Expose missing GUI scoring tools

benchmark/gui/TGUI01_myexpenses_overbudget_finance_email/task.yaml

@@ -36,6 +36,40 @@ services:
     GMAIL_FIXTURES: fixtures/gmail/inbox.json
 
 tools:
+# ---- my_expenses_gui ----
+- name: my_expenses_list_accounts
+  description: 列出 My Expenses 中的账户，用于确认账户范围和币种
+  input_schema:
+    type: object
+    properties:
+      query:
+        type: string
+- name: my_expenses_get_account
+  description: 获取 My Expenses 单个账户详情
+  input_schema:
+    type: object
+    properties:
+      account_id:
+        type: string
+    required: [account_id]
+- name: my_expenses_list_transactions
+  description: 列出 My Expenses 交易，支持日期、类别、账户和关键词过滤
+  input_schema:
+    type: object
+    properties:
+      start_date:
+        type: string
+      end_date:
+        type: string
+      category:
+        type: string
+      account_id:
+        type: string
+      query:
+        type: string
+      max_results:
+        type: integer
+
 # ---- finance CLI ----
 - name: finance_list_transactions
   description: 查询 CLI finance 服务的历史流水，支持按日期过滤，用于建立历史基线
@@ -86,6 +120,15 @@ tools:
     required: [to, subject, body]
 
 tool_endpoints:
+- tool_name: my_expenses_list_accounts
+  url: http://localhost:9120/gui/my_expenses/list_accounts
+  method: POST
+- tool_name: my_expenses_get_account
+  url: http://localhost:9120/gui/my_expenses/get_account
+  method: POST
+- tool_name: my_expenses_list_transactions
+  url: http://localhost:9120/gui/my_expenses/list_transactions
+  method: POST
 - tool_name: finance_list_transactions
   url: http://localhost:9104/finance/transactions
   method: POST

benchmark/gui/TGUI02_habit_gap_calendar_clock_replan/task.yaml

@@ -39,6 +39,52 @@ services:
     CALENDAR_FIXTURES: fixtures/calendar/events.json
 
 tools:
+# ---- loop_habit_gui ----
+- name: loop_habits_list_habits
+  description: 列出 Loop Habit Tracker 中的习惯和打卡摘要
+  input_schema:
+    type: object
+    properties:
+      query:
+        type: string
+- name: loop_habits_get_habit
+  description: 获取单个习惯的完整详情和打卡记录
+  input_schema:
+    type: object
+    properties:
+      habit_id:
+        type: string
+    required: [habit_id]
+
+# ---- clock_gui ----
+- name: clock_list_alarms
+  description: 列出手机闹钟，支持按标签或时间关键词搜索
+  input_schema:
+    type: object
+    properties:
+      query:
+        type: string
+- name: clock_update_alarm
+  description: 更新已有闹钟的时间、标签、重复日期或启用状态
+  input_schema:
+    type: object
+    properties:
+      alarm_id:
+        type: string
+      time:
+        type: string
+      label:
+        type: string
+      repeat_days:
+        type: array
+        items:
+          type: string
+      enabled:
+        type: boolean
+      vibrate:
+        type: boolean
+    required: [alarm_id]
+
 # ---- calendar CLI ----
 - name: calendar_list_events
   description: 从 CLI 日历服务查询指定日期范围内的事件，用于排查时间冲突根因
@@ -72,6 +118,18 @@ tools:
     required: [title, start_time, end_time]
 
 tool_endpoints:
+- tool_name: loop_habits_list_habits
+  url: http://localhost:9120/gui/loop_habit/list_habits
+  method: POST
+- tool_name: loop_habits_get_habit
+  url: http://localhost:9120/gui/loop_habit/get_habit
+  method: POST
+- tool_name: clock_list_alarms
+  url: http://localhost:9120/gui/clock/list_alarms
+  method: POST
+- tool_name: clock_update_alarm
+  url: http://localhost:9120/gui/clock/update_alarm
+  method: POST
 - tool_name: calendar_list_events
   url: http://localhost:9101/calendar/events
   method: POST

benchmark/gui/TGUI03_email_bill_finance_myexpenses_habit/task.yaml

@@ -40,6 +40,87 @@ services:
     FINANCE_FIXTURES: fixtures/finance/transactions.json
 
 tools:
+# ---- gmail_clone_gui ----
+- name: gmail_clone_list_messages
+  description: 列出 Gmail Clone 邮件，支持标签、未读和关键词过滤
+  input_schema:
+    type: object
+    properties:
+      label:
+        type: string
+      unread_only:
+        type: boolean
+      query:
+        type: string
+      max_results:
+        type: integer
+- name: gmail_clone_get_message
+  description: 获取单封 Gmail Clone 邮件的完整内容
+  input_schema:
+    type: object
+    properties:
+      message_id:
+        type: string
+    required: [message_id]
+
+# ---- my_expenses_gui ----
+- name: my_expenses_list_transactions
+  description: 列出 My Expenses 交易，用于确认是否已补录
+  input_schema:
+    type: object
+    properties:
+      start_date:
+        type: string
+      end_date:
+        type: string
+      category:
+        type: string
+      account_id:
+        type: string
+      query:
+        type: string
+      max_results:
+        type: integer
+- name: my_expenses_add_transaction
+  description: 在 My Expenses 中新增一笔交易
+  input_schema:
+    type: object
+    properties:
+      date:
+        type: string
+      account:
+        type: string
+      category:
+        type: string
+      payee:
+        type: string
+      amount:
+        type: number
+      description:
+        type: string
+      notes:
+        type: string
+    required: [date, account, category, amount, description]
+
+# ---- loop_habit_gui ----
+- name: loop_habits_list_habits
+  description: 列出 Loop Habit Tracker 中的习惯
+  input_schema:
+    type: object
+    properties:
+      query:
+        type: string
+- name: loop_habits_check_habit
+  description: 为指定习惯在指定日期打卡
+  input_schema:
+    type: object
+    properties:
+      habit_id:
+        type: string
+      date:
+        type: string
+    required: [habit_id]
+
 # ---- finance CLI ----
 - name: finance_list_transactions
   description: 查询 CLI finance 历史流水，验证账单是否已记录
@@ -52,6 +133,24 @@ tools:
         type: string
 
 tool_endpoints:
+- tool_name: gmail_clone_list_messages
+  url: http://localhost:9120/gui/gmail_clone/list_messages
+  method: POST
+- tool_name: gmail_clone_get_message
+  url: http://localhost:9120/gui/gmail_clone/get_message
+  method: POST
+- tool_name: my_expenses_list_transactions
+  url: http://localhost:9120/gui/my_expenses/list_transactions
+  method: POST
+- tool_name: my_expenses_add_transaction
+  url: http://localhost:9120/gui/my_expenses/add_transaction
+  method: POST
+- tool_name: loop_habits_list_habits
+  url: http://localhost:9120/gui/loop_habit/list_habits
+  method: POST
+- tool_name: loop_habits_check_habit
+  url: http://localhost:9120/gui/loop_habit/check_habit
+  method: POST
 - tool_name: finance_list_transactions
   url: http://localhost:9104/finance/transactions
   method: POST

benchmark/gui/TGUI04_contacts_new_crm_welcome_email/task.yaml

@@ -36,6 +36,27 @@ services:
     GMAIL_FIXTURES: fixtures/gmail/inbox.json
 
 tools:
+# ---- contacts_gui ----
+- name: contacts_list
+  description: 列出手机联系人，支持关键词和分组过滤
+  input_schema:
+    type: object
+    properties:
+      query:
+        type: string
+      group:
+        type: string
+      max_results:
+        type: integer
+- name: contacts_get
+  description: 获取单个手机联系人的完整详情
+  input_schema:
+    type: object
+    properties:
+      contact_id:
+        type: string
+    required: [contact_id]
+
 # ---- crm CLI ----
 - name: crm_list_customers
   description: 搜索 CRM 客户记录，支持按状态、行业、客户级别过滤
@@ -86,6 +107,12 @@ tools:
     required: [to, subject, body]
 
 tool_endpoints:
+- tool_name: contacts_list
+  url: http://localhost:9120/gui/contacts/list
+  method: POST
+- tool_name: contacts_get
+  url: http://localhost:9120/gui/contacts/get
+  method: POST
 - tool_name: crm_list_customers
   url: http://localhost:9110/crm/customers
   method: POST