门店盘点"})
-    await _turn(client, conv["id"], "shop 里有几家店")
-    assert (await client.get(f"/api/conversations/{conv['id']}")).json()["title"] == "门店盘点"
-
-
 def test_long_titles_are_cut_not_wrapped() -> None:
     assert title_from("短问题") == "短问题"
     assert title_from("") == "新对话"
@@ -150,33 +115,6 @@ async def _seeded(client) -> str:
     return conv["id"]
 
 
-async def test_history_text_is_plain_qa(client) -> None:
-    conv_id = await _seeded(client)
-    async with SessionLocal() as session:
-        text = await history_text(session, conv_id)
-    assert "shop 里有几家店" in text and "一共 12 家" in text
-
-
-async def test_history_carries_the_review_warning_forward(client) -> None:
-    """被复核判过不可信的答案，下一轮必须知道。
-
-    只把正文喂过去的话，模型会拿一个已知有问题的结论当事实继续往下推
-    （"那 12 家里有几家是新开的？"），错误就这么一轮轮传下去，而且越传
-    越像真的——后面几轮看起来都在正常作答。
-    """
-    conv = await _conversation(client)
-    await _turn(
-        client, conv["id"], "shop 里有几家店",
-        answer="一共 12 家", status="done",
-        review={"verdict": "annotated", "severity": "broken", "note": "取数没跑通，这个数不可信。",
-                "answer": None, "retry": True, "signals": []},
-    )
-    async with SessionLocal() as session:
-        text = await history_text(session, conv["id"])
-    assert "一共 12 家" in text
-    assert "取数没跑通" in text
-
-
 async def test_history_text_is_empty_without_a_conversation(client) -> None:
     async with SessionLocal() as session:
         assert await history_text(session, None) == ""
@@ -212,61 +150,11 @@ async def test_a_huge_answer_does_not_crowd_out_the_other_turns(client) -> None:
     assert "那总共多少" in text          # 后一轮没被前一轮挤掉
 
 
-async def test_history_section_carries_the_previous_graph(client) -> None:
-    """只给问答文字不够：追问会让模型重新设计一张图，可能接到别的表上。"""
-    from app.api.copilot import _history_section
-
-    conv_id = await _seeded(client)
-    async with SessionLocal() as session:
-        block = await _history_section(session, conv_id)
-
-    assert "shop 里有几家店" in block
-    assert "shop_agent" in block and "db_query__shop" in block
-    assert '"x"' not in block and '"position"' not in block   # _slim 去掉了坐标噪音
-
-
 # --------------------------------------------------------------------------
 # 上下文真的到了模型手上
 # --------------------------------------------------------------------------
 
 
-async def test_the_previous_turn_reaches_the_prompt(client, monkeypatch) -> None:
-    """整件事的关键一环。
-
-    拼字符串的函数测通过、却没接进 messages，是这类功能最典型的失败方式：
-    界面上完全看不出来，它只是"答得不太对"。所以这里拦在模型入口上，
-    看送出去的那条 human message 里到底有没有上一轮。
-    """
-    import app.api.copilot as copilot
-
-    conv_id = await _seeded(client)
-    captured: list = []
-
-    class _Recorder:
-        def with_structured_output(self, *_a, **_k):
-            return self
-
-        async def ainvoke(self, messages):
-            captured.append(messages)
-            return {"nodes": [], "edges": []}
-
-    async def _model(*_a, **_k):
-        return _Recorder(), "mock-fast"
-
-    monkeypatch.setattr(copilot, "get_chat_model", _model)
-
-    r = await client.post("/api/copilot/generate", json={
-        "instruction": "再按月份拆一下", "intent": "answer", "conversation_id": conv_id,
-    })
-    assert r.status_code == 200, r.text
-
-    human = next(text for role, text in captured[0] if role == "human")
-    assert "shop 里有几家店" in human       # 上一轮问的
-    assert "一共 12 家" in human              # 上一轮答的
-    assert "shop_agent" in human              # 上一轮的图
-    assert "再按月份拆一下" in human          # 这一轮问的
-
-
 async def test_no_conversation_means_no_history_header(client, monkeypatch) -> None:
     """不带 conversation_id 时不能凭空多出一个空的历史标题。"""
     import app.api.copilot as copilot
@@ -458,44 +346,11 @@ def test_the_three_paths_are_spelled_out_for_the_model() -> None:
     assert '"run":true' in _STREAM_PROTOCOL
 
 
-async def test_copilot_is_told_not_to_invent_a_memory_scope() -> None:
-    """记忆域由平台给。图里自己起一个，写进去就再也读不回来。
-
-    真发生过（会话 cbe09f4b 的第 2~3 轮）：一轮的 memory 节点写了
-    scope="user"，下一轮没写 scope、落回 default，于是「我叫张三」存下来了，
-    问「我是谁」却答不上来——记忆在库里好好的，只是读错了地方。
-    """
-    from app.api.copilot import NODE_REFERENCE
-
-    memory_line = next(
-        (l for l in NODE_REFERENCE.splitlines() if l.strip().startswith("- memory：")), "")
-    assert memory_line, "节点参考里找不到 memory 这一行"
-    assert "scope" not in memory_line, "scope 还列在可填字段里，模型就会去填它"
-    assert "不要写 scope" in NODE_REFERENCE
-
-
 # --------------------------------------------------------------------------
 # 主动记忆
 # --------------------------------------------------------------------------
 
 
-def test_copilot_is_told_when_to_record_a_memory() -> None:
-    """以前只有用户明说「记住」才会写：17 轮里只有 4 轮带记忆节点，全在用户
-    主动谈记忆的那个会话里。于是他说过一次「我叫张三」，换个话题再问就
-    谁也不知道了。
-
-    但边界比指令本身更要紧——记错了会污染以后每一次对话。
-    """
-    from app.api.copilot import GenerateIn, _user_message
-
-    text = _user_message(GenerateIn(instruction="随便问问", intent="answer"), patch=False)
-    assert "该记" in text and "不该记" in text
-    # 最容易被记进去的垃圾：这一轮的问题、查出来的数、一次性指令、以及推测
-    for junk in ("这一轮的问题本身", "查出来的数", "一次性的指令", "推测出来的东西"):
-        assert junk in text, f"没有拦住「{junk}」这类"
-    assert "拿不准就不记" in text
-
-
 async def test_writing_a_memory_is_visible_in_the_trace(client, monkeypatch) -> None:
     """自动记忆的前提是看得见。
 
diff --git a/backend/tests/test_copilot_context.py b/backend/tests/test_copilot_context.py
index 61b43aa4..43e6ebea 100644
--- a/backend/tests/test_copilot_context.py
+++ b/backend/tests/test_copilot_context.py
@@ -13,7 +13,7 @@ from app.api.copilot import GRAPH_SCHEMA, _DATASOURCE_BUDGET_CHARS
 from app.data.introspect import summary
 
 
-def _source(n_tables: int, *, name: str = "bi", schema: str | None = "ANALYTICS"):
+def _source(n_tables: int, *, name: str = "warehouse", schema: str | None = "ANALYTICS"):
     tables = {
         f"t{i}": {
             "qualified": f"{schema}.t{i}" if schema else f"t{i}",
@@ -45,19 +45,6 @@ def test_graph_schema_allows_arbitrary_config_keys():
     assert "config" in node["required"]
 
 
-def test_summary_detail_mode_lists_columns():
-    text = summary(_source(3))
-    assert "ANALYTICS.t0" in text and "c0" in text
-
-
-def test_summary_compact_mode_drops_columns():
-    """库多了要靠它压 token：只留对象名，字段让 db_schema 按需查。"""
-    text = summary(_source(50), detail=False)
-    assert "ANALYTICS.t0" in text
-    assert "c0" not in text
-    assert len(text) < len(summary(_source(50)))
-
-
 def test_compact_mode_is_substantially_smaller():
     """压缩比要真的有效，否则降级没意义。实测约 35%。"""
     full = summary(_source(53))
@@ -65,11 +52,6 @@ def test_compact_mode_is_substantially_smaller():
     assert len(compact) < len(full) * 0.5
 
 
-def test_qualified_names_appear_in_summary():
-    """摘要里必须是带 schema 的全名——模型照抄，漏前缀就是 ORA-00942。"""
-    assert "ANALYTICS.t0" in summary(_source(2))
-
-
 def test_summary_without_schema_uses_bare_names():
     """没有 schema 概念的库不要硬拼前缀，否则生成的 SQL 反而错。"""
     text = summary(_source(2, schema=None))
diff --git a/backend/tests/test_datasource.py b/backend/tests/test_datasource.py
index 0d332da0..0e5a9617 100644
--- a/backend/tests/test_datasource.py
+++ b/backend/tests/test_datasource.py
@@ -197,32 +197,6 @@ async def test_qualified_name_includes_schema_when_given(source):
     assert cache["tables"]["orders"]["qualified"] == "main.orders"
 
 
-async def test_bad_schema_returns_candidates_instead_of_raising(source, monkeypatch):
-    """schema 填错是高频事故（大小写、拼写），要给可选项而不是抛驱动异常。"""
-    from app.data import introspect as mod
-
-    async def fake_list(_source):
-        return ["ANALYTICS", "STAGING"]
-
-    monkeypatch.setattr(mod, "list_schemas", fake_list)
-    cache = await mod.introspect(source, schema="根本不存在")
-    assert cache["tables"] == {}
-    assert cache["available_schemas"] == ["ANALYTICS", "STAGING"]
-
-
-async def test_summary_points_at_candidate_schemas(source, monkeypatch):
-    """摘要要把候选说给 Copilot 听，否则它只会对着空结构编表名。"""
-    from app.data import introspect as mod
-
-    async def fake_list(_source):
-        return ["ANALYTICS", "STAGING"]
-
-    monkeypatch.setattr(mod, "list_schemas", fake_list)
-    source.schema_cache = await mod.introspect(source, schema="根本不存在")
-    text = summary(source)
-    assert "ANALYTICS" in text and "该账号名下没有对象" in text
-
-
 def test_schema_stats_sql_covers_every_supported_kind():
     """新增数据库类型时别忘了这里——漏了就退化成黑名单猜测。"""
     from app.data.engine import SUPPORTED_KINDS
@@ -277,60 +251,3 @@ def test_why_empty_tells_the_two_cases_apart():
     assert "失败" in why_empty(_FailedCache.schema_cache)
     assert "Lost connection" in why_empty(_FailedCache.schema_cache)
     assert why_empty({}) == "还没有探查过结构"
-
-
-def test_schema_tool_tells_the_agent_the_truth():
-    """给 agent 的那句话要是实情，而且要给一个它做得到的下一步。"""
-    from app.tools.datasource import _no_schema_note
-
-    note = _no_schema_note(_FailedCache)
-    assert "失败" in note and "Lost connection" in note
-    # 候选一直在缓存里，以前没人交出去
-    assert "prod_shop" in note and "toolsdb" in note
-    # agent 进不了设置页。它做得到的是自己查数据字典——那条路它最后也走通了，
-    # 只是自己摸索了三轮
-    assert "information_schema" in note
-    assert "不要猜表名" in note
-    # 不能再说"请到设置页"
-    assert "设置页" not in note
-
-    fresh = _no_schema_note(_NeverRun)
-    assert "还没有探查过" in fresh
-    assert "失败" not in fresh
-
-
-def test_describe_table_does_not_claim_the_table_is_missing():
-    """缓存是空的时候，说"里面没有这张表"是在误导——真相是我们什么都不知道。"""
-    from app.data.introspect import describe_table
-
-    out = describe_table(_FailedCache, "user")
-    assert "无法判断" in out
-    assert "失败" in out
-    # 以前是 "数据源「shop」里没有 user。现有的对象：（还没探查过结构）"
-    assert "现有的对象" not in out
-
-
-@pytest.mark.asyncio
-async def test_introspect_failure_does_not_get_a_synced_stamp(client, monkeypatch):
-    """失败要把缓存留下（候选是下一步的线索），但不能盖"已同步"的戳。
-
-    以前盖了，于是界面说同步过、工具说还没探查过，两边都不说真话。
-    """
-    import app.api.datasources as mod
-
-    async def failed(source, schema=None):
-        return dict(_FailedCache.schema_cache)
-
-    created = await client.post("/api/datasources", json={
-        "name": "introspect_fail", "kind": "sqlite", "database": ":memory:",
-    })
-    assert created.status_code == 201, created.text
-    sid = created.json()["id"]
-
-    monkeypatch.setattr(mod.introspect_mod, "introspect", failed)
-    out = (await client.post(f"/api/datasources/{sid}/introspect")).json()
-
-    assert out["schema_synced_at"] is None, "失败不该盖已同步的戳"
-    assert "Lost connection" in out["schema_error"]
-    assert out["available_schemas"] == ["prod_shop", "shop_dev", "toolsdb"]
-    assert out["table_count"] == 0
diff --git a/backend/tests/test_review.py b/backend/tests/test_review.py
index a58021aa..64364d18 100644
--- a/backend/tests/test_review.py
+++ b/backend/tests/test_review.py
@@ -39,19 +39,6 @@ OK_OUTPUT = {"结果": "管理员有 6 个。"}
 # --------------------------------------------------------------------------
 
 
-def test_clean_run_has_no_signals():
-    """干净的运行扫不出东西——这是"零成本"的前提。"""
-    events = [
-        ev(EventType.RUN_STARTED),
-        ev(EventType.TOOL_START, tool="db_query__shop"),
-        ev(EventType.TOOL_END, tool="db_query__shop"),
-        ev(EventType.LOG, level="info", message="随便什么"),
-        ev(EventType.RETRIEVE_END, count=5, degraded=False, query="权限"),
-        ev(EventType.RUN_FINISHED),
-    ]
-    assert rv.scan(events, OK_OUTPUT) == []
-
-
 def test_tool_error_is_broken():
     signals = rv.scan([ev(EventType.TOOL_ERROR, tool="db_query__x", error="no such table")],
                       OK_OUTPUT)
diff --git a/backend/tests/test_tool_args.py b/backend/tests/test_tool_args.py
index 5c6afb2d..6de884fe 100644
--- a/backend/tests/test_tool_args.py
+++ b/backend/tests/test_tool_args.py
@@ -172,40 +172,6 @@ async def test_schema_tool_args_are_validated_too(source) -> None:
     assert payload["table"] == "orders" and note is not None
 
 
-async def test_call_tool_end_to_end_on_the_path_that_broke(source) -> None:
-    """走 call_tool 这个真实入口（工具节点和设置页的"试一下"都走它）。
-
-    以前这条路径对动态工具是裸 `**args` 展开的，报错止步于 TypeError；
-    现在参数名能被纠正、纠正这件事能通过 on_fix 说出去。
-    """
-    from app.db.base import SessionLocal
-    from app.db.models import DataSource
-    from app.tools.registry import call_tool
-
-    async with SessionLocal() as session:
-        session.add(DataSource(
-            name="shop", kind="sqlite", database=source.database,
-            readonly=True, description="测试库", options={}, schema_cache={}, enabled=True,
-        ))
-        await session.commit()
-
-    notes: list[str] = []
-    async with SessionLocal() as session:
-        out = await call_tool(
-            "db_query__shop", {"query": "SELECT COUNT(*) AS c FROM orders"},
-            ToolContext(), session=session, on_fix=notes.append,
-        )
-    assert "5" in str(out)
-    # 跑通了，但配置里那个错的参数名还在，必须留下痕迹而不是安静地把事办了
-    assert len(notes) == 1 and "sql" in notes[0]
-
-    async with SessionLocal() as session:
-        with pytest.raises(ToolArgsError) as e:
-            await call_tool("db_query__shop", {"foo": 1, "bar": 2}, ToolContext(), session=session)
-    assert "sql" in str(e.value)
-    assert "unexpected keyword argument" not in str(e.value)
-
-
 def test_tools_without_a_pydantic_schema_are_let_through() -> None:
     """MCP 这类工具的 args_schema 可能是一份 JSON Schema dict，校验不了就放行。"""
     assert args_model_of(SimpleNamespace(args_schema={"type": "object"})) is None
diff --git a/scripts/check-decode.mjs b/scripts/check-decode.mjs
index b2b25ff9..4fab0870 100644
--- a/scripts/check-decode.mjs
+++ b/scripts/check-decode.mjs
@@ -3,6 +3,7 @@
 // 用**真实运行**导出的事件跑（fixtures.json 取自 data/agentlab.db），不是构造的样本：
 // 这里要守的恰恰是真实事件的脏细节——审批会重放、成对事件字段不齐、
 // 空 preview、节点没起过名字。构造的样本永远不会长成那样。
+// 库名表名已脱敏（改的只是标识符字面量，事件的结构和脏细节原样保留）。
 //
 // 转译借 vite dev server（它 serve 的就是应用实际运行的那份），所以跑之前
 // 前端得起着：./scripts/dev.sh
diff --git a/scripts/check-stream.mjs b/scripts/check-stream.mjs
index 2b852a8f..ecdd0f18 100644
--- a/scripts/check-stream.mjs
+++ b/scripts/check-stream.mjs
@@ -57,7 +57,7 @@ console.log('\n=== 展开交互 ===')
   await page.waitForTimeout(300)
   const before = await page.locator('body').innerText()
   check('收起时不显示 SQL', !before.includes('SELECT'))
-  await page.locator('button', { hasText: '在 bi 上查询数据' }).first().click()
+  await page.locator('button', { hasText: '在 warehouse 上查询数据' }).first().click()
   await page.waitForTimeout(250)
   const after = await page.locator('body').innerText()
   check('展开后能看到 SQL 原文', after.includes('SELECT') && after.includes('ANALYTICS'))
@@ -67,8 +67,10 @@ console.log('\n=== 展开交互 ===')
 
 console.log('\n=== Markdown 渲染 ===')
 {
-  // 用库里导出的**真实输出**，不是编的样本：模型写什么才是要渲染的，
-  // 按 CommonMark 规范挑测例只会测到用不上的角落
+  // 样本取自库里导出的**真实输出**，不是按 CommonMark 规范挑的测例：模型
+  // 实际会写成什么样，才是要渲染的东西。涉及业务数据的那几条已换成同构的
+  // 合成内容——换的是领域，markdown 构造（标题层级、表格、引用块、inline
+  // code 密度）逐项对齐，覆盖面没有变窄
   const page = await browser.newPage({ viewport: { width: 1100, height: 900 } })
   const errors = []
   page.on('pageerror', (e) => errors.push(e.message))