fix(llm): stream structured fallback and expose fetch error cause

.github/dependabot.yml

@@ -0,0 +1,42 @@
+version: 2
+updates:
+  - package-ecosystem: "npm"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+      day: "monday"
+      time: "08:30"
+      timezone: "Asia/Shanghai"
+    open-pull-requests-limit: 5
+    labels:
+      - "dependencies"
+    commit-message:
+      prefix: "chore(deps)"
+      include: "scope"
+    groups:
+      typescript-tooling:
+        patterns:
+          - "typescript"
+          - "tsx"
+          - "tslib"
+          - "@types/*"
+      eslint:
+        patterns:
+          - "eslint"
+          - "eslint-*"
+          - "@eslint/*"
+
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"
+      day: "monday"
+      time: "08:30"
+      timezone: "Asia/Shanghai"
+    open-pull-requests-limit: 3
+    labels:
+      - "dependencies"
+      - "ci"
+    commit-message:
+      prefix: "ci(deps)"
+      include: "scope"

.github/workflows/desktop-beta-release.yml

@@ -16,12 +16,13 @@ jobs:
       AI_NOVEL_RELEASE_CHANNEL: beta
       AI_NOVEL_GITHUB_OWNER: ExplosiveCoderflome
       AI_NOVEL_GITHUB_REPO: AI-Novel-Writing-Assistant
+      AI_NOVEL_WINDOWS_PUBLISHER_NAME: AI Novel Writing Assistant Team
       GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       CSC_LINK: ${{ secrets.WINDOWS_CSC_LINK }}
       CSC_KEY_PASSWORD: ${{ secrets.WINDOWS_CSC_KEY_PASSWORD }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup pnpm
         uses: pnpm/action-setup@v4

.github/workflows/desktop-release.yml

@@ -5,26 +5,25 @@ on:
   push:
     tags:
       - "v*"
-      - "desktop-v*"
 
 permissions:
   contents: write
 
 jobs:
   publish-release:
-    if: ${{ startsWith(github.ref_name, 'desktop-v') || !contains(github.ref_name, '-') }}
+    if: ${{ !contains(github.ref_name, '-') }}
     runs-on: windows-latest
     env:
       AI_NOVEL_RELEASE_CHANNEL: release
-      AI_NOVEL_ALLOW_UNSIGNED_RELEASE: "true"
       AI_NOVEL_GITHUB_OWNER: ExplosiveCoderflome
       AI_NOVEL_GITHUB_REPO: AI-Novel-Writing-Assistant
+      AI_NOVEL_WINDOWS_PUBLISHER_NAME: AI Novel Writing Assistant Team
       GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       CSC_LINK: ${{ secrets.WINDOWS_CSC_LINK }}
       CSC_KEY_PASSWORD: ${{ secrets.WINDOWS_CSC_KEY_PASSWORD }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup pnpm
         uses: pnpm/action-setup@v4

.gitignore

@@ -1,11 +1,13 @@
 node_modules/
+.cursor/
+*.log
+tmp/
 dist/
 .turbo/
 .DS_Store
 .env
 .env.local
 dev.db
-m[[]0])
 client/.env
 client/.env.local
 server/.env

AGENTS.md

@@ -97,6 +97,26 @@
 - If a build is triggered manually or from a non-matching tag, treat it as verification or packaging only. It must not be treated as a valid public release upload.
 - If the required `vX.Y.Z` tag and `desktop/package.json` version are not aligned, stop before upload, fix the version/tag pair first, and then rerun the release flow.
 
+## Desktop Runtime Debugging Rules
+
+- For packaged desktop bugs, do not assume a source fix has reached the executable just because a rebuild or packaging command ran.
+- First verify the packaged runtime behavior from the actual desktop logs before changing more code.
+- For portable builds, inspect the runtime logs under the portable data directory beside the executable, for example `*-data/logs/desktop-main.log`.
+- When debugging packaged prompt or LLM issues, capture and compare at least:
+  - the concrete API route;
+  - the prompt id or task label when available;
+  - the structured strategy actually used, such as `json_schema`, `json_object`, or `prompt_json`;
+  - the latest error text, especially whether it changed after a rebuild.
+- If the packaged log shows the strategy or stack has changed, treat that as a new layer of failure and re-evaluate the root cause instead of continuing under the previous assumption.
+
+### Packaged Runtime Lessons
+
+- When multiple unpacked or packaged desktop instances exist, always identify the fresh runtime by its new startup timestamp and bound server port before trusting any later log line. Old `desktop-main.log` failure records can remain in the same file and must not be mistaken for the current run.
+- For live verification, pair desktop log inspection with runtime API checks such as `/api/llm/model-routes`, `/api/llm/structured-fallback`, and the active director task snapshot. This is the fastest way to prove whether route normalization, fallback policy, and the current failure layer actually changed in the packaged app.
+- For OpenAI-compatible structured prompts, the real packaged path may start at `json_schema` even when the stored route format is `prompt_json`. Shared transport resilience must therefore cover `json_schema -> json_object -> prompt_json -> direct chat fallback`, not only direct `prompt_json` failures.
+- Timeout-governance tests for structured calls must use real abort semantics. Test doubles that immediately throw `fetch failed` can hide timeout-specific behavior and produce false conclusions about whether direct fallback should run.
+- When shared structured fallback has demonstrably taken over a transport path, prefer removing duplicated service-local retries while preserving only the final prompt-family-specific direct-call hop needed for packaged-runtime parity.
+
 ## Prompt Governance
 
 - `server/src/prompting/` is the only allowed entrypoint for adding new product-level prompts.

README.md

@@ -124,25 +124,15 @@
 
 完整历史更新见 [docs/releases/release-notes.md](./docs/releases/release-notes.md)。
 
-### 2026-05-08
-
-自动导演主链完成一次完整发布整合：从新手开书、方向确认、后台执行、章节拆分、正文生成、审校修复、状态回灌到任务恢复，都更接近“一句想法持续推进成一本书”的连续体验。
-
-- 新手可以从“新手上路”、首页、小说列表和小说工作台进入更清晰的自动导演路线，按系统推荐完成模型配置、AI 开书、方向确认和章节执行。
-- 自动导演支持更完整的整本自动推进：系统会准备故事规划、角色资产、卷战略、节奏拆章、章节任务单、正文、审校、修复和状态同步，普通质量问题优先由 AI 继续处理。
-- 每本小说拥有统一的 AI 驾驶舱和任务投影，能看到当前进展、阻塞原因、下一步动作、后台队列、最近事件、产物摘要和 AI 用量，减少在小说页、任务中心和弹窗之间来回猜状态。
-- 后台执行链更稳定：候选生成、方向确认、接管继续、恢复、重试、取消、章节执行和质量修复逐步收口到统一运行事实，服务重启或任务中断后更容易从最近安全进度继续。
-- 自动导演候选生成的继续路径更可靠：书级候选方案生成中断后，后台继续和重试会沿用当前运行步骤继续生成，减少卡在 0% 并提示缺少执行器的情况。
-- 章节执行更可靠：系统按真实章节产物判断可执行范围，缺少任务单或场景拆解时先回到节奏 / 拆章补齐，指定章节范围完成后按本次范围验收，不再误报后续章节未写导致的中断。
-- 质量闭环更适合新手：章节审校、局部修复、整章重写、窗口重规划、质量预算和质量待回收会形成连续判断，能先推进整本书，再集中处理可回收质量问题。
-- 写法、提示词和上下文能力进一步统一：写法参考会先净化为可迁移指导，产品级提示词纳入提示词资产和上下文经纪层，方便后续稳定调参和排查。
-- 提示词工作台可以更清楚地查看每个提示词的上下文需求、可安全调整的表达槽位、锁定字段和资料检查结果，方便先确认小说资料是否齐全，再进入后续提示词优化。
-- 提示词工作台支持自定义补充要求：可以为章节写作、章节审校和章节修复设置全局或单本小说要求，启用后作为额外上下文参与生成，不覆盖内置提示词。
-- 提示词列表优先展示可补充的提示词，方便先处理章节写作、审校和修复这类主要生成链路。
-- 开发启动体验更顺：前端页面可以先显示本地创作服务连接中，等待服务可用后自动进入工作台，减少刚打开页面就看到加载错误。
-- 桌面客户端进入 `0.3.1` 正式发布，用于承载本次整合后的公开客户端构建和启动体验优化。
-- 桌面客户端推进到 `0.3.2`，用于发布候选生成继续路径修复后的 Windows 客户端包。
-- 桌面客户端推进到 `0.3.3`，用于发布提示词工作台和自定义补充要求能力后的 Windows 客户端包。
+### 2026-05-10
+
+桌面版自动导演的结构化调用治理和打包稳定性继续收口：OpenAI 兼容模型在候选生成、故事规划、角色准备和分卷战略里的共享降级路径更稳定，恢复中的整本流程已经能穿过旧的角色阵容故障层，继续推进到卷战略与卷骨架阶段；fresh 打包时的 Prisma 运行时装配和 electron-builder 兼容问题也做了补强。
+
+- 桌面版在 fresh runtime 上已经确认：角色阵容生成和修复都能继续推进，不再停在旧的首层传输错误。
+- 自动导演恢复后已从角色准备推进到卷战略与卷骨架生成，让长链路开书更接近持续自动完成。
+- 结构化协议与共享降级路径更集中，后续桌面版排错可以更快判断当前真实失败层。
+- 旧库存量缺字段的工作区分析和手动改文影响分析更稳，书级 framing 与故事宏观规划在 OpenAI 兼容端点上的兜底也更一致。
+- fresh 打包时更不容易被 Prisma 运行时版本漂移或 electron-builder 兼容问题卡住。
 
 ## 功能预览
 ### 功能概览中的95%以上编写都是AI完成
@@ -620,4 +610,3 @@ docs/     设计文档、阶段检查点、模块计划与历史归档
 - 请遵守开源协议条款，并在适用场景下取得相应授权。
 
 贡献说明：新贡献默认按 [CLA.md](./CLA.md) 提交，可随项目按 AGPL-3.0-only 分发，并可纳入项目维护者另行提供的商业授权；详见 [CONTRIBUTING.md](./CONTRIBUTING.md)。
-

client/package.json

@@ -47,14 +47,14 @@
     "zustand": "^5.0.11"
   },
   "devDependencies": {
-    "@types/node": "^25.3.3",
+    "@types/node": "^25.6.2",
     "@types/react": "^19.2.14",
     "@types/react-dom": "^19.2.3",
     "@vitejs/plugin-react": "^5.1.4",
     "autoprefixer": "^10.4.27",
     "postcss": "^8.5.8",
     "tailwindcss": "^3.4.17",
-    "typescript": "^5.9.3",
+    "typescript": "^6.0.3",
     "vite": "^7.3.1",
     "vite-plugin-pages": "^0.33.3"
   }

client/src/api/settings.ts

@@ -121,6 +121,7 @@ export interface ModelRoutesResponse {
     maxTokens: number | null;
     requestProtocol: ModelRouteRequestProtocol;
     structuredResponseFormat: ModelRouteStructuredResponseFormat;
+    requestHeadersText: string | null;
   }>;
 }