[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-93215":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":14,"stars7d":14,"stars30d":15,"stars90d":14,"forks30d":14,"starsTrendScore":14,"compositeScore":16,"rankGlobal":9,"rankLanguage":9,"license":17,"archived":18,"fork":18,"defaultBranch":19,"hasWiki":20,"hasPages":18,"topics":21,"createdAt":9,"pushedAt":9,"updatedAt":22,"readmeContent":23,"aiSummary":24,"trendingCount":14,"starSnapshotCount":14,"syncStatus":25,"lastSyncTime":26,"discoverSource":27},93215,"DevAgent-Studio","biheto\u002FDevAgent-Studio","biheto","面向开发者的多智能体任务工作站",null,"Python",136,14,113,0,12,44.73,"MIT License",false,"main",true,[],"2026-07-22 04:02:08","# DevAgent Studio\n\n[English](#english) | [中文](#中文)\n\n\u003Ca id=\"english\">\u003C\u002Fa>\n\nDevAgent Studio is an open-source multi-agent workbench for **software project understanding and engineering governance**. It is built with FastAPI, LangGraph, and React to help teams inspect project structure, review code risks, preserve project knowledge, govern AI and tool calls, and run traceable workflows.\n\nIt is deliberately not a code-writing IDE. Its focus is making software delivery work easier to understand, audit, evaluate, and improve.\n\n![DevAgent Studio architecture](docs\u002Fassets\u002Farchitecture.png)\n\n## Why DevAgent Studio\n\n- **Multi-agent workflows, not isolated prompts**: Planner, Project Analyzer, Code Reviewer, RAG Processor, Supervisor, and Reporter are composed through LangGraph.\n- **Governed runtime**: Harness Runtime provides task context, event timelines, artifacts, review state, persistence, deterministic policy checks, and resume support.\n- **Visual workflows that execute**: a drag-and-drop canvas is compiled into executable LangGraph workflows instead of serving as a static diagram.\n- **Observable LLM operations**: prompt versions, model configuration, traces, token and cost data, fallback records, and A\u002FB comparison are available from the UI.\n- **Extensible Skills with guardrails**: Skills are versioned, permission-scoped, testable, dependency-aware, and usable from both the console and a workflow.\n- **Plugin Marketplace**: install resource packs from built-in catalogs, local paths, URLs, GitHub-style sources, or an external `SKILL.md` file.\n- **Safe third-party code execution**: Code Skills can run in a constrained Docker sandbox with no network, read-only mounts, resource limits, and audit logs.\n\n## Product Preview \u002F 页面预览\n\n| Run workbench \u002F 运行 | Visual workflow \u002F 编排 |\n| --- | --- |\n| ![Run workbench](docs\u002Fassets\u002Frun-preview.png) | ![Visual workflow](docs\u002Fassets\u002Fworkflow-preview.png) |\n\n| Reports \u002F 报告 | Interactive chat \u002F 追问 |\n| --- | --- |\n| ![Reports](docs\u002Fassets\u002Freport-preview.png) | ![Interactive chat](docs\u002Fassets\u002Fchat-preview.png) |\n\n| Task history \u002F 历史 | LLM governance \u002F LLM |\n| --- | --- |\n| ![Task history](docs\u002Fassets\u002Fhistory-preview.png) | ![LLM console](docs\u002Fassets\u002Fllm-console-preview.png) |\n\n| MCP console \u002F MCP | Skills console \u002F Skills |\n| --- | --- |\n| ![MCP console](docs\u002Fassets\u002Fmcp-console-preview.png) | ![Skills console](docs\u002Fassets\u002Fskill-preview.png) |\n\n| Plugin Marketplace \u002F Market | Benchmark dashboard \u002F Bench |\n| --- | --- |\n| ![Plugin Marketplace](docs\u002Fassets\u002Fmarket-preview.png) | ![Benchmark dashboard](docs\u002Fassets\u002Fbenchmark-dashboard-preview.png) |\n\n## Architecture\n\n```text\nUser \u002F React workbench\n        |\n        v\nFastAPI APIs ---- Marketplace ---- Skills Console ---- MCP Console\n        |\n        v\nHarness Runtime\n  context | policy | events | artifacts | human review | persistence\n        |\n        +-----------------------+\n        |                       |\n        v                       v\nLangGraph workflows          Skill Runtime\nplanner \u002F reviewer \u002F         prompt skills \u002F code skills \u002F\nRAG \u002F supervisor \u002F reporter  dependency and permission checks\n        |                       |\n        +-----------+-----------+\n                    v\n      LLM \u002F MCP \u002F RAG \u002F SQLite \u002F pgvector \u002F Docker sandbox\n```\n\n### Typical execution flow\n\n```text\nTask request\n  -> FastAPI creates a task\n  -> Harness Runtime creates context and emits events\n  -> LangGraph invokes agents, Skills, LLMs, RAG, or MCP tools\n  -> policy and human-review checks gate sensitive operations\n  -> traces, artifacts, logs, and state are persisted\n  -> Reporter produces a governance report\n```\n\n## Core Capabilities\n\n| Area | What it provides |\n| --- | --- |\n| Project analysis | Structure scanning, technology identification, module summary, risks, and governance suggestions. |\n| Code review | Hybrid rule, call-chain, and LLM semantic review with findings and test recommendations. |\n| RAG knowledge | Document chunking, ingestion, retrieval, project notes, SQLite default storage, and optional pgvector retrieval. |\n| Learning coach | Project-oriented learning plans and interactive follow-up questions. |\n| Collaboration | Planner, analyzer, reviewer, RAG, supervisor, and reporter run as a traceable collaboration graph. |\n| Workflow | Drag, connect, configure, validate, save, and execute workflow JSON compiled to LangGraph. |\n| Human review | Node-level approval\u002Frejection, checkpoint\u002Fresume, retry, and recovery visualization. |\n| LLM governance | Per-agent model configuration, call trace, prompt versions, token\u002Fcost data, fallback display, and A\u002FB tests. |\n| MCP management | Server registration, stdio tool discovery, enable\u002Fdisable, approval, test invocation, and call logs. |\n| Benchmark | LLM, RAG, Workflow, MCP, and multi-agent evaluation with success rate, P95 latency, recall, completeness, token, and cost metrics. |\n\n## Governed Skill Plugin System\n\nA Skill is a reusable capability such as code review, RAG processing, learning coaching, security scanning, or workflow execution. A Skill can be tested in the Skills console or added to a visual workflow.\n\n### Skill governance\n\n| Capability | Purpose |\n| --- | --- |\n| Contract validation | Validates `input_schema`, `output_schema`, `permissions`, and `execution_type` during package preview and install. |\n| Permission levels | Classifies access as `safe`, `project-read`, `llm`, `workflow-write`, `network`, or `filesystem`, and calculates risk. |\n| Strict approvals | Approval is scoped by `skill_code + agent_code`; testing and workflow execution are independently approved. |\n| Version management | Keeps Skill snapshots for upgrade comparison and rollback. |\n| Dependencies | Declares MCP tools, RAG collections, prompt versions, and model requirements before execution. |\n| Built-in tests | Allows packages to provide test cases and lets users run them after installation. |\n| Trust metadata | Records source URL, author, manifest signature verification, install count, and local validation state. |\n| Workflow mapping | Maps outputs from earlier workflow nodes into a Skill node input. |\n\n### Prompt Skill and Code Skill\n\nAn external `SKILL.md` is imported as a **Prompt Skill** when there is no `plugin.json`. The system reads its instruction text and uses it as an LLM prompt. It never executes third-party code.\n\nA **Code Skill** contains an executable entry point, for example:\n\n```text\nplugin\u002F\n  plugin.json\n  skills\u002F\n    security_scan.py\n```\n\n```json\n{\n  \"code\": \"security.scan\",\n  \"execution_type\": \"python\",\n  \"entrypoint\": \"skills\u002Fsecurity_scan.py:run\",\n  \"permissions\": [\"project-read\"]\n}\n```\n\n### Docker sandbox for Code Skills\n\nCode Skills can use a Docker sandbox. The runtime starts a temporary container and removes it when execution ends. The sandbox applies:\n\n- `--network none`: no outbound network access.\n- `--read-only`: immutable container root filesystem.\n- read-only mount for the Skill package.\n- memory, CPU, PID, and execution-time limits.\n- dropped Linux capabilities and `no-new-privileges`.\n- invocation result and failure logging.\n\nThis makes third-party extensions practical without treating them as trusted local code. Docker isolation is a defense layer, not a substitute for reviewing plugin source and permissions.\n\n## Plugin Marketplace\n\nThe Marketplace installs and tracks resource packages. Supported package types include `skill_pack`, `rag_pack`, `mcp_pack`, `prompt_pack`, `workflow_pack`, and `benchmark_pack`.\n\nSupported sources:\n\n- Built-in catalog packages.\n- Local folders or a local `plugin.json`.\n- URL and GitHub-style package sources.\n- External `SKILL.md` files, automatically converted to a safe Prompt Skill.\n\nAfter installation, the UI shows installed resources, source and trust details, available Skills, approval actions, test calls, workflow insertion, and uninstall status.\n\n### Strict approval model\n\n```text\napproval key = skill_code + agent_code\n```\n\nThe two common execution identities are:\n\n| Agent code | Meaning |\n| --- | --- |\n| `skill_console` | Manual test call from the Skills page. |\n| `workflow_runner` | Automatic call from a visual workflow. |\n\nApproving `skill_console` does not approve `workflow_runner`. A Skill must be explicitly approved for the context in which it will run.\n\n## Quick Start\n\n### Requirements\n\n- Python 3.11+ (Python 3.13 is supported by the current project setup)\n- Node.js 18+\n- Docker Desktop, optional for pgvector and Docker Code Skill sandboxing\n\n### Install\n\n```powershell\ngit clone https:\u002F\u002Fgithub.com\u002Fbiheto\u002FDevAgent-Studio.git\ncd DevAgent-Studio\n\npython -m venv .venv\n.\\.venv\\Scripts\\activate\npip install -e \".[llm,vector,dev]\"\n\ncd web\nnpm install\nnpm run build\ncd ..\n\ncopy .env.example .env\n```\n\n### Start the application\n\n```powershell\n.\\.venv\\Scripts\\python.exe -m uvicorn app.main:app --host 127.0.0.1 --port 8100\n```\n\nOpen `http:\u002F\u002F127.0.0.1:8100\u002F` for the workbench and `http:\u002F\u002F127.0.0.1:8100\u002Fdocs` for the API documentation.\n\n### Optional services\n\nRun pgvector:\n\n```powershell\ndocker compose -f docker-compose.pgvector.yml up -d\n```\n\nConfigure Docker Code Skill sandboxing in `.env`:\n\n```env\nDEV_AGENT_SKILL_SANDBOX=docker\nDEV_AGENT_SKILL_SANDBOX_IMAGE=python:3.13-slim\nDEV_AGENT_SKILL_SANDBOX_MEMORY=256m\nDEV_AGENT_SKILL_SANDBOX_CPUS=0.5\nDEV_AGENT_SKILL_SANDBOX_PIDS_LIMIT=64\nDEV_AGENT_SKILL_SANDBOX_FALLBACK=false\n```\n\n`subprocess` is the default sandbox mode for local development. `docker` requires Docker Desktop to be running. The sandbox status can be checked at `GET \u002Fapi\u002Fv1\u002Fskills\u002Fsandbox\u002Fstatus`.\n\n### LLM configuration\n\nThe application works with deterministic fallback responses when no key is configured. Set an API key for real LLM calls:\n\n```env\nOPENAI_API_KEY=your_api_key\nOPENAI_BASE_URL=\nDEV_AGENT_LLM_MODEL=gpt-4o-mini\n\n# Optional per-agent overrides\nDEV_AGENT_LLM_MODEL_PLANNER=gpt-4o-mini\nDEV_AGENT_LLM_MODEL_REPORTER=gpt-4o-mini\nDEV_AGENT_LLM_MODEL_CODE_REVIEWER=gpt-4o-mini\n```\n\n## Development Checks\n\n```powershell\n# Backend compilation\n.\\.venv\\Scripts\\python.exe -m compileall -q app\n\n# Frontend production build\ncd web\nnpm run build\ncd ..\n\n# Verify the Skill sandbox configuration\n.\\.venv\\Scripts\\python.exe -c \"from app.skills.sandbox import python_skill_sandbox_status; print(python_skill_sandbox_status())\"\n```\n\n## Project Structure\n\n```text\nDevAgent Studio\u002F\n  app\u002F\n    agents\u002F              # Project, review, RAG, learning, and report logic\n    api\u002F                 # FastAPI route modules\n    graphs\u002F              # LangGraph graphs and visual workflow compiler\n    harness\u002F             # Context, events, policy, artifacts, review\u002Fresume runtime\n    marketplace\u002F         # Package preview, installer, trust, and SKILL.md compatibility\n    persistence\u002F         # Task, governance, RAG, and trace persistence\n    providers\u002F           # LLM, MCP, and RAG provider interfaces\n    skills\u002F              # Registry, contracts, versions, dependencies, sandbox runtime\n    benchmark_runner.py  # LLM\u002FRAG\u002FWorkflow\u002FMCP\u002Fcollaboration benchmarks\n  web\u002F                   # React workbench\n  scripts\u002F               # MCP launchers and test servers\n  docs\u002F                  # Architecture and implementation notes\n  examples\u002F              # API request examples\n  docker-compose.pgvector.yml\n```\n\n## Documentation\n\n- [Implementation timeline](docs\u002FIMPLEMENTATION_TIMELINE.md)\n- [Workflow production notes](docs\u002FPHASE_8_WORKFLOW_PRODUCTION.md)\n\n## Roadmap\n\n- Add a richer UI for Docker sandbox health and test invocation.\n- Add signed external plugin publishing examples and contributor tooling.\n- Expand API, workflow compiler, runtime-state, MCP contract, RAG retrieval, LLM fallback, and benchmark test coverage.\n- Add conditional branch, parallel node, and richer input\u002Foutput mapping UX for workflows.\n\n## License and Attribution\n\nThis project is licensed under the [MIT License](LICENSE).\n\nCopyright (c) 2026 biheto. When redistributing the project, preserve the original copyright notice and license text.\n\n```text\nDevAgent Studio by biheto\nhttps:\u002F\u002Fgithub.com\u002Fbiheto\u002FDevAgent-Studio\n```\n\n---\n\n\u003Ca id=\"中文\">\u003C\u002Fa>\n\n# DevAgent Studio 中文说明\n\nDevAgent Studio 是一个面向**软件项目理解与研发治理**的开源多 Agent 工作台。项目基于 FastAPI、LangGraph 和 React 构建，用于帮助团队理解项目结构、识别代码风险、沉淀项目知识、治理 LLM 与工具调用，并执行可追踪的研发工作流。\n\n它不是代码编写 IDE，核心目标是让研发过程更容易理解、审计、评估和持续改进。\n\n## 项目亮点\n\n- **多 Agent 协作而非单次 Prompt**：通过 LangGraph 编排 Planner、项目分析、代码审查、RAG、监督和报告节点。\n- **Harness Runtime 运行时治理**：统一任务上下文、事件时间线、产物、策略、人工审核、持久化与恢复执行。\n- **真正可执行的可视化 Workflow**：前端拖拽画布会编译成 LangGraph 工作流执行，而不只是展示图。\n- **LLM 可观测与可治理**：可查看 Prompt 版本、调用 Trace、token\u002Fcost、fallback、A\u002FB Test，以及按 Agent 配置模型。\n- **安全可治理的 Skill 插件体系**：Skill 有契约校验、权限分级、严格审批、版本快照、依赖检测、测试用例和执行日志。\n- **插件市场与外部兼容**：支持内置包、本地路径、URL、GitHub 风格来源，以及外部 `SKILL.md` 自动转换。\n- **Docker 代码型 Skill 沙箱**：第三方代码可以在禁网、只读、限时限资源的临时容器中运行。\n\n## 核心能力\n\n| 模块 | 能力 |\n| --- | --- |\n| 项目分析 | 扫描目录、识别技术栈、归纳模块职责、风险和治理建议。 |\n| 代码审查 | 结合规则、调用链与 LLM 语义审查，输出问题和测试建议。 |\n| RAG 知识加工 | 文档切片、入库、检索、项目笔记；默认 SQLite，可选 pgvector。 |\n| 学习陪练 | 基于项目上下文生成学习计划和追问。 |\n| 多 Agent 协作 | Planner、Analyzer、Reviewer、RAG、Supervisor、Reporter 组成协作图。 |\n| Workflow | 拖拽、连线、配置、校验、保存并执行 Workflow JSON。 |\n| 人工审核 | 支持节点级通过\u002F拒绝、checkpoint\u002Fresume、重试与恢复事件展示。 |\n| MCP | 支持 Server 配置、工具发现、启停、审批、测试调用和日志追踪。 |\n| Benchmark | 覆盖 LLM、RAG、Workflow、MCP、多 Agent 协作的指标评估。 |\n\n## Skill 插件体系\n\nSkill 是可复用能力，例如代码审查、RAG 加工、学习陪练、安全扫描或 Workflow 执行。它可以在 Skills 页面单独测试，也可以加入可视化 Workflow。\n\n已实现的治理能力：\n\n- **契约校验**：安装和预览时校验 `input_schema`、`output_schema`、`permissions`、`execution_type`，避免格式错误的包进入运行时。\n- **权限风险分级**：使用 `safe`、`project-read`、`llm`、`workflow-write`、`network`、`filesystem` 标识访问能力和风险级别。\n- **严格审批**：审批键为 `skill_code + agent_code`。手动测试和工作流执行分别审批，互不放行。\n- **版本与回滚**：安装升级会保留版本快照，可对比和回滚。\n- **依赖声明**：Skill 可声明依赖的 MCP 工具、RAG collection、Prompt 版本和 LLM 模型，运行前会检查缺失项。\n- **测试用例**：插件包可携带测试用例，安装后可一键运行并记录结果。\n- **可信来源**：记录来源 URL、作者、manifest SHA-256 签名校验、安装次数和本地校验状态。\n- **Workflow 输入输出映射**：前序节点输出可映射到 Skill 节点输入，让 Skill 参与复杂工作流。\n\n### Prompt Skill 与代码型 Skill\n\n如果外部来源没有 `plugin.json`，但包含 `SKILL.md`，系统会将其识别为 **Prompt Skill**：只读取其中的指令文本并交给 LLM，不执行第三方代码。\n\n**代码型 Skill** 则带有可执行入口，例如 Python 文件。它能力更强，但必须通过权限审批和沙箱限制后执行。\n\n### Docker 沙箱\n\n当 `.env` 中设置 `DEV_AGENT_SKILL_SANDBOX=docker` 后，代码型 Skill 会在临时 Docker 容器中执行，并使用以下限制：\n\n- 禁止网络访问。\n- 容器根文件系统只读，Skill 包只读挂载。\n- 限制内存、CPU、进程数和执行超时。\n- 移除 Linux capabilities，禁止提升权限。\n- 执行结束自动删除容器，同时保留调用结果和失败日志。\n\nDocker 沙箱是隔离层，不代表插件天然可信。安装前仍应检查来源、manifest、权限和代码内容。\n\n## 插件市场\n\nMarketplace 支持安装和管理 `skill_pack`、`rag_pack`、`mcp_pack`、`prompt_pack`、`workflow_pack`、`benchmark_pack` 等资源包。\n\n支持的来源：内置资源包、本地目录或 `plugin.json`、URL\u002FGitHub 风格地址，以及外部 `SKILL.md`。安装后可查看已安装资源、来源与信任信息、Skill 列表、权限审批、测试调用、添加到 Workflow 和卸载状态。\n\n### 审批如何隔离\n\n```text\n审批键 = skill_code + agent_code\n```\n\n| Agent code | 使用场景 |\n| --- | --- |\n| `skill_console` | 在 Skills 页面手动点击测试调用。 |\n| `workflow_runner` | 在可视化 Workflow 中自动执行。 |\n\n因此，批准 `skill_console` 不等于批准 `workflow_runner`。Skill 必须在实际运行身份下单独获得批准。\n\n## 快速启动\n\n环境要求：Python 3.11+（当前项目支持 Python 3.13）、Node.js 18+；如需 pgvector 或 Docker 沙箱，还需要 Docker Desktop。\n\n```powershell\ngit clone https:\u002F\u002Fgithub.com\u002Fbiheto\u002FDevAgent-Studio.git\ncd DevAgent-Studio\n\npython -m venv .venv\n.\\.venv\\Scripts\\activate\npip install -e \".[llm,vector,dev]\"\n\ncd web\nnpm install\nnpm run build\ncd ..\n\ncopy .env.example .env\n.\\.venv\\Scripts\\python.exe -m uvicorn app.main:app --host 127.0.0.1 --port 8100\n```\n\n打开 `http:\u002F\u002F127.0.0.1:8100\u002F` 使用工作台，打开 `http:\u002F\u002F127.0.0.1:8100\u002Fdocs` 查看 API 文档。\n\n可选配置：\n\n```env\n# LLM\nOPENAI_API_KEY=your_api_key\nDEV_AGENT_LLM_MODEL=gpt-4o-mini\n\n# pgvector RAG\nDEV_AGENT_RAG_STORE=pgvector\nPGVECTOR_DATABASE_URL=postgresql:\u002F\u002Fpostgres:postgres@localhost:5432\u002Fdev_agent_studio\n\n# Docker 代码型 Skill 沙箱\nDEV_AGENT_SKILL_SANDBOX=docker\nDEV_AGENT_SKILL_SANDBOX_IMAGE=python:3.13-slim\nDEV_AGENT_SKILL_SANDBOX_MEMORY=256m\nDEV_AGENT_SKILL_SANDBOX_CPUS=0.5\nDEV_AGENT_SKILL_SANDBOX_PIDS_LIMIT=64\nDEV_AGENT_SKILL_SANDBOX_FALLBACK=false\n```\n\n启动 pgvector：\n\n```powershell\ndocker compose -f docker-compose.pgvector.yml up -d\n```\n\n可通过 `GET \u002Fapi\u002Fv1\u002Fskills\u002Fsandbox\u002Fstatus` 查看 Skill 沙箱配置状态。\n\n## 开发验证\n\n```powershell\n.\\.venv\\Scripts\\python.exe -m compileall -q app\n\ncd web\nnpm run build\ncd ..\n\n.\\.venv\\Scripts\\python.exe -c \"from app.skills.sandbox import python_skill_sandbox_status; print(python_skill_sandbox_status())\"\n```\n\n## 项目结构\n\n```text\nDevAgent Studio\u002F\n  app\u002F\n    agents\u002F              # 项目分析、代码审查、RAG、学习、报告逻辑\n    api\u002F                 # FastAPI 路由\n    graphs\u002F              # LangGraph 图与 Workflow 编译器\n    harness\u002F             # 上下文、事件、策略、产物、审核\u002F恢复运行时\n    marketplace\u002F         # 资源包预览、安装、可信信息、SKILL.md 兼容层\n    persistence\u002F         # 任务、治理、RAG、Trace 持久化\n    providers\u002F           # LLM、MCP、RAG Provider\n    skills\u002F              # Registry、契约、版本、依赖、沙箱运行时\n  web\u002F                   # React 工作台\n  scripts\u002F               # MCP 启动和测试脚本\n  docs\u002F                  # 设计和实现说明\n```\n\n## 后续计划\n\n- 增加 Docker 沙箱状态和测试调用的前端可视化。\n- 增加外部签名插件发布示例和贡献工具。\n- 补齐 API、Workflow 编译、Harness 状态、MCP 契约、RAG 命中、LLM fallback 和 Benchmark 的自动化测试。\n- 增强 Workflow 条件分支、并行节点和输入输出映射交互。\n\n## License \u002F 引用\n\n项目使用 [MIT License](LICENSE)。分发或修改时请保留原版权和许可证文本。\n\n```text\nDevAgent Studio by biheto\nhttps:\u002F\u002Fgithub.com\u002Fbiheto\u002FDevAgent-Studio\n```\n","DevAgent Studio 是一个面向开发者的开源多智能体软件工程治理工作台，专注于项目理解、代码风险审查、知识沉淀与AI工具调用治理。它基于 FastAPI、LangGraph 和 React 构建，支持可视化编排可执行的多智能体工作流（如 Planner、Code Reviewer、RAG Processor 等），提供可观测的 LLM 操作追踪、带沙箱约束的技能执行、版本化权限管控的插件市场，以及可审计、可恢复的任务运行时（Harness Runtime）。适用于中大型研发团队进行代码质量治理、技术债务分析、新成员项目上手辅助及 AI 工程化落地过程的合规性与可追溯性管理。",2,"2026-07-13 02:30:03","CREATED_QUERY"]