@sanqianx/project-knowledge 5.0.24 → 5.0.26

package/CHANGELOG.md

@@ -1,1353 +1,1361 @@
-# Changelog
-
-## [5.0.24] - 2026-10-20
-
-- 纳入 Workbench 0.13.6：Claude Code／GLM 连续对话保留每轮输入与回答的关联，上一轮迟到的结果和完整回答不再重复出现在新问题后面，也不会把新一轮误标为空闲；主、子对话同时修复。
-- 历史重新加载按事件顺序合并，已合并的流式正文替换原片段；主对话与子代理、不同原生会话的正文不会合并到同一历史行。
-- 修复较早任务列表响应覆盖新任务状态、将已完成子代理误判为失效的问题；经过服务端验证的任务历史可恢复显示，读取失败提供只读重试，不重新派发任务。
-- 保留 5.0.23 的明确选择提交及 5.0.22 的后台启动修复。
-
-## [5.0.23] - 2026-10-20
-
-- 接入 Workbench 0.13.5：Claude Code（含 GLM 等自定义模型）的用户提问显示单选、多选及自定义答案，明确提交后才继续；绕过普通工具授权时也不能代选或空答案放行。Codex 原生用户提问同步接入选择界面。
-- 主对话与子代理分别收集并回传答案，普通答案草稿刷新可恢复；关闭标签、收起面板和刷新不会提交答案，秘密输入不写入浏览器持久存储。
-- 子代理选择卡片并入对话滚动区域，发送长消息后输入框恢复高度，确保选项、提交和停止按钮可以使用。保留并排工作区、按需知识引用及 5.0.22 的后台启动修复。
-
-## [5.0.22] - 2026-10-20
-
-- 将启动时的 Agent 接入检查和自动修复放入独立进程，避免慢客户端命令阻塞后台就绪通知、网页和实例控制；超时会记录可重试状态，停止服务同时清理自有检查进程。
-- 仅检查用户已连接的客户端；未能核实的状态不再当成未安装并触发自动修复。保留并排工作区、已有接入选择及知识数据。
-- 补充真实后台 CLI 在原生命令阻塞时的就绪、HTTP、状态查询和停止回归，以及超时、崩溃、手动修复结果优先级验证。
-
-## [5.0.21] - 2026-10-20
-
-- 开始工作采用并排工作区：从加号新建子代理或按需添加知识引用，使用统一的对话、工具步骤、文件修改和输入框；移除面向客户的任务标识、派发阶段及诊断信息。
-- 子代理首次发送时继承当前项目、Agent、模型及主对话的冻结上下文，之后独立交流；原生任务可从主对话打开，并基于结果继续创建子对话。
-- 知识引用仅保留会话内单一标签，检索不自动展开、不抢焦点；修复刷新时内容提前投递的问题，并跟随产品深浅主题。
-- 主子草稿、标签和阅读位置可恢复；关闭标签不停止任务，窄窗口自动使用抽屉。精确锁定 Workbench 0.13.3。
-
-- 默认维护模型明确报告额度不足时，保留并展示可重试的失败原因；请求限流、认证失败分别展示。同项目后续提交等待前一条恢复，失败及排队记录跨重启保留。
-- 事件消费在分析任务建单前受阻时，活动页与系统状态展示故障，不再只显示空记录；故障及恢复通过现有实时连接更新，稳定状态不触发额外数据轮询。
-
-## [5.0.19] - 2026-10-16
-
-- 工作台展示来自真实 SDK 事件的后台任务、工具活动及完成结果；主回复结束后可继续输入，后台任务继续归属原会话，切换和刷新可恢复。
-- 移除外壳列表定时轮询，复用实时连接推送摘要变化并增量更新侧栏；草稿按服务入口、项目和会话隔离，保护发送失败、迟到创建和附件读取的原始归属。
-- 修复新会话创建时的旧会话误用、追加消息 409、汇总回复重复展示，以及重启读取项目列表失败导致的重复登记与接线缓存问题。
-- 候选内容清单按 npm 实际安装行为规范 CLI shebang 的换行符，避免 Windows 工作树打包后产生错误的内容哈希告警；普通文件与脚本正文继续逐字节校验。
-- 自动长期记忆仅接受具有冻结人类原句与来源证据的跨项目长期规范；代理派单、单次修复和项目版本事实保留为项目知识。旧候选不自动改写。
-- 纳入 Workbench 0.13.2-rc.21 和经过真实向量回归的引擎修复；固定 30 条检索查询全部通过，109 处引用一致，真实修复提交完成分析、晋升、索引并由下一轮 GLM 实际采用。验收范围见 5.0.19 记录。
-
-## [5.0.17] - 2026-10-15
-
-- 修复网页首屏被慢模块拖住、开始工作 Agent 探测同步阻塞与返回页面状态不刷新的问题；独立读取、明确失败状态和缓存保留，iframe 不重建。
-- 同源实时连接合并为每标签一条 WebSocket，保留独立模块 SSE，恢复时先读取快照并按游标续接；项目接线失败不再阻塞其他项目。
-- 工作台输出时向上滚轮立即暂停跟随，完整历史独立持久化并向前分页；旧 Claude 历史只读恢复，不再把有限事件缓存冒充全会话。
-- 精确锁定 Workbench 0.13.2-rc.17 与引擎 0.2.7-rc.10；补齐诊断和任务启动的异步发现旁路，保留 ZCode 注册表安装路径，并阻止已中止或被替换的初始化迟到启动任务。最终候选已通过独立下载安装、真实 MiniMax 与向量闭环，正式包先发 next 验收后再晋升 latest，详见 5.0.17 验收记录。
-
-## [5.0.16] - 2026-10-15
-
-- 收敛 rc.8 已通过的 30 项核心门禁：修复真实主题证据遗漏、短向量子片段重复占位与小型明确提交上下文不完整，保留当前文件校验、记忆优先级及上下文预算。
-- 精确锁定 Workbench 0.13.2-rc.13、vector-hub 1.0.8-rc.8、event-bridge 0.7.2 和 vendored engine 0.2.7-rc.9；保留原生流式会话、项目双目录和 D 盘索引。
-- Codex 本地市场与 MCP 接线使用真实原生命令验证；四客户端实际接入，Claude/Codex/ZCode 首次开发请求的事实与产品记忆送达均实测通过，OpenCode 保留真实工具检索兼容，不冒称自动注入。
-- 正式包先发布 next 并重新下载安装验收，完成后才更新 latest；候选通过不等于正式包已验收。首次失败、升级保护与完整证据见版本验收记录。
-
-## [5.0.16-rc.8] - 2026-10-15
-
-- 纳入引擎 0.2.7-rc.9：经验证的短向量子范围与包含它的 Markdown 段落不再重复占用片段名额；保留较大独立向量范围与原有校验，不扩散向量分数。
-- 用户明确引用已识别提交时，在原有上下文预算内完整送达当前项目的小型条目，先复验文档哈希与归属，避免摘要和交叉引用挤掉真实证据。大条目、普通主题和跨项目参考不展开；确认记忆优先级不变。
-- 保留 rc.7 真实网页失败及源码红测试。rc.7 双 Node 外壳 139 项、安装版八项、真实记忆与并发十项通过，不抵消真实任务正文遗漏；rc.8 必须重新验 npm 安装版，未完成前不更新 latest。
-
-## [5.0.16-rc.7] - 2026-10-15
-
-- 纳入引擎 0.2.7-rc.8：修复 commit 标识与主题混合查询只返回概述、漏掉已有正文证据的问题。已识别 SHA 只限定来源，不重复参与主题打分；提交内按主题 BM25 选择片段，普通语义查询保留 RRF，并补中英单写者与引擎词汇。
-- 保持每文档两段、正文／向量区间校验、授权范围与上下文预算；同分排序稳定，不以标题中的 SHA 冒充任务事实已送达。
-- 新增真实原文遗漏的红／绿用例及 topic-context 发布门禁；真实工作台网页测试验证具体事实和实际回答，其他模块版本不变。候选未完成安装版与原生验收前不更新 latest。
-
-## [5.0.16-rc.6] - 2026-10-15
-
-- 修复 Codex 0.137 将 scoped npm 目录中的 `@` 误作 Git ref：本地目录含 `@` 时从该目录执行 `marketplace add .`，更换和回退分别使用自己的工作目录，不复制目录、不访问 GitHub。
-- CLI 子操作日志关联原操作 ID，允许继续处理的失败仍记录安全诊断；本地路径误判保留 `CODEX_LOCAL_SOURCE_REF`，不写入原始 CLI 输出。
-- 保留 rc.5 的已验证四模块版本、索引与知识功能。rc.5 本机保真升级已通过：7 项目、43 历史文件、565 索引文件保留；真实记忆闭环与双 Node 浏览器门禁通过，热查询 P95 为 693ms（1000 文档合成语料）。
-- 本机 Codex 已通过原生本地市场操作恢复，MCP 声明入口 initialize/六工具/resolve/search/get 验证通过；Hook 尚待用户审核，不计自动上下文通过。候选不更新 latest。
-
-## [5.0.16-rc.5] - 2026-10-15
-
-- 精确锁定 Workbench 0.13.2-rc.13：模型目录加载失败不再伪装成空目录；保留上次有效结果，提供明确错误与重试，打开菜单可恢复请求，关闭菜单后的迟到响应不会重开菜单。保存或删除供应商后不复用保存前的旧快照，普通菜单打开仍共享请求。
-- 保留 rc.4 的引擎与向量检索修复；模型菜单错误不显示原始异常或凭据。补充连续保存的竞态回归及测试服务器模块隔离，安装版故障恢复仍须单独验收。
-- Leader 负责关键开发和发布收尾，GLM 并行独立做故障测试与审查。候选发布到 next，不改变 latest。
-
-## [5.0.16-rc.4] - 2026-10-14
-
-- 引擎 0.2.7-rc.7 复用同一次检索刚读取的知识正文，保留晚到响应后的文件与范围校验；vector-hub 1.0.8-rc.8 将有界查询缓存延长到 10 分钟，并防止旧请求失败删除新缓存。源码真实向量热查询 P95 为 1069ms，不代表安装版验收通过。
-- 锁定 vector-hub 1.0.8-rc.8（ebf6f586a534c4b851c919c53a5536671cc8bd12），保留 Workbench 0.13.2-rc.11 与 bridge 0.7.2；四个模块集与官方 npm tarball 双向逐字节核对（vendored 引擎 40、Workbench 96、vector-hub 59、bridge 38）。
-- 候选不改变 latest；完整安装版和原生客户端验收通过后才允许晋升。
-
-## [5.0.15] - 2026-10-13
-
-- 冻结 rc.5 已验收的中文 BM25 与可校验向量片段融合、本项目进度与跨范围上下文、无可靠匹配诊断及离线 Claude 插件接入。
-- 包含 UI 属性/源码示例被知识晋升误判的修复，以及 Codex 本地市场重载修复；保留真实 Git 失败和原生 Hook 信任审核。
-- 正式包先下载和测试，再更新 latest；本机与另一台设备的结果分别报告。
-
-## [5.0.15-rc.5] - 2026-10-13
-
-- 修复 Codex 本地插件市场被强制执行 Git 刷新而无法更新的问题；识别原生明确的非 Git 市场反馈后重新加载本地插件，真实 Git/认证失败仍终止，保留可用旧插件。
-- 不写入 Hook 信任、不跳过原生审核；保留 rc.4 的知识晋升误判修复与检索改进。
-- 候选包先下载验收，不提升 latest。
-
-## [5.0.15-rc.4] - 2026-10-13
-
-- 修复知识晋升校验把 UI 的 `placeholder=` 属性及有正文解释的源码示例误判为未完成内容的问题；真实未完成标记、空正文、越界及哈希校验继续拦截。
-- 保留失败记录及项目 FIFO，不升级后自动重做失败提交；用户可通过运行详情显式重试。
-- 保留 rc.3 的检索与离线接线修复；重新进行 npm 安装版验收，暂不提升 latest。
-
-## [5.0.15-rc.3] - 2026-10-13
-
-- 当前项目没有维护历史但已有其他知识时，最近进度检索明确显示“本次无可靠匹配”，不误报整个项目知识为空。
-- 增加用户原句“我想知道我当前向量库存储的位置在哪”的回归与现有知识文件只读验收；存储规则引用与当前设备实际路径分别核验。
-- 保留 rc.2 的片段校验、中文排序、跨项目范围隔离和离线插件接入；候选版不更新 latest。
-
-## [5.0.15-rc.2] - 2026-10-13
-
-- Windows 路径边界校验改用原生规范化，保留中文/空格、缺失末级目录和目录联接越界检查。
-- 保留 rc.1 的全部检索及离线接线修复；性能验收明确区分真实向量召回与最终选择词法结果，不把词法结果标为向量命中。
-- 候选版仍不提升 latest，源码、npm 安装版、真实客户端及另一台设备的反馈分别记录。
-
-## [5.0.15-rc.1] - 2026-10-13
-
-- 候选版检索采用当前 Markdown 分块上的中文 BM25 与可验证向量片段融合；每篇最多两个互补片段，不将文档向量分数扩散到全部章节。
-- 最近进度优先本项目提交依据；开发上下文本项目事实保留预算、领域参考限量，记忆仍独立准备。无可靠匹配不再表示知识库为空。
-- Claude 插件默认从 npm 包内本地市场安装，无需 GitHub；接入状态探测放入有界后台进程，超时与未安装分开显示并记录日志。
-- 保留 v5 数据、第三方 Hook 与客户端配置。候选发布不更新正式 latest；真实模型、客户端及安装版验收分别记录。
-
-## [5.0.14] - 2026-10-13
-
-- 长期记忆页增加“添加记忆”：填写内容并选择全局、领域或项目，用户确认后直接保存为已确认记忆。
-- 手动来源独立标明，不伪造 Agent 分析、项目来源或 commit；复用引擎受控写入、持久化、共享索引和开发前上下文。
-- 校验无效目标、重复内容和越界路径；取消不保存，保存与编辑自动刷新，写入失败保留记录并报告待恢复。
-- 保留 5.0.13 的重复采集修复：已有旧 Hook 仍需在 Agent 接入中执行“接入/修复”，升级启动不等于已清理旧接线。
-
-## [5.0.13] - 2026-10-13
-
-- 修复独立 bridge 与产品 bridge 同时接入导致同一项目生成两个知识目录：接入/修复只替换明确的旧自有入口，并去除当前入口的重复项。
-- Claude Code、ZCode 重复采集不再显示为正常；保留混合分组中的第三方 Hook、知识读取、模型配置和历史日志。
-- Codex 通知链识别默认旧 bridge，包括已知通知分支中嵌套的旧入口；保留第三方通知，不恢复退役的重复采集。
-
-## [5.0.12] - 2026-10-12
-
-- 冻结 rc.3 已验收的三客户端开发前知识读取：Claude Code、Codex、ZCode 获取统一事实及适用的已确认长期记忆。
-- 当前编辑记忆优先于历史实现；指定已知源文件不会被重复历史向量结果挤出，保留引用与范围校验。
-- ZCode 入口跨 npm 安装目录更新并去重，Codex 遵循原生 Hook 信任审核，保留用户模型、第三方 Hook 和既有采集。
-- 正式包先下载验收，再更新 latest；OpenCode 保持 MCP/采集兼容，不宣称未经验证的自动注入。
-
-## [5.0.12-rc.3] - 2026-10-12
-
-- 修复点名源文件的需求被相似历史向量结果挤出的情况；已有文件标识先匹配知识依据，新文件名仍可检索相似背景。
-- 修复 ZCode 安装目录变化时残留旧上下文入口；只替换明确属于本产品的脚本，去除重复入口，保留第三方 Hook、采集及模型。
-- 保留真实模型首轮失败和再次验证记录，未完成 Codex 原生信任审核前不提升 latest。
-
-## [5.0.12-rc.2] - 2026-10-12
-
-- 修正自动上下文的当前记忆与历史事实边界，保留编辑时间及项目、领域、全局的适用优先级。
-- 新增保留旧代码及历史知识的真实编辑记忆验收；候选版仍需真实模型验证，不以注入成功代替实际采用。
-- 保持用户当前需求及安全限制优先，不因记忆增加授权；不改动用户知识文件、路径和第三方设置。
-
-## [5.0.12-rc.1] - 2026-10-12
-
-- 候选版新增 Codex、ZCode 原生需求提交前的知识读取，复用引擎同一份事实、记忆与引用。
-- 修复 Codex 旧插件存在时跳过升级；区分安装、定义完整、启用和原生信任状态，未信任不显示已就绪。
-- ZCode 使用用户配置的 process Hook，保留第三方 Hook、模型及既有采集；被禁用的 Hook 不静默重启。
-- 工作台提供上下文和原始会话采集时，仅抑制自身原生重复入口；候选实测未通过前不替换 latest。
-
-## [5.0.11] - 2026-10-12
-
-- Claude Code 的普通开发需求通过原生前置 Hook 自动取得项目/领域知识和适用的已确认长期记忆；MCP 保留用于深入读取。
-- 引擎统一准备有界上下文，工作台复用同一接口并避免重复注入；知识读取失败明确提示后继续，向量故障不冒充成功。
-- 修复旧知识插件存在即跳过更新及 Windows Git CRLF/npm LF 接线误判；插件版本、启用、文件与采集能力分别验证。
-- 保留 v5 数据、模型、第三方 Hook 和向量位置；原生模型请求与开发提交闭环分开验收。首批只承诺 Claude Code 自动读取，Codex/ZCode 后续交付。
-
-## [5.0.10] - 2026-10-12
-
-- Freeze the accepted rc.2 scoped-memory runtime: review and edit candidates,
-  choose global/domain/project applicability, and confirm in one transaction.
-- Preserve v5 knowledge, source evidence, histories and directory bindings;
-  only applicable, confirmed and synchronized memories enter task retrieval.
-- Include real MiniMax candidate extraction, 1536-dimensional embeddings,
-  project/domain/global isolation, CLI/MCP agreement and restart regression.
-- Publish first to next; formal npm installation and existing-user upgrade
-  acceptance precede stable latest promotion. Other module versions stay pinned.
-
-## [5.0.10-rc.2] - 2026-10-12
-
-- Require candidate entry IDs to name an actually promoted Markdown document;
-  record rejected candidate counts and safe reason codes instead of silently
-  losing model suggestions or inventing source citations.
-- Keep valid knowledge updates independent of optional candidate failures;
-  roll back an unpersisted candidate batch and expose diagnostics in activity.
-- Distinguish changed-memory index jobs from unchanged health probes so ready
-  memories do not remain falsely marked as indexing.
-
-## [5.0.10-rc.1] - 2026-10-12
-
-- Add global/domain/project memory applicability and a single review form for
-  editing content, selecting any existing target and confirming atomically.
-- Retain original evidence, scope-aware duplicate suppression, existing v5
-  global memories and all fixed code/knowledge directory bindings.
-- Filter exclusive memories before retrieval ranking and file reads; clear old
-  projections after scope changes and report/retry pending file synchronization.
-- Keep the accepted Workbench, vector-hub and event-bridge versions unchanged;
-  candidate package acceptance precedes stable latest promotion.
-
-## [5.0.9] - 2026-10-12
-
-- Freeze the accepted rc.19 web runtime and exact module versions; no new
-  product features, old-version migration or existing v5 data reset.
-- Include index failure diagnostics, index-only recovery, knowledge-root
-  index storage, safe client discovery/notify chaining, maintenance isolation
-  and the embedded Workbench startup/streaming fixes from the candidate series.
-- Publish initially to next. Official npm installation, Node 22/24 regression,
-  real MiniMax knowledge/embedding loops and existing-user upgrade acceptance
-  must pass before stable latest promotion.
-
-## [5.0.9-rc.19] - 2026-10-12
-
-- Pin Workbench 0.13.2-rc.7. Startup shows a loading state and keeps controls
-  inert until projects and event handlers are ready; the existing streaming
-  and session recovery paths are unchanged.
-- Preserve the real rc.18 premature-click failure and require its actual
-  usable-page ready receipt in browser acceptance. Add a deterministic
-  startup gate regression and Node 22/24 retests.
-- Include Windows test-server cleanup hardening and corrected storage-root
-  assertions. Retain existing v5 data, third-party configuration and the
-  official-package acceptance gate before stable latest promotion.
-
-## [5.0.9-rc.18] - 2026-10-12
-
-- Project resolution for MCP and CLI now reports the vector-hub storage root
-  selected in settings, not the retired shell database path. A running tool
-  rereads root changes; resolution does not move files or rebuild indexes.
-- Preserve the reproduced rc.17 diagnostic mismatch and add Node 22/24
-  coverage for configured, changed and unset roots and MCP output.
-- Retain existing v5 data and all accepted module pins. Publish to next;
-  stable latest remains behind official-package and actual-user acceptance.
-
-## [5.0.9-rc.17] - 2026-10-11
-
-- Pin Workbench 0.13.2-rc.6. Repeated embedded notifications for the current
-  project no longer abandon a newly created session; real project changes and
-  late project lookups still preserve isolation and per-project drafts.
-- Keep a healthy progress stream labelled as connected on redundant online
-  events. Verify actual offline cleanup, snapshot recovery and exactly one
-  restored SSE connection with a real browser and HTTP proxy.
-- Preserve the rc.16 real Workbench failure and require official-package
-  streaming/citation acceptance before upgrading the actual user instance.
-- No data reset or third-party configuration overwrite. Stable latest remains
-  behind the full candidate and installed-product acceptance gates.
-
-## [5.0.9-rc.16] - 2026-10-11
-
-- Pin Workbench 0.13.2-rc.5. Bound OpenCode's health request at both
-  response headers and JSON body so a stalled local server cannot exceed
-  the startup handshake budget before the existing CLI fallback.
-- Reproduce the old behavior using two live HTTP connections, retain the
-  failed rc.15 client acceptance and require a new official-package retest.
-- No project, knowledge, history, provider or third-party hook reset.
-  Candidate is published to next; stable latest still requires acceptance.
-
-## [5.0.9-rc.15] - 2026-10-11
-
-- Serialize registration, module removal and registry deletion for each project.
-  Recheck membership inside the wiring lease so delayed supervisor registration
-  cannot resurrect a project after keep-data unbind. Unrelated projects remain
-  independent; failed deletion can still be repaired and retried.
-- Reproduce the race with a delayed real HTTP registration and preserve the
-  downloaded rc.14 Node 22 UI failure. Include the regression in the package.
-- Use non-binding CLI readiness probes in tests; the old bind probe could take
-  the exact port while its own foreground child was starting.
-- Real MiniMax acceptance covers overlapping analyses, another commit and a
-  dirty developer draft during analysis, streaming, secondary development,
-  four installed clients and evidenced preference candidates. New downloaded
-  candidate acceptance and actual-user cutover still precede stable latest.
-
-## [5.0.9-rc.14] - 2026-10-11
-
-- Vendor engine 0.2.3-rc.7: isolated maintenance consumes frozen evidence even
-  while the user edits or commits in the source repository. Do not mistake live
-  development for analyzer writes; retain live-source protection for custom
-  analyzers without the isolation capability.
-- Add a paused child-worker regression covering another commit, uncommitted
-  drafts, frozen attribution, same-project FIFO and unchanged path/tool guards.
-- Existing-user npm acceptance verified real MiniMax streaming, retained v5
-  history/configuration, D-root indexes and index-only e9f6560 recovery. Preserve
-  the actual rc.13 PROJECT_BUSY failure and require a real mid-analysis commit
-  retest from downloaded rc.14 before stable promotion.
-
-## [5.0.9-rc.13] - 2026-10-11
-
-- Give native Agent session creation and input setup a bounded 90-second proxy
-  budget. Ordinary API reads/abort/permissions retain 10 seconds; SSE streaming
-  and folder-picker contracts are unchanged.
-- Record safe proxy failure categories and deadlines without raw errors or URL
-  queries. Preserve MODULE_UNREACHABLE compatibility and add optional details.
-- Reproduce the installed OpenCode cold-start failure with a real delayed HTTP
-  server; verify the targeted budget, unrelated routes and credential redaction.
-- Retain rc.12 real acceptance and failed OpenCode results separately; require
-  a newly downloaded candidate and cold-client retest before user upgrade.
-
-## [5.0.9-rc.12] - 2026-10-11
-
-- Vendor engine 0.2.3-rc.6 to address the Windows transient state-file rename
-  failure exposed by installed rc.11 fault acceptance. Use exclusive random
-  temporary files and bounded sharing retries, retain durable state on failure,
-  and never unlink live state to force replacement.
-- Add deterministic sharing-violation, permanent-lock, foreign-temp and
-  disk-full regressions; all 39 engine tests pass on Node 22 and Node 24.
-- Preserve the rc.11 failure and require new installed acceptance before any
-  existing-user upgrade or stable latest promotion.
-
-## [5.0.9-rc.11] - 2026-10-11
-
-- Pin vector-hub 1.0.8-rc.3 and vendor engine 0.2.3-rc.5. Raw commit
-  conversations and internal backups are analysis evidence, not knowledge.
-  Exclude them from managed indexes, search and get; retain source evidence.
-- Prune stale derived records during sync and rebuild once on source-policy
-  upgrade, without repeating paid whole-index work on ordinary restarts.
-- Keep standalone vector-hub general document support unchanged.
-- Record both original failing regressions and successful Node 22/24 reruns;
-  installed, real model and existing-v5 gates still precede latest promotion.
-
-## [5.0.9-rc.10] - 2026-10-11
-
-- Preserve acknowledged brain/domain knowledge-directory bindings when changing
-  the knowledge root. Only derived indexes and future imports/shared scopes use
-  the new root; existing knowledge and confirmed memories stay in place.
-- Persist bindings in shell metadata, reject implicit rebinding, and report
-  shared-scope connection failures in UI and safe structured logs.
-- Add real-engine regression for confirmed memory, domain retrieval, restart,
-  new domain registration, offline retry and protected bindings on Node 22/24.
-- Candidate acceptance still precedes the existing-v5 upgrade and stable latest.
-
-## [5.0.9-rc.9] - 2026-10-11
-
-- Vendor engine 0.2.3-rc.4. Freeze existing target paths and hashes before
-  analysis so missing retrieval snippets cannot be mistaken for absent files.
-  Preserve full old documents when only partial context is available; illegal
-  create operations still fail safely instead of silently overwriting data.
-- Keep the failed rc.8 secondary-development maintenance evidence and verify
-  that exact commit after downloading this candidate, not patched source bytes.
-- Render long index-directory paths as a full-width block in System Status.
-- latest remains unchanged until the real closed-loop and existing-v5 gates pass.
-
-## [5.0.9-rc.8] - 2026-10-11
-
-- Pin Workbench 0.13.2-rc.4: atomically replace all native Agent histories,
-  retaining the previous complete record on disk-full or locked-file errors.
-  CLI fallback no longer deletes durable history before a failed rename.
-- Keep the original disk-full loss and separate successful background recovery
-  records. Candidate acceptance includes faults against downloaded module bytes.
-- Pin vector-hub 1.0.8-rc.2. Managed indexes now follow the knowledge root
-  under `.indexes`; show the acknowledged actual directory in System Status.
-  Rebuild on a root/model change while retaining the old index directory;
-  serialize queued writes and avoid a full paid rebuild on normal restart.
-- Do not promote latest until installed and existing-v5 acceptance pass.
-
-## [5.0.9-rc.7] - 2026-10-11
-
-- Vendor engine 0.2.3-rc.3: recognize promoted commit identifiers inside
-  natural-language tasks, including short and multiple SHAs. Raw conversation
-  mentions cannot replace the current promoted knowledge references.
-- Preserve failed installed Workbench/context checks from rc.6; latest remains
-  unchanged until fresh registry runtime and existing-v5 acceptance pass.
-
-## [5.0.9-rc.6] - 2026-10-11
-
-- Vendor engine 0.2.3-rc.2. Unassigned projects no longer share an implicit
-  domain in retrieval or Workbench task context. Explicit domains and whole
-  brain queries retain their inclusive scopes.
-- Keep the failed rc.5 real Workbench acceptance as a separate record. This
-  candidate must pass fresh downloaded-runtime checks before promotion.
-
-## [5.0.9-rc.5] - 2026-10-11
-
-- Pin Workbench 0.13.2-rc.3: retain OpenCode final answers across tool rounds,
-  isolate CLI history by product data directory and recover only owned v5
-  session records without moving or deleting originals.
-- Support current ZCode model-selection and failure events. Shared provider
-  credentials stay in memory behind a temporary loopback transport; client
-  shutdown cleans up the transport and settles rejected sends correctly.
-- Candidate publication does not promote latest or imply complete acceptance.
-  Real client execution, installed UI regression and existing-data upgrade
-  remain separately recorded gates.
-
-## [5.0.9-rc.4] - 2026-10-11
-
-- Pin Workbench 0.13.2-rc.2: resolve the installed ZCode provider resource
-  layout without rewriting desktop or personal configuration, and label
-  runtime failures with the correct agent.
-- Disable the project selector until the first registry snapshot. This fixes
-  the startup interaction race exposed by the Node 22 packaged UI regression.
-- Latest is unchanged; real installed acceptance and client authentication are
-  independent gates, not implied by a successful version probe.
-
-## [5.0.9-rc.3] - 2026-10-11
-
-- Route engine-wired CLI and MCP reads through the same authoritative engine
-  as the web UI, preserving commit references and project/domain/brain scopes.
-  Verify the selected runtime identity and reject unavailable, mismatched or
-  malformed services instead of silently querying a different local index.
-- Preserve the MCP `space` alias and expose explicit scope selection. Bound
-  responses and never forward profile or credential objects to read routes.
-- Drain isolated regression servers before Windows fixture cleanup. Candidate
-  publication still does not promote latest or claim complete acceptance.
-
-## [5.0.9-rc.2] - 2026-10-11
-
-- Pin the verified Workbench 0.13.2-rc.1 and vector-hub 1.0.8-rc.1 candidates.
-  Normalize known official MiniMax embedding route aliases at runtime without
-  changing existing v5 configuration bytes or private gateway paths.
-- Queue same-project index writes across embedding-model rebuilds. Failed writes
-  do not poison retries, and unrelated projects remain independent.
-- Restore safe relative Markdown document navigation and verify the standalone
-  retrieval UI at 1920x1080. Tests never kill unrelated port owners.
-- Isolate regression-suite ports and tolerate Windows fixture file release.
-  Candidate publication is not full product acceptance; latest remains unchanged.
-
-## [5.0.9-rc.1] - 2026-10-11
-
-- Retain classified, bounded and redacted index errors in run details, activity
-  and local log/v2 exports. Index-only retries preserve promoted files and do not
-  invoke the analysis model again.
-- Accept ZCode conversation evidence, isolate per-project journal failures and
-  report the exact blocked sequence without dropping evidence or flooding logs.
-- Synchronize production asset hashes, including the shared favicon. Permit
-  exact rc module pins but reject tags, ranges, links and source overrides.
-- Prepare a web-only candidate; latest is not promoted by this release. Existing
-  v5 data and third-party integrations are preserved; system acceptance remains
-  a separate gate from publication.
-
-## [5.0.8] - 2026-10-10
-
-- Repair managed Codex notify chains after scoped npm package or Bridge-home
-  changes. Preserve live desktop wrappers and valid third-party notifications;
-  unwrap only missing, known Codex runtime wrappers. Unknown notifiers remain
-  conflicts, never overwritten. Repeat setup is idempotent; removal unwraps
-  capture without deleting existing notifications or evidence.
-- Accept Codex's documented JSON argument as well as legacy stdin notifications.
-  Preserve the third-party payload and transport, adapt only the Bridge wake-up
-  aliases, and isolate child-process errors between notification branches.
-- Place new notify entries at the TOML root. Back up only the modified setting,
-  and refuse stale writes when another program changes the config after preflight.
-- Add real Git/Bridge/journal capture regression with isolated rollout fixtures,
-  duplicate-notify and missing-session checks; no model calls or user-data reset.
-
-## [5.0.7] - 2026-10-10
-
-- Normalize CLI command case and reject unknown commands, extra arguments and
-  invalid ports before any service or data is created.
-- Replace the shared temporary PID and blind orphan scan with per-data-directory
-  local control channels. Stop waits for the backend and owned module shutdown;
-  stale or unrelated PIDs are never signaled.
-- Add verified local instance inventory (`status --all`), explicit targeted stop
-  (`stop --port`) and data-directory selection (`--data`). Display the actual
-  port, version and data directory; refuse unverifiable or mismatched targets.
-
-## [5.0.6] - 2026-10-10
-
-- Do not mistake unregistered, abandoned legacy chat records for a knowledge
-  database on first launch. Preserve their bytes without restoring sessions or
-  importing projects; continue rejecting actual old registries and maintenance.
-- Wait for the background server's real HTTP readiness before reporting success
-  or opening the browser. Surface startup failures in the invoking terminal with
-  a nonzero exit code, and name the exact conflicting data entries.
-- Isolate regression data from the user's default data directory. Add a detached
-  npm CLI startup regression from an unrelated working directory.
-
-## [5.0.5] - 2026-10-10
-
-- Invalidate in-flight registry snapshots after domain creation, project import,
-  ownership changes and wiring repair. A read begun before the write cannot
-  hide the newly registered data until the next polling cycle.
-- Includes all knowledge-maintenance, Claude discovery and local diagnostic
-  fixes from 5.0.3, plus exact commit retrieval from 5.0.4. No schema migration
-  or automatic replay of failed model tasks.
-
-## [5.0.4] - 2026-10-10
-
-- Vendor engine 0.2.2: treat a known full/prefix commit SHA as a provenance
-  identifier. Unrelated high-scoring vector chunks cannot crowd out the matching
-  current knowledge. Preserve ordinary semantic ranking and current-file truth;
-  overwritten documents are not attributed to their earlier commit.
-- Includes the empty-knowledge and Claude discovery/diagnostic fixes from 5.0.3.
-
-## [5.0.3] - 2026-10-10
-
-- Pin Workbench 0.13.1: reject stale Claude executable overrides without hiding
-  a valid installation; discover native/PATH/package entries and probe JS/CJS/MJS
-  through Node. Discovery, doctor and runtime failures record their actual cause.
-- Vendor engine 0.2.1: an empty knowledge tree is a valid first-use state. Require
-  evidence-backed nonempty Markdown; normalize only missing titles in validated
-  staging output and reseal hashes before promotion. Empty/invalid output still
-  fails safely with a concrete reason, without overwriting existing knowledge.
-- Route module diagnostics through the supervisor to local product logs and the
-  diagnostic page; redact runtime credentials before IPC. Preserve failed runs
-  for explicit retry rather than silently replaying model tasks on upgrade.
-
-## [5.0.2] - 2026-10-10
-
-- Add ZCode capture to the unified CLI and web System Status connector setup,
-  including repair, status and safe removal. Keep unverified ZCode knowledge
-  plugin support explicitly separate from working capture hooks.
-- Correct command/config discovery labels and stop treating consumer
-  registration or disabled hooks as proof of capture. Report partial failures
-  with their operation ID and persist connector errors from both web and CLI.
-- Pin event-bridge 0.7.2: external hooks resolve their own Bridge home without
-  inherited environment variables; ZCode events preserve their source. Repair
-  older entries in place and retain third-party hooks within mixed groups.
-
-## [5.0.1] - 2026-10-10
-
-- Fix project import on Git versions before 2.31: query hooks and common Git
-  directories without `--path-format`, then resolve relative paths against the
-  Git command's working directory. Reject empty, multi-line or echoed-option
-  responses before writing hooks or repository identities.
-- Apply the same correction to commit capture and pin event-bridge console
-  0.7.1 so importing a project and capturing its next commit use consistent
-  workspace metadata. No previous data is migrated.
-- Keep import failure feedback visible with its error code and operation ID.
-  Persist the actual failed transaction stage; system diagnostics default to
-  all projects so failures of not-yet-registered imports are visible/exportable.
-
-## [5.0.0] - 2026-10-10
-
-- Fresh-data-only web product. Default state moves to `~/.project-knowledge-v5`;
-  no previous knowledge, project registry, unfinished claims or model profiles
-  are imported. Explicitly selected old data directories are rejected unchanged.
-- Remove shell migration/cutover and engine `/api/migrate`. Engine 0.2.0 uses
-  schema version 3; validate its data generation and runtime authentication
-  before registering projects or enabling consumption. System status reports
-  safe connection stage/error codes instead of silently waiting for takeover.
-- Isolate Workbench, bridge journals and vector configuration under the new
-  product directory. No automatic embedding-credential migration.
-- Queued, unstarted tasks are cancelled on unbind; active knowledge/index write
-  leases still block removal. The visible confirmation dialog updates project
-  selection immediately and rejects stale registry responses without reloading.
-- Reimport can explicitly replace a leftover knowledge-managed Git Hook. Failed
-  imports restore the prior Hook; third-party hooks and existing knowledge files
-  are never overwritten. Import tracks HEAD without analyzing historical commits.
-
-
-## [4.8.0] - 2026-10-10
-
-- Unified brain web product: six real pages for work, retrieval, memories,
-  domains/projects and activity. Existing module UIs retain their streaming
-  and recovery behavior with one embedded theme and project context.
-- Project import wires all four modules and preserves code/knowledge paths;
-  domains and confirmed memories add shared scopes without moving knowledge.
-- The vendored knowledge-engine owns journal consumption, isolated analysis,
-  staging promotion and retrieval. Per-project FIFO and two concurrent projects
-  persist across recovery; promotion and indexing have separate failure states.
-- Pin the validated module releases: Workbench 0.13.0, vector-hub 1.0.7,
-  event-bridge console 0.7.0 and vendored knowledge-engine 0.1.0.
-  Runtime file hashes are verified before release, not resolved from `latest`.
-- Web-only release. No new Windows installer or desktop auto-update delivery.
-  Regression uses controlled model/embedding doubles; real-model and real-index
-  smoke tests are reported separately. Remaining fault combinations are not
-  represented as a completed stability certification.
-
-## [4.7.0] - 2026-09-21
-
-- All four embedded module dependencies (`claude-ai-workbench`,
-  `@sanqianx/vector-hub`, `@sanqianx/ai-coding-event-bridge-console`, plus the
-  `@sanqianx/ai-coding-event-bridge` consumer inside the shell) are now pinned
-  to `latest`, so a fresh `npm install` pulls whatever the upstream modules
-  ship instead of lagging behind the project's npm releases. The legacy
-  `@sanqianx/ai-coding-event-bridge` package — which used to be the engine
-  library — is deprecated on npm and removed from this project's dependencies;
-  its capture engine now ships inside `@sanqianx/ai-coding-event-bridge-console`
-  at `src/bridge`, and the four production call sites (hook capture, requirement
-  recorder, integration installer, bridge adapter) plus six tests were rewritten
-  to import from there. The legacy package's every version carries a deprecation
-  message pointing to the new entry.
-- The package is published as `@sanqianx/project-knowledge` (moved from the
-  unscoped `project-knowledge`); the unscoped name still resolves to the old
-  versions for users who haven't migrated, but new installs land on the
-  scoped package.
-
-## [4.6.0] - 2026-09-21
-
-- The three embedded module runtimes now ship as real npm dependencies instead
-  of the `_modules/` committed snapshot: `claude-ai-workbench@0.4.1`,
-  `@sanqianx/vector-hub@1.0.1` and
-  `@sanqianx/ai-coding-event-bridge-console@0.4.0` (bringing
-  `@sanqianx/ai-coding-event-bridge@0.4.0` transitively). The `_modules/`
-  directory, its vendor script and the tarball entry are removed.
-- Why it matters: the snapshot lagged the source repo — npm consumers missed
-  the codex resident-bridge Windows fix (per-turn process respawns made every
-  turn 5-20s), claude failure-reason surfacing, `/system/doctor` diagnostics
-  and the folder-picker foreground/centering fixes. Module upgrades are now a
-  one-line version bump verified by `npm ci`, with no manual snapshot step.
-- `module-bridge` resolves each runtime entry through `require.resolve` from
-  the installed packages (same supervised spawn, reuse and env-override
-  contract as before); the `NODE_PATH` stitching is gone. `release:verify` no
-  longer gates on the vendored snapshot.
-
-## [4.5.0] - 2026-09-21
-
-- The embedded Agent Terminal's settings view gains a 会话捕获 (capture) tab.
-  Terminal conversations remain exempt from Development Conversation capture by
-  default — the terminal is internal AI runtime, not the developer's CLI
-  record. Flipping the toggle persists the scope, applies it to the runner
-  immediately, and marks only NEW terminal sessions with
-  `AI_CODING_EVENT_BRIDGE_CAPTURE=1`, so they reach the Bridge journal and the
-  会话浏览 explorer like any CLI tool conversation. Existing sessions are
-  unaffected, and nothing changes for standalone terminals that never touch
-  the toggle. Ships the updated vendored Agent Terminal runtime.
-
-## [4.4.0] - 2026-09-21
-
-- The ai-coding-event-bridge conversation console joins the Agent Terminal and
-  vector-hub as the third embedded Control Center module. The vendored runtime
-  (console 0.4.0 + its core) ships inside the npm package and starts with the
-  shell on 127.0.0.1:8790; the console's dependency on
-  `@sanqianx/ai-coding-event-bridge` resolves from a node_modules directory
-  vendored inside the console package, so the host package's own older npm
-  dependency can never shadow it. The global `node_modules/` gitignore rule
-  carries a negation for that directory, and `vendor:modules --check` now
-  verifies the runtime is present and git-tracked — the 4.3.0 "runtime missing
-  from the tarball" incident class cannot recur for this module either.
-- Module integration follows the established contract: same-origin reverse
-  proxy under `/api/eventbridge/`, supervised spawn with reuse of an already
-  listening standalone console (`KB_EVENTBRIDGE_COMMAND` override,
-  `AI_CODING_EVENT_BRIDGE_HOME` passthrough so the explorer shows the same
-  journal the in-process BridgeAdapter writes), idempotent import-time
-  registration keyed by Git work tree path using the recommended
-  `<knowledge>/dev-conversations` store, removal that keeps journals and sealed
-  commit documents, aggregated sidebar projection and module health.
-- The shell UI gains the 会话浏览 view (connect probe, 独立打开 link, per-project
-  桥 badge, import/remove copy) — and the conversation explorer itself learned
-  the `kb:sidebar` embed protocol (`?embed=1`, `kb:ready` announcement,
-  loopback-only origins), so its project sidebar folds via the ▤ control exactly
-  like the terminal and vector-hub embeds. Standalone console opens are
-  unaffected.
-
-## [4.3.1] - 2026-09-20
-
-- The 4.3.0 npm package could never start the vector-hub module: the global
-  `dist/` gitignore rule kept the vendored runtime
-  (`_modules/vectorhub/dist/`, 25 files) out of git, so the CI-packed tarball
-  shipped without `bin.js` and every fresh install logged "Vendored vector-hub
-  runtime not present". The runtime is committed now, and `release:verify`
-  gained a vendored-modules gate that fails when a required runtime file is
-  missing from disk or not tracked by git — the check that was previously
-  wired into nothing.
-- The Control Center's embedded Agent Terminal linked to no project on first
-  load. The shell sends `kb:select-project` 150ms after `kb:ready`, but the
-  terminal loads its project list only after health + agent discovery, so the
-  one-shot selection landed on an empty list: the linkage chip stuck on
-  "未登记" and the embed had no project selected, making it impossible to open
-  a terminal from the shell. The terminal now re-announces `kb:ready` once its
-  projects are loaded, and the shell retries a failed linkage with backoff.
-- Module status probing no longer reports a false "未检测到 … 界面将显示为
-  空白" forever: cold-starting module services are re-probed with backoff
-  (up to 10 attempts) before the status is marked down, and the top bar
-  终端/检索 indicator updates when a late probe succeeds.
-
-## [4.3.0] - 2026-09-20
-
-- The Agent Terminal and vector-hub module services now ship inside the npm
-  package and start with the shell. `npm install -g project-knowledge` is
-  self-contained: the Control Center, the embedded terminal (:5760) and the
-  vector-hub console (:8787) all come up together on a fresh machine — no
-  sibling checkouts required. KB_MODULES_AUTOSTART=0 disables spawning;
-  a service already listening on its URL is never double-spawned, so dev
-  machines running standalone module instances are unaffected. vector-hub
-  needs Node >= 22 — on older runtimes the shell skips spawning it and logs
-  a warning. Also fixed a latent spawn bug: quoted paths in module commands
-  reached node with the quotes embedded, which crash-looped the (previously
-  default-off) autostart path.
-
-## [4.2.20] - 2026-09-20
-
-- Startup no longer relocates legacy v4.1.x package-root runtime data. This is
-  a fresh knowledge-base system by decision: old data under the npm package
-  dir is ignored, not migrated. Previously a conflict between stale
-  package-root files and the user data dir aborted startup with "Legacy
-  runtime relocation failed", bricking machines whose install carried
-  v4.1.x-era files.
-
-## [4.2.19] - 2026-09-20
-
-- Background launch is diagnosable and no longer lies about the URL. The
-  detached child's stdout/stderr now append to `<data-dir>/launcher.log`, so
-  port fallbacks ("Port 5757 was busy, using 5758 instead") and startup
-  failures are visible after the fact instead of being silently discarded
-  (`stdio: 'ignore'`). The parent no longer prints a guessed URL or races a
-  browser open 1.2s after spawn — the child opens the browser itself once it
-  is actually listening, with the real port. The free-port probe timeout went
-  1s → 3s to survive slow listen() under antivirus/cold-cache load (a false
-  "busy" previously cascaded into "No free port found").
-
-## [4.2.18] - 2026-09-20
-
-- Module-embed shell (T0-T4 of docs/ui-integration-repair-plan.md): the Control
-  Center UI now embeds the Agent Terminal (:5760) and vector-hub console
-  (:8787) as original UIs behind a same-origin proxy (`/api/terminal/*`,
-  `/api/vectorhub/*`). Project import registers both modules (wiring table in
-  project state); the sidebar aggregates module registrations with per-project
-  session history (kb:list-sessions / kb:open-session) — session lifecycle
-  (archive/restore) belongs to the terminal module only, the shell keeps a
-  read-only projection; per-module sidebar fold toggles on the nav buttons
-  (kb:sidebar); project selection drives both modules (kb:select-project)
-  with live linkage chips.
-  Settings drawer is now 开发对话/日志 only (model settings live in the
-  terminal UI). Module removal keeps local knowledge and terminal session
-  records (upstream `removeProject?keepSessions=1`). Optional module service
-  supervision behind `KB_MODULES_AUTOSTART=1`. Old workbench chat panel and
-  its UI tests retired; full suite 115/115 green.
-
-## [4.2.17] - 2026-08-25
-
-- Recognized a healthy Project Knowledge fan-out nested inside Codex Desktop's
-  computer-use `--previous-notify` wrapper. Integration status and one-click
-  repair no longer report a false third-party conflict or attempt to wrap the
-  notifier again.
-
-## [4.2.16] - 2026-08-25
-
-- Upgraded AI Coding Event Bridge to 0.1.2 so one Codex task can be captured
-  across every continuation rollout. Existing single-file cursors migrate in
-  place without replaying consumed history, and real nested turn identities
-  keep commentary and final responses attached to the correct user prompt.
-- Added a joint Codex / Claude Code / OpenCode restart-and-recovery gate that
-  verifies durable capture, control-message exclusion, exact workspace
-  attribution, turn grouping, and duplicate-notification idempotency.
-- Kept Windows CDP failure tests fast and reliable by restricting detached
-  profile cleanup to actual Chrome/Edge launchers.
-
-## [4.2.15] - 2026-08-25
-
-- Repaired Codex Development Conversation turns without rewriting durable
-  history: generated control/context records are excluded, orphan assistant
-  responses are projected onto the correct user turn within the same session,
-  and stale open-turn facts no longer duplicate commit associations.
-- Removed synthetic empty-user and pending-assistant UI messages, with focused
-  projection, binding, API, and real-history regression coverage.
-
-## [4.2.14] - 2026-08-25
-
-- Recovered commit analysis safely after a server restart by making leftover
-  claims terminal without running startup analysis or replaying historical or
-  offline commits; the next explicit Git Hook event processes only its own SHA.
-- Restored v4.1.21 Claude Workbench history through the migrated project's
-  legacy slug while preserving the original session files and preferring the
-  canonical runtime record on duplicate session IDs.
-- Restored historical session selection and live Commit automation visibility
-  in the Claude Code UI, including prompts, thinking, tools, permissions,
-  retries, results, stop controls, and real browser regression coverage.
-
-## [4.2.13] - 2026-08-25
-
-- Restored the globally installed web UI by including `ui/i18n.js` in the
-  npm package; a missing or incomplete localization runtime can no longer
-  abort project, AI profile, Knowledge Root, and conversation loading.
-- Added package-asset closure validation plus a real
-  `npm pack -> npm install -> server -> Chrome` release smoke test, so CI
-  rejects tarballs whose UI differs from the tested source tree.
-
-## [4.2.9] - 2026-08-24
-
-- Made persisted UI-language selection fully reliable across page reloads and
-  slower browser runners: translated labels now re-render, and the document
-  language is initialized from the saved preference.
-
-## [4.2.8] - 2026-08-24
-
-- Fixed release metadata drift by deriving plugin manifests, marketplace
-  metadata, MCP package pins, and both lockfiles from the root package
-  version during the standard npm version lifecycle.
-- Added release and tag-version guards to CI and the npm publishing workflow,
-  so a mismatched release is rejected before packaging or publishing.
-- Stabilized Windows browser E2E tests by handling Chrome launcher detachment
-  and cleaning every test-specific Chrome process by its isolated profile.
-
-## [4.2.7] - 2026-08-24
-
-- Made runtime data migration fail closed: existing legacy data is now
-  discovered and relocated safely, migration failures prevent startup,
-  and settings-only legacy data is preserved instead of being treated as
-  a fresh install.
-- Restricted Knowledge Analysis to explicit post-commit hook events.
-  Startup, bridge synchronization, historical catch-up, and offline
-  commits no longer trigger implicit analysis.
-- Preserved conversation and commit-boundary facts across offline periods
-  without allowing those facts to contaminate a later online commit.
-
-## [4.2.6] - 2026-08-20
-
-- Hotfix for the v4.2.5 commit-reconciler regression: the post-batch
-  re-scan was unconditionally writing `analysis.status='idle'` even
-  after a successful batch of commits, clobbering the
-  `'state.advanced'` value that `advanceState()` had just set. The CI
-  suite surfaced this on a fast runner where the test's state read
-  raced ahead of the trailing re-scan; on slower machines the bug
-  passed silently. runSweep() now only marks `'idle'` when this call
-  never processed any commits, leaving a freshly-promoted batch's
-  `state.advanced` visible to the UI.
-- Bumped `plugins/project-knowledge/.mcp.json` from
-  `project-knowledge@4.2.4` to `project-knowledge@4.2.6` so the
-  integration-adapters MCP-pin assertion matches the package version.
-  (The v4.2.5 release had bumped everything except this MCP pin.)
-
-105 root + 2 desktop tests pass.
-
-## [4.2.5] - 2026-08-20
-
-- Restored the v4.1.22 → v4.2.x product contract that the recent
-  refactor had broken. The shared ` `resolveEffectiveAiProfile` resolver
-  is the single source of truth for "which AI profile does this project
-  use right now" (project → default → first-usable → AI_PROFILE_REQUIRED),
-  and the Workbench / analyzer / CommitReconciler paths no longer
-  carry their own drift-prone fallback logic.
-- Hook runtime: post-commit hooks now prefix `ELECTRON_RUN_AS_NODE=1`
-  in the shell body so a packaged Desktop install no longer tries to
-  launch the Electron GUI from a Git hook invocation. Node ignores the
-  env var so CLI installs are unaffected.
-- Hook migration state machine: a missing legacy hook is no longer
-  silently marked migrationVersion=2. The new state machine only
-  advances migrationVersion when install + readback verify succeed,
-  persists third-party conflicts in state.hook.lastConflict for UI
-  consumption, and reports status=verified|conflict|pending|failed per
-  project.
-- Hook health API: GET /api/projects/:id/hook-status and POST
-  /api/projects/:id/hook-repair expose every required field
-  (installed / managed / managedVersion / runtimeTarget / lastVerifiedAt
-  / conflict / reason / repairAvailable) and refuse to overwrite
-  third-party hooks.
-- Import UX: a new preflight endpoint and UI flow (Desktop folder
-  picker, knowledge language selector, AI profile selector, optional
-  Team Knowledge binding, per-problem action hints, auto preflight on
-  every input change, Import button stays disabled until ready=true)
-  replace the path-only dialog.
-- Knowledge language control: import + per-project PATCH both accept
-  knowledgeLanguage; changing it affects only future generated
-  knowledge, never historical files.
-- Project Goal editor: GET/PUT /api/projects/:id/goal atomically reads
-  and writes the project's GOAL.md, scoped to its own knowledgePath.
-- UI language toggle (zh-CN / en-US) persists in localStorage and
-  re-renders the most-visible labels without a reload.
-- P0 E2E gate proves the four primary contract scenarios (fresh /
-  legacy / offline / no-conversation) against the live server.
-  105 root + 2 desktop tests pass; the G00–G08 acceptance gates are
-  green.
-
-## [4.2.4] - 2026-08-20
-
-- Fixed the Bridge consumer never receiving connector wake-ups on a fresh
-  install: `BridgeConsumerService.start(notifyUrl)` ignored its argument, so
-  the `project-knowledge` consumer registered with empty metadata. Captured
-  conversations piled up in the journal until the next server restart instead
-  of appearing in Development Conversation immediately. The wake-up URL now
-  lands in the consumer registration and drains fire per event.
-- Projects imported before v4.2.3 (no canonical `repo-identity/v1`) are
-  upgraded automatically: on the first consumer drain after updating, each
-  legacy project with an existing repository path gets its workspace identity
-  resolved and persisted, its conversation baseline established, and its
-  external conversations start projecting — no manual migration step. A
-  workspace already claimed by another project is never written twice.
-- The Bridge dependency now resolves to `@sanqianx/ai-coding-event-bridge`
-  0.1.1, which accepts Claude Code's real `hook_event_name` payload field
-  (previously such payloads were misrouted to the Codex connector and lost as
-  capture gaps).
-
-## [4.2.3] - 2026-08-20
-
-- Development Conversation now captures real external Claude Code / Codex /
-  OpenCode sessions through `ai-coding-event-bridge` (repo-identity/v1 with
-  workspaceId as the primary key, durable Bridge-owned turn identity, and
-  exact-workspace project matching — unimported workspaces are skipped and
-  ACKed, never written into another project).
-- The internal Workbench capture path was removed: sending a Workbench chat
-  message no longer records requirements or conversation events (explicit
-  `record_requirement` still works), every internal SDK session runs with
-  `AI_CODING_EVENT_BRIDGE_CAPTURE=0`, and legacy embedded Workbench pairs are
-  excluded from presentation and new commit binding via
-  `conversation-exclusions/v1` without rewriting history.
-- Commit conversation snapshots are frozen strictly after
-  append-boundary -> drainThrough: no event beyond `boundaryEndCursor`
-  (assistant replies included) can enter a snapshot, and imports establish a
-  conversation baseline so pre-import history is never backfilled.
-- One Integration Setup action installs Knowledge Integration and Development
-  Capture for Claude Code / Codex / OpenCode with separate status reporting,
-  third-party config preservation, Codex notify conflict detection, and a
-  host-level Bridge consumer that survives per-connector uninstalls; new
-  `/api/bridge/*` and `/api/integrations/*` endpoints expose capture health.
-- Windows CI browser tests now fail fast: CDP launch failures clean up the
-  spawned browser tree and profile directory, report exit/stderr diagnostics,
-  use a dynamically discovered DevTools port with per-request timeouts, and
-  the test runner reports `TIMEOUT` explicitly under a configurable budget
-  (`PK_TEST_TIMEOUT_MS`, default 90s) instead of hanging for 300 seconds.
-
-## [4.2.2] - 2026-08-19
-
-- Rebuilt the web Control Center visual layer to faithfully match the v13
-  reference design (`project-knowledge-base-faithful-repair-ui-v10.html`):
-  sidebar brand block with the knowledge root path, multi-line project cards
-  (name, id, repository path, status dot) with right-click removal, the
-  in-main-area header, full-height Workbench panel with chat bubbles, tool
-  cards, circular send button and composer toolbar, the reference design
-  tokens including the complete dark theme, and responsive breakpoints.
-- Restored the reference Settings drawer structure with separate AI model
-  editing (add/save profiles via the API), project model assignment,
-  knowledge relations, mature development-conversation turn cards with
-  commit labels, and the V7 run-records list geometry.
-- Import view now shows the reference split layout with a post-import
-  placement preview; UI contract tests were updated to the reference DOM.
-
-## [4.2.1] - 2026-08-19
-
-- Fixed the legacy layout migration failing with `Legacy migration source
-  JSON is corrupt` (and the server refusing to start) when any legacy source
-  file carries a UTF-8 BOM, which PowerShell and several editors commonly
-  write. BOM-prefixed sources are now parsed normally.
-
-## [4.2.0] - 2026-08-19
-
-- Completed the v13 refactor: conversation capture contracts, atomic Bridge
-  commit boundaries with capture-gap persistence, and frozen
-  `CommitConversationSnapshot` binding so post-commit prompts never bind to
-  old commits and claims freeze their analyzer evidence.
-- Replaced `patch:null` handling for large diffs with an exact, hash-verified
-  patch chunk evidence bundle; commit analysis now carries frozen evidence,
-  retrieval, and conversation manifests end to end.
-- Unified user search and commit context selection into one
-  `KnowledgeRetrievalService` (hybrid recall, deterministic rerank, chunk
-  budgets) with an index source manifest and a Markdown delta overlay so
-  fresh knowledge is visible while the derived index is dirty.
-- Logging is now permanent per-day JSONL under `logs/system` and
-  `logs/projects/<id>` with reverse-chunk cursors, SSE streaming without
-  gaps, and immutable capture policy (no retention/capacity settings).
-- Background work registers per operation with one root `operationId`
-  propagated across HTTP, Hook, startup, and reconciler logs, and shutdown
-  drains every registered task.
-- Project delete is transactional with a journal, tombstone, and per-stage
-  recovery; migration no longer converts legacy logging knobs into settings.
-- Restored the full Control Center web shell (Workbench, Import, Settings)
-  with a mature run-records log view and a read-only development-conversation
-  page driven by frozen commit snapshots; logs left the standalone console.
-- Added non-release CI (Linux Node 18/24, Windows Node 24) for every push;
-  desktop packaging validation is paused for this web-only cycle.
-
-## [4.1.23] - 2026-08-17
-
-- Replaced legacy whole-registry project state with schema-v2 settings, a
-  minimal project registry, stable per-project config/state, and locked
-  requirement JSONL records under one canonical `StorageLayout`.
-- Reduced knowledge analysis to the managed post-commit Hook and startup
-  recovery, both routed through one per-project reconciler and one Commit
-  prompt. Import now establishes a baseline without AI or placeholder files.
-- Added manifest-validated AI staging, journaled Markdown promotion, crash
-  recovery, dirty-index semantics, and a single serialized `IndexService` for
-  incremental writes and atomic full rebuilds.
-- Rebuilt logging around six-level `log/v2` JSONL records, recursive secret
-  redaction, daily/size rotation, retention/capacity cleanup, cursor queries,
-  exports, health state, and orphaned-operation detection.
-- Replaced the duplicate dashboard implementations with one responsive
-  light/dark logging console, including operation flows, structured error
-  details, pause/refresh, filtering, cursor pagination, and export.
-- Unified server, CLI, MCP, Hook, index, and project path resolution. Read-only
-  knowledge search falls back to scoped Markdown when the derived index is
-  dirty or unavailable.
-- Added staged `layout-v2` migration with centralized recovery backups,
-  source/hash validation, last-step completion, fault recovery, and preservation
-  of legacy paths, secrets, Commit pointers, indexes, and logs.
-- Hardened the API by removing wildcard CORS, raw-file access, generic project
-  replacement, manual Hook/analysis routes, and stack responses; non-loopback
-  binding now requires authentication and Origin validation.
-- Removed CLAUDE.md management, the second automation queue, configurable log
-  and index locations, legacy direct database writers, and unused Vue/Tailwind
-  browser runtimes.
-- Added Windows end-to-end coverage for online Hook processing, offline Commit
-  catch-up, requirement binding, promotion/index completion, structured log
-  chains, paths with spaces/non-ASCII text, and immediate recovery of locks
-  owned by a crashed process.
-- Hardened release packaging by explicitly mirroring core runtime dependencies
-  into the Electron app and allowlisting shipped documentation, preventing
-  local review artifacts from leaking into npm or desktop packages.
-
-## [4.1.22] - 2026-07-27
-
-- Added a read-only `project-knowledge-mcp` server for OpenCode, Codex, and
-  other MCP-compatible coding agents. It resolves the current Git project,
-  exposes scoped search/ask/get/history tools, and includes the search-first
-  workflow without depending on `CLAUDE.md`.
-- Added a safe keyword-search fallback for legacy Markdown knowledge bases and
-  kept routine knowledge writes exclusively in post-commit automation.
-- Added a shared Claude Code/Codex plugin with MCP configuration and a
-  search-first `project-knowledge` Skill, plus a native Claude marketplace and
-  Codex marketplace manifest.
-- Added an OpenCode adapter and a universal, idempotent integration manager for
-  installation, updates, removal, dry-runs, and status checks across Claude
-  Code, OpenCode, and Codex.
-
-## [4.1.21] - 2026-07-27
-
-- Fixed Windows desktop Claude sessions failing with “executable exists but
-  failed to launch” when a global Claude Code installation exposes `cli.js`
-  but `node` is unavailable in the desktop process PATH. JavaScript Claude
-  entry points now use the application's own Node/Electron runtime.
-- Added slow-backend and packaged-client regressions covering desktop startup
-  readiness and Claude script launch through the installed Windows executable.
-
-## [4.1.20] - 2026-07-27
-
-- Rebuilt the daily workspace around Claude Code while keeping GitHub/Gitea
-  sign-in, project import, AI setup, project knowledge relations, knowledge
-  storage, embedding model setup, logs, and desktop updates available from one
-  settings drawer.
-- Moved project removal to a right-click project-list action with confirmation.
-  Projects with a running Claude session or queued automation task now cannot
-  be removed.
-
-## [4.1.19] - 2026-07-24
-
-- Stabilized release tests: expected EventSource cancellation during a page
-  reload is no longer treated as a failed request, and the Windows CLAUDE.md
-  refresh test no longer removes a just-used temporary server directory.
-
-## [4.1.18] - 2026-07-23
-
-- Simplified the daily workspace into a full-width Claude Code interface.
-- Removed the project-supervision summary, pending-commit and issue panels,
-  dashboard status/goal cards, and manual automation preview/simulation UI.
-
-## [4.1.17] - 2026-07-23
-
-- Rebuilt commit-triggered knowledge automation around persistent per-commit
-  state, ordered recovery, and exactly one AI task per Git commit.
-- Removed manual draft review, auto-apply controls, and scheduled-task paths;
-  generated knowledge is now written directly through the automated pipeline.
-
-## [4.1.16] - 2026-07-22
-
-- Fixed the desktop client timing out / hanging on the loading screen for large
-  knowledge bases (many projects, multi-GB DB) while the web dashboard opened
-  fine. Root cause was several compounding issues:
-  - Desktop startup probe was too impatient: raised the per-request `/api/state`
-    timeout (1.5 s → 30 s) and the overall backend-ready budgets (owned backend
-    45 s → 180 s, existing backend 20 s → 60 s) so it waits like a browser does.
-  - Server startup no longer blocks the event loop: the CLAUDE.md audit and
-    orphaned-run cleanup now run in `setImmediate`, so `/api/state` can answer
-    immediately after `listen()`.
-  - `getScheduleInfo()` (which spawns `schtasks`) is now cached for 5 s so
-    repeated `/api/state` hits don't each pay process-spawn latency.
-  - `isProcessAlive` now treats `EPERM` as alive, so a running backend's
-    runtime-endpoint file is no longer deleted (and a duplicate cold-started)
-    when the desktop and CLI run at different privilege levels on Windows.
-  - The dashboard now retries the initial connection every 3 s until it first
-    succeeds, instead of waiting up to 60 s, so a slow-starting backend no
-    longer leaves the page spinning.
-- Stopped the knowledge-storage relocation from logging
-  `target ... is not empty` on every startup: relocation is now skipped (with a
-  single warning) when the configured storage directory already holds unrelated
-  content, and the existing storage location keeps being used.
-
-## [4.1.15] - 2026-07-21
-
-- Replaced the dashboard's fixed-interval polling with a server-sent-events
-  (SSE) push model. The backend now emits a `state/changed` event over a new
-  `/api/state-stream` channel whenever project state is written or a job
-  starts/finishes, and the frontend refreshes only in response (300 ms
-  debounced). A slow 60 s poll remains as a fallback. Result: changes such as
-  the sidebar pending count and summary cards update within ~1-2 s of the
-  per-commit analysis cycle, while an idle dashboard makes almost no requests
-  instead of hitting the server every 5 s.
-
-## [4.1.14] - 2026-07-21
-
-- Fixed `runCommitAnalysis` to mutate the projects map entry directly
-  (`projects[slug].lastAnalyzedCommit`) instead of a copy. The v4.1.11
-  one-by-one refactor accepted `project = { slug, ...projects[slug],
-  kbPath }`, so `project.lastAnalyzedCommit = ...` mutated a throwaway
-  copy and the `writeProjects` callback persisted the unchanged value.
-  Symptom: the "待分析" count never decreases after analysis; the
-  pending count appears stuck or oscillates between runs because the
-  scan counter and the manual analysis counter diverge.
-
-## [4.1.13] - 2026-07-21
-
-- Bumped the GitHub OAuth HTTP timeout from 20s to 30s in
-  `requestJson` and `requestFormJson` (github-team-store.js) and made
-  the timeout error message actionable: it now names the host that
-  failed and suggests using a Personal Access Token as a workaround
-  when GitHub is unreachable from the host network.
-
-## [4.1.12] - 2026-07-21
-
-- Hid the "Runs / Drafts" sidebar nav entry. All registered projects
-  use `knowledgeMode: autoApply` so manual review is not part of the
-  workflow; the runs view itself is preserved for programmatic access
-  (uncomment the one line in `navItems` to restore the button).
-
-## [4.1.11] - 2026-07-21
-
-- Refactored `runCommitAnalysis` from batch to one-by-one: each pending
-  commit now gets its own run record and its own draft, analyzed in
-  chronological order (FIFO). After every successful single-commit
-  run, `project.lastAnalyzedCommit` advances to that commit's hash so
-  the "待分析" count drops by one per analysis, matching the UI label.
-  Previously a 10-commit batch produced 1 run + 10 drafts and the
-  pending count did not decrease until the user clicked Apply (which
-  could also advance the checkpoint past commits the user had not
-  actually applied drafts for).
-
-## [4.1.10] - 2026-07-21
-
-- Made startup-recovery's staging copy (`prepareKnowledgeWorkspace`) and
-  Claude-executable lookup (`findClaudeExecutableForSdk`) asynchronous
-  so the server's HTTP handlers (notably `/api/state`) stay responsive
-  while pending-commit workers are being dispatched. Previously the
-  desktop client would time out with `backend request timed out` when
-  several projects had pending commits on startup, because the sync
-  `fs.cpSync` and `spawnSync` calls were freezing the event loop for
-  several seconds during `/api/state` polling.
-
-## [4.1.9] - 2026-07-21
-
-- Refused to import or upsert a project whose `kbPath` resolves to the
-  same directory as `gitPath` / `localPath`. Writing the knowledge base
-  inside the source tree pollutes the repo (and trips Windows symlink
-  errors in the staging workspace), so both entry points now return 400.
-
-## [4.1.8] - 2026-07-21
-
-- Stopped the post-commit automation worker from being auto-paused when a
-  user aborts the Claude session in the terminal. Pending commits queued
-  before the abort are now picked up again on the next desktop launch.
-
-## [4.1.7] - 2026-07-19
-
-- Restored GitHub and Gitea login in the Windows desktop client through an
-  acknowledged Electron bridge that opens OAuth pages in the system browser.
-  Browser launch failures are now reported immediately instead of appearing
-  later as an authorization timeout.
-- Added support for HTTP authorization URLs used by private, intranet Gitea
-  installations while continuing to reject local-file and script URLs at the
-  desktop security boundary.
-- Forwarded Electron's Windows system-proxy resolution to the backend and made
-  GitHub OAuth plus remote repository discovery proxy-aware, while preserving
-  loopback and `NO_PROXY` bypasses for private Gitea installations.
-- Verified the full team-knowledge path after desktop migration: GitHub device
-  authorization, Gitea callback/token persistence, remote repository scanning,
-  manifest discovery, cached store loading, and team knowledge-base binding.
-- Added desktop bridge/security regressions and an end-to-end Gitea OAuth plus
-  remote team-store discovery test.
-
-## [4.1.6] - 2026-07-19
-
-- Replaced overlapping post-commit batches with a single per-project worker
-  that treats Git history as the durable queue. Live hooks and desktop startup
-  recovery now use the same path and process exactly one commit at a time,
-  strictly from oldest to newest.
-- Preserved the original knowledge contract: every analyzed commit must create
-  an independent `changes/` Markdown record containing its full hash, including
-  test-only, documentation, and infrastructure commits.
-- Added transactional knowledge workspaces. Claude edits a temporary KB copy;
-  validated Markdown is applied to the live KB only after the session succeeds,
-  and LanceDB must finish before `lastAnalyzedCommit` advances.
-- Added resumable vector finalization and automatic startup recovery. Index
-  failures retry without rerunning Claude or duplicating Markdown, while an
-  interrupted desktop session resumes from the oldest unfinished commit.
-- Made Stop pause the project's automatic commit worker, discard uncommitted
-  staged knowledge, and prevent queued work from immediately restarting. The
-  Claude Code UI now provides a dedicated Resume automatic analysis action.
-- Added end-to-end regressions for duplicate Hook delivery, strict commit order,
-  per-commit changes records, staged Markdown, abort behavior, and vector-index
-  recovery. The complete core suite now contains 53 passing test files.
-
-## [4.1.5] - 2026-07-19
-
-- Added a desktop-client card in Settings that always shows the installed
-  application version and provides Check for updates / Restart and install
-  actions without opening a browser.
-- Integrated Electron's Squirrel updater with a stable GitHub Latest Release
-  feed. Release automation now uploads `RELEASES` and the full `.nupkg` beside
-  the installer so future desktop versions download and install in-app.
-- Hardened the PowerShell folder-picker fallback: scripts are unpacked from the
-  desktop ASAR, stdout uses UTF-8, and only an existing absolute directory can
-  be returned. PowerShell banners and error text can never populate a path
-  field.
-- Added updater lifecycle and folder-output regressions, bringing the complete
-  core suite to 52 passing test files while retaining the native Electron
-  folder dialog introduced in 4.1.4.
-
-## [4.1.4] - 2026-07-19
-
-- Fixed Windows Claude Code discovery so extensionless npm shell shims and
-  command wrappers are never passed to the Agent SDK. The resolver now prefers
-  the installed package's native `bin/claude.exe` and retains `cli.js` only as
-  a legacy fallback.
-- Changed the post-commit Hook endpoint to acknowledge valid events with HTTP
-  202 before commit inspection and Claude startup. Dispatch continues in the
-  background, while the Hook accepts every successful 2xx response instead of
-  recording a false timeout/error after a run was already created.
-- Restored both desktop directory-selection flows with a narrow Electron IPC
-  bridge and native directory dialog. Project import and Team Knowledge folder
-  selection no longer execute a PowerShell script from inside the packaged
-  ASAR or mistake PowerShell banner text for a selected path.
-
-## [4.1.3] - 2026-07-19
-
-- Removed the bundled 218 MiB Claude Code executable from the Windows desktop
-  application. The small Agent SDK control layer now discovers the user's
-  installed Claude Code through `CLAUDE_CODE_EXECPATH`, `PATH`, or the npm
-  global installation and reports a clear error without disabling knowledge
-  management when Claude Code is unavailable.
-- Removed LanceDB's unused optional embedding-provider stack and its duplicate
-  Transformers.js 3.0.2 / ONNX Runtime 1.19.2 dependency tree while retaining
-  the application's supported Transformers.js 3.8.1 embedding service.
-- Pruned non-Windows and non-x64 ONNX binaries, browser-only ONNX Web runtimes,
-  Transformers browser builds, source maps, and unused platform packages from
-  the Windows x64 bundle.
-- Added packaged runtime probes for LanceDB, Transformers.js, ONNX Runtime, and
-  the real 512-dimensional `bge-small-zh-v1.5` embedding model, plus release
-  size auditing that rejects bundled Claude binaries, duplicate runtimes, and
-  oversized artifacts.
-- Reduced the measured Windows installer from 476.0 MiB to 221.1 MiB and the
-  installed application bundle from 1,491.1 MiB to 617.5 MiB without changing
-  existing model caches, Markdown sources, LanceDB data, or storage paths.
-
-## [4.1.2] - 2026-07-16
-
-- Made the desktop dependency lock reproducible under the newer npm version on
-  GitHub's Windows runner by explicitly locking the encoding compatibility
-  dependency used by the packaging tree.
-- Made the desktop release workflow fail immediately when a native npm command
-  fails instead of allowing PowerShell to continue into a misleading later
-  step.
-
-## [4.1.1] - 2026-07-16
-
-- Fixed Windows GitHub Actions browser discovery for the `windows-2025`
-  runner by detecting system-level Chrome and Microsoft Edge installations.
-- Restored the release gate for all three real-browser UI suites so the
-  Windows desktop installer can be built, smoke-tested, and uploaded only
-  after the complete 50-file core regression suite passes.
-
-## [4.1.0] - 2026-07-16
-
-- Added a Windows x64 Electron desktop application and Squirrel installer that
-  reuse the existing configuration, configured knowledge root, model cache,
-  Markdown sources, Team Knowledge Mode, and LanceDB without copying or
-  re-vectorizing user data.
-- Added single-instance and system-tray lifecycle behavior: closing the window
-  keeps Git hooks available, while an explicit tray Quit stops only a backend
-  owned by the desktop application.
-- Added an atomic, local-only runtime endpoint record so CLI, desktop, and
-  existing Git hooks discover the active fallback port dynamically and never
-  open competing LanceDB writers.
-- Made the desktop application safely attach to an already-running npm CLI and
-  leave that external backend alive when the desktop window exits.
-- Added hardened Electron navigation, context isolation, sandboxing, denied
-  permission requests, bounded desktop backend logs, Squirrel lifecycle event
-  handling, and explicit native-module ASAR unpacking.
-- Added Windows release automation that runs all 50 core regression files,
-  desktop unit tests, a packaged EXE/API smoke test, and packaged LanceDB native
-  read/write tests before uploading the installer and SHA-256 to GitHub
-  Releases. The existing npm provenance publish workflow remains unchanged.
-
-## [4.0.4] - 2026-07-16
-
-- Added a Settings-based Markdown knowledge health audit for every registered
-  personal and team knowledge base, with per-project file size, deterministic
-  fixes, structural errors, and semantic-review findings.
-- Added safe batch or per-project optimization with retained backups under the
-  configured knowledge root, atomic writes, formatting repair, compact index
-  regeneration, and automatic LanceDB refresh for migrated projects.
-- Reduced generated change indexes to the latest 40 entries with bounded cells
-  and generated-file markers; a real 125 KiB legacy index now rebuilds to about
-  9 KiB without deleting its source change documents.
-- Prevented background AI automation from writing `00-index.md`, injected
-  mandatory in-place-update hygiene rules even for custom prompts, and rebuilt
-  compatibility indexes after every successful automatic update.
-- Preserved Team Knowledge Mode by de-duplicating shared physical KB paths
-  during maintenance while refreshing every aliased vector space.
-- Expanded the complete regression suite to 48 passing test files.
-
-## [4.0.3] - 2026-07-15
-
-- Added an embedding-model setup panel with one-click download and real
-  inference verification, configurable Hugging Face endpoint/mirror, local
-  model directory, offline mode, and copyable PowerShell environment-variable
-  examples.
-- Stop a migration batch before changing any project when the shared embedding
-  model cannot load, instead of reporting the same model failure once for every
-  project.
-- Surface migration failures through a prominent settings panel, a one-time
-  browser alert, detailed per-project errors, and direct retry actions.
-- Persist model settings and verified download state while rejecting partial
-  cache files or verification results from a different model configuration.
-- Expanded the complete regression suite to 46 passing test files.
-
-## [4.0.2] - 2026-07-15
-
-- Store LanceDB, its maintenance metadata, and retained database backups under
-  `<configured knowledge root>/.project-knowledge/` instead of a fixed app-data
-  location.
-- Automatically relocate v4.0.0/v4.0.1 databases on startup to the knowledge
-  root already selected in Settings.
-- Safely relocate and verify the database before saving a changed knowledge
-  root, including cross-volume copies, target-conflict protection, and retained
-  backup path rebasing.
-- Display the effective vector database path directly below the Knowledge Store
-  root setting.
-- Expanded the complete regression suite to 45 passing test files, including
-  real LanceDB startup and live root-change relocation tests.
-
-## [4.0.1] - 2026-07-15
-
-- Excluded generated `00-index.md` files from embedding and removed stale
-  derived-index rows while retaining the Markdown files for legacy and Team
-  Knowledge Mode compatibility.
-- Replaced unbounded Tags and change lists in generated Markdown indexes with
-  compact, capped navigation metadata; all module links remain available.
-- Stopped running LanceDB compaction after every commit. Maintenance now runs
-  only after bounded mutation thresholds and never rewrites the oversized
-  v4.0.0 full-text index in place.
-- Added a compact full-text schema that indexes `chunk_text` once with Chinese
-  bigrams instead of storing and indexing a second full document copy.
-- Added one-click, atomic database rebuilding with disk statistics, vector and
-  keyword verification, optional retained backup, and manual rollback.
-- Added upgrade and regression coverage for old v4.0.0 databases; the complete
-  suite now contains 43 passing test files.
-
-## [4.0.0] - 2026-07-15
-
-- Added a local LanceDB knowledge store containing original text, metadata,
-  and 512-dimensional `Xenova/bge-small-zh-v1.5` embeddings.
-- Added incremental Markdown indexing with heading-aware chunks, content-hash
-  skips, stale-chunk deletion, Chinese n-gram/BM25 search, and hybrid RRF.
-- Added scoped `search`, `get`, `ask`, and `history` HTTP/CLI tools that return
-  readable source text without exposing raw vectors.
-- Added explicit, bidirectional, non-transitive related-project search scopes;
-  writes remain restricted to each project's primary space.
-- Preserved Team Knowledge Mode: Git remains the portable Markdown transport,
-  stable team `kbId` values map to local spaces, and vectors stay local.
-- Added resumable one-click migration for all legacy KBs with backups,
-  verification probes, per-project atomic cutover, retry, and rollback.
-- Centralized Claude Code instructions in one user-relative file and added a
-  safe bulk refresh for all managed `CLAUDE.md` blocks independent of Hooks.
-- Enforced development-time read-only KB access; post-commit automation owns
-  routine updates and re-indexes migrated projects after successful runs.
-- Added configurable mirrored/offline embedding model locations through
-  `KB_EMBEDDING_REMOTE_HOST` and `KB_EMBEDDING_LOCAL_PATH`.
-- Expanded the regression suite to 41 test files.
-
-## [2.4.7] - 2026-07-01
-
-- Initial public Apache-2.0 release of the local-first `project-knowledge`
-  core.
-- Includes the CLI, local dashboard, Git scanner, project registry,
-  knowledge-base workflow, AI profile support, prompt registry, templates,
-  and regression tests.
-- Keeps runtime data, local project knowledge bases, API keys, generated AI
-  drafts, logs, and local maintenance notes outside the public repository.
+# Changelog
+
+## [5.0.25] - 2026-10-24
+
+- 纳入 Workbench 0.13.8：大体积历史切换采用严格单趟扫描与恢复索引旁路，损坏历史如实上报 HISTORY_CORRUPT；会话打开失败区分服务端判定、超时与网络错误并保留选中。
+- 模型管理新增当前 Key 的余额、订阅窗口与刷新时间展示；提供同源只读接口。
+- 商标悬停或键盘聚焦时显示运行中的知识库版本。
+- 合入引擎重试与晚到提交顺序保护、Hook 拒绝的故障隔离、Windows 事件采集路径修复。
+- 补齐桌面依赖、模块来源及发布验证；本次桌面沿用 npm 轻量壳，安装替换以正式制品验收和备份为前提。
+
+## [5.0.24] - 2026-10-20
+
+- 纳入 Workbench 0.13.6：Claude Code／GLM 连续对话保留每轮输入与回答的关联，上一轮迟到的结果和完整回答不再重复出现在新问题后面，也不会把新一轮误标为空闲；主、子对话同时修复。
+- 历史重新加载按事件顺序合并，已合并的流式正文替换原片段；主对话与子代理、不同原生会话的正文不会合并到同一历史行。
+- 修复较早任务列表响应覆盖新任务状态、将已完成子代理误判为失效的问题；经过服务端验证的任务历史可恢复显示，读取失败提供只读重试，不重新派发任务。
+- 保留 5.0.23 的明确选择提交及 5.0.22 的后台启动修复。
+
+## [5.0.23] - 2026-10-20
+
+- 接入 Workbench 0.13.5：Claude Code（含 GLM 等自定义模型）的用户提问显示单选、多选及自定义答案，明确提交后才继续；绕过普通工具授权时也不能代选或空答案放行。Codex 原生用户提问同步接入选择界面。
+- 主对话与子代理分别收集并回传答案，普通答案草稿刷新可恢复；关闭标签、收起面板和刷新不会提交答案，秘密输入不写入浏览器持久存储。
+- 子代理选择卡片并入对话滚动区域，发送长消息后输入框恢复高度，确保选项、提交和停止按钮可以使用。保留并排工作区、按需知识引用及 5.0.22 的后台启动修复。
+
+## [5.0.22] - 2026-10-20
+
+- 将启动时的 Agent 接入检查和自动修复放入独立进程，避免慢客户端命令阻塞后台就绪通知、网页和实例控制；超时会记录可重试状态，停止服务同时清理自有检查进程。
+- 仅检查用户已连接的客户端；未能核实的状态不再当成未安装并触发自动修复。保留并排工作区、已有接入选择及知识数据。
+- 补充真实后台 CLI 在原生命令阻塞时的就绪、HTTP、状态查询和停止回归，以及超时、崩溃、手动修复结果优先级验证。
+
+## [5.0.21] - 2026-10-20
+
+- 开始工作采用并排工作区：从加号新建子代理或按需添加知识引用，使用统一的对话、工具步骤、文件修改和输入框；移除面向客户的任务标识、派发阶段及诊断信息。
+- 子代理首次发送时继承当前项目、Agent、模型及主对话的冻结上下文，之后独立交流；原生任务可从主对话打开，并基于结果继续创建子对话。
+- 知识引用仅保留会话内单一标签，检索不自动展开、不抢焦点；修复刷新时内容提前投递的问题，并跟随产品深浅主题。
+- 主子草稿、标签和阅读位置可恢复；关闭标签不停止任务，窄窗口自动使用抽屉。精确锁定 Workbench 0.13.3。
+
+- 默认维护模型明确报告额度不足时，保留并展示可重试的失败原因；请求限流、认证失败分别展示。同项目后续提交等待前一条恢复，失败及排队记录跨重启保留。
+- 事件消费在分析任务建单前受阻时，活动页与系统状态展示故障，不再只显示空记录；故障及恢复通过现有实时连接更新，稳定状态不触发额外数据轮询。
+
+## [5.0.19] - 2026-10-16
+
+- 工作台展示来自真实 SDK 事件的后台任务、工具活动及完成结果；主回复结束后可继续输入，后台任务继续归属原会话，切换和刷新可恢复。
+- 移除外壳列表定时轮询，复用实时连接推送摘要变化并增量更新侧栏；草稿按服务入口、项目和会话隔离，保护发送失败、迟到创建和附件读取的原始归属。
+- 修复新会话创建时的旧会话误用、追加消息 409、汇总回复重复展示，以及重启读取项目列表失败导致的重复登记与接线缓存问题。
+- 候选内容清单按 npm 实际安装行为规范 CLI shebang 的换行符，避免 Windows 工作树打包后产生错误的内容哈希告警；普通文件与脚本正文继续逐字节校验。
+- 自动长期记忆仅接受具有冻结人类原句与来源证据的跨项目长期规范；代理派单、单次修复和项目版本事实保留为项目知识。旧候选不自动改写。
+- 纳入 Workbench 0.13.2-rc.21 和经过真实向量回归的引擎修复；固定 30 条检索查询全部通过，109 处引用一致，真实修复提交完成分析、晋升、索引并由下一轮 GLM 实际采用。验收范围见 5.0.19 记录。
+
+## [5.0.17] - 2026-10-15
+
+- 修复网页首屏被慢模块拖住、开始工作 Agent 探测同步阻塞与返回页面状态不刷新的问题；独立读取、明确失败状态和缓存保留，iframe 不重建。
+- 同源实时连接合并为每标签一条 WebSocket，保留独立模块 SSE，恢复时先读取快照并按游标续接；项目接线失败不再阻塞其他项目。
+- 工作台输出时向上滚轮立即暂停跟随，完整历史独立持久化并向前分页；旧 Claude 历史只读恢复，不再把有限事件缓存冒充全会话。
+- 精确锁定 Workbench 0.13.2-rc.17 与引擎 0.2.7-rc.10；补齐诊断和任务启动的异步发现旁路，保留 ZCode 注册表安装路径，并阻止已中止或被替换的初始化迟到启动任务。最终候选已通过独立下载安装、真实 MiniMax 与向量闭环，正式包先发 next 验收后再晋升 latest，详见 5.0.17 验收记录。
+
+## [5.0.16] - 2026-10-15
+
+- 收敛 rc.8 已通过的 30 项核心门禁：修复真实主题证据遗漏、短向量子片段重复占位与小型明确提交上下文不完整，保留当前文件校验、记忆优先级及上下文预算。
+- 精确锁定 Workbench 0.13.2-rc.13、vector-hub 1.0.8-rc.8、event-bridge 0.7.2 和 vendored engine 0.2.7-rc.9；保留原生流式会话、项目双目录和 D 盘索引。
+- Codex 本地市场与 MCP 接线使用真实原生命令验证；四客户端实际接入，Claude/Codex/ZCode 首次开发请求的事实与产品记忆送达均实测通过，OpenCode 保留真实工具检索兼容，不冒称自动注入。
+- 正式包先发布 next 并重新下载安装验收，完成后才更新 latest；候选通过不等于正式包已验收。首次失败、升级保护与完整证据见版本验收记录。
+
+## [5.0.16-rc.8] - 2026-10-15
+
+- 纳入引擎 0.2.7-rc.9：经验证的短向量子范围与包含它的 Markdown 段落不再重复占用片段名额；保留较大独立向量范围与原有校验，不扩散向量分数。
+- 用户明确引用已识别提交时，在原有上下文预算内完整送达当前项目的小型条目，先复验文档哈希与归属，避免摘要和交叉引用挤掉真实证据。大条目、普通主题和跨项目参考不展开；确认记忆优先级不变。
+- 保留 rc.7 真实网页失败及源码红测试。rc.7 双 Node 外壳 139 项、安装版八项、真实记忆与并发十项通过，不抵消真实任务正文遗漏；rc.8 必须重新验 npm 安装版，未完成前不更新 latest。
+
+## [5.0.16-rc.7] - 2026-10-15
+
+- 纳入引擎 0.2.7-rc.8：修复 commit 标识与主题混合查询只返回概述、漏掉已有正文证据的问题。已识别 SHA 只限定来源，不重复参与主题打分；提交内按主题 BM25 选择片段，普通语义查询保留 RRF，并补中英单写者与引擎词汇。
+- 保持每文档两段、正文／向量区间校验、授权范围与上下文预算；同分排序稳定，不以标题中的 SHA 冒充任务事实已送达。
+- 新增真实原文遗漏的红／绿用例及 topic-context 发布门禁；真实工作台网页测试验证具体事实和实际回答，其他模块版本不变。候选未完成安装版与原生验收前不更新 latest。
+
+## [5.0.16-rc.6] - 2026-10-15
+
+- 修复 Codex 0.137 将 scoped npm 目录中的 `@` 误作 Git ref：本地目录含 `@` 时从该目录执行 `marketplace add .`，更换和回退分别使用自己的工作目录，不复制目录、不访问 GitHub。
+- CLI 子操作日志关联原操作 ID，允许继续处理的失败仍记录安全诊断；本地路径误判保留 `CODEX_LOCAL_SOURCE_REF`，不写入原始 CLI 输出。
+- 保留 rc.5 的已验证四模块版本、索引与知识功能。rc.5 本机保真升级已通过：7 项目、43 历史文件、565 索引文件保留；真实记忆闭环与双 Node 浏览器门禁通过，热查询 P95 为 693ms（1000 文档合成语料）。
+- 本机 Codex 已通过原生本地市场操作恢复，MCP 声明入口 initialize/六工具/resolve/search/get 验证通过；Hook 尚待用户审核，不计自动上下文通过。候选不更新 latest。
+
+## [5.0.16-rc.5] - 2026-10-15
+
+- 精确锁定 Workbench 0.13.2-rc.13：模型目录加载失败不再伪装成空目录；保留上次有效结果，提供明确错误与重试，打开菜单可恢复请求，关闭菜单后的迟到响应不会重开菜单。保存或删除供应商后不复用保存前的旧快照，普通菜单打开仍共享请求。
+- 保留 rc.4 的引擎与向量检索修复；模型菜单错误不显示原始异常或凭据。补充连续保存的竞态回归及测试服务器模块隔离，安装版故障恢复仍须单独验收。
+- Leader 负责关键开发和发布收尾，GLM 并行独立做故障测试与审查。候选发布到 next，不改变 latest。
+
+## [5.0.16-rc.4] - 2026-10-14
+
+- 引擎 0.2.7-rc.7 复用同一次检索刚读取的知识正文，保留晚到响应后的文件与范围校验；vector-hub 1.0.8-rc.8 将有界查询缓存延长到 10 分钟，并防止旧请求失败删除新缓存。源码真实向量热查询 P95 为 1069ms，不代表安装版验收通过。
+- 锁定 vector-hub 1.0.8-rc.8（ebf6f586a534c4b851c919c53a5536671cc8bd12），保留 Workbench 0.13.2-rc.11 与 bridge 0.7.2；四个模块集与官方 npm tarball 双向逐字节核对（vendored 引擎 40、Workbench 96、vector-hub 59、bridge 38）。
+- 候选不改变 latest；完整安装版和原生客户端验收通过后才允许晋升。
+
+## [5.0.15] - 2026-10-13
+
+- 冻结 rc.5 已验收的中文 BM25 与可校验向量片段融合、本项目进度与跨范围上下文、无可靠匹配诊断及离线 Claude 插件接入。
+- 包含 UI 属性/源码示例被知识晋升误判的修复，以及 Codex 本地市场重载修复；保留真实 Git 失败和原生 Hook 信任审核。
+- 正式包先下载和测试，再更新 latest；本机与另一台设备的结果分别报告。
+
+## [5.0.15-rc.5] - 2026-10-13
+
+- 修复 Codex 本地插件市场被强制执行 Git 刷新而无法更新的问题；识别原生明确的非 Git 市场反馈后重新加载本地插件，真实 Git/认证失败仍终止，保留可用旧插件。
+- 不写入 Hook 信任、不跳过原生审核；保留 rc.4 的知识晋升误判修复与检索改进。
+- 候选包先下载验收，不提升 latest。
+
+## [5.0.15-rc.4] - 2026-10-13
+
+- 修复知识晋升校验把 UI 的 `placeholder=` 属性及有正文解释的源码示例误判为未完成内容的问题；真实未完成标记、空正文、越界及哈希校验继续拦截。
+- 保留失败记录及项目 FIFO，不升级后自动重做失败提交；用户可通过运行详情显式重试。
+- 保留 rc.3 的检索与离线接线修复；重新进行 npm 安装版验收，暂不提升 latest。
+
+## [5.0.15-rc.3] - 2026-10-13
+
+- 当前项目没有维护历史但已有其他知识时，最近进度检索明确显示“本次无可靠匹配”，不误报整个项目知识为空。
+- 增加用户原句“我想知道我当前向量库存储的位置在哪”的回归与现有知识文件只读验收；存储规则引用与当前设备实际路径分别核验。
+- 保留 rc.2 的片段校验、中文排序、跨项目范围隔离和离线插件接入；候选版不更新 latest。
+
+## [5.0.15-rc.2] - 2026-10-13
+
+- Windows 路径边界校验改用原生规范化，保留中文/空格、缺失末级目录和目录联接越界检查。
+- 保留 rc.1 的全部检索及离线接线修复；性能验收明确区分真实向量召回与最终选择词法结果，不把词法结果标为向量命中。
+- 候选版仍不提升 latest，源码、npm 安装版、真实客户端及另一台设备的反馈分别记录。
+
+## [5.0.15-rc.1] - 2026-10-13
+
+- 候选版检索采用当前 Markdown 分块上的中文 BM25 与可验证向量片段融合；每篇最多两个互补片段，不将文档向量分数扩散到全部章节。
+- 最近进度优先本项目提交依据；开发上下文本项目事实保留预算、领域参考限量，记忆仍独立准备。无可靠匹配不再表示知识库为空。
+- Claude 插件默认从 npm 包内本地市场安装，无需 GitHub；接入状态探测放入有界后台进程，超时与未安装分开显示并记录日志。
+- 保留 v5 数据、第三方 Hook 与客户端配置。候选发布不更新正式 latest；真实模型、客户端及安装版验收分别记录。
+
+## [5.0.14] - 2026-10-13
+
+- 长期记忆页增加“添加记忆”：填写内容并选择全局、领域或项目，用户确认后直接保存为已确认记忆。
+- 手动来源独立标明，不伪造 Agent 分析、项目来源或 commit；复用引擎受控写入、持久化、共享索引和开发前上下文。
+- 校验无效目标、重复内容和越界路径；取消不保存，保存与编辑自动刷新，写入失败保留记录并报告待恢复。
+- 保留 5.0.13 的重复采集修复：已有旧 Hook 仍需在 Agent 接入中执行“接入/修复”，升级启动不等于已清理旧接线。
+
+## [5.0.13] - 2026-10-13
+
+- 修复独立 bridge 与产品 bridge 同时接入导致同一项目生成两个知识目录：接入/修复只替换明确的旧自有入口，并去除当前入口的重复项。
+- Claude Code、ZCode 重复采集不再显示为正常；保留混合分组中的第三方 Hook、知识读取、模型配置和历史日志。
+- Codex 通知链识别默认旧 bridge，包括已知通知分支中嵌套的旧入口；保留第三方通知，不恢复退役的重复采集。
+
+## [5.0.12] - 2026-10-12
+
+- 冻结 rc.3 已验收的三客户端开发前知识读取：Claude Code、Codex、ZCode 获取统一事实及适用的已确认长期记忆。
+- 当前编辑记忆优先于历史实现；指定已知源文件不会被重复历史向量结果挤出，保留引用与范围校验。
+- ZCode 入口跨 npm 安装目录更新并去重，Codex 遵循原生 Hook 信任审核，保留用户模型、第三方 Hook 和既有采集。
+- 正式包先下载验收，再更新 latest；OpenCode 保持 MCP/采集兼容，不宣称未经验证的自动注入。
+
+## [5.0.12-rc.3] - 2026-10-12
+
+- 修复点名源文件的需求被相似历史向量结果挤出的情况；已有文件标识先匹配知识依据，新文件名仍可检索相似背景。
+- 修复 ZCode 安装目录变化时残留旧上下文入口；只替换明确属于本产品的脚本，去除重复入口，保留第三方 Hook、采集及模型。
+- 保留真实模型首轮失败和再次验证记录，未完成 Codex 原生信任审核前不提升 latest。
+
+## [5.0.12-rc.2] - 2026-10-12
+
+- 修正自动上下文的当前记忆与历史事实边界，保留编辑时间及项目、领域、全局的适用优先级。
+- 新增保留旧代码及历史知识的真实编辑记忆验收；候选版仍需真实模型验证，不以注入成功代替实际采用。
+- 保持用户当前需求及安全限制优先，不因记忆增加授权；不改动用户知识文件、路径和第三方设置。
+
+## [5.0.12-rc.1] - 2026-10-12
+
+- 候选版新增 Codex、ZCode 原生需求提交前的知识读取，复用引擎同一份事实、记忆与引用。
+- 修复 Codex 旧插件存在时跳过升级；区分安装、定义完整、启用和原生信任状态，未信任不显示已就绪。
+- ZCode 使用用户配置的 process Hook，保留第三方 Hook、模型及既有采集；被禁用的 Hook 不静默重启。
+- 工作台提供上下文和原始会话采集时，仅抑制自身原生重复入口；候选实测未通过前不替换 latest。
+
+## [5.0.11] - 2026-10-12
+
+- Claude Code 的普通开发需求通过原生前置 Hook 自动取得项目/领域知识和适用的已确认长期记忆；MCP 保留用于深入读取。
+- 引擎统一准备有界上下文，工作台复用同一接口并避免重复注入；知识读取失败明确提示后继续，向量故障不冒充成功。
+- 修复旧知识插件存在即跳过更新及 Windows Git CRLF/npm LF 接线误判；插件版本、启用、文件与采集能力分别验证。
+- 保留 v5 数据、模型、第三方 Hook 和向量位置；原生模型请求与开发提交闭环分开验收。首批只承诺 Claude Code 自动读取，Codex/ZCode 后续交付。
+
+## [5.0.10] - 2026-10-12
+
+- Freeze the accepted rc.2 scoped-memory runtime: review and edit candidates,
+  choose global/domain/project applicability, and confirm in one transaction.
+- Preserve v5 knowledge, source evidence, histories and directory bindings;
+  only applicable, confirmed and synchronized memories enter task retrieval.
+- Include real MiniMax candidate extraction, 1536-dimensional embeddings,
+  project/domain/global isolation, CLI/MCP agreement and restart regression.
+- Publish first to next; formal npm installation and existing-user upgrade
+  acceptance precede stable latest promotion. Other module versions stay pinned.
+
+## [5.0.10-rc.2] - 2026-10-12
+
+- Require candidate entry IDs to name an actually promoted Markdown document;
+  record rejected candidate counts and safe reason codes instead of silently
+  losing model suggestions or inventing source citations.
+- Keep valid knowledge updates independent of optional candidate failures;
+  roll back an unpersisted candidate batch and expose diagnostics in activity.
+- Distinguish changed-memory index jobs from unchanged health probes so ready
+  memories do not remain falsely marked as indexing.
+
+## [5.0.10-rc.1] - 2026-10-12
+
+- Add global/domain/project memory applicability and a single review form for
+  editing content, selecting any existing target and confirming atomically.
+- Retain original evidence, scope-aware duplicate suppression, existing v5
+  global memories and all fixed code/knowledge directory bindings.
+- Filter exclusive memories before retrieval ranking and file reads; clear old
+  projections after scope changes and report/retry pending file synchronization.
+- Keep the accepted Workbench, vector-hub and event-bridge versions unchanged;
+  candidate package acceptance precedes stable latest promotion.
+
+## [5.0.9] - 2026-10-12
+
+- Freeze the accepted rc.19 web runtime and exact module versions; no new
+  product features, old-version migration or existing v5 data reset.
+- Include index failure diagnostics, index-only recovery, knowledge-root
+  index storage, safe client discovery/notify chaining, maintenance isolation
+  and the embedded Workbench startup/streaming fixes from the candidate series.
+- Publish initially to next. Official npm installation, Node 22/24 regression,
+  real MiniMax knowledge/embedding loops and existing-user upgrade acceptance
+  must pass before stable latest promotion.
+
+## [5.0.9-rc.19] - 2026-10-12
+
+- Pin Workbench 0.13.2-rc.7. Startup shows a loading state and keeps controls
+  inert until projects and event handlers are ready; the existing streaming
+  and session recovery paths are unchanged.
+- Preserve the real rc.18 premature-click failure and require its actual
+  usable-page ready receipt in browser acceptance. Add a deterministic
+  startup gate regression and Node 22/24 retests.
+- Include Windows test-server cleanup hardening and corrected storage-root
+  assertions. Retain existing v5 data, third-party configuration and the
+  official-package acceptance gate before stable latest promotion.
+
+## [5.0.9-rc.18] - 2026-10-12
+
+- Project resolution for MCP and CLI now reports the vector-hub storage root
+  selected in settings, not the retired shell database path. A running tool
+  rereads root changes; resolution does not move files or rebuild indexes.
+- Preserve the reproduced rc.17 diagnostic mismatch and add Node 22/24
+  coverage for configured, changed and unset roots and MCP output.
+- Retain existing v5 data and all accepted module pins. Publish to next;
+  stable latest remains behind official-package and actual-user acceptance.
+
+## [5.0.9-rc.17] - 2026-10-11
+
+- Pin Workbench 0.13.2-rc.6. Repeated embedded notifications for the current
+  project no longer abandon a newly created session; real project changes and
+  late project lookups still preserve isolation and per-project drafts.
+- Keep a healthy progress stream labelled as connected on redundant online
+  events. Verify actual offline cleanup, snapshot recovery and exactly one
+  restored SSE connection with a real browser and HTTP proxy.
+- Preserve the rc.16 real Workbench failure and require official-package
+  streaming/citation acceptance before upgrading the actual user instance.
+- No data reset or third-party configuration overwrite. Stable latest remains
+  behind the full candidate and installed-product acceptance gates.
+
+## [5.0.9-rc.16] - 2026-10-11
+
+- Pin Workbench 0.13.2-rc.5. Bound OpenCode's health request at both
+  response headers and JSON body so a stalled local server cannot exceed
+  the startup handshake budget before the existing CLI fallback.
+- Reproduce the old behavior using two live HTTP connections, retain the
+  failed rc.15 client acceptance and require a new official-package retest.
+- No project, knowledge, history, provider or third-party hook reset.
+  Candidate is published to next; stable latest still requires acceptance.
+
+## [5.0.9-rc.15] - 2026-10-11
+
+- Serialize registration, module removal and registry deletion for each project.
+  Recheck membership inside the wiring lease so delayed supervisor registration
+  cannot resurrect a project after keep-data unbind. Unrelated projects remain
+  independent; failed deletion can still be repaired and retried.
+- Reproduce the race with a delayed real HTTP registration and preserve the
+  downloaded rc.14 Node 22 UI failure. Include the regression in the package.
+- Use non-binding CLI readiness probes in tests; the old bind probe could take
+  the exact port while its own foreground child was starting.
+- Real MiniMax acceptance covers overlapping analyses, another commit and a
+  dirty developer draft during analysis, streaming, secondary development,
+  four installed clients and evidenced preference candidates. New downloaded
+  candidate acceptance and actual-user cutover still precede stable latest.
+
+## [5.0.9-rc.14] - 2026-10-11
+
+- Vendor engine 0.2.3-rc.7: isolated maintenance consumes frozen evidence even
+  while the user edits or commits in the source repository. Do not mistake live
+  development for analyzer writes; retain live-source protection for custom
+  analyzers without the isolation capability.
+- Add a paused child-worker regression covering another commit, uncommitted
+  drafts, frozen attribution, same-project FIFO and unchanged path/tool guards.
+- Existing-user npm acceptance verified real MiniMax streaming, retained v5
+  history/configuration, D-root indexes and index-only e9f6560 recovery. Preserve
+  the actual rc.13 PROJECT_BUSY failure and require a real mid-analysis commit
+  retest from downloaded rc.14 before stable promotion.
+
+## [5.0.9-rc.13] - 2026-10-11
+
+- Give native Agent session creation and input setup a bounded 90-second proxy
+  budget. Ordinary API reads/abort/permissions retain 10 seconds; SSE streaming
+  and folder-picker contracts are unchanged.
+- Record safe proxy failure categories and deadlines without raw errors or URL
+  queries. Preserve MODULE_UNREACHABLE compatibility and add optional details.
+- Reproduce the installed OpenCode cold-start failure with a real delayed HTTP
+  server; verify the targeted budget, unrelated routes and credential redaction.
+- Retain rc.12 real acceptance and failed OpenCode results separately; require
+  a newly downloaded candidate and cold-client retest before user upgrade.
+
+## [5.0.9-rc.12] - 2026-10-11
+
+- Vendor engine 0.2.3-rc.6 to address the Windows transient state-file rename
+  failure exposed by installed rc.11 fault acceptance. Use exclusive random
+  temporary files and bounded sharing retries, retain durable state on failure,
+  and never unlink live state to force replacement.
+- Add deterministic sharing-violation, permanent-lock, foreign-temp and
+  disk-full regressions; all 39 engine tests pass on Node 22 and Node 24.
+- Preserve the rc.11 failure and require new installed acceptance before any
+  existing-user upgrade or stable latest promotion.
+
+## [5.0.9-rc.11] - 2026-10-11
+
+- Pin vector-hub 1.0.8-rc.3 and vendor engine 0.2.3-rc.5. Raw commit
+  conversations and internal backups are analysis evidence, not knowledge.
+  Exclude them from managed indexes, search and get; retain source evidence.
+- Prune stale derived records during sync and rebuild once on source-policy
+  upgrade, without repeating paid whole-index work on ordinary restarts.
+- Keep standalone vector-hub general document support unchanged.
+- Record both original failing regressions and successful Node 22/24 reruns;
+  installed, real model and existing-v5 gates still precede latest promotion.
+
+## [5.0.9-rc.10] - 2026-10-11
+
+- Preserve acknowledged brain/domain knowledge-directory bindings when changing
+  the knowledge root. Only derived indexes and future imports/shared scopes use
+  the new root; existing knowledge and confirmed memories stay in place.
+- Persist bindings in shell metadata, reject implicit rebinding, and report
+  shared-scope connection failures in UI and safe structured logs.
+- Add real-engine regression for confirmed memory, domain retrieval, restart,
+  new domain registration, offline retry and protected bindings on Node 22/24.
+- Candidate acceptance still precedes the existing-v5 upgrade and stable latest.
+
+## [5.0.9-rc.9] - 2026-10-11
+
+- Vendor engine 0.2.3-rc.4. Freeze existing target paths and hashes before
+  analysis so missing retrieval snippets cannot be mistaken for absent files.
+  Preserve full old documents when only partial context is available; illegal
+  create operations still fail safely instead of silently overwriting data.
+- Keep the failed rc.8 secondary-development maintenance evidence and verify
+  that exact commit after downloading this candidate, not patched source bytes.
+- Render long index-directory paths as a full-width block in System Status.
+- latest remains unchanged until the real closed-loop and existing-v5 gates pass.
+
+## [5.0.9-rc.8] - 2026-10-11
+
+- Pin Workbench 0.13.2-rc.4: atomically replace all native Agent histories,
+  retaining the previous complete record on disk-full or locked-file errors.
+  CLI fallback no longer deletes durable history before a failed rename.
+- Keep the original disk-full loss and separate successful background recovery
+  records. Candidate acceptance includes faults against downloaded module bytes.
+- Pin vector-hub 1.0.8-rc.2. Managed indexes now follow the knowledge root
+  under `.indexes`; show the acknowledged actual directory in System Status.
+  Rebuild on a root/model change while retaining the old index directory;
+  serialize queued writes and avoid a full paid rebuild on normal restart.
+- Do not promote latest until installed and existing-v5 acceptance pass.
+
+## [5.0.9-rc.7] - 2026-10-11
+
+- Vendor engine 0.2.3-rc.3: recognize promoted commit identifiers inside
+  natural-language tasks, including short and multiple SHAs. Raw conversation
+  mentions cannot replace the current promoted knowledge references.
+- Preserve failed installed Workbench/context checks from rc.6; latest remains
+  unchanged until fresh registry runtime and existing-v5 acceptance pass.
+
+## [5.0.9-rc.6] - 2026-10-11
+
+- Vendor engine 0.2.3-rc.2. Unassigned projects no longer share an implicit
+  domain in retrieval or Workbench task context. Explicit domains and whole
+  brain queries retain their inclusive scopes.
+- Keep the failed rc.5 real Workbench acceptance as a separate record. This
+  candidate must pass fresh downloaded-runtime checks before promotion.
+
+## [5.0.9-rc.5] - 2026-10-11
+
+- Pin Workbench 0.13.2-rc.3: retain OpenCode final answers across tool rounds,
+  isolate CLI history by product data directory and recover only owned v5
+  session records without moving or deleting originals.
+- Support current ZCode model-selection and failure events. Shared provider
+  credentials stay in memory behind a temporary loopback transport; client
+  shutdown cleans up the transport and settles rejected sends correctly.
+- Candidate publication does not promote latest or imply complete acceptance.
+  Real client execution, installed UI regression and existing-data upgrade
+  remain separately recorded gates.
+
+## [5.0.9-rc.4] - 2026-10-11
+
+- Pin Workbench 0.13.2-rc.2: resolve the installed ZCode provider resource
+  layout without rewriting desktop or personal configuration, and label
+  runtime failures with the correct agent.
+- Disable the project selector until the first registry snapshot. This fixes
+  the startup interaction race exposed by the Node 22 packaged UI regression.
+- Latest is unchanged; real installed acceptance and client authentication are
+  independent gates, not implied by a successful version probe.
+
+## [5.0.9-rc.3] - 2026-10-11
+
+- Route engine-wired CLI and MCP reads through the same authoritative engine
+  as the web UI, preserving commit references and project/domain/brain scopes.
+  Verify the selected runtime identity and reject unavailable, mismatched or
+  malformed services instead of silently querying a different local index.
+- Preserve the MCP `space` alias and expose explicit scope selection. Bound
+  responses and never forward profile or credential objects to read routes.
+- Drain isolated regression servers before Windows fixture cleanup. Candidate
+  publication still does not promote latest or claim complete acceptance.
+
+## [5.0.9-rc.2] - 2026-10-11
+
+- Pin the verified Workbench 0.13.2-rc.1 and vector-hub 1.0.8-rc.1 candidates.
+  Normalize known official MiniMax embedding route aliases at runtime without
+  changing existing v5 configuration bytes or private gateway paths.
+- Queue same-project index writes across embedding-model rebuilds. Failed writes
+  do not poison retries, and unrelated projects remain independent.
+- Restore safe relative Markdown document navigation and verify the standalone
+  retrieval UI at 1920x1080. Tests never kill unrelated port owners.
+- Isolate regression-suite ports and tolerate Windows fixture file release.
+  Candidate publication is not full product acceptance; latest remains unchanged.
+
+## [5.0.9-rc.1] - 2026-10-11
+
+- Retain classified, bounded and redacted index errors in run details, activity
+  and local log/v2 exports. Index-only retries preserve promoted files and do not
+  invoke the analysis model again.
+- Accept ZCode conversation evidence, isolate per-project journal failures and
+  report the exact blocked sequence without dropping evidence or flooding logs.
+- Synchronize production asset hashes, including the shared favicon. Permit
+  exact rc module pins but reject tags, ranges, links and source overrides.
+- Prepare a web-only candidate; latest is not promoted by this release. Existing
+  v5 data and third-party integrations are preserved; system acceptance remains
+  a separate gate from publication.
+
+## [5.0.8] - 2026-10-10
+
+- Repair managed Codex notify chains after scoped npm package or Bridge-home
+  changes. Preserve live desktop wrappers and valid third-party notifications;
+  unwrap only missing, known Codex runtime wrappers. Unknown notifiers remain
+  conflicts, never overwritten. Repeat setup is idempotent; removal unwraps
+  capture without deleting existing notifications or evidence.
+- Accept Codex's documented JSON argument as well as legacy stdin notifications.
+  Preserve the third-party payload and transport, adapt only the Bridge wake-up
+  aliases, and isolate child-process errors between notification branches.
+- Place new notify entries at the TOML root. Back up only the modified setting,
+  and refuse stale writes when another program changes the config after preflight.
+- Add real Git/Bridge/journal capture regression with isolated rollout fixtures,
+  duplicate-notify and missing-session checks; no model calls or user-data reset.
+
+## [5.0.7] - 2026-10-10
+
+- Normalize CLI command case and reject unknown commands, extra arguments and
+  invalid ports before any service or data is created.
+- Replace the shared temporary PID and blind orphan scan with per-data-directory
+  local control channels. Stop waits for the backend and owned module shutdown;
+  stale or unrelated PIDs are never signaled.
+- Add verified local instance inventory (`status --all`), explicit targeted stop
+  (`stop --port`) and data-directory selection (`--data`). Display the actual
+  port, version and data directory; refuse unverifiable or mismatched targets.
+
+## [5.0.6] - 2026-10-10
+
+- Do not mistake unregistered, abandoned legacy chat records for a knowledge
+  database on first launch. Preserve their bytes without restoring sessions or
+  importing projects; continue rejecting actual old registries and maintenance.
+- Wait for the background server's real HTTP readiness before reporting success
+  or opening the browser. Surface startup failures in the invoking terminal with
+  a nonzero exit code, and name the exact conflicting data entries.
+- Isolate regression data from the user's default data directory. Add a detached
+  npm CLI startup regression from an unrelated working directory.
+
+## [5.0.5] - 2026-10-10
+
+- Invalidate in-flight registry snapshots after domain creation, project import,
+  ownership changes and wiring repair. A read begun before the write cannot
+  hide the newly registered data until the next polling cycle.
+- Includes all knowledge-maintenance, Claude discovery and local diagnostic
+  fixes from 5.0.3, plus exact commit retrieval from 5.0.4. No schema migration
+  or automatic replay of failed model tasks.
+
+## [5.0.4] - 2026-10-10
+
+- Vendor engine 0.2.2: treat a known full/prefix commit SHA as a provenance
+  identifier. Unrelated high-scoring vector chunks cannot crowd out the matching
+  current knowledge. Preserve ordinary semantic ranking and current-file truth;
+  overwritten documents are not attributed to their earlier commit.
+- Includes the empty-knowledge and Claude discovery/diagnostic fixes from 5.0.3.
+
+## [5.0.3] - 2026-10-10
+
+- Pin Workbench 0.13.1: reject stale Claude executable overrides without hiding
+  a valid installation; discover native/PATH/package entries and probe JS/CJS/MJS
+  through Node. Discovery, doctor and runtime failures record their actual cause.
+- Vendor engine 0.2.1: an empty knowledge tree is a valid first-use state. Require
+  evidence-backed nonempty Markdown; normalize only missing titles in validated
+  staging output and reseal hashes before promotion. Empty/invalid output still
+  fails safely with a concrete reason, without overwriting existing knowledge.
+- Route module diagnostics through the supervisor to local product logs and the
+  diagnostic page; redact runtime credentials before IPC. Preserve failed runs
+  for explicit retry rather than silently replaying model tasks on upgrade.
+
+## [5.0.2] - 2026-10-10
+
+- Add ZCode capture to the unified CLI and web System Status connector setup,
+  including repair, status and safe removal. Keep unverified ZCode knowledge
+  plugin support explicitly separate from working capture hooks.
+- Correct command/config discovery labels and stop treating consumer
+  registration or disabled hooks as proof of capture. Report partial failures
+  with their operation ID and persist connector errors from both web and CLI.
+- Pin event-bridge 0.7.2: external hooks resolve their own Bridge home without
+  inherited environment variables; ZCode events preserve their source. Repair
+  older entries in place and retain third-party hooks within mixed groups.
+
+## [5.0.1] - 2026-10-10
+
+- Fix project import on Git versions before 2.31: query hooks and common Git
+  directories without `--path-format`, then resolve relative paths against the
+  Git command's working directory. Reject empty, multi-line or echoed-option
+  responses before writing hooks or repository identities.
+- Apply the same correction to commit capture and pin event-bridge console
+  0.7.1 so importing a project and capturing its next commit use consistent
+  workspace metadata. No previous data is migrated.
+- Keep import failure feedback visible with its error code and operation ID.
+  Persist the actual failed transaction stage; system diagnostics default to
+  all projects so failures of not-yet-registered imports are visible/exportable.
+
+## [5.0.0] - 2026-10-10
+
+- Fresh-data-only web product. Default state moves to `~/.project-knowledge-v5`;
+  no previous knowledge, project registry, unfinished claims or model profiles
+  are imported. Explicitly selected old data directories are rejected unchanged.
+- Remove shell migration/cutover and engine `/api/migrate`. Engine 0.2.0 uses
+  schema version 3; validate its data generation and runtime authentication
+  before registering projects or enabling consumption. System status reports
+  safe connection stage/error codes instead of silently waiting for takeover.
+- Isolate Workbench, bridge journals and vector configuration under the new
+  product directory. No automatic embedding-credential migration.
+- Queued, unstarted tasks are cancelled on unbind; active knowledge/index write
+  leases still block removal. The visible confirmation dialog updates project
+  selection immediately and rejects stale registry responses without reloading.
+- Reimport can explicitly replace a leftover knowledge-managed Git Hook. Failed
+  imports restore the prior Hook; third-party hooks and existing knowledge files
+  are never overwritten. Import tracks HEAD without analyzing historical commits.
+
+
+## [4.8.0] - 2026-10-10
+
+- Unified brain web product: six real pages for work, retrieval, memories,
+  domains/projects and activity. Existing module UIs retain their streaming
+  and recovery behavior with one embedded theme and project context.
+- Project import wires all four modules and preserves code/knowledge paths;
+  domains and confirmed memories add shared scopes without moving knowledge.
+- The vendored knowledge-engine owns journal consumption, isolated analysis,
+  staging promotion and retrieval. Per-project FIFO and two concurrent projects
+  persist across recovery; promotion and indexing have separate failure states.
+- Pin the validated module releases: Workbench 0.13.0, vector-hub 1.0.7,
+  event-bridge console 0.7.0 and vendored knowledge-engine 0.1.0.
+  Runtime file hashes are verified before release, not resolved from `latest`.
+- Web-only release. No new Windows installer or desktop auto-update delivery.
+  Regression uses controlled model/embedding doubles; real-model and real-index
+  smoke tests are reported separately. Remaining fault combinations are not
+  represented as a completed stability certification.
+
+## [4.7.0] - 2026-09-21
+
+- All four embedded module dependencies (`claude-ai-workbench`,
+  `@sanqianx/vector-hub`, `@sanqianx/ai-coding-event-bridge-console`, plus the
+  `@sanqianx/ai-coding-event-bridge` consumer inside the shell) are now pinned
+  to `latest`, so a fresh `npm install` pulls whatever the upstream modules
+  ship instead of lagging behind the project's npm releases. The legacy
+  `@sanqianx/ai-coding-event-bridge` package — which used to be the engine
+  library — is deprecated on npm and removed from this project's dependencies;
+  its capture engine now ships inside `@sanqianx/ai-coding-event-bridge-console`
+  at `src/bridge`, and the four production call sites (hook capture, requirement
+  recorder, integration installer, bridge adapter) plus six tests were rewritten
+  to import from there. The legacy package's every version carries a deprecation
+  message pointing to the new entry.
+- The package is published as `@sanqianx/project-knowledge` (moved from the
+  unscoped `project-knowledge`); the unscoped name still resolves to the old
+  versions for users who haven't migrated, but new installs land on the
+  scoped package.
+
+## [4.6.0] - 2026-09-21
+
+- The three embedded module runtimes now ship as real npm dependencies instead
+  of the `_modules/` committed snapshot: `claude-ai-workbench@0.4.1`,
+  `@sanqianx/vector-hub@1.0.1` and
+  `@sanqianx/ai-coding-event-bridge-console@0.4.0` (bringing
+  `@sanqianx/ai-coding-event-bridge@0.4.0` transitively). The `_modules/`
+  directory, its vendor script and the tarball entry are removed.
+- Why it matters: the snapshot lagged the source repo — npm consumers missed
+  the codex resident-bridge Windows fix (per-turn process respawns made every
+  turn 5-20s), claude failure-reason surfacing, `/system/doctor` diagnostics
+  and the folder-picker foreground/centering fixes. Module upgrades are now a
+  one-line version bump verified by `npm ci`, with no manual snapshot step.
+- `module-bridge` resolves each runtime entry through `require.resolve` from
+  the installed packages (same supervised spawn, reuse and env-override
+  contract as before); the `NODE_PATH` stitching is gone. `release:verify` no
+  longer gates on the vendored snapshot.
+
+## [4.5.0] - 2026-09-21
+
+- The embedded Agent Terminal's settings view gains a 会话捕获 (capture) tab.
+  Terminal conversations remain exempt from Development Conversation capture by
+  default — the terminal is internal AI runtime, not the developer's CLI
+  record. Flipping the toggle persists the scope, applies it to the runner
+  immediately, and marks only NEW terminal sessions with
+  `AI_CODING_EVENT_BRIDGE_CAPTURE=1`, so they reach the Bridge journal and the
+  会话浏览 explorer like any CLI tool conversation. Existing sessions are
+  unaffected, and nothing changes for standalone terminals that never touch
+  the toggle. Ships the updated vendored Agent Terminal runtime.
+
+## [4.4.0] - 2026-09-21
+
+- The ai-coding-event-bridge conversation console joins the Agent Terminal and
+  vector-hub as the third embedded Control Center module. The vendored runtime
+  (console 0.4.0 + its core) ships inside the npm package and starts with the
+  shell on 127.0.0.1:8790; the console's dependency on
+  `@sanqianx/ai-coding-event-bridge` resolves from a node_modules directory
+  vendored inside the console package, so the host package's own older npm
+  dependency can never shadow it. The global `node_modules/` gitignore rule
+  carries a negation for that directory, and `vendor:modules --check` now
+  verifies the runtime is present and git-tracked — the 4.3.0 "runtime missing
+  from the tarball" incident class cannot recur for this module either.
+- Module integration follows the established contract: same-origin reverse
+  proxy under `/api/eventbridge/`, supervised spawn with reuse of an already
+  listening standalone console (`KB_EVENTBRIDGE_COMMAND` override,
+  `AI_CODING_EVENT_BRIDGE_HOME` passthrough so the explorer shows the same
+  journal the in-process BridgeAdapter writes), idempotent import-time
+  registration keyed by Git work tree path using the recommended
+  `<knowledge>/dev-conversations` store, removal that keeps journals and sealed
+  commit documents, aggregated sidebar projection and module health.
+- The shell UI gains the 会话浏览 view (connect probe, 独立打开 link, per-project
+  桥 badge, import/remove copy) — and the conversation explorer itself learned
+  the `kb:sidebar` embed protocol (`?embed=1`, `kb:ready` announcement,
+  loopback-only origins), so its project sidebar folds via the ▤ control exactly
+  like the terminal and vector-hub embeds. Standalone console opens are
+  unaffected.
+
+## [4.3.1] - 2026-09-20
+
+- The 4.3.0 npm package could never start the vector-hub module: the global
+  `dist/` gitignore rule kept the vendored runtime
+  (`_modules/vectorhub/dist/`, 25 files) out of git, so the CI-packed tarball
+  shipped without `bin.js` and every fresh install logged "Vendored vector-hub
+  runtime not present". The runtime is committed now, and `release:verify`
+  gained a vendored-modules gate that fails when a required runtime file is
+  missing from disk or not tracked by git — the check that was previously
+  wired into nothing.
+- The Control Center's embedded Agent Terminal linked to no project on first
+  load. The shell sends `kb:select-project` 150ms after `kb:ready`, but the
+  terminal loads its project list only after health + agent discovery, so the
+  one-shot selection landed on an empty list: the linkage chip stuck on
+  "未登记" and the embed had no project selected, making it impossible to open
+  a terminal from the shell. The terminal now re-announces `kb:ready` once its
+  projects are loaded, and the shell retries a failed linkage with backoff.
+- Module status probing no longer reports a false "未检测到 … 界面将显示为
+  空白" forever: cold-starting module services are re-probed with backoff
+  (up to 10 attempts) before the status is marked down, and the top bar
+  终端/检索 indicator updates when a late probe succeeds.
+
+## [4.3.0] - 2026-09-20
+
+- The Agent Terminal and vector-hub module services now ship inside the npm
+  package and start with the shell. `npm install -g project-knowledge` is
+  self-contained: the Control Center, the embedded terminal (:5760) and the
+  vector-hub console (:8787) all come up together on a fresh machine — no
+  sibling checkouts required. KB_MODULES_AUTOSTART=0 disables spawning;
+  a service already listening on its URL is never double-spawned, so dev
+  machines running standalone module instances are unaffected. vector-hub
+  needs Node >= 22 — on older runtimes the shell skips spawning it and logs
+  a warning. Also fixed a latent spawn bug: quoted paths in module commands
+  reached node with the quotes embedded, which crash-looped the (previously
+  default-off) autostart path.
+
+## [4.2.20] - 2026-09-20
+
+- Startup no longer relocates legacy v4.1.x package-root runtime data. This is
+  a fresh knowledge-base system by decision: old data under the npm package
+  dir is ignored, not migrated. Previously a conflict between stale
+  package-root files and the user data dir aborted startup with "Legacy
+  runtime relocation failed", bricking machines whose install carried
+  v4.1.x-era files.
+
+## [4.2.19] - 2026-09-20
+
+- Background launch is diagnosable and no longer lies about the URL. The
+  detached child's stdout/stderr now append to `<data-dir>/launcher.log`, so
+  port fallbacks ("Port 5757 was busy, using 5758 instead") and startup
+  failures are visible after the fact instead of being silently discarded
+  (`stdio: 'ignore'`). The parent no longer prints a guessed URL or races a
+  browser open 1.2s after spawn — the child opens the browser itself once it
+  is actually listening, with the real port. The free-port probe timeout went
+  1s → 3s to survive slow listen() under antivirus/cold-cache load (a false
+  "busy" previously cascaded into "No free port found").
+
+## [4.2.18] - 2026-09-20
+
+- Module-embed shell (T0-T4 of docs/ui-integration-repair-plan.md): the Control
+  Center UI now embeds the Agent Terminal (:5760) and vector-hub console
+  (:8787) as original UIs behind a same-origin proxy (`/api/terminal/*`,
+  `/api/vectorhub/*`). Project import registers both modules (wiring table in
+  project state); the sidebar aggregates module registrations with per-project
+  session history (kb:list-sessions / kb:open-session) — session lifecycle
+  (archive/restore) belongs to the terminal module only, the shell keeps a
+  read-only projection; per-module sidebar fold toggles on the nav buttons
+  (kb:sidebar); project selection drives both modules (kb:select-project)
+  with live linkage chips.
+  Settings drawer is now 开发对话/日志 only (model settings live in the
+  terminal UI). Module removal keeps local knowledge and terminal session
+  records (upstream `removeProject?keepSessions=1`). Optional module service
+  supervision behind `KB_MODULES_AUTOSTART=1`. Old workbench chat panel and
+  its UI tests retired; full suite 115/115 green.
+
+## [4.2.17] - 2026-08-25
+
+- Recognized a healthy Project Knowledge fan-out nested inside Codex Desktop's
+  computer-use `--previous-notify` wrapper. Integration status and one-click
+  repair no longer report a false third-party conflict or attempt to wrap the
+  notifier again.
+
+## [4.2.16] - 2026-08-25
+
+- Upgraded AI Coding Event Bridge to 0.1.2 so one Codex task can be captured
+  across every continuation rollout. Existing single-file cursors migrate in
+  place without replaying consumed history, and real nested turn identities
+  keep commentary and final responses attached to the correct user prompt.
+- Added a joint Codex / Claude Code / OpenCode restart-and-recovery gate that
+  verifies durable capture, control-message exclusion, exact workspace
+  attribution, turn grouping, and duplicate-notification idempotency.
+- Kept Windows CDP failure tests fast and reliable by restricting detached
+  profile cleanup to actual Chrome/Edge launchers.
+
+## [4.2.15] - 2026-08-25
+
+- Repaired Codex Development Conversation turns without rewriting durable
+  history: generated control/context records are excluded, orphan assistant
+  responses are projected onto the correct user turn within the same session,
+  and stale open-turn facts no longer duplicate commit associations.
+- Removed synthetic empty-user and pending-assistant UI messages, with focused
+  projection, binding, API, and real-history regression coverage.
+
+## [4.2.14] - 2026-08-25
+
+- Recovered commit analysis safely after a server restart by making leftover
+  claims terminal without running startup analysis or replaying historical or
+  offline commits; the next explicit Git Hook event processes only its own SHA.
+- Restored v4.1.21 Claude Workbench history through the migrated project's
+  legacy slug while preserving the original session files and preferring the
+  canonical runtime record on duplicate session IDs.
+- Restored historical session selection and live Commit automation visibility
+  in the Claude Code UI, including prompts, thinking, tools, permissions,
+  retries, results, stop controls, and real browser regression coverage.
+
+## [4.2.13] - 2026-08-25
+
+- Restored the globally installed web UI by including `ui/i18n.js` in the
+  npm package; a missing or incomplete localization runtime can no longer
+  abort project, AI profile, Knowledge Root, and conversation loading.
+- Added package-asset closure validation plus a real
+  `npm pack -> npm install -> server -> Chrome` release smoke test, so CI
+  rejects tarballs whose UI differs from the tested source tree.
+
+## [4.2.9] - 2026-08-24
+
+- Made persisted UI-language selection fully reliable across page reloads and
+  slower browser runners: translated labels now re-render, and the document
+  language is initialized from the saved preference.
+
+## [4.2.8] - 2026-08-24
+
+- Fixed release metadata drift by deriving plugin manifests, marketplace
+  metadata, MCP package pins, and both lockfiles from the root package
+  version during the standard npm version lifecycle.
+- Added release and tag-version guards to CI and the npm publishing workflow,
+  so a mismatched release is rejected before packaging or publishing.
+- Stabilized Windows browser E2E tests by handling Chrome launcher detachment
+  and cleaning every test-specific Chrome process by its isolated profile.
+
+## [4.2.7] - 2026-08-24
+
+- Made runtime data migration fail closed: existing legacy data is now
+  discovered and relocated safely, migration failures prevent startup,
+  and settings-only legacy data is preserved instead of being treated as
+  a fresh install.
+- Restricted Knowledge Analysis to explicit post-commit hook events.
+  Startup, bridge synchronization, historical catch-up, and offline
+  commits no longer trigger implicit analysis.
+- Preserved conversation and commit-boundary facts across offline periods
+  without allowing those facts to contaminate a later online commit.
+
+## [4.2.6] - 2026-08-20
+
+- Hotfix for the v4.2.5 commit-reconciler regression: the post-batch
+  re-scan was unconditionally writing `analysis.status='idle'` even
+  after a successful batch of commits, clobbering the
+  `'state.advanced'` value that `advanceState()` had just set. The CI
+  suite surfaced this on a fast runner where the test's state read
+  raced ahead of the trailing re-scan; on slower machines the bug
+  passed silently. runSweep() now only marks `'idle'` when this call
+  never processed any commits, leaving a freshly-promoted batch's
+  `state.advanced` visible to the UI.
+- Bumped `plugins/project-knowledge/.mcp.json` from
+  `project-knowledge@4.2.4` to `project-knowledge@4.2.6` so the
+  integration-adapters MCP-pin assertion matches the package version.
+  (The v4.2.5 release had bumped everything except this MCP pin.)
+
+105 root + 2 desktop tests pass.
+
+## [4.2.5] - 2026-08-20
+
+- Restored the v4.1.22 → v4.2.x product contract that the recent
+  refactor had broken. The shared ` `resolveEffectiveAiProfile` resolver
+  is the single source of truth for "which AI profile does this project
+  use right now" (project → default → first-usable → AI_PROFILE_REQUIRED),
+  and the Workbench / analyzer / CommitReconciler paths no longer
+  carry their own drift-prone fallback logic.
+- Hook runtime: post-commit hooks now prefix `ELECTRON_RUN_AS_NODE=1`
+  in the shell body so a packaged Desktop install no longer tries to
+  launch the Electron GUI from a Git hook invocation. Node ignores the
+  env var so CLI installs are unaffected.
+- Hook migration state machine: a missing legacy hook is no longer
+  silently marked migrationVersion=2. The new state machine only
+  advances migrationVersion when install + readback verify succeed,
+  persists third-party conflicts in state.hook.lastConflict for UI
+  consumption, and reports status=verified|conflict|pending|failed per
+  project.
+- Hook health API: GET /api/projects/:id/hook-status and POST
+  /api/projects/:id/hook-repair expose every required field
+  (installed / managed / managedVersion / runtimeTarget / lastVerifiedAt
+  / conflict / reason / repairAvailable) and refuse to overwrite
+  third-party hooks.
+- Import UX: a new preflight endpoint and UI flow (Desktop folder
+  picker, knowledge language selector, AI profile selector, optional
+  Team Knowledge binding, per-problem action hints, auto preflight on
+  every input change, Import button stays disabled until ready=true)
+  replace the path-only dialog.
+- Knowledge language control: import + per-project PATCH both accept
+  knowledgeLanguage; changing it affects only future generated
+  knowledge, never historical files.
+- Project Goal editor: GET/PUT /api/projects/:id/goal atomically reads
+  and writes the project's GOAL.md, scoped to its own knowledgePath.
+- UI language toggle (zh-CN / en-US) persists in localStorage and
+  re-renders the most-visible labels without a reload.
+- P0 E2E gate proves the four primary contract scenarios (fresh /
+  legacy / offline / no-conversation) against the live server.
+  105 root + 2 desktop tests pass; the G00–G08 acceptance gates are
+  green.
+
+## [4.2.4] - 2026-08-20
+
+- Fixed the Bridge consumer never receiving connector wake-ups on a fresh
+  install: `BridgeConsumerService.start(notifyUrl)` ignored its argument, so
+  the `project-knowledge` consumer registered with empty metadata. Captured
+  conversations piled up in the journal until the next server restart instead
+  of appearing in Development Conversation immediately. The wake-up URL now
+  lands in the consumer registration and drains fire per event.
+- Projects imported before v4.2.3 (no canonical `repo-identity/v1`) are
+  upgraded automatically: on the first consumer drain after updating, each
+  legacy project with an existing repository path gets its workspace identity
+  resolved and persisted, its conversation baseline established, and its
+  external conversations start projecting — no manual migration step. A
+  workspace already claimed by another project is never written twice.
+- The Bridge dependency now resolves to `@sanqianx/ai-coding-event-bridge`
+  0.1.1, which accepts Claude Code's real `hook_event_name` payload field
+  (previously such payloads were misrouted to the Codex connector and lost as
+  capture gaps).
+
+## [4.2.3] - 2026-08-20
+
+- Development Conversation now captures real external Claude Code / Codex /
+  OpenCode sessions through `ai-coding-event-bridge` (repo-identity/v1 with
+  workspaceId as the primary key, durable Bridge-owned turn identity, and
+  exact-workspace project matching — unimported workspaces are skipped and
+  ACKed, never written into another project).
+- The internal Workbench capture path was removed: sending a Workbench chat
+  message no longer records requirements or conversation events (explicit
+  `record_requirement` still works), every internal SDK session runs with
+  `AI_CODING_EVENT_BRIDGE_CAPTURE=0`, and legacy embedded Workbench pairs are
+  excluded from presentation and new commit binding via
+  `conversation-exclusions/v1` without rewriting history.
+- Commit conversation snapshots are frozen strictly after
+  append-boundary -> drainThrough: no event beyond `boundaryEndCursor`
+  (assistant replies included) can enter a snapshot, and imports establish a
+  conversation baseline so pre-import history is never backfilled.
+- One Integration Setup action installs Knowledge Integration and Development
+  Capture for Claude Code / Codex / OpenCode with separate status reporting,
+  third-party config preservation, Codex notify conflict detection, and a
+  host-level Bridge consumer that survives per-connector uninstalls; new
+  `/api/bridge/*` and `/api/integrations/*` endpoints expose capture health.
+- Windows CI browser tests now fail fast: CDP launch failures clean up the
+  spawned browser tree and profile directory, report exit/stderr diagnostics,
+  use a dynamically discovered DevTools port with per-request timeouts, and
+  the test runner reports `TIMEOUT` explicitly under a configurable budget
+  (`PK_TEST_TIMEOUT_MS`, default 90s) instead of hanging for 300 seconds.
+
+## [4.2.2] - 2026-08-19
+
+- Rebuilt the web Control Center visual layer to faithfully match the v13
+  reference design (`project-knowledge-base-faithful-repair-ui-v10.html`):
+  sidebar brand block with the knowledge root path, multi-line project cards
+  (name, id, repository path, status dot) with right-click removal, the
+  in-main-area header, full-height Workbench panel with chat bubbles, tool
+  cards, circular send button and composer toolbar, the reference design
+  tokens including the complete dark theme, and responsive breakpoints.
+- Restored the reference Settings drawer structure with separate AI model
+  editing (add/save profiles via the API), project model assignment,
+  knowledge relations, mature development-conversation turn cards with
+  commit labels, and the V7 run-records list geometry.
+- Import view now shows the reference split layout with a post-import
+  placement preview; UI contract tests were updated to the reference DOM.
+
+## [4.2.1] - 2026-08-19
+
+- Fixed the legacy layout migration failing with `Legacy migration source
+  JSON is corrupt` (and the server refusing to start) when any legacy source
+  file carries a UTF-8 BOM, which PowerShell and several editors commonly
+  write. BOM-prefixed sources are now parsed normally.
+
+## [4.2.0] - 2026-08-19
+
+- Completed the v13 refactor: conversation capture contracts, atomic Bridge
+  commit boundaries with capture-gap persistence, and frozen
+  `CommitConversationSnapshot` binding so post-commit prompts never bind to
+  old commits and claims freeze their analyzer evidence.
+- Replaced `patch:null` handling for large diffs with an exact, hash-verified
+  patch chunk evidence bundle; commit analysis now carries frozen evidence,
+  retrieval, and conversation manifests end to end.
+- Unified user search and commit context selection into one
+  `KnowledgeRetrievalService` (hybrid recall, deterministic rerank, chunk
+  budgets) with an index source manifest and a Markdown delta overlay so
+  fresh knowledge is visible while the derived index is dirty.
+- Logging is now permanent per-day JSONL under `logs/system` and
+  `logs/projects/<id>` with reverse-chunk cursors, SSE streaming without
+  gaps, and immutable capture policy (no retention/capacity settings).
+- Background work registers per operation with one root `operationId`
+  propagated across HTTP, Hook, startup, and reconciler logs, and shutdown
+  drains every registered task.
+- Project delete is transactional with a journal, tombstone, and per-stage
+  recovery; migration no longer converts legacy logging knobs into settings.
+- Restored the full Control Center web shell (Workbench, Import, Settings)
+  with a mature run-records log view and a read-only development-conversation
+  page driven by frozen commit snapshots; logs left the standalone console.
+- Added non-release CI (Linux Node 18/24, Windows Node 24) for every push;
+  desktop packaging validation is paused for this web-only cycle.
+
+## [4.1.23] - 2026-08-17
+
+- Replaced legacy whole-registry project state with schema-v2 settings, a
+  minimal project registry, stable per-project config/state, and locked
+  requirement JSONL records under one canonical `StorageLayout`.
+- Reduced knowledge analysis to the managed post-commit Hook and startup
+  recovery, both routed through one per-project reconciler and one Commit
+  prompt. Import now establishes a baseline without AI or placeholder files.
+- Added manifest-validated AI staging, journaled Markdown promotion, crash
+  recovery, dirty-index semantics, and a single serialized `IndexService` for
+  incremental writes and atomic full rebuilds.
+- Rebuilt logging around six-level `log/v2` JSONL records, recursive secret
+  redaction, daily/size rotation, retention/capacity cleanup, cursor queries,
+  exports, health state, and orphaned-operation detection.
+- Replaced the duplicate dashboard implementations with one responsive
+  light/dark logging console, including operation flows, structured error
+  details, pause/refresh, filtering, cursor pagination, and export.
+- Unified server, CLI, MCP, Hook, index, and project path resolution. Read-only
+  knowledge search falls back to scoped Markdown when the derived index is
+  dirty or unavailable.
+- Added staged `layout-v2` migration with centralized recovery backups,
+  source/hash validation, last-step completion, fault recovery, and preservation
+  of legacy paths, secrets, Commit pointers, indexes, and logs.
+- Hardened the API by removing wildcard CORS, raw-file access, generic project
+  replacement, manual Hook/analysis routes, and stack responses; non-loopback
+  binding now requires authentication and Origin validation.
+- Removed CLAUDE.md management, the second automation queue, configurable log
+  and index locations, legacy direct database writers, and unused Vue/Tailwind
+  browser runtimes.
+- Added Windows end-to-end coverage for online Hook processing, offline Commit
+  catch-up, requirement binding, promotion/index completion, structured log
+  chains, paths with spaces/non-ASCII text, and immediate recovery of locks
+  owned by a crashed process.
+- Hardened release packaging by explicitly mirroring core runtime dependencies
+  into the Electron app and allowlisting shipped documentation, preventing
+  local review artifacts from leaking into npm or desktop packages.
+
+## [4.1.22] - 2026-07-27
+
+- Added a read-only `project-knowledge-mcp` server for OpenCode, Codex, and
+  other MCP-compatible coding agents. It resolves the current Git project,
+  exposes scoped search/ask/get/history tools, and includes the search-first
+  workflow without depending on `CLAUDE.md`.
+- Added a safe keyword-search fallback for legacy Markdown knowledge bases and
+  kept routine knowledge writes exclusively in post-commit automation.
+- Added a shared Claude Code/Codex plugin with MCP configuration and a
+  search-first `project-knowledge` Skill, plus a native Claude marketplace and
+  Codex marketplace manifest.
+- Added an OpenCode adapter and a universal, idempotent integration manager for
+  installation, updates, removal, dry-runs, and status checks across Claude
+  Code, OpenCode, and Codex.
+
+## [4.1.21] - 2026-07-27
+
+- Fixed Windows desktop Claude sessions failing with “executable exists but
+  failed to launch” when a global Claude Code installation exposes `cli.js`
+  but `node` is unavailable in the desktop process PATH. JavaScript Claude
+  entry points now use the application's own Node/Electron runtime.
+- Added slow-backend and packaged-client regressions covering desktop startup
+  readiness and Claude script launch through the installed Windows executable.
+
+## [4.1.20] - 2026-07-27
+
+- Rebuilt the daily workspace around Claude Code while keeping GitHub/Gitea
+  sign-in, project import, AI setup, project knowledge relations, knowledge
+  storage, embedding model setup, logs, and desktop updates available from one
+  settings drawer.
+- Moved project removal to a right-click project-list action with confirmation.
+  Projects with a running Claude session or queued automation task now cannot
+  be removed.
+
+## [4.1.19] - 2026-07-24
+
+- Stabilized release tests: expected EventSource cancellation during a page
+  reload is no longer treated as a failed request, and the Windows CLAUDE.md
+  refresh test no longer removes a just-used temporary server directory.
+
+## [4.1.18] - 2026-07-23
+
+- Simplified the daily workspace into a full-width Claude Code interface.
+- Removed the project-supervision summary, pending-commit and issue panels,
+  dashboard status/goal cards, and manual automation preview/simulation UI.
+
+## [4.1.17] - 2026-07-23
+
+- Rebuilt commit-triggered knowledge automation around persistent per-commit
+  state, ordered recovery, and exactly one AI task per Git commit.
+- Removed manual draft review, auto-apply controls, and scheduled-task paths;
+  generated knowledge is now written directly through the automated pipeline.
+
+## [4.1.16] - 2026-07-22
+
+- Fixed the desktop client timing out / hanging on the loading screen for large
+  knowledge bases (many projects, multi-GB DB) while the web dashboard opened
+  fine. Root cause was several compounding issues:
+  - Desktop startup probe was too impatient: raised the per-request `/api/state`
+    timeout (1.5 s → 30 s) and the overall backend-ready budgets (owned backend
+    45 s → 180 s, existing backend 20 s → 60 s) so it waits like a browser does.
+  - Server startup no longer blocks the event loop: the CLAUDE.md audit and
+    orphaned-run cleanup now run in `setImmediate`, so `/api/state` can answer
+    immediately after `listen()`.
+  - `getScheduleInfo()` (which spawns `schtasks`) is now cached for 5 s so
+    repeated `/api/state` hits don't each pay process-spawn latency.
+  - `isProcessAlive` now treats `EPERM` as alive, so a running backend's
+    runtime-endpoint file is no longer deleted (and a duplicate cold-started)
+    when the desktop and CLI run at different privilege levels on Windows.
+  - The dashboard now retries the initial connection every 3 s until it first
+    succeeds, instead of waiting up to 60 s, so a slow-starting backend no
+    longer leaves the page spinning.
+- Stopped the knowledge-storage relocation from logging
+  `target ... is not empty` on every startup: relocation is now skipped (with a
+  single warning) when the configured storage directory already holds unrelated
+  content, and the existing storage location keeps being used.
+
+## [4.1.15] - 2026-07-21
+
+- Replaced the dashboard's fixed-interval polling with a server-sent-events
+  (SSE) push model. The backend now emits a `state/changed` event over a new
+  `/api/state-stream` channel whenever project state is written or a job
+  starts/finishes, and the frontend refreshes only in response (300 ms
+  debounced). A slow 60 s poll remains as a fallback. Result: changes such as
+  the sidebar pending count and summary cards update within ~1-2 s of the
+  per-commit analysis cycle, while an idle dashboard makes almost no requests
+  instead of hitting the server every 5 s.
+
+## [4.1.14] - 2026-07-21
+
+- Fixed `runCommitAnalysis` to mutate the projects map entry directly
+  (`projects[slug].lastAnalyzedCommit`) instead of a copy. The v4.1.11
+  one-by-one refactor accepted `project = { slug, ...projects[slug],
+  kbPath }`, so `project.lastAnalyzedCommit = ...` mutated a throwaway
+  copy and the `writeProjects` callback persisted the unchanged value.
+  Symptom: the "待分析" count never decreases after analysis; the
+  pending count appears stuck or oscillates between runs because the
+  scan counter and the manual analysis counter diverge.
+
+## [4.1.13] - 2026-07-21
+
+- Bumped the GitHub OAuth HTTP timeout from 20s to 30s in
+  `requestJson` and `requestFormJson` (github-team-store.js) and made
+  the timeout error message actionable: it now names the host that
+  failed and suggests using a Personal Access Token as a workaround
+  when GitHub is unreachable from the host network.
+
+## [4.1.12] - 2026-07-21
+
+- Hid the "Runs / Drafts" sidebar nav entry. All registered projects
+  use `knowledgeMode: autoApply` so manual review is not part of the
+  workflow; the runs view itself is preserved for programmatic access
+  (uncomment the one line in `navItems` to restore the button).
+
+## [4.1.11] - 2026-07-21
+
+- Refactored `runCommitAnalysis` from batch to one-by-one: each pending
+  commit now gets its own run record and its own draft, analyzed in
+  chronological order (FIFO). After every successful single-commit
+  run, `project.lastAnalyzedCommit` advances to that commit's hash so
+  the "待分析" count drops by one per analysis, matching the UI label.
+  Previously a 10-commit batch produced 1 run + 10 drafts and the
+  pending count did not decrease until the user clicked Apply (which
+  could also advance the checkpoint past commits the user had not
+  actually applied drafts for).
+
+## [4.1.10] - 2026-07-21
+
+- Made startup-recovery's staging copy (`prepareKnowledgeWorkspace`) and
+  Claude-executable lookup (`findClaudeExecutableForSdk`) asynchronous
+  so the server's HTTP handlers (notably `/api/state`) stay responsive
+  while pending-commit workers are being dispatched. Previously the
+  desktop client would time out with `backend request timed out` when
+  several projects had pending commits on startup, because the sync
+  `fs.cpSync` and `spawnSync` calls were freezing the event loop for
+  several seconds during `/api/state` polling.
+
+## [4.1.9] - 2026-07-21
+
+- Refused to import or upsert a project whose `kbPath` resolves to the
+  same directory as `gitPath` / `localPath`. Writing the knowledge base
+  inside the source tree pollutes the repo (and trips Windows symlink
+  errors in the staging workspace), so both entry points now return 400.
+
+## [4.1.8] - 2026-07-21
+
+- Stopped the post-commit automation worker from being auto-paused when a
+  user aborts the Claude session in the terminal. Pending commits queued
+  before the abort are now picked up again on the next desktop launch.
+
+## [4.1.7] - 2026-07-19
+
+- Restored GitHub and Gitea login in the Windows desktop client through an
+  acknowledged Electron bridge that opens OAuth pages in the system browser.
+  Browser launch failures are now reported immediately instead of appearing
+  later as an authorization timeout.
+- Added support for HTTP authorization URLs used by private, intranet Gitea
+  installations while continuing to reject local-file and script URLs at the
+  desktop security boundary.
+- Forwarded Electron's Windows system-proxy resolution to the backend and made
+  GitHub OAuth plus remote repository discovery proxy-aware, while preserving
+  loopback and `NO_PROXY` bypasses for private Gitea installations.
+- Verified the full team-knowledge path after desktop migration: GitHub device
+  authorization, Gitea callback/token persistence, remote repository scanning,
+  manifest discovery, cached store loading, and team knowledge-base binding.
+- Added desktop bridge/security regressions and an end-to-end Gitea OAuth plus
+  remote team-store discovery test.
+
+## [4.1.6] - 2026-07-19
+
+- Replaced overlapping post-commit batches with a single per-project worker
+  that treats Git history as the durable queue. Live hooks and desktop startup
+  recovery now use the same path and process exactly one commit at a time,
+  strictly from oldest to newest.
+- Preserved the original knowledge contract: every analyzed commit must create
+  an independent `changes/` Markdown record containing its full hash, including
+  test-only, documentation, and infrastructure commits.
+- Added transactional knowledge workspaces. Claude edits a temporary KB copy;
+  validated Markdown is applied to the live KB only after the session succeeds,
+  and LanceDB must finish before `lastAnalyzedCommit` advances.
+- Added resumable vector finalization and automatic startup recovery. Index
+  failures retry without rerunning Claude or duplicating Markdown, while an
+  interrupted desktop session resumes from the oldest unfinished commit.
+- Made Stop pause the project's automatic commit worker, discard uncommitted
+  staged knowledge, and prevent queued work from immediately restarting. The
+  Claude Code UI now provides a dedicated Resume automatic analysis action.
+- Added end-to-end regressions for duplicate Hook delivery, strict commit order,
+  per-commit changes records, staged Markdown, abort behavior, and vector-index
+  recovery. The complete core suite now contains 53 passing test files.
+
+## [4.1.5] - 2026-07-19
+
+- Added a desktop-client card in Settings that always shows the installed
+  application version and provides Check for updates / Restart and install
+  actions without opening a browser.
+- Integrated Electron's Squirrel updater with a stable GitHub Latest Release
+  feed. Release automation now uploads `RELEASES` and the full `.nupkg` beside
+  the installer so future desktop versions download and install in-app.
+- Hardened the PowerShell folder-picker fallback: scripts are unpacked from the
+  desktop ASAR, stdout uses UTF-8, and only an existing absolute directory can
+  be returned. PowerShell banners and error text can never populate a path
+  field.
+- Added updater lifecycle and folder-output regressions, bringing the complete
+  core suite to 52 passing test files while retaining the native Electron
+  folder dialog introduced in 4.1.4.
+
+## [4.1.4] - 2026-07-19
+
+- Fixed Windows Claude Code discovery so extensionless npm shell shims and
+  command wrappers are never passed to the Agent SDK. The resolver now prefers
+  the installed package's native `bin/claude.exe` and retains `cli.js` only as
+  a legacy fallback.
+- Changed the post-commit Hook endpoint to acknowledge valid events with HTTP
+  202 before commit inspection and Claude startup. Dispatch continues in the
+  background, while the Hook accepts every successful 2xx response instead of
+  recording a false timeout/error after a run was already created.
+- Restored both desktop directory-selection flows with a narrow Electron IPC
+  bridge and native directory dialog. Project import and Team Knowledge folder
+  selection no longer execute a PowerShell script from inside the packaged
+  ASAR or mistake PowerShell banner text for a selected path.
+
+## [4.1.3] - 2026-07-19
+
+- Removed the bundled 218 MiB Claude Code executable from the Windows desktop
+  application. The small Agent SDK control layer now discovers the user's
+  installed Claude Code through `CLAUDE_CODE_EXECPATH`, `PATH`, or the npm
+  global installation and reports a clear error without disabling knowledge
+  management when Claude Code is unavailable.
+- Removed LanceDB's unused optional embedding-provider stack and its duplicate
+  Transformers.js 3.0.2 / ONNX Runtime 1.19.2 dependency tree while retaining
+  the application's supported Transformers.js 3.8.1 embedding service.
+- Pruned non-Windows and non-x64 ONNX binaries, browser-only ONNX Web runtimes,
+  Transformers browser builds, source maps, and unused platform packages from
+  the Windows x64 bundle.
+- Added packaged runtime probes for LanceDB, Transformers.js, ONNX Runtime, and
+  the real 512-dimensional `bge-small-zh-v1.5` embedding model, plus release
+  size auditing that rejects bundled Claude binaries, duplicate runtimes, and
+  oversized artifacts.
+- Reduced the measured Windows installer from 476.0 MiB to 221.1 MiB and the
+  installed application bundle from 1,491.1 MiB to 617.5 MiB without changing
+  existing model caches, Markdown sources, LanceDB data, or storage paths.
+
+## [4.1.2] - 2026-07-16
+
+- Made the desktop dependency lock reproducible under the newer npm version on
+  GitHub's Windows runner by explicitly locking the encoding compatibility
+  dependency used by the packaging tree.
+- Made the desktop release workflow fail immediately when a native npm command
+  fails instead of allowing PowerShell to continue into a misleading later
+  step.
+
+## [4.1.1] - 2026-07-16
+
+- Fixed Windows GitHub Actions browser discovery for the `windows-2025`
+  runner by detecting system-level Chrome and Microsoft Edge installations.
+- Restored the release gate for all three real-browser UI suites so the
+  Windows desktop installer can be built, smoke-tested, and uploaded only
+  after the complete 50-file core regression suite passes.
+
+## [4.1.0] - 2026-07-16
+
+- Added a Windows x64 Electron desktop application and Squirrel installer that
+  reuse the existing configuration, configured knowledge root, model cache,
+  Markdown sources, Team Knowledge Mode, and LanceDB without copying or
+  re-vectorizing user data.
+- Added single-instance and system-tray lifecycle behavior: closing the window
+  keeps Git hooks available, while an explicit tray Quit stops only a backend
+  owned by the desktop application.
+- Added an atomic, local-only runtime endpoint record so CLI, desktop, and
+  existing Git hooks discover the active fallback port dynamically and never
+  open competing LanceDB writers.
+- Made the desktop application safely attach to an already-running npm CLI and
+  leave that external backend alive when the desktop window exits.
+- Added hardened Electron navigation, context isolation, sandboxing, denied
+  permission requests, bounded desktop backend logs, Squirrel lifecycle event
+  handling, and explicit native-module ASAR unpacking.
+- Added Windows release automation that runs all 50 core regression files,
+  desktop unit tests, a packaged EXE/API smoke test, and packaged LanceDB native
+  read/write tests before uploading the installer and SHA-256 to GitHub
+  Releases. The existing npm provenance publish workflow remains unchanged.
+
+## [4.0.4] - 2026-07-16
+
+- Added a Settings-based Markdown knowledge health audit for every registered
+  personal and team knowledge base, with per-project file size, deterministic
+  fixes, structural errors, and semantic-review findings.
+- Added safe batch or per-project optimization with retained backups under the
+  configured knowledge root, atomic writes, formatting repair, compact index
+  regeneration, and automatic LanceDB refresh for migrated projects.
+- Reduced generated change indexes to the latest 40 entries with bounded cells
+  and generated-file markers; a real 125 KiB legacy index now rebuilds to about
+  9 KiB without deleting its source change documents.
+- Prevented background AI automation from writing `00-index.md`, injected
+  mandatory in-place-update hygiene rules even for custom prompts, and rebuilt
+  compatibility indexes after every successful automatic update.
+- Preserved Team Knowledge Mode by de-duplicating shared physical KB paths
+  during maintenance while refreshing every aliased vector space.
+- Expanded the complete regression suite to 48 passing test files.
+
+## [4.0.3] - 2026-07-15
+
+- Added an embedding-model setup panel with one-click download and real
+  inference verification, configurable Hugging Face endpoint/mirror, local
+  model directory, offline mode, and copyable PowerShell environment-variable
+  examples.
+- Stop a migration batch before changing any project when the shared embedding
+  model cannot load, instead of reporting the same model failure once for every
+  project.
+- Surface migration failures through a prominent settings panel, a one-time
+  browser alert, detailed per-project errors, and direct retry actions.
+- Persist model settings and verified download state while rejecting partial
+  cache files or verification results from a different model configuration.
+- Expanded the complete regression suite to 46 passing test files.
+
+## [4.0.2] - 2026-07-15
+
+- Store LanceDB, its maintenance metadata, and retained database backups under
+  `<configured knowledge root>/.project-knowledge/` instead of a fixed app-data
+  location.
+- Automatically relocate v4.0.0/v4.0.1 databases on startup to the knowledge
+  root already selected in Settings.
+- Safely relocate and verify the database before saving a changed knowledge
+  root, including cross-volume copies, target-conflict protection, and retained
+  backup path rebasing.
+- Display the effective vector database path directly below the Knowledge Store
+  root setting.
+- Expanded the complete regression suite to 45 passing test files, including
+  real LanceDB startup and live root-change relocation tests.
+
+## [4.0.1] - 2026-07-15
+
+- Excluded generated `00-index.md` files from embedding and removed stale
+  derived-index rows while retaining the Markdown files for legacy and Team
+  Knowledge Mode compatibility.
+- Replaced unbounded Tags and change lists in generated Markdown indexes with
+  compact, capped navigation metadata; all module links remain available.
+- Stopped running LanceDB compaction after every commit. Maintenance now runs
+  only after bounded mutation thresholds and never rewrites the oversized
+  v4.0.0 full-text index in place.
+- Added a compact full-text schema that indexes `chunk_text` once with Chinese
+  bigrams instead of storing and indexing a second full document copy.
+- Added one-click, atomic database rebuilding with disk statistics, vector and
+  keyword verification, optional retained backup, and manual rollback.
+- Added upgrade and regression coverage for old v4.0.0 databases; the complete
+  suite now contains 43 passing test files.
+
+## [4.0.0] - 2026-07-15
+
+- Added a local LanceDB knowledge store containing original text, metadata,
+  and 512-dimensional `Xenova/bge-small-zh-v1.5` embeddings.
+- Added incremental Markdown indexing with heading-aware chunks, content-hash
+  skips, stale-chunk deletion, Chinese n-gram/BM25 search, and hybrid RRF.
+- Added scoped `search`, `get`, `ask`, and `history` HTTP/CLI tools that return
+  readable source text without exposing raw vectors.
+- Added explicit, bidirectional, non-transitive related-project search scopes;
+  writes remain restricted to each project's primary space.
+- Preserved Team Knowledge Mode: Git remains the portable Markdown transport,
+  stable team `kbId` values map to local spaces, and vectors stay local.
+- Added resumable one-click migration for all legacy KBs with backups,
+  verification probes, per-project atomic cutover, retry, and rollback.
+- Centralized Claude Code instructions in one user-relative file and added a
+  safe bulk refresh for all managed `CLAUDE.md` blocks independent of Hooks.
+- Enforced development-time read-only KB access; post-commit automation owns
+  routine updates and re-indexes migrated projects after successful runs.
+- Added configurable mirrored/offline embedding model locations through
+  `KB_EMBEDDING_REMOTE_HOST` and `KB_EMBEDDING_LOCAL_PATH`.
+- Expanded the regression suite to 41 test files.
+
+## [2.4.7] - 2026-07-01
+
+- Initial public Apache-2.0 release of the local-first `project-knowledge`
+  core.
+- Includes the CLI, local dashboard, Git scanner, project registry,
+  knowledge-base workflow, AI profile support, prompt registry, templates,
+  and regression tests.
+- Keeps runtime data, local project knowledge bases, API keys, generated AI
+  drafts, logs, and local maintenance notes outside the public repository.