dsh-openwrite 0.2.23 → 0.2.25

package/README.md

@@ -29,12 +29,12 @@
 
 ## 快速安装
 
-**0.2.23 标准插件包**兼容精确 dsh 版本 **0.1.2-rc.1 / 0.1.5-alpha.1 / 0.2.0-rc.2 / 0.2.1-rc.2**。四个版本均已完成 Web 工作台渲染与 DoG 数据读取验证，Windows Electron 桌面端启动/恢复链路尚未实机复测。可下载版本、各平台验收结果和 SHA-256 以 [GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/latest) 附件为准；验收方法见 [发布验收](docs/RELEASE_ACCEPTANCE.md)。默认使用 npm `latest` 渠道；商城所需的精确预览版兼容性、验证范围和安装限制见 [DSH STORE 接入](docs/DSH_STORE.md)。
+**0.2.24 标准插件包**兼容精确 dsh 版本 **0.1.2-rc.1 / 0.1.5-alpha.1 / 0.2.0-rc.2 / 0.2.1-rc.2**。四个版本均已完成 Web 工作台渲染与 DoG 数据读取验证，Windows Electron 桌面端启动/恢复链路尚未实机复测。可下载版本、各平台验收结果和 SHA-256 以 [GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/latest) 附件为准；验收方法见 [发布验收](docs/RELEASE_ACCEPTANCE.md)。商城所需的精确预览版兼容性、验证范围和安装限制见 [DSH STORE 接入](docs/DSH_STORE.md)。
 
-npm 包和对应 GitHub Release 使用同一份验收产物。也可安装精确版本的 Release 包：
+npm `dsh-openwrite@0.2.24` 与对应 GitHub Release 均已发布，`latest` 已更新到 0.2.24，使用同一份验收产物。也可安装精确版本的 Release 包：
 
 ```sh
-dsh plugin --profile web add -w https://github.com/LiPu-jpg/Openwrite/releases/download/v0.2.23/dsh-openwrite-0.2.23.tgz
+dsh plugin --profile web add -w https://github.com/LiPu-jpg/Openwrite/releases/download/v0.2.24/dsh-openwrite-0.2.24.tgz
 dsh web
 ```
 
@@ -64,7 +64,11 @@ dsh web
 
 ### dsh 0.2 桌面端升级
 
-`0.2.23` 修复 [#83](https://github.com/LiPu-jpg/Openwrite/issues/83)/[#84](https://github.com/LiPu-jpg/Openwrite/issues/84)/[#85](https://github.com/LiPu-jpg/Openwrite/issues/85)：根包与全部内嵌子包的兼容声明加入精确宿主 `0.2.1-rc.2`，安装和激活不再被宿主版本闸门拒绝，无需版本豁免或手工编辑 `compatibility.json`。已在 `0.2.1-rc.2` 完成 Web 客户端验收（启动、OpenWrite 入口、工作台三页签、DoG 面板与其他插件共存），`0.1.2-rc.1`、`0.2.0-rc.2` 回归通过；Windows Electron 桌面端未实机复测。升级后停止并重新启动 dsh，再刷新浏览器页面。
+`0.2.25`（Core 5.8.12）为本分支的未发布候选，新增连载体检、分段续写、共享术语/范本、章组审计和漂移裁决，并修复验收与任务可观测性。范围、离线升级工具及宿主侧限制见 [#98–#100 实现说明](docs/ISSUES_98_100.md)。不要把候选版本当作已上架版本。
+
+`0.2.24`（Core 5.8.11）修复 [#88–#95](docs/ISSUES_88_95.md)：删章验收账本、结算降级、长文本流式生成、后台写章进度、统一字数门限、草稿发现与恢复及诊断/路由语义。[GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.24) 与 [npm](https://www.npmjs.com/package/dsh-openwrite/v/0.2.24) 均已发布，完整验收见 [发布验收](docs/RELEASE_ACCEPTANCE.md#024)。升级后停止并重新启动 dsh，再刷新浏览器页面。
+
+`0.2.23` 修复 [#83](https://github.com/LiPu-jpg/Openwrite/issues/83)/[#84](https://github.com/LiPu-jpg/Openwrite/issues/84)/[#85](https://github.com/LiPu-jpg/Openwrite/issues/85)：根包与全部内嵌子包的兼容声明加入精确宿主 `0.2.1-rc.2`，安装和激活不再被宿主版本闸门拒绝，无需版本豁免或手工编辑 `compatibility.json`。已在 `0.2.1-rc.2` 完成 Web 客户端验收（启动、OpenWrite 入口、工作台三页签、DoG 面板与其他插件共存），`0.1.2-rc.1`、`0.2.0-rc.2` 回归通过；Windows Electron 桌面端未实机复测。该版本完整验收见 [发布验收](docs/RELEASE_ACCEPTANCE.md#023)。升级后停止并重新启动 dsh，再刷新浏览器页面。
 
 `0.2.21` 修复 [#79](https://github.com/LiPu-jpg/Openwrite/issues/79) 的 Windows Desktop 目录选择：宿主接口失败后使用桌面提供的系统文件夹对话框，覆盖侧栏进入与工作台内目录选择。Core 保持 5.8.10。排查与手填路径方式见 [安装指南](docs/INSTALL.md#windows-desktop-目录选择)。
 

package/docs/COST_MEASUREMENT.md

@@ -11,7 +11,7 @@ npm --prefix packages/openwrite-bridge run build
 node --test packages/openwrite-bridge/scripts/tool-policy.test.mjs
 ```
 
-`real dsh tool registry isolates 90 tools` 会打印两行 JSON：
+`real dsh tool registry isolates 95 tools` 会打印两行 JSON：
 
 | 字段 | 含义 | 不是什么 |
 |---|---|---|
@@ -20,7 +20,7 @@ node --test packages/openwrite-bridge/scripts/tool-policy.test.mjs
 | `sdkBytes` | 注入到 system prompt 的 SDK 文本字节 | 不是费用 |
 | `tokenizer: not-measured` | 本仓库不内置某家 tokenizer | 不得改写成已测 token |
 
-同一套检查还断言：普通预设 `novel_*` 为 0；创作预设为 90 且不随 PTC 另一 scope 改变。
+同一套检查还断言：普通预设 `novel_*` 为 0；创作预设为 95 且不随 PTC 另一 scope 改变。
 
 要对比「请求大小」，把一次真实 `tools.schemas(agent)` 或 PTC `wireSchemas` 的 JSON 写入文件后对体积做 `wc -c` / SHA-256。这仍是请求体字节，不是 token。
 

package/docs/DEVELOPMENT.md

@@ -18,7 +18,7 @@
 
 | 路径 | 说明 |
 |---|---|
-| `packages/openwrite-bridge/` | dsh 插件：90 个 `novel_*` 工具，覆盖 Studio HTTP 动作面并提供隔离模型横评与脱敏 trace |
+| `packages/openwrite-bridge/` | dsh 插件：95 个 `novel_*` 工具，覆盖 Studio HTTP 动作面并提供隔离模型横评与脱敏 trace |
 | `packages/studio-panel/` | dsh web 原生工作台：创作 / 资料 / 任务，含规范章节导航、连续审读、场景双序、正文编辑器、双 DAG、模型测试台与工具卡 |
 | `presets/openwrite/` | OpenWrite 全流程 agent 预设（规划/资产/写章/评审/修订） |
 | `scripts/install.sh` | 构建插件、安装预设到 `~/.dsh/.agent-presets/`、插件装进 dsh profile |

package/docs/DSH_STORE.md

@@ -23,7 +23,7 @@ OpenWrite 是一个完整的小说创作插件套件。商城安装目标是仓
 
 alpha 浏览器验收进一步发现 persona 配置由 `text` 改成了必填 `prefix`；预设使用 YAML 锚点将同一份身份文本提供给两个字段，并以两版实际 Config / apply 验证身份只注册一次。
 
-构建基线仍是 `0.1.2-rc.1`。兼容声明覆盖精确版本 `0.1.2-rc.1`、`0.1.5-alpha.1`、`0.2.0-rc.2`、`0.2.1-rc.2`，并分别保留已完成的 Web 客户端验收；`0.1.3-alpha.1`、`0.1.3-alpha.2` 保持未知。这不自动升级用户宿主，也不承诺整个 alpha 或 0.2 范围都兼容。`0.2.23` 对 `0.2.1-rc.2` 的本机验证覆盖安装闸门、两个客户端激活、预设选择、工作台入口与三页签、DoG 面板及其他插件共存；完整平台矩阵和发布产物仍以下述 CI 报告为准。
+构建基线仍是 `0.1.2-rc.1`。兼容声明覆盖精确版本 `0.1.2-rc.1`、`0.1.5-alpha.1`、`0.2.0-rc.2`、`0.2.1-rc.2`，并分别保留已完成的 Web 客户端验收；`0.1.3-alpha.1`、`0.1.3-alpha.2` 保持未知。这不自动升级用户宿主，也不承诺整个 alpha 或 0.2 范围都兼容。`0.2.23` 已发布到 npm `latest` 与 GitHub Release，完整平台矩阵、源码安装和四份宿主客户端报告见[发布验收](RELEASE_ACCEPTANCE.md#023)；本机另复验了 `0.2.1-rc.2` 的安装闸门、客户端激活、预设选择、工作台入口与三页签、DoG 面板及其他插件共存。
 
 2026-09-09，从官方 npm Registry 直接安装 `@deepseek-ai/dsh@0.1.5-alpha.1` 时，上游范围依赖选中了正在发布的 alpha.2，其中 `@deepseek-ai/dsh-tool-pwsh-persistent@^0.1.5-alpha.2` 当时缺失，安装失败（ETARGET）。兼容测试使用 npm 的 `--before=2026-09-09T00:00:00Z` 解析 alpha.1 发布时的依赖，并保留解析锁。没有修改官方包源码或用 replacements/overrides 冒充官方组件。此证据仅适用于报告中的宿主依赖图，不代表后续浮动依赖安装也通过。
 

package/docs/ISSUES_88_95.md

@@ -0,0 +1,40 @@
+# Core 5.8.11：写作与恢复修复
+
+插件 0.2.24 / Core 5.8.11 修复 #88–#95，已发布到 [GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.24) 与 [npm](https://www.npmjs.com/package/dsh-openwrite/v/0.2.24)；分发状态和完整验收见 [发布验收](RELEASE_ACCEPTANCE.md#024)。测试使用假模型和隔离目录，不调用付费模型。
+
+| Issue | 行为 |
+| --- | --- |
+| #88 | 删章在项目写锁内清除验收 head、递增 revision、保存删除 operation；未完成旧 operation 不再允许 resume。 |
+| #89 | 账本有 head 而文件不存在时返回 `missing` 和 `CHAPTER_FILE_MISSING` warning，不阻断事实门禁；原文恢复后自动回到 current。 |
+| #90 | 结算格式错误最多重试两次，收紧提示且保留全部正文。耗尽后保存带降级标记的摘要，reconcile task 返回 `completed`、`degraded: true`。网络、认证等异常仍返回实际失败。 |
+| #91 | 写作 Agent、修订、语料分析/打包与结算使用 `OW_AGENT_STREAM`（默认 1）；流式请求使用读空闲超时，保留已接收的短片段。`OW_LLM_TIMEOUT` 可覆盖超时，默认 300 秒，默认 SDK 重试一次。 |
+| #92 | HTTP `POST /api/write` 提交 `chapter_write` 后立即返回 `task_id`。任务详情保留原有 `phase`，另提供 context/plan/draft/fact_extract/settle/validate/commit 的 `stage`、`run_id_v2`。后台任务保留原始工具调用追踪。锁每 15 秒更新心跳，死进程锁在下一次获取时回收。 |
+| #93 | 上下文、writer 和 reviewer 共用 `chapter_word_bounds`（目标 -20%/+50%）和 `count_writing_units`。4200/5000 单位无需补写；真正超限先尝试两次唯一文本锚定修改，再允许一次整章调整。 |
+| #94 | 删章派生数据移入 `data/chapter_deletions/<operation_id>/`，备份地址指向归档。未提交草稿有列表和显式恢复接口；恢复使用原 run id、逐字节复制，不生成新正文。 |
+| #95 | 任务写入 `finished_at`，读端兼容旧 `completed_at` 并返回 `duration_ms`。导出预检使用阅读顺序的缺章判定；诊断默认只显示近 24 小时的历史事件，历史数量单独返回，成功任务重置失败连续计数。未知 API 路由返回 404 `ROUTE_NOT_FOUND`，方法错误返回 405 `METHOD_NOT_ALLOWED` 和 `Allow`。 |
+
+## 草稿恢复
+
+1. `GET /api/manuscript/drafts` 返回 chapter_id、run_id_v2、draft_path、draft_artifact、word_count、created_at、draft_revision 和 restorable。
+2. `POST /api/manuscript/drafts/restore` 传入 `run_id_v2`、`confirm: true`。覆盖已有正文还必须传入当前正文的 `expected_revision`；缺失正文传空字符串。
+3. 响应返回 `operation` 和 `requires_reconcile: true`。原始草稿已经写回正文；运行 acceptance reconcile 后才会宣告事实 current。恢复正文这一步不调用模型。
+
+同步 Python 应用方法 `write_next_chapter()` 保持可用。HTTP 写作的结果通过 `/api/tasks/<task_id>` 读取，错误中的 `details` 包含已落盘草稿路径。取消或流中断也保留可发现的正文片段。
+
+诊断支持 GET/POST `/api/diagnostics` 的 `since`（ISO 时间）与 `window_hours`；CLI 为 `openwrite diagnose --window 24 --since <ISO>`。缺章、当前设定冲突等仍存在的状态问题持续报告；时间窗口只淘汰已发生的任务、运行和传输事件。
+
+## 验证范围
+
+回归覆盖五轮删章/重建、缺失文件恢复、三次非 mapping 降级、原字节/CRLF 草稿恢复、取消与流中断保稿、后台进度、错误恢复路径、HTTP 404/405、过期诊断及统一字数门限。没有将真实服务商连续 20 次请求或 Windows Electron 实机验收标记为通过。
+
+## dsh 工具入口
+
+`novel_manuscript_edit_action(action="drafts")` 列出未提交草稿；`action="restore_draft"` 传 `run_id_v2`、`confirm: true`，覆盖已有正文时用 `revision` 传当前正文 SHA。工具将其映射到 Core 的 `expected_revision`。恢复属于本地写入操作；不会按生成任务执行。随后使用现有验收调和工具处理返回的 operation。
+
+#86、#87 的精确宿主 `0.2.1-rc.2` 声明已经包含在 0.2.23，本版继续保留该兼容范围。
+
+插件验证：`npm run check:plugin` 通过（构建、doctor、36 项生命周期检查、预设与双 profile、bridge/panel smoke、14 项 epoch、298 项组件）；`npm run test:managed` 23 项通过，跨仓库 contracts 通过。
+
+核心全量 pytest：1409 passed、32 skipped；新 wheel 另做打包内容与针对性回归检查。发布版本为插件 0.2.24 / Core 5.8.11。
+
+新 wheel 解包后隔离导入并执行针对性回归：11 passed。已合并核心 [PR #96](https://github.com/LiPu-jpg/Openwrite/pull/96) 与插件 [PR #97](https://github.com/LiPu-jpg/Openwrite/pull/97)。

package/docs/ISSUES_98_100.md

@@ -0,0 +1,59 @@
+# #98–#100：连载写作与运行可靠性
+
+本分支为 dsh-openwrite 0.2.25 / Core 5.8.12 候选，尚未发布。Core 代码位于同仓库 native-core 分支；安装包将捆绑对应 wheel。完整数据契约见 Core `docs/SERIAL_WRITING.md`。
+
+## 覆盖范围
+
+| Issue | 实现 |
+| --- | --- |
+| #98.1 插件升级破坏 bundle | 新增停机离线旁路升级工具，复制整个 profile 后安装、校验、切换，保留旧目录和日志；支持修复不完整 OpenWrite bundle 与中断恢复。原生 dsh 安装器和宿主启动降级未修改，见下方限制。 |
+| #98.2 外部验收失败 | `no_model=true` / `mode=fast` 承认正文版本，派生事实明确 stale，可稍后完整对账。 |
+| #98.3 误连实例 | status/响应头标识 root、instance、port；Bridge 校验实际根，不一致返回 409。 |
+| #98.4、#99.1 长章 | 按 beats 分段，携带 1,500 字尾文续写，拼接去重；不足时续写，预算耗尽保留草稿。 |
+| #98.5 已定稿章横评 | 显式选择历史隔离改写，产物不写正式正文/事实账本；标注当前正典上下文限制。 |
+| #98.6 自留存档 | `personal_archive` 档位将质量门禁降为警告并写 hash 回执，结构错误仍阻断。 |
+| #98.7 心跳与超时 | 活动、worker 心跳、当前模型调用时长分别显示；流式总期限有界，超时保留部分输出。 |
+| #98.8 读写互斥 | GET 与纯读诊断不被写任务统一阻断；修改仍持项目写锁。 |
+| #98.9 章级体检 | 汉字/段落、句长和段长 P50/P90、对白占比、项目词典名词密度，写后自动保存。 |
+| #99.2–#99.6 写作资产 | anchor 范本、共享术语表及例句、最低次数、review 开关、12 字窗口重合/重复、角色属性分组轮换和明确已用标记。 |
+| #99.7 用量 | reasoning_tokens 从上游嵌套 usage 提升，真实 0 与未知分别保留。 |
+| #100 章节组审计 | 章纲四字段、只读计数与缺失报告、声明伏笔连边、漂移预算与冻结、吸收/回收/冻结裁决、不可变基线和事务恢复。 |
+
+工作台「操作 → 连载体检」提供资产编辑、范本导入、章节体检、章纲元数据、漂移记账与裁决。相关工具为 `novel_writing_assets`、`novel_manuscript_metrics`、`novel_outline_audit`、`novel_drift_entry`、`novel_arc_reconcile`；当前预设共 95 个领域工具。大纲吸收要求真实版本化更新；界面里的承诺计数是声明检查，不能当作已验证的故事推进。
+
+## 离线升级和损坏修复
+
+先完整退出 DeepSeek Harness / DSH Desktop 并停止 dsh、OpenWrite managed runtime。整个升级过程中保持停机。工具会在安装前及切换前检查进程；未知进程状态也拒绝修改。宿主原生命令不受此检查保护。
+
+从解压的候选/发行包运行（或已安装版本提供的 `openwrite-safe-upgrade`）：
+
+```sh
+node /path/to/package/scripts/safe-upgrade.mjs --profile desktop --package /absolute/path/dsh-openwrite-0.2.25.tgz
+```
+
+候选版本未上架前不要使用 `dsh-openwrite@0.2.25` 下载命令。`DSH_HOME` 必须指向实际宿主的 home；默认 `~/.dsh`。Windows 另设 `DSH_BIN` 为已验收宿主运行时的绝对 `lib/bin.js` 路径，使用 Node 直接执行而非 `.cmd` shell。其余平台可通过 DSH_BIN 指定同一受信任 CLI。
+
+安装在旁路 profile 中进行；整个旧 profile（包括其他插件、锁文件及设置）在切换前保持可用。只有 OpenWrite 的必要文件和 manifest 校验成功才进入目录切换。切换为两次同盘 rename，**不是整个目录树的单次原子交换**；停机状态避免读取窗口，中断由持久事务记录恢复。锁覆盖本工具的并发操作，不能阻止未经协调的原生安装器或用户重新启动宿主。
+
+记录保存在 `<DSH_HOME>/openwrite/upgrades/upgrade_*/`：`operation.json`、`pnpm.log`、成功切换后的 `previous-profile/`。stdout/stderr 实时落盘；安装 180 秒超时终止子进程。保留目录可能占用较多空间，确认升级稳定后再人工清理。
+
+```sh
+# 中断恢复，只恢复同一批准操作；若后来修改过 profile 则拒绝覆盖
+node /path/to/package/scripts/safe-upgrade.mjs --recover upgrade_<实际操作ID>
+
+# 损坏 bundle 修复：先检查，再明确应用。仅禁用不完整的 OpenWrite bundle。
+node /path/to/package/scripts/safe-upgrade.mjs --profile desktop --repair
+node /path/to/package/scripts/safe-upgrade.mjs --profile desktop --repair --apply
+```
+
+修复会备份 profile manifest，不删除正文、其他插件或设置；不会假装重装已经成功。完整 OpenWrite bundle 不会被该修复入口禁用。之后使用旁路升级重新安装。
+
+## 验收与边界
+
+自动化覆盖分段续写/去重、版本冲突、术语与 Unicode 位置、快验收及后续完整对账、持双重写锁时纯读、心跳不伪造进度、流式期限、漂移超预算及中断恢复、存档回执、升级失败时旧 profile 不变和非空日志。前端测试覆盖连载体检读取/资产保存/版本化追认；手工浏览器验收仅使用 `~/my_novel`。
+
+这些测试不调用付费模型。真实模型达到 10,000 字的稳定率仍由模型决定；系统保证有限续写与草稿留存，不保证模型质量。具象名词密度依赖作者词典；承诺是否兑现未做语义判断。历史横评使用当前正典，不声称复原当时章前状态。伏笔声明边与 DAG 已确认状态分别报告，不自动篡改已收状态。
+
+Windows Desktop 的文件句柄冲突与恢复需 Windows 实机补验；本次在 macOS 完成实现和本地回归。原生 `dsh plugin add/update` 与「损坏 bundle 自动跳过并提供修复入口」属于 deepseek-harness 宿主代码，未在这个仓库修复，因此 #98 不应标为全部解决。现有宿主用户必须使用本工具的停机路径，或等宿主更新后再走原生更新。
+
+本地回归记录（2026-10-23，macOS arm64）：Core 全量 1,424 passed / 32 skipped；后续超时配置校验等定向回归 38 passed、规划转换回归 36 passed；前端 301 passed；managed runtime/策略 24 passed；离线升级 5 passed；跨端契约检查通过。Core wheel 已从独立解压目录验证版本、体检模块及 HTTP 路由。真实服务商调用、Windows 实机和发布操作均未执行。

package/docs/RELEASE_ACCEPTANCE.md

@@ -6,13 +6,23 @@
 
 新建作品的验收请求最多等待 60 秒（原为 15 秒），保留成功耗时和失败诊断；不重试写请求，超时仍失败。此调整容纳慢速运行器，不代表初始化耗时问题已被修复。
 
-## 0.2.23（候选）
+## 0.2.24
 
-修复 [#83](https://github.com/LiPu-jpg/Openwrite/issues/83)/[#84](https://github.com/LiPu-jpg/Openwrite/issues/84)/[#85](https://github.com/LiPu-jpg/Openwrite/issues/85)：根包与全部内嵌子包的 `peerDependencies`、`dsh.compatibility.dsh` 和 `dshReleases` 加入精确宿主 `0.2.1-rc.2`。宿主版本闸门读取的是 dsh peer 范围，因此只修改展示用的 `dsh.compatibility` 不能让旧包通过；本版四个包同步声明后，安装与激活均不需要版本豁免。
+[GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.24) 与 [npm `dsh-openwrite@0.2.24`](https://www.npmjs.com/package/dsh-openwrite/v/0.2.24) 均已发布，npm `latest` 已更新到 0.2.24。直接使用[完整发布 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/39716542888) 验收通过的原始产物，未重新打包。来源提交 `d16623c6b0d2a84d45dffe23c24e641df790175f`，源码 tree `e0b2fd0647114e46e6fd188f4fb32b13fadc6b3d` 与主分支合并提交 `32166de6a3f66d8ebb2e3b5bd3912c1c1236f07b` 一致。npm 包 SHA-256 为 `a9fb766e8277b0f7dfecd9ea2284720512583e766e82515b70a529c967aa0170`；从官方 registry 无认证重新下载后，包字节、SHA-1、SHA-512 integrity 与 CI 产物和 registry 元数据均一致。
 
-本机 macOS arm64 / Node v24.15 已使用官方 registry 的精确 `@deepseek-ai/dsh@0.2.1-rc.2` 完成候选包验证：tarball 经 pnpm 安装进隔离 profile，宿主启动不再出现 `incompatible with dsh` / `skipping profile bundle`；浏览器中两个客户端入口激活，OpenWrite 入口和作品启动器可用。`client-boot-smoke` 在 `0.2.1-rc.2` 上通过版本闸门、预设选择、工作台三页签、DoG 面板与其他插件共存检查；`0.1.2-rc.1`、`0.2.0-rc.2` 回归通过。真实模型调用为 0。
+8 份原生安装/源码报告与 4 份客户端契约报告全部通过，覆盖 Linux、macOS arm64/x64、Windows、Node 22.19/24/26、源码安装、alpha 兼容与精确宿主 `0.1.2-rc.1` / `0.1.5-alpha.1` / `0.2.0-rc.2` / `0.2.1-rc.2`。GitHub 四份公开附件重新下载后均与 CI 原文件逐字节一致；合并后的[主分支完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/39717124199) 也已通过。源码安装命令的超时从两分钟调整为五分钟，保留原有成功条件，并补充退出状态、信号和错误码诊断。
 
-完整六平台/Node、源码安装、CI 产物及 npm 发布尚未由本提交的 `Release artifact validation` 运行确认。在这些报告和同字节产物通过前，0.2.23 只作为候选版本，不计为已发布；Windows Electron 桌面端原生启动与恢复仍须对应平台实机验收。
+修复 #88–#95，合并核心 [PR #96](https://github.com/LiPu-jpg/Openwrite/pull/96) 与插件 [PR #97](https://github.com/LiPu-jpg/Openwrite/pull/97)。捆绑 Core 5.8.11，wheel 来源提交 `596709e3f4bbe4eab7ce59513c9f36c4390ad367` 已包含在 `native-core` 分支，wheel SHA-256 为 `7937914eee6b99cd64fee2ead96a5e6d63e807acae0a610c1c4abf863e814188`。行为与草稿恢复步骤见[修复说明](ISSUES_88_95.md)。
+
+核心全量 pytest：1409 passed、32 skipped；wheel 解包后隔离回归：11 passed。插件构建、doctor、生命周期、双 profile、预设、bridge/panel smoke、14 项 epoch、298 项组件、23 项受管理运行时检查及跨仓库 contracts 通过。真实服务商连续请求与 Windows Electron 实机验收未运行，不计入通过。
+
+## 0.2.23
+
+[GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.23) 与 npm `dsh-openwrite@0.2.23` 已发布，npm `latest` 已更新到 0.2.23。使用[主分支完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/39219394324) 验收的原始产物，来源提交 `341408b80075fe4cc429f30a3327bcbdf2e69903`，npm 包 SHA-256 为 `8eb13e3de31b2144f1d5a57094abff38951c0329058520c96331aed03c735470`。8 份安装/源码报告与 4 份宿主客户端报告（含精确 `0.2.1-rc.2`）全部通过；GitHub 四份附件逐一与 CI 原文件比对一致，官方 npm registry 无认证下载包的字节与 SHA-256 也一致。
+
+修复 [#83](https://github.com/LiPu-jpg/Openwrite/issues/83)/[#84](https://github.com/LiPu-jpg/Openwrite/issues/84)/[#85](https://github.com/LiPu-jpg/Openwrite/issues/85)：根包与全部内嵌子包的 `peerDependencies`、`dsh.compatibility.dsh` 和 `dshReleases` 加入精确宿主 `0.2.1-rc.2`。宿主版本闸门读取的是 dsh peer 范围，因此只修改展示用的 `dsh.compatibility` 不能让旧包通过；本版四个包同步声明后，安装与激活均不需要版本豁免。发布矩阵新增 `0.2.1-rc.2` 客户端契约，并与 `0.1.2-rc.1`、`0.1.5-alpha.1`、`0.2.0-rc.2` 一起通过。
+
+本机 macOS arm64 / Node v24.15 另使用官方 registry 的精确 `@deepseek-ai/dsh@0.2.1-rc.2` 完成安装级复验：tarball 经 pnpm 安装进隔离 profile，宿主启动不再出现 `incompatible with dsh` / `skipping profile bundle`；浏览器中两个客户端入口激活，OpenWrite 入口和作品启动器可用。真实模型调用为 0。Windows Electron 桌面端原生启动与恢复仍须对应平台实机验收，本版未计为通过。
 
 ## 0.2.22
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-openwrite",
-  "version": "0.2.23",
+  "version": "0.2.25",
   "description": "OpenWrite 长篇小说领域后端接入 DeepSeek Harness：统一创作 Agent 预设 + 原生创作工作台",
   "type": "module",
   "scripts": {
@@ -14,7 +14,7 @@
     "install-dsh": "bash scripts/install.sh",
     "dev": "bash scripts/dev.sh",
     "doctor": "node scripts/plugin-doctor.mjs",
-    "test:maintenance": "node --test scripts/plugin-doctor.test.mjs scripts/lifecycle.test.mjs packages/openwrite-bridge/scripts/runtime-lifecycle.test.mjs scripts/maintenance.test.mjs scripts/preset-lifecycle.test.mjs",
+    "test:maintenance": "node --test scripts/plugin-doctor.test.mjs scripts/lifecycle.test.mjs packages/openwrite-bridge/scripts/runtime-lifecycle.test.mjs scripts/maintenance.test.mjs scripts/preset-lifecycle.test.mjs scripts/safe-upgrade.test.mjs",
     "check": "npm run check:plugin && npm run test:contracts",
     "test:lifecycle": "node --test scripts/lifecycle.test.mjs",
     "test:runtime": "node --test packages/openwrite-bridge/scripts/runtime-lifecycle.test.mjs",
@@ -95,6 +95,7 @@
     "packages/studio-panel/package.json",
     "scripts/dog",
     "scripts/maintenance.mjs",
+    "scripts/safe-upgrade.mjs",
     "packages/openwrite-bridge/LICENSE",
     "packages/studio-panel/LICENSE",
     "release/licenses",
@@ -119,7 +120,8 @@
     "@deepseek-ai/dsh-tools": "0.1.2-rc.1 || 0.1.5-alpha.1 || 0.2.0-rc.2 || 0.2.1-rc.2"
   },
   "bin": {
-    "openwrite-maintenance": "scripts/maintenance.mjs"
+    "openwrite-maintenance": "scripts/maintenance.mjs",
+    "openwrite-safe-upgrade": "scripts/safe-upgrade.mjs"
   },
   "repository": {
     "type": "git",

package/packages/openwrite-bridge/lib/client.d.ts

@@ -1,4 +1,6 @@
 import type { BackendConnection } from './managed-runtime.js';
+/** Windows roots are case-insensitive and may use either separator. */
+export declare function workspaceRootsEqual(left: string, right: string, platform?: NodeJS.Platform): boolean;
 /**
  * Shared HTTP client for the OpenWrite Studio action surface.
  *

package/packages/openwrite-bridge/lib/client.js

@@ -1,6 +1,14 @@
 import { mkdir, writeFile } from 'node:fs/promises';
-import { join } from 'node:path';
+import { join, resolve, win32 } from 'node:path';
 import { randomUUID } from 'node:crypto';
+import { realpathSync } from 'node:fs';
+/** Windows roots are case-insensitive and may use either separator. */
+export function workspaceRootsEqual(left, right, platform = process.platform) {
+    const normalize = platform === 'win32'
+        ? (value) => win32.normalize(value).replace(/[\\/]+$/, '').toLowerCase()
+        : (value) => resolve(value);
+    return normalize(left) === normalize(right);
+}
 /** Header OpenWrite requires on every write request (POST/PUT). */
 const WRITE_HEADER = 'X-OpenWrite-Studio';
 /** HTTP header values are bytes, so Unicode paths need explicit wire encoding.
@@ -183,9 +191,21 @@ export class StudioClient {
             const connection = await this.connect?.();
             const target = connection ? new URL(url.pathname + url.search, connection.baseUrl) : url;
             const headers = connection?.token ? { ...init.headers, Authorization: `Bearer ${connection.token}` } : init.headers;
-            return await fetch(target, { ...init, headers, signal: AbortSignal.any(signals) });
+            const response = await fetch(target, { ...init, headers, signal: AbortSignal.any(signals) });
+            const actual = response.headers.get('X-OpenWrite-Project-Path');
+            if (response.ok && actual && this.context?.workspaceRoot) {
+                const expected = realpathSync(this.context.workspaceRoot);
+                const received = decodeURIComponent(actual);
+                if (!workspaceRootsEqual(received, expected)) {
+                    await response.body?.cancel();
+                    throw new StudioError('Connected workspace differs from requested project', 409, 'WORKSPACE_CONTEXT_MISMATCH', { requested_path: expected, project_path: received });
+                }
+            }
+            return response;
         }
         catch (error) {
+            if (error instanceof StudioError)
+                throw error;
             if (this.context?.workspaceRoot) {
                 try {
                     const root = join(this.context.workspaceRoot, '.openwrite', 'transport-errors');

package/packages/openwrite-bridge/lib/domain.js

@@ -1,5 +1,5 @@
 import { Service } from '@deepseek-ai/cordis';
-import { StudioClient, workspaceRootHeaders, mutationAffectsState } from './client.js';
+import { StudioClient, workspaceRootHeaders, workspaceRootsEqual, mutationAffectsState } from './client.js';
 import { workspaceContextFromExec } from './tools.js';
 export const CONFIG_ROUTE = '/studio-panel/config.json';
 export const API_PROXY_ROUTE = '/studio-panel/api';
@@ -9,6 +9,8 @@ const PROXY_TIMEOUT_MS = 60_000;
 /** Cap for the best-effort upstream context_epoch read; never blocks a snapshot longer. */
 const CONTEXT_EPOCH_TIMEOUT_MS = 5_000;
 const WRITABLE_PATHS = new Set([
+    'writing-assets',
+    'arc-reconcile',
     'assets',
     'assets/update',
     'assets/package/import',
@@ -37,6 +39,7 @@ const WRITABLE_PATHS = new Set([
     'document/change-plan',
     'structured/change-plan',
     'manuscript-editing',
+    'manuscript/drafts/restore',
     'manuscript/acceptance/baseline',
     'manuscript/acceptance/external',
     'manuscript/acceptance/reconcile',
@@ -84,7 +87,7 @@ function resourceForPath(path) {
         return 'tasks';
     if (path.startsWith('/api/project-archives'))
         return 'workspace';
-    if (path.startsWith('/api/document') || path.startsWith('/api/write') || path.startsWith('/api/manuscript/acceptance'))
+    if (path.startsWith('/api/document') || path.startsWith('/api/write') || path.startsWith('/api/manuscript/'))
         return 'manuscript';
     if (path.startsWith('/api/outline'))
         return 'outline';
@@ -473,6 +476,12 @@ export class NovelDomainService extends Service {
                 });
                 if (lifetime.aborted)
                     return;
+                const actualRoot = upstream.headers.get('X-OpenWrite-Project-Path');
+                if (upstream.ok && actualRoot && !workspaceRootsEqual(decodeURIComponent(actualRoot), resolved.root)) {
+                    await upstream.body?.cancel();
+                    sendJson(res, 409, { error: 'Connected workspace differs from requested project', code: 'WORKSPACE_CONTEXT_MISMATCH', details: { requested_path: resolved.root, project_path: decodeURIComponent(actualRoot) } });
+                    return;
+                }
                 const upstreamBytes = Buffer.from(await upstream.arrayBuffer());
                 const bytes = compactTaskList(upstreamBytes, pathPart, method);
                 const responseHeaders = {
@@ -481,6 +490,11 @@ export class NovelDomainService extends Service {
                     'content-length': String(bytes.length),
                 };
                 const disposition = upstream.headers.get('content-disposition');
+                const instanceId = upstream.headers.get('X-OpenWrite-Instance-Id');
+                if (instanceId)
+                    responseHeaders['X-OpenWrite-Instance-Id'] = instanceId;
+                if (actualRoot)
+                    responseHeaders['X-OpenWrite-Project-Path'] = actualRoot;
                 if (disposition !== null)
                     responseHeaders['content-disposition'] = disposition;
                 res.writeHead(upstream.status, responseHeaders);

package/packages/openwrite-bridge/lib/tool-policy.js

@@ -6,6 +6,7 @@ const reads = new Set([
     'novel_revision_get', 'novel_tasks_list', 'novel_continuity', 'novel_diagnostics',
     'novel_research_status', 'novel_research_report', 'novel_model_profiles',
     'novel_manuscript_acceptance', 'novel_export_preflight',
+    'novel_manuscript_metrics', 'novel_outline_audit',
 ]);
 const generates = new Set([
     'novel_write_chapter', 'novel_review_chapter', 'novel_multi_write', 'novel_task_create',
@@ -14,11 +15,12 @@ const generates = new Set([
     'novel_model_test', 'novel_model_embedding_test',
 ]);
 const readActions = {
+    novel_writing_assets: ['get'],
     novel_model_benchmark: ['list', 'get', 'options'],
     novel_chapter_run_action: ['list', 'get'],
     novel_rolling_plan_action: ['list', 'get'],
     novel_narrative_forecast_action: ['list', 'get'],
-    novel_manuscript_edit_action: ['versions', 'version', 'compare', 'annotations'],
+    novel_manuscript_edit_action: ['drafts', 'versions', 'version', 'compare', 'annotations'],
     novel_source_action: ['status_v2', 'profile_v2'],
     novel_reference_library_action: ['list', 'status', 'profile'],
     novel_runtime_skill_action: ['list', 'diagnose', 'resolve'],

package/packages/openwrite-bridge/lib/tools.js

@@ -178,6 +178,11 @@ function resolveReviewPath(args) {
 function objectValue(value) {
     return value !== null && typeof value === 'object' && !Array.isArray(value) ? value : {};
 }
+function requireRecord(value, name) {
+    if (value === null || typeof value !== 'object' || Array.isArray(value))
+        throw new Error(`${name} must be a JSON object`);
+    return value;
+}
 async function deliveryForResponse(client, response, chapterId, signal, threshold) {
     const embedded = objectValue(response)['workspace'];
     const workspace = embedded !== null && typeof embedded === 'object' && !Array.isArray(embedded)
@@ -979,7 +984,7 @@ export function registerNovelTools(ctx, clientFactory, options) {
             'so the downloaded file is bound to the manuscript state that was inspected.',
         parameters: {
             format: { type: 'string', enum: ['md', 'txt', 'epub'], description: 'Export format (default "md").' },
-            purpose: { type: 'string', enum: ['backup', 'delivery'], description: 'Export purpose (default "delivery"). Delivery applies publication gates; backup preserves the current work.' },
+            purpose: { type: 'string', enum: ['backup', 'delivery', 'personal_archive'], description: 'Export purpose (default "delivery"). personal_archive preserves author copies with gate findings recorded as warnings.' },
         },
         output: { schema: JSON_OUTPUT_SCHEMA, render: (_args, value) => renderJson(value) },
         timeoutMs,
@@ -998,7 +1003,7 @@ export function registerNovelTools(ctx, clientFactory, options) {
             'preflight_revision unchanged to bind the download to that inspected manuscript state. Returns the saved file path.',
         parameters: {
             format: { type: 'string', enum: ['md', 'txt', 'epub'], description: 'Export format (default "md").' },
-            purpose: { type: 'string', enum: ['backup', 'delivery'], description: 'Export purpose (default "delivery").' },
+            purpose: { type: 'string', enum: ['backup', 'delivery', 'personal_archive'], description: 'Export purpose (default "delivery").' },
             preflight_revision: { type: 'string', description: 'Revision returned by novel_export_preflight for the same format and purpose.' },
         },
         output: {
@@ -2117,7 +2122,8 @@ export function registerNovelTools(ctx, clientFactory, options) {
     }));
     ctx.tools.register(defineTool({
         name: 'novel_manuscript_edit_action',
-        description: 'Chapter version snapshots and annotations. Actions: "versions" (chapter_id), "version" (chapter_id + ' +
+        description: 'Chapter snapshots, annotations and saved drafts. Actions: "drafts" (list recoverable uncommitted prose), ' +
+            '"restore_draft" (run_id_v2 + revision + confirm: true; copy the exact saved bytes without generation, then reconcile facts). Other actions: "versions" (chapter_id), "version" (chapter_id + ' +
             'version_id — returns the snapshot content), "checkpoint" (chapter_id, optional label — save a snapshot), ' +
             '"compare" (chapter_id + version_id — preview current-to-snapshot differences without writing), ' +
             '"restore" (chapter_id + version_id + revision — REVERTS the chapter to a snapshot; revision-gated and ' +
@@ -2125,8 +2131,9 @@ export function registerNovelTools(ctx, clientFactory, options) {
             'revision + quote + note, optional start_hint/end_hint and color amber|rose|sky|lime|violet), ' +
             '"resolve_annotation" (chapter_id + annotation_id).',
         parameters: {
-            action: { type: 'string', required: true, enum: ['versions', 'version', 'compare', 'checkpoint', 'restore', 'annotations', 'annotate', 'resolve_annotation'], description: 'The manuscript-editing operation.' },
-            chapter_id: { type: 'string', required: true, description: 'Chapter id (ch_<digits>).' },
+            action: { type: 'string', required: true, enum: ['drafts', 'restore_draft', 'versions', 'version', 'compare', 'checkpoint', 'restore', 'annotations', 'annotate', 'resolve_annotation'], description: 'The manuscript-editing operation.' },
+            chapter_id: { type: 'string', description: 'Chapter id (ch_<digits>), required except for drafts/restore_draft.' },
+            run_id_v2: { type: 'string', description: 'Saved run id from drafts (restore_draft).' },
             version_id: { type: 'string', description: 'Snapshot id (version/restore).' },
             label: { type: 'string', description: 'Snapshot label (checkpoint).' },
             revision: { type: 'string', description: 'Current chapter revision (restore/annotate).' },
@@ -2140,10 +2147,19 @@ export function registerNovelTools(ctx, clientFactory, options) {
         },
         output: { schema: JSON_OUTPUT_SCHEMA, render: (_args, value) => renderJson(value) },
         timeoutMs,
-        isConcurrencySafe: (args) => args.action === 'versions' || args.action === 'version' || args.action === 'compare' || args.action === 'annotations',
+        isConcurrencySafe: (args) => args.action === 'drafts' || args.action === 'versions' || args.action === 'version' || args.action === 'compare' || args.action === 'annotations',
         async execute(args, exec) {
             const client = clientFor(exec);
-            if (!CHAPTER_ID_PATTERN.test(args.chapter_id))
+            if (args.action === 'drafts')
+                return await client.getJson('/api/manuscript/drafts', {}, exec.signal);
+            if (args.action === 'restore_draft') {
+                if (!args.run_id_v2)
+                    throw new Error('run_id_v2 is required; list drafts first');
+                return await client.postJson('/api/manuscript/drafts/restore', {
+                    run_id_v2: args.run_id_v2, expected_revision: args.revision ?? '', confirm: args.confirm === true,
+                }, exec.signal);
+            }
+            if (!args.chapter_id || !CHAPTER_ID_PATTERN.test(args.chapter_id))
                 throw new Error(`chapter_id must match ch_<digits>, got "${args.chapter_id}"`);
             const body = { action: args.action, chapter_id: args.chapter_id };
             for (const key of ['version_id', 'label', 'revision', 'confirm', 'quote', 'note', 'start_hint', 'end_hint', 'color', 'annotation_id']) {
@@ -2163,6 +2179,8 @@ export function registerNovelTools(ctx, clientFactory, options) {
             'acknowledge are confirmation-gated by the server.',
         parameters: {
             action: { type: 'string', required: true, enum: ['baseline', 'external', 'resume', 'acknowledge', 'abandon'], description: 'The reconciliation action.' },
+            mode: { type: 'string', enum: ['full', 'fast'], description: 'resume: fast accepts frozen manuscript without calling a model; derived facts remain stale and can be reconciled later.' },
+            no_model: { type: 'boolean', description: 'external: accept manuscript bytes without extracting facts; does not clear the facts guard.' },
             chapter_id: { type: 'string', description: 'Optional chapter id (ch_<digits>); omit for the project-level operation.' },
             operation_id: { type: 'string', description: 'Durable operation id to resume or acknowledge.' },
             domains: { type: 'array', items: { type: 'string' }, description: 'Affected domains to acknowledge or retry; acknowledge defaults to outline and foreshadowing.' },
@@ -2175,7 +2193,7 @@ export function registerNovelTools(ctx, clientFactory, options) {
                 throw new Error(`chapter_id must match ch_<digits>, got "${args.chapter_id}"`);
             }
             const body = {};
-            for (const key of ['chapter_id', 'operation_id', 'domains', 'confirm']) {
+            for (const key of ['chapter_id', 'operation_id', 'domains', 'confirm', 'mode', 'no_model']) {
                 const value = args[key];
                 if (value !== undefined)
                     body[key] = value;
@@ -2186,6 +2204,89 @@ export function registerNovelTools(ctx, clientFactory, options) {
             return await clientFor(exec).postJson(`/api/manuscript/acceptance/${route}`, body, exec.signal);
         },
     }));
+    ctx.tools.register(defineTool({
+        name: 'novel_writing_assets',
+        description: 'Read or save shared project terminology, review_terms toggle, concrete noun lexicon, overlap_threshold and character_groups. Anchors are Markdown files in the novel anchor/ directory. Save requires the revision returned by get. Character groups are optional descriptive material; retain identity and canonical state in character cards.',
+        parameters: {
+            action: { type: 'string', required: true, enum: ['get', 'save', 'save_anchor', 'mark_usage'], description: 'Asset operation.' },
+            assets: { type: 'json', description: 'Object: terms [{term, forbidden?, replacement?, min_occurrences?, example?}], character_groups {name: [{id,text}]}, review_terms?, concrete_nouns?, overlap_threshold?.' },
+            expected_revision: { type: 'string', description: 'Revision from get; empty for a new asset table.' },
+            name: { type: 'string', description: 'save_anchor: simple .md filename.' },
+            content: { type: 'string', description: 'save_anchor: Markdown reference prose.' },
+            chapter_id: { type: 'string', description: 'mark_usage: chapter whose optional character details were used.' },
+            groups: { type: 'json', description: 'mark_usage: character name to group IDs used in this chapter.' },
+        },
+        output: { schema: JSON_OUTPUT_SCHEMA, render: (_args, value) => renderJson(value) },
+        timeoutMs,
+        isConcurrencySafe: args => args.action === 'get',
+        async execute(args, exec) {
+            const client = clientFor(exec);
+            if (args.action === 'get')
+                return await client.getJson('/api/writing-assets', {}, exec.signal);
+            if (args.action === 'save_anchor')
+                return await client.postJson('/api/writing-assets', { action: args.action, name: args.name ?? '', content: args.content ?? '', expected_revision: args.expected_revision ?? '' }, exec.signal);
+            if (args.action === 'mark_usage')
+                return await client.postJson('/api/writing-assets', { action: args.action, chapter_id: args.chapter_id ?? '', groups: requireRecord(args.groups, 'groups') }, exec.signal);
+            const assets = requireRecord(args.assets, 'assets');
+            return await client.postJson('/api/writing-assets', { assets: assets, expected_revision: args.expected_revision ?? '' }, exec.signal);
+        },
+    }));
+    ctx.tools.register(defineTool({
+        name: 'novel_manuscript_metrics',
+        description: 'Read deterministic chapter metrics, terminology findings, overlap against anchors and the previous three chapters, duplicate paragraphs and selected character groups. Overlap spans use Unicode codepoint offsets. No model call.',
+        parameters: { chapter_id: { type: 'string', required: true, description: 'Canonical chapter ID.' } },
+        output: { schema: JSON_OUTPUT_SCHEMA, render: (_args, value) => renderJson(value) },
+        timeoutMs, isConcurrencySafe: () => true,
+        async execute(args, exec) {
+            if (!CHAPTER_ID_PATTERN.test(args.chapter_id))
+                throw new Error('Invalid chapter_id');
+            return await clientFor(exec).getJson(`/api/manuscript/${encodeURIComponent(args.chapter_id)}/metrics`, {}, exec.signal);
+        },
+    }));
+    ctx.tools.register(defineTool({
+        name: 'novel_outline_audit',
+        description: 'Read countable chapter-group facts: budgets/actual words, declared promises and function streaks, declared/confirmed payoffs, overdue hooks and drift ledger. Never generates a score or claims prose fulfilled a promise. No model call.',
+        parameters: {
+            group_id: { type: 'string', description: 'Optional sec_* or arc_* group.' },
+            start: { type: 'string', description: 'Optional first chapter ID.' },
+            end: { type: 'string', description: 'Optional last chapter ID.' },
+        },
+        output: { schema: JSON_OUTPUT_SCHEMA, render: (_args, value) => renderJson(value) },
+        timeoutMs, isConcurrencySafe: () => true,
+        async execute(args, exec) {
+            return await clientFor(exec).getJson('/api/outline/audit', { group_id: args.group_id ?? '', start: args.start ?? '', end: args.end ?? '' }, exec.signal);
+        },
+    }));
+    ctx.tools.register(defineTool({
+        name: 'novel_drift_entry',
+        description: 'Record author-observed chapter drift. Saves an immutable outline/manuscript baseline first. More than two unresolved rewrite drifts in a section freezes generation until explicit reconciliation and replanning.',
+        parameters: {
+            group_id: { type: 'string', required: true, description: 'Section ID sec_*.' },
+            chapter_id: { type: 'string', required: true, description: 'Chapter in this section.' },
+            kind: { type: 'string', required: true, enum: ['growth', 'escape', 'dilution', 'snowball'], description: 'Drift kind.' },
+            impact: { type: 'string', required: true, enum: ['local', 'section', 'arc'], description: 'Impact scope.' },
+            rewrite: { type: 'boolean', description: 'Counts toward rewrite budget; default true.' },
+            note: { type: 'string', required: true, description: 'Observed drift and supporting evidence.' },
+            expected_revision: { type: 'integer', required: true, description: 'ledger.revision from outline audit.' },
+        },
+        output: { schema: JSON_OUTPUT_SCHEMA, render: (_args, value) => renderJson(value) },
+        timeoutMs,
+        async execute(args, exec) {
+            return await clientFor(exec).postJson('/api/arc-reconcile', { ...args, action: 'drift_entry' }, exec.signal);
+        },
+    }));
+    ctx.tools.register(defineTool({
+        name: 'novel_arc_reconcile',
+        description: 'Resolve selected drift entries by absorbing authored outline metadata, scheduling recovery in a later chapter, or freezing. Absorption requires concrete outline_updates and outline_revision; baseline is saved before edits. resume=true explicitly resumes a replanned group within budget. Does not rewrite manuscript.',
+        parameters: {
+            data: { type: 'json', required: true, description: 'Object: group_id, expected_revision, entry_ids[], decision absorb/recover/freeze, confirm:true. absorb: outline_revision, outline_updates[{node_id,metadata}], resume?. recover: plan, target_chapter_id.' },
+        },
+        output: { schema: JSON_OUTPUT_SCHEMA, render: (_args, value) => renderJson(value) },
+        timeoutMs,
+        async execute(args, exec) {
+            return await clientFor(exec).postJson('/api/arc-reconcile', { ...requireRecord(args.data, 'data'), action: 'arc_reconcile' }, exec.signal);
+        },
+    }));
     // ── style sources, reference library, runtime skills, rules ──────────────
     ctx.tools.register(defineTool({
         name: 'novel_source_action',
@@ -2382,13 +2483,14 @@ export function registerNovelTools(ctx, clientFactory, options) {
             run_id: { type: 'string', description: 'Benchmark run id for get.' },
             task_type: { type: 'string', enum: ['chapter', 'outline'], description: 'Run target: chapter (default) or outline. Outline supports framework mode only.' },
             chapter_id: { type: 'string', description: 'Chapter task target: next (default) or an available canonical chapter_id from options, e.g. ch_003. Unavailable targets are rejected without fallback; inspect availability/reason first.' },
+            allow_historical: { type: 'boolean', description: 'Explicit isolated rewrite of an already written chapter using current canon. This is not a reconstructed pre-chapter baseline. Source manuscript and ledgers remain unchanged.' },
             outline_start_chapter: { type: 'integer', description: 'Outline task first chapter number, inclusive, 1-100000. Omit to continue after the last outline chapter.' },
             outline_chapter_count: { type: 'integer', description: 'Outline task range length including its starting chapter, 1-20 (default 3).' },
             writer_profile_ids: { type: 'array', items: { type: 'string' }, description: 'Writing profiles to compare (run).' },
             reviewer_profile_ids: { type: 'array', items: { type: 'string' }, description: 'Independent blind-review profiles (run).' },
             execution_mode: { type: 'string', enum: ['framework', 'creative'], description: 'Execution path: full framework (default); creative diagnostic is available only for chapter tasks. Read task pipelines with options.' },
             repeats: { type: 'integer', description: 'Repeats per writing profile, 1-5.' },
-            target_words: { type: 'integer', description: 'Chapter candidate target length, 200-12000; use outline_chapter_count for outline range.' },
+            target_words: { type: 'integer', description: 'Chapter candidate target length, 200-50000; use outline_chapter_count for outline range.' },
             concurrency: { type: 'integer', description: 'Concurrent model jobs, 1-4.' },
             limit: { type: 'integer', description: 'List limit, default 20.' },
         },
@@ -2418,7 +2520,7 @@ export function registerNovelTools(ctx, clientFactory, options) {
                 writer_profile_ids: args.writer_profile_ids,
                 reviewer_profile_ids: args.reviewer_profile_ids,
             };
-            for (const key of ['task_type', 'chapter_id', 'outline_start_chapter', 'outline_chapter_count', 'execution_mode', 'repeats', 'target_words', 'concurrency']) {
+            for (const key of ['task_type', 'chapter_id', 'allow_historical', 'outline_start_chapter', 'outline_chapter_count', 'execution_mode', 'repeats', 'target_words', 'concurrency']) {
                 const value = args[key];
                 if (value !== undefined)
                     body[key] = value;