中文版

版本定位

ScienceDiscovery 全面拥抱 jiuwen-swarm 后端的第一个 Beta 版本：保留科研工作台体验，以 jiuwen-swarm 支撑 Agent 执行，在新架构上整合科研专属能力。

本版本为 0.3.0-beta，上一版本为 0.2.0。开发主线为 main，本次发布基于 releases/v0.3.0.beta 的提交 6d3c49a8。

Release Note

重点支持场景

本版本重点保障以下四类场景，不追求一次性迁移全部历史能力或覆盖所有边界情况。

场景	主要流程与交付	Domain 文档
文献调研	检索与获取文献、提取证据、多专家协作，交付带可追溯引用的调研报告	文献调研教程
数据分析	读取数据、准备环境、执行分析，交付可复现代码、图表与结论	脓毒症数据分析教程
抗体设计	结合领域 Skill、专业模型和科研环境完成设计与筛选；依赖相应模型资源与算力，计算结果不等同于实验验证	抗体设计教程
算法优化	通过科研产物 RSI / Evolve 生成候选、执行评估与迭代，保留结果与过程记录	算法优化教程

核心功能与改进

统一 JiuwenSwarm 默认后端：源码、Docker 和发行包启动器均默认选择 JiuwenSwarm；保留显式 native 切换。主子 Agent 默认通过平台 task 委派、由 Swarm 执行，并非全面切换为 Swarm 原生多 Agent 编排。
科研执行环境与工作区：支持本地/远程执行、托管环境、后台任务、文件传递、工作区交接及产物版本关联。
记忆图谱：记录科研任务、文件与产物关系，支持引用追溯；源码及 Docker 部署提供默认本地存储，可选 Neo4j；Linux 二进制发行包也已集成并托管可选的本地记忆图谱服务（#217、#218）。
科研 Skill、MCP 与 Specialist：支持内置能力和自定义扩展，组合专业专家完成科研任务。
科研产物 RSI 与 Idea Tree：支持产物迭代优化、候选评估和研究思路的树状探索。
可观测性与测试：完善执行轨迹、子任务状态和错误诊断，补充 Mock E2E、真实科研 Benchmark 与质量评估记录。
文档与入门体验：整理中英文文档和四类场景教程，明确 Swarm/native 在上下文、规划、工具、Skill 等方面的行为差异。

这些功能包含已有能力的迁移与增强，并非全部首次引入。

官网、社区与交付配套（本版本新增）

统一对外网站：提供官网与安装入口，集中展示安装指引、使用文档与科研场景教程；网站内容随 main 分支同步更新。
用户交流社区：面向海外与中国用户，分别建立 Slack 和飞书用户交流群，用于使用交流、问题反馈与社区协作；加入方式见仓库首页。
双分支每日 CI：构建每日质量检查，覆盖 main 和 releases/v0.3.0.beta。由 main 的定时任务触发两条分支各自的检查，分别保留结果、日志和构建产物，为主线演进与版本发布提供持续质量反馈。
多平台、多种安装方式：支持 Linux、macOS，Windows 可通过 WSL 2 使用；提供源码、Linux 二进制发行包及 Docker 等部署方式。不同系统可用的安装方式与前置条件不同，详见官网安装页。

主要 Bugfix

记忆图谱内存与查询性能：修复折叠查询的路径爆炸，使用共享路径及累计集合预算，避免已复现的高内存分配；超预算返回结构化错误。优化会话索引查询，避免大量无关历史数据拖累新会话（#212，修复 #210）。
启动行为一致性：修复源码与 Docker/发行包默认后端不一致，确保命令行后端选项优先于 .env；Docker 转发 Swarm 配置，E2E 显式选择后端（#214）。
Swarm/MCP 隔离与恢复：改进模型、会话、工具绑定隔离，工具刷新、共享连接和权限生命周期（#173、#181、#182）。
子任务与远程执行可靠性：改进并发排队、超时、取消、工作区交接和子 Agent 恢复；远程执行状态查询增加瞬时错误重试（#182、#187、#191）。
模型输出与事件流：改进截断输出恢复，避免执行不完整工具参数，增强工具事件流存活与诊断（#190、#194）。
科研数据与交付：改进 PDF 校验、arXiv 接入兼容性、产物版本关联，以及界面状态、快照作用域和取消诊断（#171、#180、#192、#179、#189、#193）。

Beta 边界与升级说明

Nightly Real E2E 尚不稳定：自动真实模型 E2E CI 仍在稳定化过程中，当前真实科研场景的验收以本地实际执行并保留记录的结果为准，需注明被测提交、模型和配置。Nightly 结果作为辅助诊断，不单独作为版本质量已达标的依据；跳过或隔离的用例不视为通过。UT/ST/Mock E2E 门禁仍是独立的质量依据。
部分安装路径为实验特性：macOS Docker 以及所有 Windows 安装路径（包括 WSL 2 和 Docker Desktop）尚未全面验证 Agent 执行效果。当前预编译单文件包面向 Linux，不提供原生 macOS/Windows 二进制；macOS 优先使用源码方式，Windows 的 Linux 二进制/源码方式需在 WSL 2 内运行。具体支持范围与依赖见官网安装页。
官网与发行版的差异：官网跟随 main，可能包含尚未进入本次 Beta 的更新；本版本的准确行为与限制以发行分支文档和最终 Release Note 为准。
场景验收与效果边界：四类场景是重点验收范围，效果仍受模型、网络、数据源和算力影响；测试已加入不代表所有真实场景已稳定通过，不应将本 Beta 视为所有科研任务均已稳定通过的保证。
Swarm 初始化与后端切换：源码首次使用 Swarm 前执行 scripts/jiuwenswarm.sh setup。保留原生循环可使用 --no-jiuwenswarm 或 SCIENCE_AGENT_EXECUTOR=native。
记忆图谱升级与故障处理：升级不会覆盖已保存的记忆功能设置；此前手动关闭的用户需在系统设置中重新开启。已有图谱可通过 SCIENCE_AGENT_MEMORY_GRAPH_DATA_DIR 指向原目录，不自动迁移。设置 SCIENCE_AGENT_MEMORY_GRAPH_AVAILABLE=0 可跳过服务；该可选服务启动失败时会报告原因并在禁用图谱的情况下继续启动，修正后需重启。
外部数据源限制：arXiv 406/429、排队超时等外部数据源问题仍由 #208 跟踪。

问题反馈

欢迎通过 GitHub Issues、Slack 或飞书反馈使用问题，请附配置、复现步骤及日志，并注意移除 API Key 等敏感信息。

English version

Release positioning

This is ScienceDiscovery’s first Beta release embracing the jiuwen-swarm backend. It retains the scientific workbench experience, uses jiuwen-swarm for Agent execution, and integrates science-specific capabilities into the new architecture.

This release is 0.3.0-beta, following 0.2.0. Development continues on main; this release is built from commit 6d3c49a8 on releases/v0.3.0.beta.

Release notes

Priority scenarios

This Beta focuses on four scientific workflows rather than migrating every historical capability or covering every corner case.

Scenario	Workflow and deliverables	Domain documentation
Literature review	Search and retrieve literature, extract evidence, collaborate across specialists, and deliver a report with traceable citations	Literature research tutorial
Data analysis	Inspect data, prepare an environment, execute analysis, and deliver reproducible code, figures, and conclusions	Sepsis data analysis tutorial
Antibody design	Combine domain Skills, specialized models, and scientific environments for design and screening; requires appropriate model resources and compute, and computational results do not constitute experimental validation	Antibody design tutorial
Algorithm optimization	Generate, evaluate, and iterate on candidates through scientific-artifact RSI / Evolve, retaining results and execution records	Evolve a solution tutorial

Core capabilities and improvements

Consistent JiuwenSwarm default: source, Docker, and packaged launchers now select JiuwenSwarm by default, with an explicit native opt-out. Subagents default to platform task delegation with Swarm execution—not a complete switch to Swarm-native multi-agent orchestration.
Scientific environments and workspaces: local/remote execution, managed environments, background tasks, file transfers, workspace handoffs, and artifact-version associations.
Memory graph: research-task, file, and artifact relationships with traceable citations; source and Docker deployments provide local storage by default, with optional Neo4j. Linux binary packages also bundle and supervise the optional local memory-graph service (#217, #218).
Scientific Skills, MCP, and Specialists: built-in capabilities and custom extensions for domain-expert collaboration.
Scientific-artifact RSI and Idea Tree: iterative artifact optimization, candidate evaluation, and tree-based exploration of research ideas.
Observability and testing: improved trajectories, subtask status, and diagnostics, alongside Mock E2E, real scientific benchmarks, and quality-evaluation records.
Documentation and onboarding: reorganized English/Chinese documentation and tutorials for the four workflows, with explicit Swarm/native differences in context, planning, tools, and Skills.

These include migrations and enhancements of existing capabilities, not exclusively new features.

Website, community, and delivery infrastructure (new in this release)

Unified public website: the website and installation hub bring together installation guidance, documentation, and scientific workflow tutorials. Website content is kept in sync with main.
User communities: Slack and Feishu groups have been established for international and China-based users respectively, supporting usage discussions, feedback, and community collaboration. See the repository homepage for joining instructions.
Daily CI for both branches: daily quality checks cover main and releases/v0.3.0.beta. The scheduled main workflow triggers independent runs for both branches, retaining separate results, logs, and build artifacts to support development and release validation.
Multiple platforms and installation options: Linux and macOS are supported, with Windows access through WSL 2. Options include source installation, Linux binary packages, and Docker. Availability and prerequisites vary by platform; see the installation hub.

Major fixes

Memory-graph query performance and memory use: eliminate path explosion in folded queries, introduce shared paths and cumulative collection budgets for reproduced allocation failures, and return structured budget errors. Indexed session queries avoid scanning unrelated historical data (#212, fixes #210).
Consistent startup behavior: align source defaults with Docker/packages, make CLI backend selection override .env, forward Swarm settings through Docker, and select E2E backends explicitly (#214).
Swarm/MCP isolation and recovery: improve model/session/tool binding isolation, tool refresh, shared connections, and permission lifecycle (#173, #181, #182).
Subtask and remote-execution reliability: improve admission queues, timeouts, cancellation, workspace handoffs, and subagent resumption; retry transient remote execution-status failures (#182, #187, #191).
Model output and event streams: recover from output truncation without executing incomplete tool arguments, and improve tool-stream liveness and diagnostics (#190, #194).
Scientific data and delivery: improve PDF validation, arXiv compatibility, artifact-version associations, UI state, snapshot scope, and cancellation diagnostics (#171, #180, #192, #179, #189, #193).

Beta limitations and upgrade notes

Nightly Real E2E is not yet stable: automated real-model E2E CI is still being stabilized. For now, acceptance of real scientific workflows relies on recorded local runs, identifying the tested commit, model, and configuration. Nightly results provide supplementary diagnostics, not standalone proof of release readiness. Skipped or quarantined c...