AI-native programmer
- 57+ merged upstream PRs, including Openclaw, PyTorch, vLLM, Mem0, LightRAG, AstrBot, Qdrant, Numpy, AgentScope, Qwen Code, DeepTutor, QwenPaw, Google ADK, Pydantic AI, Ktransformers, Microsoft Agent Framework, LiveKit Agents, Mooncake, Pymilvus, HiveMind.
- 11 public projects, led by LLM-Wiki, KnowFoundry-RAG-Console, OpenSense, vectorpeak-blogs, carbon-tower-predictor.
| Area | Project | Stars | Notes |
|---|---|---|---|
| Applied agents | KnowFoundry-RAG-Console | 29+ | A console project for knowledge-base and RAG workflows, used to capture engineering practice for retrieval-augmented generation apps |
| Applied agents | vectorpeak-blogs | new+ | Personal blog and knowledge site for technical learning, deep learning practice, algorithm notes, and staged reflections |
| Applied agents | Codex-Pet-Live | new+ | A Windows desktop Codex pet runtime for hosting, displaying, and driving interactive desktop companions |
| Applied agents | Fill-Contributions-Graph | new+ | A utility for backfilling and maintaining GitHub contribution graph activity |
| Applied agents | lingban-ai-chat | new+ | Lingban AI chat Android client and backend API repository |
| Coding agents | OpenSense | new+ | Daily PR opportunity finder for known open-source repositories |
| Coding agents | vectorpeak-agent-skills | new+ | A reusable Agent and Codex skill library for software engineering, knowledge work, and professional workflow automation |
| LLM tooling | LLM-Wiki | 30+ | A structured knowledge base for large language model technologies, covering LLMs, Agents, RAG, model training, evaluation methods, and AI engineering practices |
| Research | carbon-tower-predictor | new+ | An industrial carbonization tower outlet temperature prediction project using historical DCS sensor data for trend forecasting, anomaly warning, and predictive maintenance |
| Research | industrial-fault-text-classifier | new+ | BERT-based industrial equipment fault text classification for repair-ticket analysis, fault recognition, and risk warning |
| Research | kaggle-tabular-forge | new+ | A goal-driven Kaggle tabular experiment factory covering feature mining, OOF discipline, deep stacking, and AI review workflows |
| Project | PR | What I Fixed |
|---|---|---|
| #188961 | Fix Windows target determination for POSIX changed-file paths |
|
| #188830 | Fixed cross-platform CI test target detection by normalizing Git-style and Windows-style test paths |
|
| #47099 | [Bugfix] Align OpenCV video metadata timeline |
|
| #31906 | BUG: support PathLike paths in DataSource methods |
|
| #2071 | fix(cli): detect bound ports before launch |
|
| #2770 | [Bugfix] Reject empty keys in HTTP metadata server |
|
| #2587 | Preserved empty values in the Mooncake Store REST GET API so clients can reliably read cache metadata and blank fields |
| Project | PR | What I Fixed |
|---|---|---|
| #101815 | Split encoded video URLs from captions |
|
| #99180 | Reject tools/call requests with non-object arguments |
|
| #97972 | CDP auth fails with percent-encoded credentials |
|
| #97630 | Normalize Windows inbound paths case-insensitively |
|
| #9175 | Validate dashboard account username updates |
|
| #9141 | Enforce ownership when reading ChatUI sessions |
|
| #9085 | Reject non-200 download responses |
|
| #9084 | Enable only synced ModelScope MCP servers |
|
| #8971 | Fixed a path traversal vulnerability in knowledge base upload filenames by validating filename boundaries and blocking directory escape |
|
| #8968 | Fixed a path traversal vulnerability in plugin upload filenames by validating filename boundaries and blocking directory escape |
|
| #6483 | Reject Windows-style workspace artifact paths |
|
| #6455 | Reject fractional LSP limit inputs |
|
| #6413 | Align monitor limit parameter schemas |
|
| #6381 | Require integer ReadFile pagination params |
|
| #6344 | Preserve glued automation history records |
|
| #6323 | Avoid null OpenAPI schema types |
|
| #6317 | Preserve partial remote input JSONL records |
|
| #6243 | Preserve descriptionless tools |
|
| #6033 | Parse tagged thinking for GLM responses |
|
| #6029 | Support Windows-style tilde paths |
|
| #5914 | Harden remaining source path validation |
|
| #5911 | Normalize source slug validation errors |
|
| #5829 | Reject unsafe source slugs before deletion |
|
| #5783 | Reject userinfo URLs in WebFetch validation |
|
| #602 | Preserve zero LlamaIndex chunk overlap |
|
| #5766 | Recover JSON objects after leading whitespace |
|
| #5750 | Route plugin market details through link guard |
|
| #6328 | Preserve image format for Bedrock images |
|
| #6306 | Preserve provider status errors |
|
| #6305 | Avoid mutating raw tool schemas for Responses |
| Project | PR | What I Fixed |
|---|---|---|
| #5963 | Encode dynamic URL path segments |
|
| #1954 | Reject negative grep pagination values |
|
| #6314 | Avoid mutating event validation input |
|
| #6327 | Preserve HTTPX event hooks |
|
| #6282 | Give a clear error for non-positive concurrency max_running |
|
| #6267 | Reject non-positive OnlineEvaluator.max_concurrency |
|
| #6251 | Write relative schema refs for AgentSpec files |
|
| #6228 | Reject non-positive max_concurrency |
|
| #6143 | Fix absolute schema_path handling in Dataset.to_file |
|
| #6142 | Avoid duplicating /v1 in base_url |
|
| #6140 | Clarify end_strategy values in agent spec |
|
| #6903 | Python: normalize single Anthropic tools |
|
| #6818 | Python: accept A2A data URI media parameters |
|
| #28 | Improved event connection handling and reconnection logic, strengthening disconnect recovery and long-running connection stability |
| Project | PR | What I Fixed |
|---|---|---|
| #3324 | Prevent task ID URL path injection |
|
| #3316 | Fixed a path traversal vulnerability in resource file generation by validating directory boundaries and preventing unauthorized writes or overwrites |
|
| #9723 | Parse Windows snapshot mappings |
|
| #9677 | Preserve empty result shape for zero-limit batches |
|
| #3673 | Allow null auto_id primary column in DataFrame insert |
|
| #3658 | Preserve remote upload prefixes on Windows |
AI-native programmer
- 57+ merged upstream PRs，包括 Openclaw, PyTorch, vLLM, Mem0, LightRAG, AstrBot, Qdrant, Numpy, AgentScope, Qwen Code, DeepTutor, QwenPaw, Google ADK, Pydantic AI, Ktransformers, Microsoft Agent Framework, LiveKit Agents, Mooncake, Pymilvus, HiveMind。
- 11 个公开项目，代表项目包括 LLM-Wiki, KnowFoundry-RAG-Console, OpenSense, vectorpeak-blogs, carbon-tower-predictor。
| Area | Project | Stars | Notes |
|---|---|---|---|
| Applied agents | KnowFoundry-RAG-Console | 29+ | 面向知识库和 RAG 工作流的控制台项目，用于沉淀检索增强生成应用的工程实践 |
| Applied agents | vectorpeak-blogs | new+ | 个人博客与知识站点，用于记录技术学习、深度学习实践、算法题解与阶段性思考 |
| Applied agents | Codex-Pet-Live | new+ | 面向 Windows 桌面的 Codex 宠物运行时，用于承载、展示和驱动交互式桌面宠物 |
| Applied agents | Fill-Contributions-Graph | new+ | GitHub Contribution Graph 荒漠绿化治理工程 |
| Applied agents | lingban-ai-chat | new+ | 灵伴 AI 聊天 Android 客户端与后端 API 仓库 |
| Coding agents | OpenSense | new+ | 面向已知开源项目的每日 PR 机会发现工具 |
| Coding agents | vectorpeak-agent-skills | new+ | 可复用的 Agent / Codex 技能库，覆盖软件工程、知识工作和职业工作流自动化 |
| LLM tooling | LLM-Wiki | 30+ | 大语言模型知识库，覆盖 LLM、Agent、RAG、训练、评测方法和 AI 工程实践 |
| Research | carbon-tower-predictor | new+ | 工业碳化塔取出液温度预测项目，用历史 DCS 传感器数据进行趋势预测、异常预警与预测性维护 |
| Research | industrial-fault-text-classifier | new+ | 基于 BERT 的化工设备故障文本分类方案，用于报修工单分析、故障类别识别和风险预警 |
| Research | kaggle-tabular-forge | new+ | 面向 Kaggle 表格竞赛的实验工厂，覆盖特征挖掘、OOF 纪律、深度 stacking 和 AI 评审工作流 |
| Project | PR | What I Fixed |
|---|---|---|
| #188961 | Fix Windows target determination for POSIX changed-file paths |
|
| #188830 | Fixed cross-platform CI test target detection by normalizing Git-style and Windows-style test paths |
|
| #47099 | 对齐 OpenCV 视频元数据时间线。 |
|
| #31906 | 支持 DataSource 方法中的 PathLike 路径。 |
|
| #2071 | fix(cli): detect bound ports before launch |
|
| #2770 | 在 HTTP metadata server 中拒绝空 key。 |
|
| #2587 | 保留 Mooncake Store REST GET API 中的空值，使客户端能够可靠读取缓存元数据和空字段。 |
| Project | PR | What I Fixed |
|---|---|---|
| #101815 | 将编码后的视频 URL 与 captions 分离。 |
|
| #99180 | 拒绝参数不是对象的 tools/call 请求。 |
|
| #97972 | CDP 认证会因百分号编码的凭据而失败。 |
|
| #97630 | 以不区分大小写的方式规范化 Windows 入站路径。 |
|
| #9175 | 校验仪表盘账号用户名更新。 |
|
| #9141 | 读取 ChatUI 会话时强制校验所有权。 |
|
| #9085 | 拒绝非 200 的下载响应。 |
|
| #9084 | 仅启用已同步的 ModelScope MCP 服务。 |
|
| #8971 | 通过验证文件名边界并阻止目录逃逸，修复知识库上传文件名中的路径穿越漏洞。 |
|
| #8968 | 通过验证文件名边界并阻止目录逃逸，修复插件上传文件名中的路径穿越漏洞。 |
|
| #6483 | 拒绝 Windows 风格的 workspace artifact 路径。 |
|
| #6455 | 拒绝小数形式的 LSP limit 输入。 |
|
| #6413 | 对齐 monitor limit 参数 schema。 |
|
| #6381 | 要求 ReadFile 分页参数为整数。 |
|
| #6344 | 保留粘连的 automation 历史记录。 |
|
| #6323 | 避免空的 OpenAPI schema 类型。 |
|
| #6317 | 保留不完整的远程输入 JSONL 记录。 |
|
| #6243 | 保留没有 description 的工具。 |
|
| #6033 | 解析 GLM 响应中的标记化 thinking。 |
|
| #6029 | 支持 Windows 风格的波浪号路径。 |
|
| #5914 | 强化其余 source path 校验。 |
|
| #5911 | 统一 source slug 校验错误。 |
|
| #5829 | 删除前拒绝不安全的 source slug。 |
|
| #5783 | 在 WebFetch 校验中拒绝 userinfo URL。 |
|
| #602 | 保留为零的 LlamaIndex chunk overlap。 |
|
| #5766 | 在前导空白后恢复 JSON 对象。 |
|
| #5750 | 让插件市场详情经过 link guard 路由。 |
|
| #6328 | 保留 Bedrock images 的图像格式。 |
|
| #6306 | 保留 provider status errors。 |
|
| #6305 | 避免为 Responses 修改原始 tool schemas。 |
| Project | PR | What I Fixed |
|---|---|---|
| #5963 | 对动态 URL 路径片段进行编码。 |
|
| #1954 | 拒绝负数 grep 分页值。 |
|
| #6314 | Avoid mutating event validation input |
|
| #6327 | 保留 HTTPX event hooks。 |
|
| #6282 | 为非正数 concurrency max_running 给出清晰错误。 |
|
| #6267 | 拒绝非正数的 OnlineEvaluator.max_concurrency。 |
|
| #6251 | 为 AgentSpec 文件写入相对 schema refs。 |
|
| #6228 | 拒绝非正数的 max_concurrency。 |
|
| #6143 | 修复 Dataset.to_file 中绝对 schema_path 的处理。 |
|
| #6142 | 避免在 base_url 中重复拼接 /v1。 |
|
| #6140 | 明确 agent spec 中的 end_strategy 取值。 |
|
| #6903 | Python: normalize single Anthropic tools |
|
| #6818 | Python：接受 A2A data URI 媒体参数。 |
|
| #28 | 改进事件连接处理和重连逻辑，增强断线恢复和长连接稳定性。 |
| Project | PR | What I Fixed |
|---|---|---|
| #3324 | 阻止任务 ID URL 路径注入。 |
|
| #3316 | 通过验证目录边界并防止未授权写入或覆盖，修复资源文件生成中的路径穿越漏洞。 |
|
| #9723 | 解析 Windows snapshot mappings。 |
|
| #9677 | 保留 zero-limit batches 的空结果形状。 |
|
| #3673 | Allow null auto_id primary column in DataFrame insert |
|
| #3658 | 在 Windows 上保留远程上传 prefixes。 |