dsh-openwrite 0.2.11 → 0.2.14

package/README.md

@@ -29,12 +29,14 @@
 
 ## 快速安装
 
-**0.2.11 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**。0.2.0-rc.2 已完成 Web 客户端启动验证，Windows Electron 桌面端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)。
+**0.2.13 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**。0.2.0-rc.2 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)。
 
 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：
 
+**0.2.13 的 GitHub 安装包已发布，npm `latest` 暂时仍为 0.2.12，等待本次发布的 2FA 验证后同步发布。** 0.2.12 不包含 #57 的创作页刷新循环修复；请使用下方已发布的 0.2.13 Release 包。
+
 ```sh
-dsh plugin --profile web add -w https://github.com/LiPu-jpg/Openwrite/releases/download/v0.2.11/dsh-openwrite-0.2.11.tgz
+dsh plugin --profile web add -w https://github.com/LiPu-jpg/Openwrite/releases/download/v0.2.13/dsh-openwrite-0.2.13.tgz
 dsh web
 ```
 
@@ -64,9 +66,15 @@ dsh web
 
 ### dsh 0.2 桌面端升级
 
+`0.2.14` 候选包修复 [#60](https://github.com/LiPu-jpg/Openwrite/issues/60) 和 [#61](https://github.com/LiPu-jpg/Openwrite/issues/61)：系统代理下 `NO_PROXY` 含 `[::1]` 不再让所有模型连接测试误报"服务商拒绝了请求"（桥接层消毒 + Core 5.8.5 归类为本地配置错误）；Embedding 连接测试的依赖预热、硬超时与收尾兜底，加上桥接层的健康探活与 stderr 落盘，不再出现"后端活着但不服务、只能重启 dsh"的僵尸态。正式可下载版本以 npm `latest` 和 GitHub Release 为准。升级后需停止并重新启动 dsh，使桥接服务端加载新代码，再刷新浏览器页面。
+
+`0.2.13` 修复 [#57](https://github.com/LiPu-jpg/Openwrite/issues/57) 的创作页刷新循环：批注与版本读取不再作为写操作通知刷新，Core 的只读请求也不再增加工作区变更序号。正式可下载版本以 npm `latest` 和 GitHub Release 为准。升级后需停止并重新启动 dsh，使桥接服务端加载新代码，再刷新浏览器页面。
+
+`0.2.12` 进一步修复 [#50](https://github.com/LiPu-jpg/Openwrite/issues/50)、[#51](https://github.com/LiPu-jpg/Openwrite/issues/51)、[#52](https://github.com/LiPu-jpg/Openwrite/issues/52) 的预设注册、导航、工作台挂载和 DoG 接口。连续审读按内容修订刷新，避免轮询闪烁（[#54](https://github.com/LiPu-jpg/Openwrite/issues/54)）。
+
 `0.2.11` 修复 [#45](https://github.com/LiPu-jpg/Openwrite/issues/45)，适配精确宿主版本 `0.2.0-rc.2`：工作台兼容新的 turn-tail list 槽位，DoG 调试面板使用新的 `sessionStatus`；根包和所有子包的兼容声明同步更新。已验证 0.2.0-rc.2 的 Web 客户端启动、OpenWrite 入口、DoG 面板及其他插件共存；Windows Electron 桌面端恢复链路尚未实机复测。
 
-**`dsh-openwrite@0.2.10` 不包含这项修复，请升级到 `0.2.11`。** 不要通过给旧子包授权来代替升级：`dsh-openwrite` 的版本豁免不会覆盖 `@dsh-novel/studio-panel@<ver>` 和 `@dsh-external/dsh-dog@<ver>`，而旧客户端在 0.2.0-rc.2 上激活失败可能触发桌面端的 profile 恢复。
+**`dsh-openwrite@0.2.10` 不包含这项修复，请升级到 `0.2.13` 的 GitHub Release 包。** 不要通过给旧子包授权来代替升级：`dsh-openwrite` 的版本豁免不会覆盖 `@dsh-novel/studio-panel@<ver>` 和 `@dsh-external/dsh-dog@<ver>`，而旧客户端在 0.2.0-rc.2 上激活失败可能触发桌面端的 profile 恢复。
 
 若已发生恢复，先关闭桌面端并备份当前 profile，再从自己的 `cordis.patch.yml.bak-*` 备份恢复 patch，同时恢复原 `package.json` 的 `dsh.profile.bundles` 列表。安装包和作品目录通常仍在；不要用出厂模板覆盖其他插件配置。详细排查见 [安装指南](docs/INSTALL.md#dsh-02-客户端兼容排查)。
 

package/cordis.patch.yml

@@ -1,3 +1,12 @@
+# 0.2's carrier-neutral Connection does not require webServer itself, but its
+# custom RPC channel registrar reads that service from its declaring context.
+# OpenWrite targets Web/Desktop, where the carrier is present; declare it here
+# so authenticated DoG RPC channels can register without changing transport.
+- id: connection
+  inject:
+    - webRuntime
+    - webServer
+
 - insert:
     - id: openwrite-bundle
       name: dsh-openwrite

package/docs/INSTALL.md

@@ -29,6 +29,10 @@ Git 是可选的作品归档能力。未安装 Git 或 Git 初始化失败时，
 
 dsh 对话模型在宿主中配置。小说生成、评审与测试模型在「任务 → 模型」配置；安装健康检查不调用这些模型。
 
+托管模式的模型档案、路由和凭据读取 `$DSH_HOME/openwrite/state/config/`，不读取独立 Core 的 `~/.config/openwrite/`。优先使用上述界面配置；需要迁移时先停机、备份该目录，再核对「模型 → 路由」指向的档案 ID。新建自定义名称的档案后请核对路由，不能只改文件名。不要将配置或凭据提交进作品仓库。
+
+0.2.12 的连接测试与生成请求使用同一个已配置的输出 token 上限，推理模型可在该额度内完成思考并返回正文。连接测试会发起真实模型请求；安装与自动验收使用替身，不产生模型费用。
+
 ## 安装影响
 
 本轮原生验收覆盖 macOS 15、Ubuntu 和 Windows GitHub 原生运行器。
@@ -129,7 +133,13 @@ node scripts/client-boot-smoke.mjs --host-cli=/absolute/path/to/dsh-0.2.0-rc.2/n
 node scripts/client-boot-smoke.mjs dsh-openwrite-<version>.tgz --host-cli=/absolute/path/to/dsh-0.2.0-rc.2/node_modules/@deepseek-ai/dsh/lib/bin.js
 ```
 
-脚本需要 Playwright Chromium，创建临时 Web profile，检查无豁免版本闸门、两个客户端加载、OpenWrite 入口、DoG 面板及其他插件共存，不准备 Python 环境、不调用模型。它验证 Web 客户端兼容性；Windows Electron 的原生启动和恢复流程仍须在对应平台完成发布验收。
+脚本需要 Playwright Chromium，创建临时 Web profile，检查无豁免版本闸门、两个客户端加载、真实预设选择、会话导航、三个工作台页签出现和创作视图激活、DoG 面板及其他插件共存，后端就绪与未初始化响应由测试夹具提供，不准备 Python 环境、不调用模型；实际安装与后端动作由独立发布验收覆盖。它验证 Web 客户端兼容性；Windows Electron 的原生启动和恢复流程仍须在对应平台完成发布验收。
+
+## 0.2.12 的预设与审读修复
+
+0.2.11 仅补齐了客户端启动与版本声明，仍存在 #50–#52 的预设注册和导航断点。0.2.12 在新宿主中注册同一份版本化预设，并保留旧宿主的目录副本；不要通过内联手工补丁或版本豁免替代升级。升级不重绑已有会话的预设，新的作品入口使用对应版本的预设。
+
+连续审读在后台工作区轮询和任务失效时保持正文与滚动位置，仅在章节身份/顺序/修订变化或手动刷新时重载。重载期间保留上次正文，失败时显示错误。切换作品会清空旧作品的阅读状态。
 
 ## 作品信息与推理模型预算
 

package/docs/RELEASE_ACCEPTANCE.md

@@ -6,8 +6,36 @@
 
 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。
 
+## 0.2.14 候选包
+
+修复 #60 系统代理场景下 `NO_PROXY` 含 `[::1]` 导致所有模型连接测试误报"服务商拒绝了请求"：桥接层 `buildChildEnv()` 转发前剥掉带方括号的 IPv6 字面量，Core 5.8.5 把 httpx 客户端构造期异常归类为本地配置错误（HTTP 412），`Studio request failed` 附带 `error.cause`。
+
+修复 #61 Embedding 连接测试在请求线程里懒加载 numpy 把后端变成"活着但不服务"的僵尸：Core 5.8.5 启动主线程预热 numpy、后台线程预热 fastembed，`run_embedding_probe()` 硬超时 `timeout_seconds + 30 s`，收尾限时 10 秒超时强制退出；桥接层复用就绪连接前做带超时的 `/api/health` 探活，探活失败的僵尸进程强制终止并自动恢复，子进程 stderr 落盘 `state/logs/backend.log`（末尾 64 KiB）。
+
+新增回归覆盖 no_proxy 消毒矩阵、僵尸连接探活恢复、stderr 落盘、错误分类矩阵与探测硬超时；wheel 重打包后全量 RECORD 哈希校验（498 文件）与 manifest 一致性检查通过。完整发布验收正在运行，正式发布将使用通过门禁的同一份 CI 产物。Windows 上的依赖预热与僵尸恢复链路本机未复测，保留 0.2.13 的同一限制。
+
+## 0.2.13
+
+修复 #57 的只读 POST 失效循环。Core 5.8.4 在 HTTP action 契约声明是否改变公开状态，通过 `X-OpenWrite-Mutated` 统一驱动浏览器代理与 Agent 客户端；旧 Core 的正文版本和批注查询保留兼容处理。保存、批注写入及恢复版本按 manuscript 资源通知，批注读取不再监听无关的 workspace 失效。
+
+新增回归覆盖连续只读请求不增加 invalidation revision、SSE 和 Core context_epoch；真实写入仍增加序号。创作页验证重复 workspace 事件不会反复请求批注，实际编辑会刷新，切换作品清空旧批注并忽略迟到响应。[完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/37483398522) 全部通过，六组原生平台/Node、源码安装、alpha 兼容及三代宿主客户端共 11 份报告均绑定同一产物。来源提交 `72d19c1a1494da13b5b89d652746e13f13e90e1a`，SHA-256 为 `091dcb7201a88240d8282ef31526d7d88f6d63a390d0813a39f580f9be990a58`。[GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.13) 已发布并核对四份附件的校验值；npm 本次发布认证已过期（404），等待重新验证，`latest` 暂时仍为 0.2.12。
+
+#50、#57 已随 [PR #59](https://github.com/LiPu-jpg/Openwrite/pull/59) 合并关闭。#50 报告的预设注册、主视图挂载、会话导航及托管模型配置问题已由 0.2.12 修复，并经三代精确 Web 宿主和原生 Windows 安装检查验证。Windows Electron 的实机启动/恢复尚未复测，该限制继续保留；本版不将其计为已通过的验收。
+
+## 0.2.12
+
+[GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.12) 已发布，直接使用[完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/37389772004)验收的同一份包。来源提交 a349ea964e43d75f9fc98d1f83354a237bc2ff51，SHA-256 为 2545031292d80689f80d4c0d8c99e007c18ff0c477d662e2813cf67bc876246c。六组原生平台/Node、源码安装、alpha 兼容及三代宿主客户端共 11 份报告全部通过；详见 Release 的 release-acceptance.json。npm `dsh-openwrite@0.2.12` 和 GitHub Release 均已发布，`latest` 已更新到 0.2.12；从公共 npm registry 无认证重新下载后，SHA-256 与上述 CI / Release 包一致。合并后的[主分支完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/37390503895)也已通过。
+
+本版修复 #50–#52 的 0.2 宿主兼容问题及 #54 阅读模式刷新闪烁，纳入 #53 的导航和会话作用域改动，并保留旧宿主的预设插件。
+
+- Core 5.8.3：模型连接测试使用所选配置的输出预算；108 项 Core 回归通过。
+- 本地 297 项组件测试、50 项 DoG 测试通过。三代精确宿主的浏览器流程纳入正式发布门禁：选择目录、预设选择、会话导航、工作台标签、点击创作后实际渲染、DoG 面板与插件共存。该客户端检查使用后端就绪及未初始化响应替身；原生后端安装和作品初始化由独立平台任务验证。
+- 完整 CI 和唯一产物校验已通过；合并提交 60949df412f10be004fe6bb5b4fb992dd65c3c60 的 tree 与 CI 来源一致。Windows Electron 桌面端尚未实机复测，真实模型验证未运行。
+
 ## 0.2.11
 
+[Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.11) 使用 [完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/37012171411) 验收的同一份包，来源提交为 39454dbcd14ea672c6d9fa25eddea26f8ea71591，SHA-256 为 d11f2a3e9abbd0aaaaa42d5ffb4268e9c192f608236178a48cd112d09e03f1f9。六组原生平台、源码安装及精确 alpha 宿主验收均通过；对该产物另完成精确 0.2.0-rc.2 Web 客户端启动验证。npm `dsh-openwrite@0.2.11` 和 GitHub Release 均已发布，`latest` 已更新到 0.2.11；从公共 npm registry 重新下载后，SHA-256 与上述 Release 包一致。
+
 - 修复 #45 的子包兼容声明和两个客户端 API 差异，增加初始化异常隔离、回滚与错误诊断；包含 #44 的 minimap CSS 修复。
 - 精确 `0.2.0-rc.2` Web 宿主已验证两个客户端激活、工作台入口、DoG 面板和其他插件共存，无需版本豁免。
 - Core 仍为 5.8.2 / contract 1。正式发布直接使用本版本完整 CI 验收通过的同一份产物，SHA-256、来源提交和原生平台结果以 Release 附件为准。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-openwrite",
-  "version": "0.2.11",
+  "version": "0.2.14",
   "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": {
@@ -28,8 +28,7 @@
   "license": "Apache-2.0",
   "devDependencies": {
     "@deepseek-ai/cordis-plugin-include": "1.0.7",
-    "@deepseek-ai/dsh": "0.1.2-rc.1",
-    "js-yaml": "4.3.1"
+    "@deepseek-ai/dsh": "0.1.2-rc.1"
   },
   "engines": {
     "node": ">=22.19.0"
@@ -71,6 +70,7 @@
   },
   "files": [
     "plugin.mjs",
+    "preset-definition.mjs",
     "plugin.d.ts",
     "cordis.patch.yml",
     "packages/studio-panel/vendor",
@@ -108,7 +108,8 @@
     "ajv-formats": "^3.0.1",
     "fflate": "0.8.2",
     "proper-lockfile": "4.1.2",
-    "tar": "7.5.7"
+    "tar": "7.5.7",
+    "js-yaml": "4.3.1"
   },
   "peerDependencies": {
     "@deepseek-ai/cordis": "4.0.2 || ~4.0.4",

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

@@ -60,6 +60,8 @@ export interface StudioDownload {
 }
 /** Query parameters; array values become repeated `key=v1&key=v2` entries. */
 export type QueryParams = Record<string, string | string[]>;
+/** Legacy Core fallback; newer servers declare mutation semantics in a response header. */
+export declare function mutationAffectsState(path: string, body: unknown, headers?: Headers): boolean;
 export declare class StudioClient {
     private readonly connect;
     private readonly baseUrl;

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

@@ -55,6 +55,30 @@ function applyParams(url, params) {
         }
     }
 }
+/** Legacy Core fallback; newer servers declare mutation semantics in a response header. */
+export function mutationAffectsState(path, body, headers) {
+    const declared = headers?.get('X-OpenWrite-Mutated');
+    if (declared === '0' || declared === '1')
+        return declared === '1';
+    const endpoint = path.split('?')[0]?.replace(/^\/+/, '').replace(/^api\//, '');
+    if (endpoint !== 'manuscript-editing')
+        return true;
+    let payload = body;
+    if (Buffer.isBuffer(payload))
+        payload = payload.toString('utf8');
+    if (typeof payload === 'string') {
+        try {
+            payload = JSON.parse(payload);
+        }
+        catch {
+            return true;
+        }
+    }
+    if (payload === null || typeof payload !== 'object' || Array.isArray(payload))
+        return true;
+    const action = payload['action'];
+    return typeof action !== 'string' || !['versions', 'version', 'compare', 'annotations'].includes(action);
+}
 export class StudioClient {
     connect;
     baseUrl;
@@ -112,7 +136,8 @@ export class StudioClient {
             body: JSON.stringify(body),
         }, signal);
         const result = await this.readJson(response);
-        this.onMutation?.(path, this.context);
+        if (mutationAffectsState(path, body, response.headers))
+            this.onMutation?.(path, this.context);
         return result;
     }
     /** PUT a JSON object with the Studio write protocol header. */
@@ -123,7 +148,8 @@ export class StudioClient {
             body: JSON.stringify(body),
         }, signal);
         const result = await this.readJson(response);
-        this.onMutation?.(path, this.context);
+        if (mutationAffectsState(path, body, response.headers))
+            this.onMutation?.(path, this.context);
         return result;
     }
     /** GET a file-download endpoint (export), returning the raw bytes and filename. */
@@ -156,7 +182,9 @@ export class StudioClient {
         catch (error) {
             if (signal?.aborted)
                 throw error;
-            throw new StudioError(`Studio request failed: ${error instanceof Error ? error.message : String(error)}`, 0, 'NETWORK_ERROR');
+            const cause = error instanceof Error ? error.cause : undefined;
+            const detail = cause instanceof Error ? ` (${cause.message})` : cause === undefined ? '' : ` (${String(cause)})`;
+            throw new StudioError(`Studio request failed: ${error instanceof Error ? error.message : String(error)}${detail}`, 0, 'NETWORK_ERROR');
         }
     }
     /**

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

@@ -1,5 +1,5 @@
 import { Service } from '@deepseek-ai/cordis';
-import { StudioClient, workspaceRootHeaders } from './client.js';
+import { StudioClient, workspaceRootHeaders, 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';
@@ -98,6 +98,8 @@ function resourceForPath(path) {
         return 'models';
     if (path.startsWith('/api/research'))
         return 'research';
+    if (path.startsWith('/api/manuscript-editing'))
+        return 'manuscript';
     if (path.startsWith('/api/revisions'))
         return 'revisions';
     return 'workspace';
@@ -481,7 +483,7 @@ export class NovelDomainService extends Service {
                     responseHeaders['content-disposition'] = disposition;
                 res.writeHead(upstream.status, responseHeaders);
                 res.end(bytes);
-                if (isWrite && upstream.ok) {
+                if (isWrite && upstream.ok && mutationAffectsState(pathPart, body, upstream.headers)) {
                     this.notifyMutation(`/api/${pathPart}`, { workspaceRoot: resolved.root, workspaceId: resolved.id });
                 }
             }

package/packages/openwrite-bridge/lib/managed-runtime.d.ts

@@ -52,6 +52,14 @@ export declare const DEFAULT_BACKOFF_MS: number[];
 export declare function sha256(file: string): Promise<string>;
 /** Diagnostics must not echo provider credentials or authenticated download URLs. */
 export declare function sanitizeDiagnostic(message: string): string;
+/**
+ * Drop bracketed IPv6 literals (e.g. `[::1]`) from a no_proxy value. DSH's
+ * loopback bypass list is written for undici as well as generic consumers, but
+ * httpx parses each entry as host:port and fails at client construction with
+ * `Invalid port: ':1]'`, so the request never leaves the process. Plain
+ * entries such as `::1` stay — httpx handles them fine.
+ */
+export declare function sanitizeNoProxy(value: string): string;
 /** Windows needs SystemRoot/PATH to start; proxy and CA vars stay when set. Credential names never copy. */
 export declare function buildChildEnv(source: NodeJS.ProcessEnv, extra?: Record<string, string | undefined>): NodeJS.ProcessEnv;
 /** Stop only a child we spawned. POSIX groups and Windows trees include grandchildren. */
@@ -78,6 +86,16 @@ export declare class ManagedRuntime {
     status(): RuntimeStatus;
     private update;
     ensure(): Promise<BackendConnection>;
+    /**
+     * The child process can outlive its listening socket (a stuck request thread
+     * during shutdown leaves it alive but refusing connections), and no exit is
+     * ever observed in that state, so the auto-restart logic never fires. Probe
+     * the cached connection before handing it out; a dead backend is torn down
+     * forcefully and pushed through the normal recovery path.
+     */
+    private verifyConnection;
+    /** Short-timeout liveness probe; false covers exit, refusal, hang, and auth drift. */
+    private probeHealth;
     cancel(): Promise<void>;
     dispose(): Promise<void>;
     private exitDiagnostic;

package/packages/openwrite-bridge/lib/managed-runtime.js

@@ -10,6 +10,47 @@ import { unzipSync } from 'fflate';
 import { lock as acquireLock } from 'proper-lockfile';
 export const DEFAULT_RESTART_LIMIT = 5;
 export const DEFAULT_BACKOFF_MS = [1_000, 2_000, 4_000, 8_000, 16_000];
+/** Budget for the per-request liveness probe of a previously-ready backend. */
+const HEALTH_PROBE_TIMEOUT_MS = 1_500;
+/** The backend stderr log on disk keeps only the newest chunk. */
+const STDERR_LOG_LIMIT = 64 * 1024;
+/**
+ * Keep the managed backend's stderr on disk (last 64 KiB) so a hung or zombie
+ * process can be diagnosed after the fact; the in-memory tail alone is lost
+ * with the host. Writes are debounced and serialized, and never throw —
+ * diagnostics must not break the backend lifecycle.
+ */
+class BackendStderrLog {
+    path;
+    tail = '';
+    timer;
+    pending = Promise.resolve();
+    constructor(path) {
+        this.path = path;
+    }
+    write(text) {
+        this.tail = (this.tail + text).slice(-STDERR_LOG_LIMIT);
+        if (this.timer !== undefined)
+            return;
+        this.timer = setTimeout(() => { this.timer = undefined; void this.flush(); }, 200);
+    }
+    tailText() { return this.tail; }
+    flush() {
+        if (this.timer !== undefined) {
+            clearTimeout(this.timer);
+            this.timer = undefined;
+        }
+        const tail = this.tail;
+        this.pending = this.pending.then(async () => {
+            try {
+                await mkdir(dirname(this.path), { recursive: true });
+                await writeFile(this.path, tail);
+            }
+            catch { /* best-effort diagnostics */ }
+        });
+        return this.pending;
+    }
+}
 const CREDENTIAL_ENV = /(?:API[_-]?KEY|ACCESS[_-]?TOKEN|SECRET|PASSWORD|CREDENTIAL|AUTHORIZATION)/i;
 const ALLOWED_ENV = new Set([
     'PATH', 'Path', 'PATHEXT',
@@ -36,6 +77,17 @@ export function sanitizeDiagnostic(message) {
         .replace(/((?:api[_-]?key|authorization|password|secret|token)\s*[:=]\s*)[^\s,;]+/gi, '$1[redacted]')
         .replace(/\b[a-f0-9]{32,}\b/gi, '[redacted]');
 }
+/**
+ * Drop bracketed IPv6 literals (e.g. `[::1]`) from a no_proxy value. DSH's
+ * loopback bypass list is written for undici as well as generic consumers, but
+ * httpx parses each entry as host:port and fails at client construction with
+ * `Invalid port: ':1]'`, so the request never leaves the process. Plain
+ * entries such as `::1` stay — httpx handles them fine.
+ */
+export function sanitizeNoProxy(value) {
+    return value.split(',').map(entry => entry.trim())
+        .filter(entry => entry !== '' && !/^\[.*\]$/.test(entry)).join(',');
+}
 /** Windows needs SystemRoot/PATH to start; proxy and CA vars stay when set. Credential names never copy. */
 export function buildChildEnv(source, extra = {}) {
     const env = {};
@@ -52,6 +104,12 @@ export function buildChildEnv(source, extra = {}) {
             continue;
         env[key] = value;
     }
+    // The Python backend reads NO_PROXY through httpx; bracketed IPv6 literals
+    // break client construction there even though Node's undici accepts them.
+    for (const key of Object.keys(env)) {
+        if (key.toLowerCase() === 'no_proxy' && typeof env[key] === 'string')
+            env[key] = sanitizeNoProxy(env[key]);
+    }
     return env;
 }
 /** Stop only a child we spawned. POSIX groups and Windows trees include grandchildren. */
@@ -119,7 +177,7 @@ export class ManagedRuntime {
         if (this.closed)
             return Promise.reject(new Error('OpenWrite 已卸载'));
         if (this.connection)
-            return Promise.resolve(this.connection);
+            return this.verifyConnection(this.connection);
         if (this.pending)
             return this.pending;
         this.allowRevive = true;
@@ -137,6 +195,52 @@ export class ManagedRuntime {
         }).finally(() => { this.pending = undefined; this.controller = undefined; });
         return this.pending;
     }
+    /**
+     * The child process can outlive its listening socket (a stuck request thread
+     * during shutdown leaves it alive but refusing connections), and no exit is
+     * ever observed in that state, so the auto-restart logic never fires. Probe
+     * the cached connection before handing it out; a dead backend is torn down
+     * forcefully and pushed through the normal recovery path.
+     */
+    async verifyConnection(connection) {
+        if (await this.probeHealth(connection)) {
+            if (this.closed)
+                throw new Error('OpenWrite 已卸载');
+            const current = this.connection;
+            if (current)
+                return current;
+            return this.queueRestart();
+        }
+        if (this.connection !== connection)
+            return this.queueRestart();
+        this.connection = undefined;
+        const child = this.child;
+        this.child = undefined;
+        if (child)
+            await stopOwnedProcess(child).catch(() => { });
+        if (this.closed)
+            throw new Error('OpenWrite 已卸载');
+        if (!this.allowRevive)
+            throw new Error('准备已取消，可重试');
+        this.update('recovering', '写作后端无响应，正在恢复');
+        return this.queueRestart();
+    }
+    /** Short-timeout liveness probe; false covers exit, refusal, hang, and auth drift. */
+    async probeHealth(connection) {
+        const child = this.child;
+        if (!child || child.exitCode !== null || child.signalCode !== null)
+            return false;
+        try {
+            const response = await fetch(connection.baseUrl + '/api/health', {
+                headers: { Authorization: `Bearer ${connection.token}` },
+                signal: AbortSignal.timeout(HEALTH_PROBE_TIMEOUT_MS),
+            });
+            return response.ok;
+        }
+        catch {
+            return false;
+        }
+    }
     async cancel() {
         this.allowRevive = false;
         this.controller?.abort();
@@ -434,8 +538,16 @@ export class ManagedRuntime {
             env: buildChildEnv(process.env, { PYTHONNOUSERSITE: '1', PYTHONUTF8: '1' }),
         });
         this.child = child;
+        // Mirror stderr to state/logs/backend.log (last 64 KiB): a stuck backend
+        // prints nothing and its in-memory tail dies with the host, leaving no way
+        // to diagnose "alive but not serving" failures after the fact.
+        const stderrLog = new BackendStderrLog(join(this.root, 'state', 'logs', 'backend.log'));
         let stderrTail = '';
-        child.stderr?.on('data', bytes => { stderrTail = (stderrTail + String(bytes)).slice(-500); });
+        child.stderr?.on('data', bytes => {
+            stderrLog.write(String(bytes));
+            stderrTail = stderrLog.tailText().slice(-500);
+        });
+        child.once('exit', () => { void stderrLog.flush(); });
         const lines = createInterface({ input: child.stdout });
         let timer;
         let started = false;