Loading ToolBench
A Go-based backend server for an AI agent platform with desktop automation, web search, arXiv paper retrieval, video generation, and browser preview capabilities
Yosuri provides 12 tools with mostly complete schemas and descriptions in Chinese. Naming follows verb_noun patterns (arxiv_search, web_search, computer_*). Descriptions are detailed (100-300 chars) and explain WHEN to use each tool. However, several critical gaps: (1) computer_* tools lack explicit output schemas, responses are inferred; (2) no error handling guidance in any tool description; (3) mambo_video has 7 optional parameters with minimal constraints; (4) no tool annotations (readOnlyHint/destructiveHint) despite clear risk levels; (5) parameter descriptions sometimes lack format/range constraints (e.g., 'x' and 'y' coordinates have no bounds). Computer interaction tools are well-named but lack defensive validation hints.
搜索 arXiv 论文并返回结构化元数据（标题/作者/日期/分类/摘要/PDF 链接），可用于论文检索、文献调研、追踪最新研究。按 ID 抓取、按关键词检索、按分类过滤都可以。
按下指定按键或组合键。可用于回车、Tab、方向键、Ctrl+C 等。
在当前位置或指定坐标执行鼠标点击。默认左键单击。
从当前位置或指定起点拖拽鼠标到目标坐标。
将鼠标移动到指定屏幕坐标 (x, y)。坐标从屏幕左上角 (0,0) 开始。
在系统浏览器（Edge）中打开指定网址，用于开始网页任务（点外卖/抢票/填表/搜索等）。打开后需 computer_screenshot 查看页面再操作。
获取当前屏幕分辨率和主显示器信息。返回宽、高和可用显示器数量。
Computer interaction tools (computer_mouse_move, computer_mouse_click, computer_scroll, computer_type) lack output schema documentation. Responses are inferred but not formally specified. LLMs cannot plan downstream actions without knowing return structure.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Tools marked 'WRITE' (computer_mouse_click, computer_type, computer_open_url) should declare destructiveHint=true. READ_ONLY tools should declare readOnlyHint=true. This prevents accidental misuse by agents.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
截取当前前台窗口（默认）或整屏，落盘成图片文件并把**路径**回给你。这是你看清画面的唯一方式：拿到路径后必须用 view_image(path=该路径) 才能真正读到画面内容。只在「不知道下一步该点哪里」时才截图；同一画面不要重复截图确认，那只会刷用户桌面。截图属内部感知步骤，不会也不需要作为交付物展示给用户。
在当前位置滚动鼠标滚轮。
在当前焦点窗口或指定位置输入文本。支持中文等 Unicode 字符。
曼波视频一键生成：输入主题（可选分段文案、素材池目录），自动完成 配音（edge-tts 中文快节奏曼波音色）→ 精确字幕（SRT，与语音对齐）→ 按每句关键词匹配素材（本地素材池，无命中用动态渐变背景+大字兜底）→ ffmpeg 竖屏 1080x1920 合成。返回视频/字幕/manifest 的绝对路径 JSON。适合做口播短视频、梗视频、卡点视频素材。
联网搜索：用关键词搜索互联网，返回带标题/链接/摘要的结果列表。引擎自动选择用户配置的搜索来源（默认免 key 的 Bing，或 Firecrawl/自定义模型/MCP 工具）。当任务需要最新信息、实时数据、或你知识范围之外的网络内容时调用，也可用于核实旧知识的时效性。搜索词用中文或英文均可。
Coordinate parameters (x, y, start_x, start_y, end_x, end_y) lack bounds or format constraints. Descriptions do not specify valid ranges (e.g., 0 to screen_width). LLMs may pass invalid coordinates without validation guidance.
mambo_video has 7 optional parameters with minimal constraints. 'rate' accepts strings like '+30%' but no enum or pattern validation. 'voice' has no enum constraint despite being a fixed set (zh-CN-XiaoxiaoNeural, zh-CN-XiaoyiNeural, zh-CN-YunxiNeural). LLMs may hallucinate invalid values.
No error handling guidance in any tool description. Tools do not explain what to do if a call fails (e.g., 'If screenshot fails, check if display is locked' or 'If web_search returns no results, try a simpler query'). Agents have no recovery path.