diff --git a/backend/app/engine/launch.py b/backend/app/engine/launch.py
index 9bdbf3fa..4ba3897e 100644
--- a/backend/app/engine/launch.py
+++ b/backend/app/engine/launch.py
@@ -87,7 +87,8 @@ async def launch_run(
     参数和 POST /api/runs 的请求体一一对应（input 叫 input_payload）；actor 是记在运行上的发起人。workflow_name 是没挂在
     工作流上的探索运行记在运行上的名字（蓝图的试运行写蓝图名），不给就是「未保存的工作流」；挂在工作流上的用工作流的名字。
     schedule_id 是发起它的定时任务，options 是整次运行沿用的选项快照，都原样交给 run_manager.start；图里有双盲复算、
-    反方审稿节点时 options 另加核查设置的快照（_with_verify_snapshot）。
+    反方审稿节点时 options 另加核查设置的快照（_with_verify_snapshot），有分析卡、或 Agent / 协作成员绑了分析工具时另加
+    分析设置的快照（_with_analysis_snapshot）。
     """
     # 没挂在工作流上（画布没保存就发起、问数据当场搭的）。老数据里是「临时图」，前端两种都认
     name = workflow_name or "未保存的工作流"
@@ -182,6 +183,7 @@ async def launch_run(
         raise LaunchRefused(400, "没有要运行的内容：请求中既没有工作流 ID，也没有工作流内容")
     await _refuse_missing_tools(session, graph)
     options = await _with_verify_snapshot(session, graph, options)
+    options = await _with_analysis_snapshot(session, graph, options)
     options, capability_changes = await _with_capability_snapshot(
         session, graph, options, published=published_capabilities)
 
@@ -311,7 +313,43 @@ async def _with_verify_snapshot(session: AsyncSession, graph: dict[str, Any],
 
 
 async def _has_verify_nodes(session: AsyncSession, spec: GraphSpec) -> bool:
-    """图里有没有双盲复算、反方审稿节点，子工作流里的也算。
+    """图里有没有双盲复算、反方审稿节点，子工作流里的也算。"""
+    return await _any_graph(session, spec, lambda graph: any(node.type in _VERIFY_NODES for node in graph.nodes))
+
+
+async def _with_analysis_snapshot(session: AsyncSession, graph: dict[str, Any],
+                                  options: dict[str, Any] | None) -> dict[str, Any] | None:
+    """图里（含子工作流）有分析卡、或 Agent / 协作成员绑了分析工具时，把分析设置（api/settings.py 的 analysis 一组：
+    判定用的阈值、次数上限）快照进 options["analysis"]（docs/design/analysis-round/00-总方案.md §3.10）。
+
+    道理同 _with_verify_snapshot：分析在运行中途读阈值，审批停下期间有人改了设置，恢复后的那一段就按另一套阈值判，正式
+    运行也说不清当时按什么规则判的「主要来自」。分析工具、分析卡只读这一份（app/tools/analysis.py 的 run_snapshot），
+    没有快照的现读设置。别的图 options 原样不动，run.started 和以前逐字相同；调用方已经给了 options["analysis"] 的照它的。
+    """
+    from app.api.settings import ANALYSIS_GROUP, analysis_settings
+
+    if isinstance(options, dict) and ANALYSIS_GROUP in options:
+        return options
+    try:
+        spec = GraphSpec.model_validate(graph)
+    except Exception:  # noqa: BLE001 - 图本身不成形交给 run_manager.start 报
+        return options
+    if not await _any_graph(session, spec, _analyzes):
+        return options
+    return {**(options or {}), ANALYSIS_GROUP: await analysis_settings(session)}
+
+
+def _analyzes(graph: GraphSpec) -> bool:
+    """这一张图（不含子工作流）会用到分析设置：有分析卡，或者有绑了分析工具的 Agent、协作成员（tool_bindings 的取法，
+    节点上没写就取图级 defaults）。"""
+    from app.tools.analysis import is_analysis_tool
+
+    return any(node.type == NodeType.ANALYSIS for node in graph.nodes) \
+        or any(is_analysis_tool(name) for _, name in tool_bindings(graph))
+
+
+async def _any_graph(session: AsyncSession, spec: GraphSpec, test: Any) -> bool:
+    """这张图或它（一层层）套着的子工作流里，有没有一张让 test(图) 为真。
 
     子工作流照运行时的取法读（同 runner._bound_sources）：固定了版本的读那一版，没固定的读工作流现在的图。读不到、
     图不成形的跳过——那个节点运行时自己会报错。同一个工作流（同一版）只读一次，互相引用也不会转圈。
@@ -320,7 +358,7 @@ async def _has_verify_nodes(session: AsyncSession, spec: GraphSpec) -> bool:
     queue: list[GraphSpec] = [spec]
     while queue:
         graph = queue.pop(0)
-        if any(node.type in _VERIFY_NODES for node in graph.nodes):
+        if test(graph):
             return True
         for node in graph.nodes:
             if node.type != NodeType.SUBGRAPH:
diff --git a/backend/tests/test_analysis_settings.py b/backend/tests/test_analysis_settings.py
index 229dddd3..0b782c5a 100644
--- a/backend/tests/test_analysis_settings.py
+++ b/backend/tests/test_analysis_settings.py
@@ -1,20 +1,31 @@
-"""分析设置（设置 → 偏好设置 → 分析，docs/design/analysis-round/01-算子与判定.md §10）。
+"""分析设置（设置 → 偏好设置 → 分析，docs/design/analysis-round/01-算子与判定.md §10）和发起时的快照。
 
 - 设置接口的 analysis 一组：键、缺省、范围、叫法全取 app/analysis/contract.py；GET 按生效的样子给（存坏了的项按缺省），
   PUT 先查后写（不认识的键、类型、范围、两两之间的大小）
+- 发起运行时快照进 run.options["analysis"]：只在图（含子工作流）里有分析卡、或 Agent / 协作成员绑了分析工具时写；
+  别的图的 run.started 一字不变
 """
 from __future__ import annotations
 
+import asyncio
+
 import pytest
 from httpx import ASGITransport, AsyncClient
+from sqlalchemy import select
 
 from app.analysis import contract
 from app.api.settings import ANALYSIS_GROUP, analysis_problem, analysis_values
 from app.db.base import SessionLocal
-from app.db.models import Setting
+from app.db.models import Run, RunEvent, Setting, Workflow
+from app.engine.launch import launch_run
 from app.engine.runner import run_manager
 from app.main import app
 
+#: 加 options 之前 run.started 就有的键（同 tests/test_run_options.py）
+LEGACY_STARTED_KEYS = {"nodes", "resumed", "memory_scope", "collection", "approval_default", "tool_trust",
+                       "agent_limits", "data_versions", "replay_protocol"}
+
+
 @pytest.fixture(autouse=True)
 async def engine_up():
     await run_manager.setup()
@@ -38,6 +49,14 @@ async def clean_group():
             await session.commit()
 
 
+def node(nid, ntype, **config):
+    return {"id": nid, "type": ntype, "position": {"x": 0, "y": 0}, "data": {"label": nid, "config": config}}
+
+
+def chain(*nodes):
+    return {"nodes": list(nodes), "edges": [{"source": a["id"], "target": b["id"]} for a, b in zip(nodes, nodes[1:])]}
+
+
 # --------------------------------------------------------------------------
 # 设置接口
 # --------------------------------------------------------------------------
@@ -83,3 +102,78 @@ async def test_put_refuses_bad_values_and_stores_nothing(client):
     r = await client.put("/api/settings", json={"values": {ANALYSIS_GROUP: {"main_share": 0.6, "k_alert": 1.5}}})
     assert r.status_code == 422 and contract.LABELS["k_alert"] in r.json()["detail"]
     assert (await client.get("/api/settings")).json()[ANALYSIS_GROUP] == contract.DEFAULTS
+
+
+# --------------------------------------------------------------------------
+# 发起时的快照
+# --------------------------------------------------------------------------
+
+
+async def _started(run_id: str) -> dict:
+    for _ in range(200):
+        async with SessionLocal() as session:
+            found = (await session.execute(select(RunEvent).where(
+                RunEvent.run_id == run_id, RunEvent.type == "run.started"))).scalars().first()
+        if found is not None:
+            await run_manager.wait_idle(run_id)
+            return found.data
+        await asyncio.sleep(0.05)
+    raise AssertionError("没有 run.started")
+
+
+async def _launch(graph, **kw) -> tuple[Run, dict]:
+    async with SessionLocal() as session:
+        run = await launch_run(session, graph=graph, actor="测试", **kw)
+    return run, await _started(run.id)
+
+
+CARD = node("card", "analysis", items=[{"id": "wow", "op": "baseline", "from": "q",
+                                        "params": {"period": "week", "measure": "gmv"}}])
+QUERY = node("q", "tool", tool="calculator", args={"expression": "1 + 1"})
+OUT = node("out", "output", fields=[{"name": "r", "value": "完成"}])
+
+
+async def test_graphs_without_analysis_start_exactly_as_before():
+    run, started = await _launch(chain(node("start", "input"), node("t", "transform", mode="template", template="x"),
+                                       OUT))
+    assert set(started) == LEGACY_STARTED_KEYS, "没有分析的图，run.started 一个键都不多"
+    async with SessionLocal() as session:
+        assert (await session.get(Run, run.id)).options in (None, {})
+    # 绑了别的工具的 Agent：有能力治理的快照（以前就有），没有分析这一组
+    _, started = await _launch(chain(node("start", "input"),
+                                     node("bot", "agent", prompt="说一句", tools=["calculator"]), OUT))
+    assert ANALYSIS_GROUP not in (started.get("options") or {})
+
+
+async def test_an_analysis_card_snapshots_the_settings():
+    async with SessionLocal() as session:
+        session.add(Setting(key=ANALYSIS_GROUP, value={"main_share": 0.6}))
+        await session.commit()
+    run, started = await _launch(chain(node("start", "input"), QUERY, CARD, OUT))
+    assert started["options"][ANALYSIS_GROUP] == {**contract.DEFAULTS, "main_share": 0.6}
+    async with SessionLocal() as session:
+        assert (await session.get(Run, run.id)).options[ANALYSIS_GROUP]["main_share"] == 0.6
+
+
+@pytest.mark.parametrize("graph", [
+    chain(node("start", "input"), node("bot", "agent", prompt="为什么", tools=[contract.TOOL_NAMES["trend"]]), OUT),
+    chain(node("start", "input"), node("team", "supervisor", goal="为什么", agents=[
+        {"name": "分析员", "tools": [contract.TOOL_NAMES["differ"]]}]), OUT),
+])
+async def test_agents_and_members_with_analysis_tools_snapshot_the_settings(graph):
+    _, started = await _launch(graph)
+    assert started["options"][ANALYSIS_GROUP] == contract.DEFAULTS
+
+
+async def test_a_card_inside_a_subgraph_counts_and_given_options_win():
+    async with SessionLocal() as session:
+        inner = Workflow(name="异动分析（子流程）", graph=chain(node("start", "input"), QUERY, CARD, OUT))
+        session.add(inner)
+        await session.commit()
+        inner_id = inner.id
+    outer = chain(node("start", "input"), node("sub", "subgraph", workflow_id=inner_id), OUT)
+    _, started = await _launch(outer)
+    assert started["options"][ANALYSIS_GROUP] == contract.DEFAULTS
+    # 调用方已经给了这一组的，照它的
+    _, started = await _launch(outer, options={ANALYSIS_GROUP: {"max_calls": 2}})
+    assert started["options"][ANALYSIS_GROUP] == {"max_calls": 2}