表助手：SDK 不再自己重试（同起草，网络错误、429、5xx 只由表助手在还没输出时重试，退避 3、10 秒）；测试补上取模型的规格（起草设置、保留思考、额度 32768 / 8192、单次时限、不重试）和两个工具在 OpenAI 兼容、Anthropic 模型上都绑得上；query_snapshot_tables 不进 catalog 的 __all__（按模块属性取用，少和起草那边的改动撞在一起）

backend/app/data/catalog.py

@@ -2351,8 +2351,7 @@ __all__ = [
     "fk_relations", "frozen_catalog", "infer_name_relations", "initial_status", "is_date_column", "is_profile_note",
     "join_paths",
     "make_item", "merge_notes", "name_words", "now_iso", "parse_path", "plan_patch", "profile_relation_holds",
-    "query_snapshot_meta", "query_snapshot_tables", "read_catalog", "read_entry", "relation_graph", "relation_id",
-    "render_table_index",
+    "query_snapshot_meta", "read_catalog", "read_entry", "relation_graph", "relation_id", "render_table_index",
     "render_table_notes", "resolve_draft_model", "resolve_table_name", "review_entry", "review_item", "same_notes",
     "same_profile_note", "save_human_edit", "save_patch", "status_counts", "structure_relation_ids",
     "system_notes_source", "table_usage", "unreviewed_status", "validate_notes", "visible_notes", "write_entry",

backend/app/data/catalog_assistant.py

@@ -75,8 +75,8 @@ QUERY_ROWS = 50
 SHOWN_ROWS = 10
 CELL_MAX = 300
 TOOL_CHARS = 12_000
-#: 网络错误、429、5xx 重试前等几秒（依次取用，个数就是最多重试几次）
-RETRY_DELAYS_S = (2.0, 5.0)
+#: 网络错误、429、5xx 重试前等几秒（依次取用，个数就是最多重试几次）。在对话里等，比起草（10、30 秒）短
+RETRY_DELAYS_S = (3.0, 10.0)
 #: 等模型的时候每隔几秒发一行 SSE 注释
 HEARTBEAT_S = 10.0
 #: 背景里「目录里还没有中文名和含义的列」最多列几个；别的表指过来的关联关系最多列几条
@@ -257,8 +257,9 @@ async def resolve_model(session: AsyncSession) -> tuple[Any, str]:
         provider, choice = await resolve_draft_provider(session, none_reason=ERROR_NO_MODEL, mock_reason=ERROR_NO_MODEL)
     except DraftModelUnavailable as e:
         raise catalog.CatalogModelUnavailable(str(e)) from e
+    # SDK 自己不重试（同起草）：网络错误、429、5xx 由这里在还没输出时重试，两层叠起来实际次数就对不上了
     spec = ModelSpec(provider=choice.provider, model=choice.model, max_tokens=catalog.MODEL_MAX_TOKENS, temperature=0,
-                     timeout=int(catalog.MODEL_TIMEOUT_S))
+                     timeout=int(catalog.MODEL_TIMEOUT_S), max_retries=0)
     try:
         big = build_chat_model(provider, spec)
         small = build_chat_model(provider, spec.model_copy(update={"max_tokens": catalog.MODEL_FALLBACK_MAX_TOKENS}))

backend/tests/test_catalog_assistant.py

@@ -228,6 +228,69 @@ async def test_unavailable_draft_model_is_409_with_the_draft_wording(client, tmp
     assert r.json()["code"] == "catalog_draft_model_unavailable" and "模型接入" in r.json()["detail"]
 
 
+@pytest.fixture
+async def mem(monkeypatch):
+    """单独的内存库：接入和设置只在这个用例里。"""
+    from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine
+    from sqlalchemy.pool import StaticPool
+
+    from app.db.base import Base
+
+    for key in ("ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN", "OPENAI_API_KEY", "ANTHROPIC_BASE_URL"):
+        monkeypatch.delenv(key, raising=False)
+    engine = create_async_engine("sqlite+aiosqlite:///:memory:", poolclass=StaticPool)
+    async with engine.begin() as conn:
+        await conn.run_sync(Base.metadata.create_all)
+    session = async_sessionmaker(engine, expire_on_commit=False, class_=AsyncSession)()
+    yield session
+    await session.close()
+    await engine.dispose()
+
+
+async def _provider(session, kind: str) -> None:
+    from app.core.crypto import encrypt
+    from app.data import draft_model
+    from app.db.models import Provider, Setting
+
+    session.add(Provider(name="起草网关", kind=kind, base_url="http://gw.example.test/v1" if kind != "anthropic" else None,
+                         api_key=encrypt("sk-test"), models=[{"id": "模型甲"}], default_model="模型甲", enabled=True,
+                         extra={}))
+    session.add(Setting(key=draft_model.CATALOG_DRAFT_SETTING_KEY, value={"provider": "起草网关", "model": None}))
+    await session.commit()
+
+
+async def test_model_follows_the_draft_setting_keeps_thinking_and_a_big_budget(mem, monkeypatch):
+    from app.providers import factory
+
+    await _provider(mem, "openai_compatible")
+    specs: list[Any] = []
+    real = factory.build_chat_model
+
+    def spy(provider, spec):
+        specs.append(spec)
+        return real(provider, spec)
+
+    monkeypatch.setattr(factory, "build_chat_model", spy)
+    model, model_id = await catalog_assistant.resolve_model(mem)
+    assert model_id == "模型甲" and model.max_tokens == catalog.MODEL_MAX_TOKENS == 32768
+    mine = specs[-2:]                      # 前两份是 catalog.resolve_draft_model 查能不能用时建的
+    assert [s.max_tokens for s in mine] == [catalog.MODEL_MAX_TOKENS, catalog.MODEL_FALLBACK_MAX_TOKENS]
+    assert all(s.thinking is None and s.temperature == 0 and s.max_retries == 0 for s in mine)
+    assert all(s.timeout == catalog.MODEL_TIMEOUT_S for s in mine)
+
+
+@pytest.mark.parametrize("kind", ["openai_compatible", "anthropic"])
+async def test_tools_bind_on_real_chat_models(mem, kind):
+    """两个工具的定义各家的模型都收（只绑定，不发请求）。"""
+    await _provider(mem, kind)
+    model, _ = await catalog_assistant.resolve_model(mem)
+    tools = [catalog_assistant.QUERY_TOOL_PROMPT, catalog_assistant.PROPOSE_TOOL_PROMPT]
+    bound = model.bind_tools(tools)
+    given = bound.model.kwargs["tools"]
+    names = [t.get("name") or t["function"]["name"] for t in given]
+    assert names == ["run_query", "propose_changes"]
+
+
 # --------------------------------------------------------------------------
 # 背景
 # --------------------------------------------------------------------------