Assets Library
面向内部团队的多模态素材中枢，统一处理图片与视频的上传、分镜、视觉分析、私有存储和语义检索。A multimodal asset hub unifying ingestion, scene detection, visual analysis, private storage, and semantic search.
基于上游项目深度改造Extensively adapted from upstream
AI 基础设施与开发者工具AI infrastructure & developer tools
让 AI 可控，也可理解。Build AI you can trust.
连接多模态内容、Agent 执行、故障诊断与开放网络。Connecting multimodal content, agent execution, failure diagnostics, and the open web.
重点项目Featured systems
从内容进入系统，到工具执行、故障取证和开放网络访问，每个项目都解决一段真实工作流。From content ingestion to execution, failure evidence, and web access, each project solves a real part of the workflow.
DSH Failure Capsule
DeepSeek Harness 的本地优先故障证据插件，将失败上下文整理为经过脱敏、可离线检查的证据包。A local-first DeepSeek Harness plugin that packages failures into redacted, offline-inspectable evidence capsules.
Song Agent
安全、可插拔的 Intent 执行框架，让 API、Agent 工具、操作确认和调度任务共享同一个执行内核。A secure intent-execution framework sharing one kernel across APIs, agent tools, confirmations, and scheduled jobs.
Search Engine Tool MCP
面向 AI 客户端的 Python MCP 搜索服务，提供零配置网页搜索、安全内容提取和多提供商接入。A Python MCP server for zero-config web search, safety-bounded extraction, and optional multi-provider access.
一个系统，而不是一组演示One system, not a set of demos
当前工作围绕四个相互连接的工程问题展开。The current work connects four engineering problems that belong together.
内容进入Content in
处理图片、视频、网页与附件，同时保留来源、边界和作用域。Ingest images, video, web pages, and attachments while preserving provenance and scope.
意图执行Intent execution
已知意图直接进入统一内核，高风险操作有确认、幂等和恢复路径。Known intents enter one kernel with confirmation, idempotency, and recovery for risky actions.
失败可见Failures visible
把运行环境、时间线和错误身份收进可复查的本地证据包。Capture runtime, timeline, and failure identity in a reviewable local evidence package.
结果可追溯Results traceable
搜索、提取与模型调用都需要明确来源和可验证的安全边界。Search, extraction, and model calls keep explicit sources and verifiable safety boundaries.
查看代码、边界与工程选择。See the code, constraints, and engineering decisions.
公开仓库保留实现细节、验证方式和当前限制。Public repositories document implementation details, verification, and current limitations.