Skip to content

Releases

OppenCouncil v5.0.1

Choose a tag to compare

@HaobinZhou HaobinZhou released this 16 Oct 13:22

English

5.0.1 lets you review, edit, and confirm project definitions directly in Codex, and improves how you read them alongside their supporting literature in the web workbench. Discuss one group at a time in chat, or open the webpage to compare the wording with its sources. Both use the same project records.

Highlights

  • Review and confirm definitions in Codex. The AI still checks the facts, appraises the literature, and prepares complete proposals before bringing them to you. With your explicit authorization, Codex can save discussion, edit proposals, record your preferred option, confirm or reconfirm a definition, or withdraw a proposal. The corresponding records also appear in the workbench. By default, each exchange covers one related group, or up to three standalone questions when no groups remain, then waits for your feedback. The proposed wording and the decision you need to make come first, with short sentences and clearly separated points.

  • Read definitions and literature side by side, with evidence that follows your selected option. Select A to see only the evidence cited by A; switch to B to see B’s evidence. The full literature reader is available for both standalone definitions and group members. Citation numbers are generated automatically in order of first appearance within the current review. Click a number to locate the supporting evidence, then open the original source.

  • More screen space for the material you are reviewing. A compact header leaves more room for reading, while groups support both individual review and option comparison. Definition text appears before the recommendation’s rationale. AI discussion is collapsed by default when viewing evidence and can be reopened whenever needed. Darker text, clearer paragraphs, and more distinct spacing make definitions, findings, and applicability easier to read, with better use of large displays.

  • Source cards explain the findings before applying them to your decision. A dedicated source summary now holds the substance of the evidence. The Skill asks the agent to describe the relevant population, study design, comparisons, and results, retaining the numbers and uncertainty needed for interpretation before explaining applicability. Summaries are separate from notes on which passages were checked, and substantive evidence is separate from project confirmation, migration, and save records. When preparation runs in parallel, each subagent owns a complete set of related questions, including source checks, proposal preparation, and self-review.

  • Research reviews include a deliberate handoff for literature research. By default, the local agent checks project facts, prepares shared research questions, and returns control to you so you can arrange literature contributions from a web AI. It checks the contributed material, synthesizes it, and prepares the review only after you explicitly ask it to continue. You can also explicitly assign the entire round to the local agent.

This release also fixes the home page’s combined listing of projects on different governance versions, adds the browser tab icon, and corrects source links that included trailing citation punctuation.

Installation and upgrades

OppenCouncil-5.0.1.zip is the complete bundle, including the workbench, skills, and optional MCP component. The wheel and tar.gz packages contain the Council Python component. See the README or Chinese README for installation and update instructions.

Existing projects can stay on v3/v4. Software updates and project migrations are separate steps; when you want to change a project’s format, follow the migration guide.


简体中文

5.0.1 增加了直接在 Codex 对话中审阅、修改和确认口径的方式,也重新整理了网页的文献查阅和阅读布局。你可以在对话里逐组讨论,也可以打开网页,把口径与原文依据放在一起看;两边使用同一份项目记录。

重点更新

  • 在 Codex 里完成口径审阅。 AI 仍先完成事实核查、文献评估和完整候选准备,再与你讨论。你明确授权后,Codex 可以保存讨论、修改候选、选择方案、确认口径、重新确认或撤回候选,网页也能看到对应记录。对话默认每轮一个关联组;只剩独立问题时,每轮最多三条,等你反馈后再继续。口径和需要你决定的内容放在前面,解释使用短句和清楚的分项。

  • 口径和文献并排看,切换方案就切换依据。 选择 A 时只显示 A 引用的依据,选择 B 时切换到 B 的依据。独立 F 和组内 F 都能使用完整的文献查阅区。引用编号由代码按当前审阅内容的首次引用顺序生成;点击编号即可定位依据,再打开原文。

  • 把更多屏幕空间留给正在审阅的内容。 页面压缩了顶部说明,组内支持逐条阅读和方案比较。口径正文排在推荐理由之前;查看依据时默认收起 AI 讨论,需要时可以手动展开。正文、文献发现和适用条件使用更深的文字、更清楚的段落与间距,大屏幕上的阅读区也更充分。

  • 文献卡先讲研究发现,再讲它与当前决定的关系。 新增独立的来源摘要。Skill 要求说明相关研究的人群、设计、比较和结果,保留判断所需的数值与不确定性,再解释适用条件。原文核查位置与摘要分开,项目的确认、迁移和保存记录也与实质依据分开。多个问题需要并行准备时,每个子代理负责一组完整问题,完成查证、候选整理和自查。

  • 科研审阅留出真正的文献交接。 默认由本地 Agent 核查项目事实、整理共享问题,然后交回控制权,让你安排网页 AI 补充文献。收到你明确的继续指令后,本地 Agent 才核对资料、汇总并准备送审。你也可以明确指定由本地 Agent 完成整轮工作。

此外,修复了新旧治理版本的项目无法同时出现在首页的问题,补齐浏览器标签页图标,并修正了部分文献链接把句末标点带入地址的问题。

安装与升级

完整安装包是 OppenCouncil-5.0.1.zip,包含工作台、skills 和可选 MCP;wheel 和 tar.gz 是 Council Python 组件。安装与更新步骤见 README/中文版。

已有项目可以继续使用 v3/v4。软件更新与项目迁移分别进行;需要转换项目格式时,按迁移指南处理。

OppenCouncil v5.0.0

Choose a tag to compare

@HaobinZhou HaobinZhou released this 15 Oct 04:35

5.0.0 重构了从 AI 查证、准备方案,到你审阅确认、再继续执行的工作流程。AI 先把事实查清,把需要你决定的事集中放到网页里。你可以看推荐依据、修改候选文字和图表规格,再让 AI 从这一轮反馈接着做。

重点更新

  • 先筛问题,再交给你。 新的调查流程要求 AI 先判断问题是否存在、已有口径是否覆盖、能否从代码或数据中查清。网页按“需要选择”“快速确认”“补查中/阻塞”“已排除”组织事项。通行做法有依据,就给推荐答案;确实存在取舍,再说明各方案对当前项目的影响。排除理由也会留下,出现新情况时可以重新调查。

  • 推荐有出处,也说明为什么适用。 文献、代码核查、数据事实和测量结果可以独立保存为证据,并关联到具体口径。科研方案会说明通行做法、支持依据、适用条件,以及偏离通行做法的理由。通过 MCP 接入的 AI 助手可以补充资料、反例或意见,由本地 Agent 核对后纳入审阅材料。

  • 反馈可以逐条接着处理。 留言、已保存的候选修改和表格编辑都会进入待处理反馈。Agent 按新增输入继续工作,为每条意见记录回应;暂时解决不了的事项保留问题、原因和继续条件。下一轮可以接着处理新反馈和已经具备条件的事项,不用每次重新整理整场讨论。

  • 在网页里直接审表、改表、看图。 正式图表可以先用模拟样例确认规格。表头、行列、单元格和脚注都能在审阅页编辑,保存后交给 AI 继续。图片样例保留当时的副本,后续重画不会改变历史版本。正式产出关联已确认的规格和生成脚本;规格没变的数据重跑沿用原确认。

  • 依据变了,可以重新确认原口径。 证据更新后,可以标明受影响的口径并重新审阅。如果正文和依赖不变,只是在新依据下再次确认,就把确认记录追加到当前版本。需要修改口径正文时,再提出候选修订并确认新版本。

  • 科研和普通项目使用同一个 skill。 新的 oppencouncil 入口统一管理两类项目,科研流程可以使用 R、Python 或其他语言。项目类型决定需要哪些依据、如何组织正式产出;oppen-project-steward 和 stepwise-r-project 两个旧名称继续可用。

安装与升级

完整安装包是 OppenCouncil-5.0.0.zip,包含工作台、skills 和可选 MCP;wheel 和 tar.gz 仅包含 Council Python 组件。安装步骤见 README/中文版。

可以先升级软件,让已有项目继续留在 v3/v4。 要使用 v5 项目流程时,再单独生成并审阅迁移计划。迁移保留原有口径及历史、讨论、Memory 和产出登记,未解决的 Attention 转为调查事项继续保留。具体步骤见升级与迁移指南。

OppenCouncil v4.3.1

Choose a tag to compare

@HaobinZhou HaobinZhou released this 11 Oct 13:42

本次更新

  • 修复划词引用、填写和保存共同讨论时跳回其他主题的问题,保留当前问题、地址和阅读位置。
  • 精简 Stepwise R 与 Project Steward 的技能入口和 Council 核心说明,按任务加载操作细节,基础阅读量减少约 65%–66%。
  • 强化新增或修订口径前的必要性核实;每条 F 聚焦可独立审阅的决定,减少无依据的审核事项。
  • 统一新版候选权限说明:AI 可直接提出修订,正式生效仍需人类确认;Memory 规则保持不变。

下载与安装

完整安装请选择 OppenCouncil-4.3.1.zip,其中包含工作台、两个 skill 和可选 MCP。解压到固定目录后,按 README 的快速安装命令部署;skill 通过软链接接入全局目录。

仅安装 Council Python 组件可使用 oppencouncil-4.3.1-py3-none-any.whl;oppencouncil-4.3.1.tar.gz 是该组件的源码包。SHA256SUMS 提供附件校验值。

发布内容不包含本地项目口径、Memory、科研数据、密码或运行配置。

验证说明

发布前运行完整产品测试,并检查源码及附件的公开内容范围。新增 7 个 Skill 行为验收案例;它们是后续模型评估素材,不代表已经完成独立模型行为验收。

本地验收通过 604 项 Python 测试、12 项子测试和 76 项网页单元测试。GitHub Actions 在 Windows、macOS、Linux 上的 Python 3.12/3.13 六种组合均通过;main 与 v4.3.1 标签均核对到发布提交 a871bea。

OppenCouncil v4.3.0

Choose a tag to compare

@HaobinZhou HaobinZhou released this 11 Oct 07:06

本次更新

  • 冻结后仍可继续修订:AI 可以为已有正式口径提交完整的新候选。确认前原版本继续有效;人类确认后,新版本成为唯一当前版本,旧版本保留在历史中。
  • 前序升级后可以继续确认:候选页面展示前序的原版本、新版本及完整正文对照。点击“确认并采用新版前序”会保存已审阅的引用并冻结本条口径。审阅期间前序再次变化时,系统会要求刷新重审。
  • 每条口径独立确认:同组其他条目的草稿与讨论不会阻止本条确认;真实前序尚未确认或仍待复核时会说明原因。
  • 隐藏滚动条:工作台、项目目录和登录页面隐藏滚动条外观,保留原有滚动操作。
  • 两个 skill、Council 和可选 MCP 统一版本为 4.3.0,相关指引同步更新。

下载与安装

一般用户下载 OppenCouncil-4.3.0.zip。完整包包含工作台、两个 skill、可选 MCP 和安装脚本。解压到固定目录,进入该目录后执行:

uv run --python 3.12 scripts/install.py --python 3.12 --skill-dir "$HOME/.agents/skills"

需要 MCP 时在命令末尾追加 --with-mcp。安装器以软链接接入全局 skills,请保留解压后的目录。Windows 创建软链接需要开发者模式或相应权限。已有安装继续指定原来的 skill 全局目录,避免重复安装。

其他附件:

  • oppencouncil-4.3.0-py3-none-any.whl:仅 Council Python 包。
  • oppencouncil-4.3.0.tar.gz:仅 Council 源码发行包。
  • SHA256SUMS:以上三个包的 SHA-256 校验值。

版本升级后请重启正在运行的 Council / MCP 服务,并刷新浏览器。现有项目记录继续保存在原项目目录中。

验证

全量本地验收通过:189 项 Council 测试、273 项 skill/集成测试和 12 个子测试、73 项网页测试、142 项 MCP 测试(5 项跳过)。格式检查、真实 Chrome 合成项目确认与刷新持久化、发布 wheel 隔离安装检查通过。

发布包仅包含产品源码与合成测试文件,不包含科研项目数据、项目口径、skill 管理记录、运行配置或凭据。

GitHub 跨平台验收全部通过:Windows、macOS、Linux,各自覆盖 Python 3.12 和 3.13。