[Contribution] Proposal from JD.com Competition Winners: Video Tools & Robust Search Workflows #113
Description
Dear Oxygent Maintainers,
尊敬的Oxygent维护人员：
I am writing to you on behalf of my team. We recently participated in the "Jingdong (JD.com) AI Agent Competition" which was based on the oxygent framework. We are proud to share that our project achieved 4th place nationally, winning the National Third Prize.
我谨代表我的团队致函。我们近期参加了基于 Oxygent 框架的 “京东多智能体大赛” 。其中我们的项目荣获全国第四名 ，并获得全国三等奖 。
During the competition, we developed several extensions and workflows to handle complex tasks. We believe these contributions could be valuable to the community and would like to propose merging them into the upstream project.
在比赛期间，我们开发了多个扩展程序和工作流程来处理复杂任务。我们相信这些贡献对社区很有价值，因此建议将它们合并到上游项目中。
Here are the two main modules we plan to contribute:
以下是我们希望贡献的两个主要模块：
- Video Analysis Tools & Workflow (Core of our winning entry)
- 视频分析工具及工作流程（我们获奖作品的核心）
Context: This was a key component of our competition submission.
背景： 这是我们参赛作品的关键组成部分。
Description: A set of tools specifically for video platforms (e.g., Bilibili), enabling search and download capabilities. We also implemented a video_analysis_workflow that orchestrates Search -> Download -> Local Video Understanding (VLM) in a seamless loop.
描述： 一套专为视频平台（例如 Bilibili）设计的工具，支持搜索和下载功能。我们还实现了一个视频分析工作流 ，将搜索、下载和本地视频理解 (VLM) 无缝衔接。此外还包含多个与视频理解等多模态任务相关的tool。
Value: It demonstrates oxygent's multimodal capabilities in real-world scenarios, proven effective in the competition.
价值： 它展示了 Oxygent 在真实场景中的多模式能力，并在比赛中证明了其有效性。
- Robust Web-Search Workflow with Rollback (Proposed Enhancement)
- 具有回滚功能的稳健网络搜索工作流程（建议的改进）
Context: Based on our experience in the competition, we identified the need for better error recovery in complex tasks.
背景： 根据我们在比赛中的经验，我们发现需要在复杂任务中更好地完成搜索任务。
Description: We propose a specialized workflow for "Deep Search" tasks. Unlike traditional linear search chains which often fail when hitting dead ends, this workflow implements a state-aware stack mechanism. It mimics a Depth-First Search (DFS) algorithm allowing the agent to backtrack.
描述： 我们提出了一种专门用于“深度搜索”任务的工作流程。与传统的线性搜索链经常在遇到死胡同时失效不同，该工作流程实现了一种状态感知栈机制 。它模拟了深度优先搜索（DFS） 算法，允许智能体回溯。
Functionality: The system maintains a history stack. If the Agent encounters a dead end (404, irrelevant content, or paywall), it automatically triggers a "Rollback" to pop the current state and resume exploration from the previous valid branch.
功能： 系统维护一个历史记录栈。如果代理遇到死胡同（404 错误、无关内容或付费墙），它会自动触发“回滚”，清除当前状态并从之前的有效分支继续探索。
Value: This architecture drastically enhances fault tolerance and robustness, transforming fragile script-like execution into intelligent, adaptive exploration.
价值： 这种架构极大地增强了容错性和鲁棒性 ，将脆弱的脚本式执行转变为智能的、自适应的探索。
My Questions: 我的问题：
Are you interested in integrating features from our competition-winning codebase?
您有兴趣将我们屡获殊荣的代码库中的功能集成到您的系统中吗？
For the "Video Workflow," would you prefer it as a core feature or within the examples/ directory?
对于“视频工作流程”，您希望将其作为核心功能还是放在 examples/ 目录中？
Do you have any specific guidelines for implementing the rollback mechanism?
对于实施回滚机制，你们有什么具体的指导原则吗？
We are eager to contribute back to the framework that helped us succeed. Looking forward to your feedback!
我们渴望回馈帮助我们取得成功的框架。期待您的反馈！
祝好
anon_tokyo