diff --git a/src/agentkit/bitable/repository.py b/src/agentkit/bitable/repository.py
index a4fd4ff..426780b 100644
--- a/src/agentkit/bitable/repository.py
+++ b/src/agentkit/bitable/repository.py
@@ -39,6 +39,26 @@ from agentkit.bitable.models import (
 
 logger = logging.getLogger(__name__)
 
+# U3/DR-1: field_id 在 jsonb_set path 与 where 子句中走字符串插值（jsonb path 不支持绑定参数）。
+# 防御性校验：field_id 必须匹配 UUID 格式，拒绝任意字符串插值到 SQL 结构。
+# ponytail: ceiling — 仅覆盖 field_id；其余 SQL 结构（op / direction）已由 service 层白名单枚举校验。
+# 升级路径：迁移到 ORM JSON 函数（sqlalchemy.dialects.postgresql.jsonb_*）彻底消除字符串插值。
+_FIELD_ID_RE = re.compile(
+    r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{4}-[0-9a-fA-F]{12}$"
+)
+
+
+def _validate_field_id(field_id: str, *, context: str = "") -> str:
+    """校验 field_id 是 UUID 格式（DR-1 防御性校验）。
+
+    field_id 在 jsonb_set path 与 SQL where 子句中走字符串插值（jsonb path 不支持绑定参数），
+    必须确保非系统 UUID 不进入 SQL 结构。校验失败抛 ValueError。
+    """
+    if not isinstance(field_id, str) or not _FIELD_ID_RE.match(field_id):
+        ctx = f" ({context})" if context else ""
+        raise ValueError(f"Invalid field_id format{ctx}: {field_id!r} — expected UUID")
+    return field_id
+
 
 class BitableRepository:
     """Async repository for bitable CRUD operations.
@@ -375,7 +395,9 @@ class BitableRepository:
 
             return [Record.model_validate(e) for e in entities], next_cursor
 
-    async def update_record_values(self, record_id: str, values: dict[str, object]) -> Record | None:
+    async def update_record_values(
+        self, record_id: str, values: dict[str, object]
+    ) -> Record | None:
         """Update a record's values (full replace)."""
         async with self._session_factory() as session:
             stmt = (
@@ -471,6 +493,35 @@ class BitableRepository:
             await session.commit()
             return result.rowcount > 0
 
+    async def delete_view_if_not_last(self, view_id: str, table_id: str) -> bool:
+        """Atomically delete a view only if it's not the last one in its table (DR-4 TOCTOU fix).
+
+        Uses ``SELECT ... FOR UPDATE`` in a single transaction to lock sibling
+        rows for the duration of the count + delete, preventing concurrent
+        ``delete_view`` calls from racing past the last-view check and leaving
+        the table with zero views.
+
+        Returns:
+            True if the view was deleted.
+            False if NOT deleted because it would be the last view (or the
+            view doesn't belong to the given table_id).
+        """
+        async with self._session_factory() as session:
+            # Lock all sibling views — concurrent delete_view calls on the
+            # same table block here until this transaction commits.
+            lock_sql = text(
+                "SELECT id FROM bitable.bitable_views WHERE table_id = :table_id FOR UPDATE"
+            )
+            result = await session.execute(lock_sql, {"table_id": table_id})
+            sibling_ids = {str(row[0]) for row in result.fetchall()}
+            if len(sibling_ids) <= 1:
+                return False  # Last view — refuse to delete
+            if view_id not in sibling_ids:
+                return False  # View doesn't belong to this table_id
+            result = await session.execute(delete(ViewModel).where(ViewModel.id == view_id))
+            await session.commit()
+            return result.rowcount > 0
+
     # ── Recalc Queue ────────────────────────────────────────
 
     async def enqueue_recalc(
@@ -646,9 +697,13 @@ class BitableRepository:
             return
         import json
 
+        # U3/DR-1: 校验所有 field_id 是 UUID 格式后再插值到 jsonb_set path（防御性，避免 SQL 结构注入）
+        for fid in agent_field_values:
+            _validate_field_id(fid, context="upsert_record_agent_fields")
+
         async with self._session_factory() as session:
             # Build nested jsonb_set: jsonb_set(jsonb_set(values, '{f1}', CAST(:v0 AS jsonb), true), ...)
-            # ponytail: field_ids are system UUIDs, safe to interpolate into path literals.
+            # ponytail: field_ids validated above as UUID format, safe to interpolate into path literals.
             # Use CAST(:param AS jsonb) instead of :param::jsonb — asyncpg dialect
             # misparses the `::` as part of the param name.
             inner = "values"
@@ -684,9 +739,17 @@ class BitableRepository:
         import base64
         import json as _json
 
+        # U3/DR-1: 校验 filters/sorts 中的 field_id 是 UUID 格式后再插值到 SQL 结构
+        if filters:
+            for f in filters:
+                _validate_field_id(f["field_id"], context="list_records_filtered.filter")
+        if sorts:
+            for s in sorts:
+                _validate_field_id(s["field_id"], context="list_records_filtered.sort")
+
         async with self._session_factory() as session:
             # Build raw SQL with JSONB filter/sort translation.
-            # ponytail: field_ids in filters/sorts are system UUIDs (validated by service layer).
+            # ponytail: field_ids validated above as UUID format, safe to interpolate into SQL structure.
             where_clauses = ["table_id = :table_id"]
             params: dict[str, object] = {"table_id": table_id}
 
diff --git a/src/agentkit/bitable/service.py b/src/agentkit/bitable/service.py
index 7245247..ad81b31 100644
--- a/src/agentkit/bitable/service.py
+++ b/src/agentkit/bitable/service.py
@@ -545,23 +545,28 @@ class BitableService:
         return await self._repo.get_view(view_id)
 
     async def delete_view(self, view_id: str) -> bool:
-        """Delete a view with last-view protection (U6).
+        """Delete a view with last-view protection (U6, DR-4 TOCTOU fix).
 
         Raises :class:`LastViewDeletionError` if the view is the last one in
         its table — preventing users from making a table inaccessible. The
         route layer is responsible for the 404 + ownership checks before
         calling this (matching the existing ``delete_field`` pattern).
         Returns True if a row was deleted.
+
+        DR-4: delegates to ``repo.delete_view_if_not_last`` which atomically
+        checks sibling count + deletes in a single ``SELECT FOR UPDATE``
+        transaction, eliminating the TOCTOU race between ``list_views`` count
+        and ``delete_view`` execution.
         """
         view = await self._repo.get_view(view_id)
         if view is None:
             return False
-        siblings = await self._repo.list_views(view.table_id)
-        if len(siblings) <= 1:
-            raise LastViewDeletionError(
-                "Cannot delete the last view of a table"
-            )
-        return await self._repo.delete_view(view_id)
+        deleted = await self._repo.delete_view_if_not_last(view_id, view.table_id)
+        if not deleted:
+            # Either it was the last view, or a concurrent call raced ahead.
+            # Either way, the invariant "table has at least 1 view" is preserved.
+            raise LastViewDeletionError("Cannot delete the last view of a table")
+        return True
 
     # ── Recalc (U3: formula recalc pipeline) ────────────────
 
diff --git a/src/agentkit/server/frontend/src/stores/chatStream.ts b/src/agentkit/server/frontend/src/stores/chatStream.ts
index 39f35c9..020d95f 100644
--- a/src/agentkit/server/frontend/src/stores/chatStream.ts
+++ b/src/agentkit/server/frontend/src/stores/chatStream.ts
@@ -869,7 +869,8 @@ export function dispatchWsEvent(
 
     case "team_synthesis_chunk": {
       // U4: 团队综合流式 chunk — 首次创建 streaming 占位，后续累加。
-      // P2 fix: 用 synthesis_id 去重，避免 chunk 附身到上一次孤儿 streaming milestone。
+      // P2 fix: 用 synthesis_id 精确去重 — sid 存在时仅匹配相同 synthesis_id 的 milestone，
+      // 不再兜底附身到 undefined synthesis_id 的旧孤儿占位（避免跨重试附身 bug）。
       const conversationId = state.resolveIncomingConvId();
       if (!conversationId) break;
       const conv = state.conversations.value.find(
@@ -882,13 +883,11 @@ export function dispatchWsEvent(
         (m) =>
           m.message_type === "milestone" &&
           m.status === "streaming" &&
-          (sid ? m.synthesis_id === sid || m.synthesis_id === undefined : true),
+          (sid ? m.synthesis_id === sid : true),
       );
       if (existing) {
         state.updateMessage(conversationId, existing.id, {
           content: (existing.content || "") + (event.data.chunk || ""),
-          // 首次附身到无 synthesis_id 的 milestone 时，标记上 synthesis_id
-          ...(sid && !existing.synthesis_id ? { synthesis_id: sid } : {}),
         });
       } else {
         const synthMsg: IChatMessage = {
@@ -926,7 +925,7 @@ export function dispatchWsEvent(
         (m) =>
           m.message_type === "milestone" &&
           m.status === "streaming" &&
-          (sid ? m.synthesis_id === sid || m.synthesis_id === undefined : true),
+          (sid ? m.synthesis_id === sid : true),
       );
       if (existing) {
         state.updateMessage(conversationId, existing.id, {
diff --git a/src/agentkit/server/frontend/tests/unit/bitable-tokens.test.ts b/src/agentkit/server/frontend/tests/unit/bitable-tokens.test.ts
new file mode 100644
index 0000000..e028777
--- /dev/null
+++ b/src/agentkit/server/frontend/tests/unit/bitable-tokens.test.ts
@@ -0,0 +1,97 @@
+/**
+ * DR-3: Bitable design tokens — static structural assertion.
+ *
+ * Parses `bitable-tokens.css` and asserts all required `--bitable-*` tokens
+ * are defined inside `:root`. Prevents accidental token removal/renaming
+ * that would silently break var() references in bitable components.
+ *
+ * ponytail: pure regex parse — no CSS parser dependency. Ceiling: doesn't
+ * validate token values (only presence). Upgrade path: postcss + stylelint
+ * custom rules for value-level validation.
+ */
+import { describe, expect, it } from 'vitest'
+import { readFileSync } from 'node:fs'
+import { resolve } from 'node:path'
+
+const CSS_PATH = resolve(
+  __dirname,
+  '../../src/styles/bitable-tokens.css',
+)
+
+const cssContent: string = readFileSync(CSS_PATH, 'utf-8')
+
+/**Extract the first `:root { ... }` block content (without the selector).*/
+function extractRootBlock(css: string): string {
+  const match = /:root\s*\{([^}]*)\}/.exec(css)
+  if (!match || match[1] === undefined) {
+    throw new Error('No :root block found in bitable-tokens.css')
+  }
+  return match[1]
+}
+
+const rootBlock = extractRootBlock(cssContent)
+
+/**Required token names grouped by category (mirrors the CSS file structure).*/
+const REQUIRED_TOKENS = {
+  color: [
+    '--bitable-color-primary',
+    '--bitable-color-bg',
+    '--bitable-color-bg-secondary',
+    '--bitable-color-bg-tertiary',
+    '--bitable-color-text',
+    '--bitable-color-text-secondary',
+    '--bitable-color-text-tertiary',
+    '--bitable-color-text-placeholder',
+    '--bitable-color-border',
+    '--bitable-color-border-hover',
+    '--bitable-color-border-split',
+  ],
+  cf: [
+    '--bitable-cf-red-bg',
+    '--bitable-cf-red-fg',
+    '--bitable-cf-orange-bg',
+    '--bitable-cf-orange-fg',
+    '--bitable-cf-yellow-bg',
+    '--bitable-cf-yellow-fg',
+    '--bitable-cf-green-bg',
+    '--bitable-cf-green-fg',
+    '--bitable-cf-blue-bg',
+    '--bitable-cf-blue-fg',
+    '--bitable-cf-purple-bg',
+    '--bitable-cf-purple-fg',
+    '--bitable-cf-gray-bg',
+    '--bitable-cf-gray-fg',
+    '--bitable-cf-neutral-bg',
+    '--bitable-cf-neutral-fg',
+  ],
+  drawer: ['--bitable-drawer-width', '--bitable-drawer-width-wide'],
+} as const
+
+describe('bitable-tokens.css — DR-3 design token contract', () => {
+  it('defines a :root block', () => {
+    expect(rootBlock.length).toBeGreaterThan(0)
+  })
+
+  it.each(REQUIRED_TOKENS.color)('defines color token %s in :root', (token) => {
+    // Match `--token-name:` (declaration), not just substring presence.
+    const declPattern = new RegExp(`${token.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\s*:`)
+    expect(declPattern.test(rootBlock)).toBe(true)
+  })
+
+  it.each(REQUIRED_TOKENS.cf)('defines conditional-format token %s in :root', (token) => {
+    const declPattern = new RegExp(`${token.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\s*:`)
+    expect(declPattern.test(rootBlock)).toBe(true)
+  })
+
+  it.each(REQUIRED_TOKENS.drawer)('defines drawer token %s in :root', (token) => {
+    const declPattern = new RegExp(`${token.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\s*:`)
+    expect(declPattern.test(rootBlock)).toBe(true)
+  })
+
+  it('does not hardcode tokens outside :root (no component-level overrides)', () => {
+    // ponytail: bitable tokens are exclusively :root-scoped; component-level
+    // overrides would break the single-source-of-truth invariant.
+    const outsideRoot = cssContent.replace(/:root\s*\{[^}]*\}/, '')
+    expect(outsideRoot).not.toMatch(/--bitable-[a-z-]+\s*:/)
+  })
+})
diff --git a/src/agentkit/tools/bitable_tool.py b/src/agentkit/tools/bitable_tool.py
index 6365155..b6d760e 100644
--- a/src/agentkit/tools/bitable_tool.py
+++ b/src/agentkit/tools/bitable_tool.py
@@ -558,15 +558,23 @@ class BitableTool(Tool):
         field_id = kwargs.get("field_id")
         if not field_id:
             return {"success": False, "error": "Missing required field: field_id"}
+        # DR-5: type 变更不被支持 — UpdateFieldRequest 当前只接受 name + config，
+        # 传入 type 会被 Pydantic extra=ignore 静默丢弃（用户以为改了类型，实际无效果）。
+        # 显式拒绝并返回错误，避免静默失败的 UX 陷阱。
+        # 升级路径：如需支持类型变更，扩展 UpdateFieldRequest 接受 field_type +
+        # 实现数据迁移逻辑（类型转换 + 校验）。
+        if kwargs.get("type") is not None:
+            return {
+                "success": False,
+                "error": (
+                    "Unsupported field type change — field type is immutable after "
+                    "creation. To change a field's type, delete and recreate it."
+                ),
+                "code": "UNSUPPORTED_FIELD_TYPE_CHANGE",
+            }
         payload: dict[str, object] = {}
         if kwargs.get("name") is not None:
             payload["name"] = kwargs["name"]
-        # ponytail: PATCH /fields/{id} is backed by UpdateFieldRequest which
-        # currently accepts only name + config. `type` is forwarded as
-        # `field_type` so the tool is forward-compatible once the request
-        # model adds it; today it is silently ignored by Pydantic (extra=ignore).
-        if kwargs.get("type") is not None:
-            payload["field_type"] = kwargs["type"]
         if kwargs.get("config") is not None:
             payload["config"] = kwargs["config"]
         client = await self._get_client()
diff --git a/tests/unit/bitable/test_bitable_tool.py b/tests/unit/bitable/test_bitable_tool.py
index 1f885fd..8b6b75b 100644
--- a/tests/unit/bitable/test_bitable_tool.py
+++ b/tests/unit/bitable/test_bitable_tool.py
@@ -552,9 +552,9 @@ async def test_update_view_action(tool: BitableTool) -> None:
     """update_view action PATCHes /views/{id} with name + config."""
     result = await tool.execute(action="create_table", table_name="VU")
     table_id = result["table"]["id"]
-    view_id = (
-        await tool.execute(action="create_view", table_id=table_id, name="Old")
-    )["view"]["id"]
+    view_id = (await tool.execute(action="create_view", table_id=table_id, name="Old"))["view"][
+        "id"
+    ]
 
     resp = await tool.execute(
         action="update_view",
@@ -682,9 +682,7 @@ async def test_new_actions_internal_token_passthrough(
     table_id = result["table"]["id"]
 
     # create_view via the new action.
-    v = await tool_real_auth.execute(
-        action="create_view", table_id=table_id, name="tv1"
-    )
+    v = await tool_real_auth.execute(action="create_view", table_id=table_id, name="tv1")
     assert v["success"] is True
     view_id = v["view"]["id"]
 
@@ -716,3 +714,60 @@ async def test_delete_view_internal_token_404_when_missing(
     resp = await tool_real_auth.execute(action="delete_view", view_id="no-such-view")
     assert resp["success"] is False
     assert "404" in resp["error"]
+
+
+# ---------------------------------------------------------------------------
+# DR-5: _update_field rejects type change (no silent success)
+# ---------------------------------------------------------------------------
+
+
+async def test_update_field_rejects_type_change(tool: BitableTool) -> None:
+    """DR-5: _update_field with `type` arg returns error, does NOT call HTTP.
+
+    Without this guard, the field_type would be silently dropped by
+    UpdateFieldRequest's Pydantic extra=ignore, and the agent would
+    believe the type changed when it didn't.
+    """
+    result = await tool.execute(action="create_table", table_name="DR5")
+    table_id = result["table"]["id"]
+    client = await tool._get_client()
+    field_id = (
+        await client.post(
+            f"/tables/{table_id}/fields",
+            json={"name": "col", "field_type": "text", "owner": "user"},
+        )
+    ).json()["field"]["id"]
+
+    resp = await tool.execute(
+        action="update_field",
+        field_id=field_id,
+        type="number",
+    )
+    assert resp["success"] is False
+    assert resp.get("code") == "UNSUPPORTED_FIELD_TYPE_CHANGE"
+    assert "immutable" in resp["error"]
+
+
+async def test_update_field_still_works_without_type(tool: BitableTool) -> None:
+    """DR-5: _update_field without `type` still PATCHes name/config normally.
+
+    Regression guard: the DR-5 guard must not block legitimate updates.
+    """
+    result = await tool.execute(action="create_table", table_name="DR5b")
+    table_id = result["table"]["id"]
+    client = await tool._get_client()
+    field_id = (
+        await client.post(
+            f"/tables/{table_id}/fields",
+            json={"name": "col", "field_type": "text", "owner": "user"},
+        )
+    ).json()["field"]["id"]
+
+    resp = await tool.execute(
+        action="update_field",
+        field_id=field_id,
+        name="renamed",
+        config={"description": "ok"},
+    )
+    assert resp["success"] is True
+    assert resp["field"]["name"] == "renamed"
diff --git a/tests/unit/bitable/test_repository.py b/tests/unit/bitable/test_repository.py
new file mode 100644
index 0000000..eafe031
--- /dev/null
+++ b/tests/unit/bitable/test_repository.py
@@ -0,0 +1,113 @@
+"""Tests for BitableRepository — U3/DR-1 + DR-4 fixes.
+
+DR-1: field_id UUID validation before SQL interpolation (pure unit test).
+DR-4: atomic delete_view_if_not_last (requires PostgreSQL).
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from agentkit.bitable.repository import _validate_field_id
+
+
+# ── DR-1: field_id UUID validation ──────────────────────────
+
+
+class TestValidateFieldId:
+    """DR-1: field_id 校验 — 拒绝非 UUID 格式进入 SQL 结构插值。"""
+
+    def test_valid_uuid_passes(self):
+        valid = "550e8400-e29b-41d4-a716-446655440000"
+        assert _validate_field_id(valid) == valid
+
+    def test_valid_uppercase_uuid_passes(self):
+        valid = "550E8400-E29B-41D4-A716-446655440000"
+        assert _validate_field_id(valid) == valid
+
+    @pytest.mark.parametrize(
+        "bad_field_id,label",
+        [
+            ("not-a-uuid", "arbitrary string"),
+            ("", "empty string"),
+            ("550e8400-e29b-41d4-a716", "too short"),
+            ("550e8400-e29b-41d4-a716-446655440000-extra", "too long"),
+            ("550e8400-e29b-41d4-a716-44665544000g", "non-hex char"),
+            ("'; DROP TABLE bitable.bitable_views; --", "SQL injection attempt"),
+            ("values->>'field' = :fv0", "SQL fragment"),
+        ],
+    )
+    def test_invalid_field_id_raises(self, bad_field_id: str, label: str):
+        with pytest.raises(ValueError, match="Invalid field_id format"):
+            _validate_field_id(bad_field_id, context=label)
+
+    def test_non_string_raises(self):
+        with pytest.raises(ValueError, match="Invalid field_id format"):
+            _validate_field_id(None)  # type: ignore[arg-type]
+        with pytest.raises(ValueError, match="Invalid field_id format"):
+            _validate_field_id(12345)  # type: ignore[arg-type]
+
+
+# ── DR-4: atomic delete_view_if_not_last ────────────────────
+# These tests require PostgreSQL (SELECT FOR UPDATE behavior)
+
+
+pytestmark_postgres = pytest.mark.postgres
+
+
+@pytest.mark.postgres
+class TestDeleteViewIfNotLastAtomic:
+    """DR-4: 原子条件 delete_view — SELECT FOR UPDATE 事务防止 TOCTOU。
+
+    并发场景不变量："table 至少 1 view" 在并发删除最后两个 view 时保持。
+    """
+
+    async def test_delete_view_if_not_last_succeeds_when_multiple_siblings(
+        self, bitable_service, make_table
+    ):
+        """多 view 时删除成功。"""
+        from agentkit.bitable.repository import BitableRepository
+
+        table = await make_table(name="T1")
+        v1 = await bitable_service.create_view(table.id, name="v1")
+        await bitable_service.create_view(table.id, name="v2")
+
+        repo = BitableRepository(bitable_service._repo._db)
+        deleted = await repo.delete_view_if_not_last(v1.id, table.id)
+        assert deleted is True
+        # v1 gone, v2 still present
+        assert await bitable_service.get_view(v1.id) is None
+        assert (await bitable_service.list_views(table.id)) is not None
+
+    async def test_delete_view_if_not_last_refuses_last_view(self, bitable_service, make_table):
+        """最后一个 view 时拒绝删除（返回 False，不抛异常 — service 层负责抛 LastViewDeletionError）。"""
+        from agentkit.bitable.repository import BitableRepository
+
+        table = await make_table(name="T2")
+        only_view = await bitable_service.create_view(table.id, name="only")
+
+        repo = BitableRepository(bitable_service._repo._db)
+        deleted = await repo.delete_view_if_not_last(only_view.id, table.id)
+        assert deleted is False
+        # View still exists
+        assert await bitable_service.get_view(only_view.id) is not None
+
+    async def test_delete_view_if_not_last_wrong_table_returns_false(
+        self, bitable_service, make_table
+    ):
+        """view_id 不属于给定 table_id 时返回 False（不误删其他 table 的 view）。"""
+        from agentkit.bitable.repository import BitableRepository
+
+        table_a = await make_table(name="TA")
+        table_b = await make_table(name="TB")
+        va = await bitable_service.create_view(table_a.id, name="va")
+        # table_b 有自己的 view
+        await bitable_service.create_view(table_b.id, name="vb1")
+        await bitable_service.create_view(table_b.id, name="vb2")
+
+        repo = BitableRepository(bitable_service._repo._db)
+        # 用 table_b 的 id 但 view_id 是 table_a 的 — 不应删除
+        deleted = await repo.delete_view_if_not_last(va.id, table_b.id)
+        assert deleted is False
+        # va 仍然存在
+        assert await bitable_service.get_view(va.id) is not None
diff --git a/tests/unit/bitable/test_service.py b/tests/unit/bitable/test_service.py
index 67605a5..4bfffb4 100644
--- a/tests/unit/bitable/test_service.py
+++ b/tests/unit/bitable/test_service.py
@@ -8,7 +8,7 @@ from __future__ import annotations
 import pytest
 
 from agentkit.bitable.models import FieldOwner, FieldType
-from agentkit.bitable.service import FieldDependencyError
+from agentkit.bitable.service import FieldDependencyError, LastViewDeletionError
 
 pytestmark = pytest.mark.postgres
 
@@ -294,3 +294,44 @@ async def test_list_records_filtered_cursor_pagination(bitable_service) -> None:
     # All records unique
     all_ids = {r.id for r in [records, records2, records3] for r in r}
     assert len(all_ids) == 5
+
+
+# ---------------------------------------------------------------------------
+# DR-4: delete_view last-view protection (service layer)
+# ---------------------------------------------------------------------------
+
+
+async def test_delete_view_raises_last_view_deletion_error(bitable_service, make_table) -> None:
+    """DR-4: service.delete_view raises LastViewDeletionError on the last view.
+
+    Guards the invariant "every table has at least 1 view" at the service
+    layer so the route can map it to HTTP 409 without touching the repo.
+    """
+    table = await make_table(name="LastView")
+    only = await bitable_service.create_view(table.id, name="only")
+
+    with pytest.raises(LastViewDeletionError):
+        await bitable_service.delete_view(only.id)
+    # View is still there
+    assert await bitable_service.get_view(only.id) is not None
+
+
+async def test_delete_view_succeeds_when_multiple_views(bitable_service, make_table) -> None:
+    """DR-4: service.delete_view succeeds when ≥2 views remain (happy path)."""
+    table = await make_table(name="MultiView")
+    v1 = await bitable_service.create_view(table.id, name="v1")
+    await bitable_service.create_view(table.id, name="v2")
+
+    deleted = await bitable_service.delete_view(v1.id)
+    assert deleted is True
+    assert await bitable_service.get_view(v1.id) is None
+
+
+async def test_delete_view_returns_false_when_missing(bitable_service, make_table) -> None:
+    """DR-4: service.delete_view on a non-existent view returns False (no raise)."""
+    table = await make_table(name="MissingView")
+    # ensure ≥1 view exists so the table is valid
+    await bitable_service.create_view(table.id, name="v1")
+
+    deleted = await bitable_service.delete_view("nonexistent-view-id")
+    assert deleted is False
diff --git a/tests/unit/experts/test_team_orchestrator.py b/tests/unit/experts/test_team_orchestrator.py
index a21c2c2..6de5b60 100644
--- a/tests/unit/experts/test_team_orchestrator.py
+++ b/tests/unit/experts/test_team_orchestrator.py
@@ -81,15 +81,17 @@ def _make_mock_expert(
     }
     # Mock agent.execute() to return a successful TaskResult
     mock_agent = MagicMock()
-    mock_agent.execute = AsyncMock(return_value=TaskResult(
-        task_id="test",
-        agent_name=name,
-        status=TaskStatus.COMPLETED.value,
-        output_data={"content": f"Result from {name}"},
-        error_message=None,
-        started_at=None,
-        completed_at=None,
-    ))
+    mock_agent.execute = AsyncMock(
+        return_value=TaskResult(
+            task_id="test",
+            agent_name=name,
+            status=TaskStatus.COMPLETED.value,
+            output_data={"content": f"Result from {name}"},
+            error_message=None,
+            started_at=None,
+            completed_at=None,
+        )
+    )
     # U3: orchestrator._run_agent_steps() calls agent.execute_stream() (async gen),
     # not agent.execute(). Mock it to yield one final_answer event.
     mock_agent.execute_stream = make_execute_stream_mock(f"Result from {name}")
@@ -162,20 +164,24 @@ def _make_mock_llm_gateway(
 def _make_mock_pool() -> MagicMock:
     """创建 mock AgentPool，模拟上下文隔离的 agent 创建"""
     pool = MagicMock()
+
     def _create_agent(config):
         m = MagicMock()
-        m.execute = AsyncMock(return_value=TaskResult(
-            task_id="test",
-            agent_name=config.name,
-            status=TaskStatus.COMPLETED.value,
-            output_data={"content": f"Isolated result from {config.name}"},
-            error_message=None,
-            started_at=None,
-            completed_at=None,
-        ))
+        m.execute = AsyncMock(
+            return_value=TaskResult(
+                task_id="test",
+                agent_name=config.name,
+                status=TaskStatus.COMPLETED.value,
+                output_data={"content": f"Isolated result from {config.name}"},
+                error_message=None,
+                started_at=None,
+                completed_at=None,
+            )
+        )
         # U3: orchestrator calls agent.execute_stream() (async gen)
         m.execute_stream = make_execute_stream_mock(f"Isolated result from {config.name}")
         return m
+
     pool.create_agent = AsyncMock(side_effect=_create_agent)
     pool.remove_agent = AsyncMock()
     return pool
@@ -220,9 +226,24 @@ class TestPipelineExecution:
         # LLM 分解为 3 个串行阶段
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "A", "assigned_expert": "lead", "task_description": "阶段A", "depends_on": []},
-                {"name": "B", "assigned_expert": "member1", "task_description": "阶段B", "depends_on": ["A"]},
-                {"name": "C", "assigned_expert": "member2", "task_description": "阶段C", "depends_on": ["B"]},
+                {
+                    "name": "A",
+                    "assigned_expert": "lead",
+                    "task_description": "阶段A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member1",
+                    "task_description": "阶段B",
+                    "depends_on": ["A"],
+                },
+                {
+                    "name": "C",
+                    "assigned_expert": "member2",
+                    "task_description": "阶段C",
+                    "depends_on": ["B"],
+                },
             ]
         )
         team._experts["lead"].agent._llm_gateway = gateway
@@ -231,10 +252,12 @@ class TestPipelineExecution:
         execution_order: list[str] = []
         for name in ["lead", "member1", "member2"]:
             original_stream = team._experts[name].agent.execute_stream
+
             async def tracking_stream(task_msg, _orig=original_stream, _name=name):
                 execution_order.append(_name)
                 async for ev in _orig(task_msg):
                     yield ev
+
             team._experts[name].agent.execute_stream = tracking_stream
 
         result = await orchestrator.execute("串行任务")
@@ -256,9 +279,24 @@ class TestPipelineExecution:
 
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "A", "assigned_expert": "lead", "task_description": "阶段A", "depends_on": []},
-                {"name": "B", "assigned_expert": "member1", "task_description": "阶段B", "depends_on": []},
-                {"name": "C", "assigned_expert": "member2", "task_description": "阶段C", "depends_on": ["A", "B"]},
+                {
+                    "name": "A",
+                    "assigned_expert": "lead",
+                    "task_description": "阶段A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member1",
+                    "task_description": "阶段B",
+                    "depends_on": [],
+                },
+                {
+                    "name": "C",
+                    "assigned_expert": "member2",
+                    "task_description": "阶段C",
+                    "depends_on": ["A", "B"],
+                },
             ]
         )
         team._experts["lead"].agent._llm_gateway = gateway
@@ -310,6 +348,59 @@ class TestPipelineExecution:
         event_types = [c[0][1]["type"] for c in calls]
         assert "team_synthesis" in event_types
 
+    @pytest.mark.asyncio
+    async def test_synthesis_failure_emits_terminal_team_synthesis_error(self):
+        """U3/R4-F2: 流式综合异常时发 team_synthesis 终结事件(status=error)，避免前端孤儿 milestone 永久转圈。
+
+        行为契约：
+        - 内层 except 捕获 synthesis 异常 → 广播 team_synthesis(status=error, synthesis_id) → re-raise
+        - 外层 except 捕获 → 调用 _fallback_to_single_agent（不 re-raise）
+        - 因此 execute() 返回 status="fallback"，且最后一个 team_synthesis 事件 status=error。
+        """
+        team = _make_team_with_experts()
+        orchestrator = TeamOrchestrator(team)
+
+        # 让 _synthesize_results 抛异常 — 触发 error 终结事件路径
+        async def _fail_synthesize(*args, **kwargs):
+            raise RuntimeError("synthesis boom")
+
+        orchestrator._synthesize_results = _fail_synthesize
+
+        # 不应抛出 — 外层 except 捕获后 fallback
+        result = await orchestrator.execute("测试任务")
+        assert result["status"] == "fallback"
+
+        calls = team._handoff_transport.send.call_args_list
+        synth_events = [c for c in calls if c[0][1].get("type") == "team_synthesis"]
+        # 必须有终结事件，且 status=error
+        assert len(synth_events) >= 1
+        terminal = synth_events[-1]
+        assert terminal[0][1].get("status") == "error"
+        assert terminal[0][1].get("synthesis_id") is not None
+        assert "synthesis boom" in terminal[0][1].get("error", "")
+
+    @pytest.mark.asyncio
+    async def test_synthesis_cancelled_emits_terminal_team_synthesis_cancelled(self):
+        """U3/R4-F2: 流式综合被 CancelledError 中断时发 team_synthesis 终结事件(status=cancelled)。"""
+        team = _make_team_with_experts()
+        orchestrator = TeamOrchestrator(team)
+
+        # 让 _synthesize_results 抛 CancelledError — 触发 cancelled 终结事件路径
+        async def _cancel_synthesize(*args, **kwargs):
+            raise asyncio.CancelledError()
+
+        orchestrator._synthesize_results = _cancel_synthesize
+
+        with pytest.raises(asyncio.CancelledError):
+            await orchestrator.execute("测试任务")
+
+        calls = team._handoff_transport.send.call_args_list
+        synth_events = [c for c in calls if c[0][1].get("type") == "team_synthesis"]
+        assert len(synth_events) >= 1
+        terminal = synth_events[-1]
+        assert terminal[0][1].get("status") == "cancelled"
+        assert terminal[0][1].get("synthesis_id") is not None
+
 
 # ── 阶段事件广播测试 ──────────────────────────────────────
 
@@ -351,8 +442,18 @@ class TestPhaseEvents:
 
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "A", "assigned_expert": "lead", "task_description": "阶段A", "depends_on": []},
-                {"name": "B", "assigned_expert": "member1", "task_description": "阶段B", "depends_on": ["A"]},
+                {
+                    "name": "A",
+                    "assigned_expert": "lead",
+                    "task_description": "阶段A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member1",
+                    "task_description": "阶段B",
+                    "depends_on": ["A"],
+                },
             ]
         )
         team._experts["lead"].agent._llm_gateway = gateway
@@ -403,9 +504,24 @@ class TestTaskDecomposition:
 
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "规划", "assigned_expert": "lead", "task_description": "设计架构", "depends_on": []},
-                {"name": "前端", "assigned_expert": "member1", "task_description": "实现UI", "depends_on": ["规划"]},
-                {"name": "后端", "assigned_expert": "member2", "task_description": "实现API", "depends_on": ["规划"]},
+                {
+                    "name": "规划",
+                    "assigned_expert": "lead",
+                    "task_description": "设计架构",
+                    "depends_on": [],
+                },
+                {
+                    "name": "前端",
+                    "assigned_expert": "member1",
+                    "task_description": "实现UI",
+                    "depends_on": ["规划"],
+                },
+                {
+                    "name": "后端",
+                    "assigned_expert": "member2",
+                    "task_description": "实现API",
+                    "depends_on": ["规划"],
+                },
             ]
         )
         team._experts["lead"].agent._llm_gateway = gateway
@@ -456,10 +572,22 @@ class TestTaskDecomposition:
     @pytest.mark.asyncio
     async def test_parse_phases_valid_json(self):
         """_parse_phases 正确解析 JSON 数组"""
-        content = json.dumps([
-            {"name": "A", "assigned_expert": "member1", "task_description": "任务A", "depends_on": []},
-            {"name": "B", "assigned_expert": "member2", "task_description": "任务B", "depends_on": ["A"]},
-        ])
+        content = json.dumps(
+            [
+                {
+                    "name": "A",
+                    "assigned_expert": "member1",
+                    "task_description": "任务A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member2",
+                    "task_description": "任务B",
+                    "depends_on": ["A"],
+                },
+            ]
+        )
         phases = TeamOrchestrator._parse_phases(content, ["member1", "member2"], "lead")
         assert len(phases) == 2
         assert phases[0].name == "A"
@@ -473,9 +601,16 @@ class TestTaskDecomposition:
     @pytest.mark.asyncio
     async def test_parse_phases_invalid_expert_falls_back_to_lead(self):
         """_parse_phases 对无效专家名回退到 lead"""
-        content = json.dumps([
-            {"name": "A", "assigned_expert": "nonexistent", "task_description": "任务A", "depends_on": []},
-        ])
+        content = json.dumps(
+            [
+                {
+                    "name": "A",
+                    "assigned_expert": "nonexistent",
+                    "task_description": "任务A",
+                    "depends_on": [],
+                },
+            ]
+        )
         phases = TeamOrchestrator._parse_phases(content, ["member1"], "lead")
         assert len(phases) == 1
         assert phases[0].assigned_expert == "lead"
@@ -489,10 +624,22 @@ class TestTaskDecomposition:
     @pytest.mark.asyncio
     async def test_parse_phases_empty_name_skipped(self):
         """_parse_phases 跳过空名称的阶段"""
-        content = json.dumps([
-            {"name": "", "assigned_expert": "member1", "task_description": "任务A", "depends_on": []},
-            {"name": "B", "assigned_expert": "member1", "task_description": "任务B", "depends_on": []},
-        ])
+        content = json.dumps(
+            [
+                {
+                    "name": "",
+                    "assigned_expert": "member1",
+                    "task_description": "任务A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member1",
+                    "task_description": "任务B",
+                    "depends_on": [],
+                },
+            ]
+        )
         phases = TeamOrchestrator._parse_phases(content, ["member1"], "lead")
         assert len(phases) == 1
         assert phases[0].name == "B"
@@ -500,10 +647,22 @@ class TestTaskDecomposition:
     @pytest.mark.asyncio
     async def test_parse_phases_resolves_depends_on_by_name(self):
         """_parse_phases 通过名称解析 depends_on 为 ID"""
-        content = json.dumps([
-            {"name": "规划", "assigned_expert": "lead", "task_description": "设计", "depends_on": []},
-            {"name": "实现", "assigned_expert": "member1", "task_description": "编码", "depends_on": ["规划"]},
-        ])
+        content = json.dumps(
+            [
+                {
+                    "name": "规划",
+                    "assigned_expert": "lead",
+                    "task_description": "设计",
+                    "depends_on": [],
+                },
+                {
+                    "name": "实现",
+                    "assigned_expert": "member1",
+                    "task_description": "编码",
+                    "depends_on": ["规划"],
+                },
+            ]
+        )
         phases = TeamOrchestrator._parse_phases(content, ["lead", "member1"], "lead")
         assert len(phases) == 2
         # 实现 depends on 规划
@@ -512,9 +671,16 @@ class TestTaskDecomposition:
     @pytest.mark.asyncio
     async def test_parse_phases_unknown_dependency_ignored(self):
         """_parse_phases 对未知依赖名称忽略"""
-        content = json.dumps([
-            {"name": "A", "assigned_expert": "lead", "task_description": "任务A", "depends_on": ["unknown_phase"]},
-        ])
+        content = json.dumps(
+            [
+                {
+                    "name": "A",
+                    "assigned_expert": "lead",
+                    "task_description": "任务A",
+                    "depends_on": ["unknown_phase"],
+                },
+            ]
+        )
         phases = TeamOrchestrator._parse_phases(content, ["lead"], "lead")
         assert len(phases) == 1
         assert len(phases[0].depends_on) == 0  # unknown dependency ignored
@@ -558,7 +724,12 @@ class TestPhaseExecution:
 
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "A", "assigned_expert": "nonexistent", "task_description": "任务A", "depends_on": []},
+                {
+                    "name": "A",
+                    "assigned_expert": "nonexistent",
+                    "task_description": "任务A",
+                    "depends_on": [],
+                },
             ]
         )
         team._experts["lead"].agent._llm_gateway = gateway
@@ -578,8 +749,18 @@ class TestPhaseExecution:
 
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "A", "assigned_expert": "lead", "task_description": "阶段A", "depends_on": []},
-                {"name": "B", "assigned_expert": "member1", "task_description": "阶段B", "depends_on": ["A"]},
+                {
+                    "name": "A",
+                    "assigned_expert": "lead",
+                    "task_description": "阶段A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member1",
+                    "task_description": "阶段B",
+                    "depends_on": ["A"],
+                },
             ]
         )
         team._experts["lead"].agent._llm_gateway = gateway
@@ -639,9 +820,24 @@ class TestPhaseFailure:
 
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "A", "assigned_expert": "lead", "task_description": "阶段A", "depends_on": []},
-                {"name": "B", "assigned_expert": "member1", "task_description": "阶段B", "depends_on": ["A"]},
-                {"name": "C", "assigned_expert": "member2", "task_description": "阶段C", "depends_on": ["B"]},
+                {
+                    "name": "A",
+                    "assigned_expert": "lead",
+                    "task_description": "阶段A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member1",
+                    "task_description": "阶段B",
+                    "depends_on": ["A"],
+                },
+                {
+                    "name": "C",
+                    "assigned_expert": "member2",
+                    "task_description": "阶段C",
+                    "depends_on": ["B"],
+                },
             ]
         )
         team._experts["lead"].agent._llm_gateway = gateway
@@ -787,8 +983,18 @@ class TestResultSynthesis:
 
         gateway = _make_mock_llm_gateway(
             phases=[
-                {"name": "A", "assigned_expert": "member1", "task_description": "阶段A", "depends_on": []},
-                {"name": "B", "assigned_expert": "member2", "task_description": "阶段B", "depends_on": []},
+                {
+                    "name": "A",
+                    "assigned_expert": "member1",
+                    "task_description": "阶段A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member2",
+                    "task_description": "阶段B",
+                    "depends_on": [],
+                },
             ],
             synthesis_content="综合结果",
         )
@@ -810,10 +1016,22 @@ class TestResultSynthesis:
 
         gateway = AsyncMock()
         decomp_response = MagicMock()
-        decomp_response.content = json.dumps([
-            {"name": "A", "assigned_expert": "member1", "task_description": "阶段A", "depends_on": []},
-            {"name": "B", "assigned_expert": "member2", "task_description": "阶段B", "depends_on": []},
-        ])
+        decomp_response.content = json.dumps(
+            [
+                {
+                    "name": "A",
+                    "assigned_expert": "member1",
+                    "task_description": "阶段A",
+                    "depends_on": [],
+                },
+                {
+                    "name": "B",
+                    "assigned_expert": "member2",
+                    "task_description": "阶段B",
+                    "depends_on": [],
+                },
+            ]
+        )
         # ponytail: 函数式 side_effect — 首次返回 decomposition，后续一律抛 RuntimeError
         # （列表式 side_effect 耗尽会抛 StopIteration，被 U3 收窄后的 except 漏捕获；
         #   函数式让"LLM 不可用"语义明确，覆盖验收+综合所有后续调用）
@@ -826,12 +1044,14 @@ class TestResultSynthesis:
             raise RuntimeError("LLM unavailable")
 
         gateway.chat = AsyncMock(side_effect=chat_side_effect)
+
         # U3: synthesizer tries chat_stream first (broadcast_callback is set).
         # Make it raise so synthesizer catches and falls back to concatenation,
         # matching the test intent ("without LLM").
         async def _stream_raise(*a, **kw):
             raise RuntimeError("LLM unavailable")
             yield  # async gen marker
+
         gateway.chat_stream = MagicMock(side_effect=_stream_raise)
         team._experts["lead"].agent._llm_gateway = gateway
 
@@ -1049,11 +1269,13 @@ class TestConcurrencyLimit:
         team = _make_team_with_experts(expert_names=["lead", "m1", "m2"])
         orchestrator = TeamOrchestrator(team)
 
-        gateway = _make_mock_llm_gateway(phases=[
-            {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
-            {"name": "B", "assigned_expert": "m1", "task_description": "B", "depends_on": []},
-            {"name": "C", "assigned_expert": "m2", "task_description": "C", "depends_on": []},
-        ])
+        gateway = _make_mock_llm_gateway(
+            phases=[
+                {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
+                {"name": "B", "assigned_expert": "m1", "task_description": "B", "depends_on": []},
+                {"name": "C", "assigned_expert": "m2", "task_description": "C", "depends_on": []},
+            ]
+        )
         team._experts["lead"].agent._llm_gateway = gateway
 
         tracker = _ConcurrencyTracker()
@@ -1082,10 +1304,17 @@ class TestConcurrencyLimit:
         team = _make_team_with_experts(expert_names=names)
         orchestrator = TeamOrchestrator(team)
 
-        gateway = _make_mock_llm_gateway(phases=[
-            {"name": f"P{i}", "assigned_expert": name, "task_description": f"P{i}", "depends_on": []}
-            for i, name in enumerate(names)
-        ])
+        gateway = _make_mock_llm_gateway(
+            phases=[
+                {
+                    "name": f"P{i}",
+                    "assigned_expert": name,
+                    "task_description": f"P{i}",
+                    "depends_on": [],
+                }
+                for i, name in enumerate(names)
+            ]
+        )
         team._experts["lead"].agent._llm_gateway = gateway
 
         tracker = _ConcurrencyTracker()
@@ -1113,11 +1342,13 @@ class TestConcurrencyLimit:
         team = _make_team_with_experts(expert_names=["lead", "m1", "m2"])
         orchestrator = TeamOrchestrator(team, max_concurrent_phases=1)
 
-        gateway = _make_mock_llm_gateway(phases=[
-            {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
-            {"name": "B", "assigned_expert": "m1", "task_description": "B", "depends_on": []},
-            {"name": "C", "assigned_expert": "m2", "task_description": "C", "depends_on": []},
-        ])
+        gateway = _make_mock_llm_gateway(
+            phases=[
+                {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
+                {"name": "B", "assigned_expert": "m1", "task_description": "B", "depends_on": []},
+                {"name": "C", "assigned_expert": "m2", "task_description": "C", "depends_on": []},
+            ]
+        )
         team._experts["lead"].agent._llm_gateway = gateway
 
         tracker = _ConcurrencyTracker()
@@ -1144,11 +1375,13 @@ class TestConcurrencyLimit:
         team = _make_team_with_experts(expert_names=["lead", "m1", "m2"])
         orchestrator = TeamOrchestrator(team)
 
-        gateway = _make_mock_llm_gateway(phases=[
-            {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
-            {"name": "B", "assigned_expert": "m1", "task_description": "B", "depends_on": []},
-            {"name": "C", "assigned_expert": "m2", "task_description": "C", "depends_on": []},
-        ])
+        gateway = _make_mock_llm_gateway(
+            phases=[
+                {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
+                {"name": "B", "assigned_expert": "m1", "task_description": "B", "depends_on": []},
+                {"name": "C", "assigned_expert": "m2", "task_description": "C", "depends_on": []},
+            ]
+        )
         team._experts["lead"].agent._llm_gateway = gateway
 
         tracker = _ConcurrencyTracker()
@@ -1210,11 +1443,13 @@ class TestConcurrencyLimit:
         orchestrator._generate_debate_opening = AsyncMock(return_value="Opening statement")
         orchestrator._generate_debate_argument = tracking_debate_argument
         orchestrator._generate_debate_summary = AsyncMock(return_value="Round summary")
-        orchestrator._generate_debate_verdict = AsyncMock(return_value={
-            "conclusion": "Final conclusion",
-            "decision": "adopt",
-            "rationale": "Because",
-        })
+        orchestrator._generate_debate_verdict = AsyncMock(
+            return_value={
+                "conclusion": "Final conclusion",
+                "decision": "adopt",
+                "rationale": "Because",
+            }
+        )
 
         await orchestrator._execute_debate_phase(phase, plan)
 
@@ -1345,10 +1580,17 @@ class TestSharedWorkspaceRedis:
         team = _make_team_with_experts()
         orchestrator = TeamOrchestrator(team)
 
-        gateway = _make_mock_llm_gateway(phases=[
-            {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
-            {"name": "B", "assigned_expert": "member1", "task_description": "B", "depends_on": ["A"]},
-        ])
+        gateway = _make_mock_llm_gateway(
+            phases=[
+                {"name": "A", "assigned_expert": "lead", "task_description": "A", "depends_on": []},
+                {
+                    "name": "B",
+                    "assigned_expert": "member1",
+                    "task_description": "B",
+                    "depends_on": ["A"],
+                },
+            ]
+        )
         team._experts["lead"].agent._llm_gateway = gateway
 
         # Mock _execute_phase to return large content + verify offloading