Files
beaver_project/skills/officebench-mcp/versions/v0001/SKILL.md
steven_li 33a9845566 ```
feat(engine): 添加技能查看工具并优化异步任务管理

- 添加SkillViewTool到引擎加载器中,增强技能管理功能
- 在AgentLoop中引入_active_direct_task来跟踪活跃任务
- 实现直接任务执行时的同步处理逻辑
- 更新工具实例化方式以支持依赖注入

feat(config): 增加智能体运行时参数配置支持

- 扩展AgentDefaultsConfig添加max_tokens和temperature字段
- 实现配置解析函数_first_config_value处理多个配置源
- 支持通过Web API动态更新智能体运行时参数
- 添加前端页面配置表单和验证逻辑

refactor(provider): 统一最大令牌数参数类型为可选整型

- 将所有LLM提供者的max_tokens参数改为int | None类型
- 为AnthropicProvider实现模型特定的最大令牌数默认值
- 调整参数传递逻辑,优先级:调用参数 > 配置文件 > 模型默认值
- 移除硬编码的默认值,改用条件判断

feat(process): 增强事件投影功能

- 添加工具调用开始/结束事件的映射逻辑
- 实现技能激活事件的识别和展示
- 添加辅助函数处理工具调用名称和参数提取
- 优化运行记录关联逻辑,提升事件匹配准确性

fix(web): 更新网络请求客户端信任环境设置

- 将WebFetchTool和WebSearchTool的trust_env参数设为True
- 确保HTTP客户端能够正确使用系统代理配置
- 修复可能的网络连接问题

test: 添加配置加载和事件投影相关测试

- 新增智能体默认参数配置测试用例
- 实现API配置持久化和重载测试
- 添加技能卡片和工具事件的投影测试
```
2026-05-27 13:37:06 +08:00

7.7 KiB

name, description, always, tools
name description always tools
officebench-mcp Guidance for OfficeBench evaluation tasks. Use the registered mcp_officebench_* tools to inspect and edit OfficeBench files, spreadsheets, documents, emails, calendars, PDFs, and answer files. true
mcp_officebench_excel_read_file
mcp_officebench_excel_set_cell
mcp_officebench_excel_delete_cell
mcp_officebench_excel_create_new_file
mcp_officebench_excel_convert_to_pdf
mcp_officebench_word_read_file
mcp_officebench_word_write_to_file
mcp_officebench_word_create_new_file
mcp_officebench_word_convert_to_pdf
mcp_officebench_email_list_emails
mcp_officebench_email_read_email
mcp_officebench_email_send_email
mcp_officebench_calendar_create_event
mcp_officebench_calendar_list_events
mcp_officebench_calendar_delete_event
mcp_officebench_pdf_read_file
mcp_officebench_pdf_convert_to_word
mcp_officebench_pdf_convert_to_image
mcp_officebench_ocr_recognize_file
mcp_officebench_shell_command
mcp_officebench_shell_list_directory
mcp_officebench_shell_read_file
mcp_officebench_shell_write_file
mcp_officebench_shell_copy_file
mcp_officebench_system_finish_task
mcp_officebench_system_get_status
mcp_officebench_image_convert_to_pdf

OfficeBench MCP Skill

Use this skill for OfficeBench evaluation runs. OfficeBench task files live in the OfficeBench MCP server, not in Beaver's local filesystem. Complete the task by calling real mcp_officebench_* tools.

Critical Rules

  1. Use actual Beaver tool calls only. Do not print XML, DSML, JSON, or markdown that describes a tool call.
  2. Never invent tool names. If you need to find files, use mcp_officebench_shell_list_directory or mcp_officebench_shell_command.
  3. Do not use Beaver local filesystem, local runtime, local terminal, or local code tools for OfficeBench files.
  4. Paths are relative to /testbed in the OfficeBench MCP container, such as data/score.xlsx.
  5. If the task context gives a workspace_id, pass that same workspace_id argument in every OfficeBench MCP tool call that supports it.
  6. Inspect files before editing them.
  7. Verify the requested output file or edited cell exists before finishing.
  8. Finish every task with mcp_officebench_system_finish_task.

Tool Names And Use

Excel

Use these for .xlsx files:

  • mcp_officebench_excel_read_file: read workbook sheets and cell values.
    • Required: file_path
    • Optional: sheet_name, workspace_id
  • mcp_officebench_excel_set_cell: write one cell.
    • Required: file_path, row, col, value
    • Optional: sheet_name, workspace_id
    • Rows and columns are 1-based.
  • mcp_officebench_excel_delete_cell: clear one cell.
    • Required: file_path, row, col
    • Optional: sheet_name, workspace_id
  • mcp_officebench_excel_create_new_file: create a workbook.
    • Required: file_path
    • Optional: workspace_id
  • mcp_officebench_excel_convert_to_pdf: convert an Excel file to PDF.
    • Required: file_path
    • Optional: workspace_id

Typical Excel sequence:

  1. Call mcp_officebench_shell_list_directory on data.
  2. Call mcp_officebench_excel_read_file on the target workbook.
  3. Identify the exact row and column.
  4. Call mcp_officebench_excel_set_cell.
  5. Read the workbook again or use status/listing to verify.
  6. Call mcp_officebench_system_finish_task.

For the common task "change Bob's midterm1 score to 100 in score.xlsx", inspect data/score.xlsx, find Bob's row and the midterm1 column, then call mcp_officebench_excel_set_cell with that row, that column, and value 100.

Word

Use these for .docx files:

  • mcp_officebench_word_read_file: read all paragraphs.
    • Required: file_path
    • Optional: workspace_id
  • mcp_officebench_word_write_to_file: overwrite or append text.
    • Required: file_path, text
    • Optional: append, workspace_id
  • mcp_officebench_word_create_new_file: create a new Word document.
    • Required: file_path
    • Optional: workspace_id
  • mcp_officebench_word_convert_to_pdf: convert Word to PDF.
    • Required: file_path
    • Optional: workspace_id

Preserve exact spelling, capitalization, punctuation, and line order from source files.

Email

Use these for email tasks:

  • mcp_officebench_email_list_emails: list available .eml messages.
    • Optional: folder, workspace_id
  • mcp_officebench_email_read_email: read one email.
    • Required: email_path
    • Optional: workspace_id
  • mcp_officebench_email_send_email: create/send an email artifact.
    • Required: to, subject, body
    • Optional: attachments, workspace_id

For email-search tasks, final answers should use plain text with literal lines like Subject: .... Do not add markdown labels.

Calendar

Use these for calendar .ics tasks:

  • mcp_officebench_calendar_list_events: inspect calendar events.
    • Optional: calendar_path, workspace_id
  • mcp_officebench_calendar_create_event: create an event.
    • Required fields depend on the task; include summary/title, start, end, and target calendar when needed.
    • Optional: workspace_id
  • mcp_officebench_calendar_delete_event: delete an event.
    • Required fields depend on the task; inspect events first.
    • Optional: workspace_id

Use the task's current date/time context when interpreting relative dates.

PDF, OCR, And Images

Use these for PDF/image tasks:

  • mcp_officebench_pdf_read_file: extract text from a PDF.
    • Required: pdf_file_path
    • Optional: workspace_id
  • mcp_officebench_pdf_convert_to_word: convert PDF to Word.
    • Required: pdf_file_path
    • Optional: workspace_id
  • mcp_officebench_pdf_convert_to_image: convert one PDF page to an image.
    • Required: pdf_file_path
    • Optional: page_number, dpi, workspace_id
  • mcp_officebench_ocr_recognize_file: OCR an image.
    • Required: image_path
    • Optional: language, workspace_id
  • mcp_officebench_image_convert_to_pdf: convert image to PDF.
    • Required: image_path
    • Optional: output_path, workspace_id

For conversion tasks, create the exact requested filename and verify it exists.

Shell And System

Use these for safe file discovery and text files:

  • mcp_officebench_shell_list_directory: list a directory.
    • Optional: path, workspace_id
  • mcp_officebench_shell_read_file: read text files such as .txt, .csv, .json, .md, .xml.
    • Required: file_path
    • Optional: workspace_id
  • mcp_officebench_shell_write_file: write text files.
    • Required: file_path, content
    • Optional: append, workspace_id
  • mcp_officebench_shell_copy_file: copy a file or directory.
    • Required: source, destination
    • Optional: workspace_id
  • mcp_officebench_shell_command: run shell commands inside the OfficeBench MCP container.
    • Required: command
    • Optional: workdir, workspace_id
  • mcp_officebench_system_get_status: inspect filesystem/git status.
    • Optional: workspace_id
  • mcp_officebench_system_finish_task: mark the task complete and optionally write an answer.
    • Optional: answer, workspace_id

Prefer dedicated Office tools for Office documents. Use shell tools for listing directories, copying/renaming files, and reading/writing plain text.

Anti-Patterns

Do not do any of the following:

  • Do not call mcp_officebench_find_in_workspace; that tool does not exist.
  • Do not output <tool_calls>, <invoke>, DSML, or pseudo tool call text.
  • Do not answer "done" without calling the required OfficeBench tools.
  • Do not edit guessed paths without first listing or reading relevant files.
  • Do not use /testbed as a literal prefix in path arguments unless a tool explicitly asks for an absolute path.
  • Do not correct misspellings found in source data. Preserve source text exactly.