dsh-openwrite 0.2.17 → 0.2.19

package/README.md

@@ -66,6 +66,10 @@ dsh web
 
 ### dsh 0.2 桌面端升级
 
+`0.2.19` 候选包修复两个加载/工具问题：部分安装下「OpenWrite 创作」预设从列表消失、DoG 工具不可用（dsh-dog 宿主包解析失败，[#74](https://github.com/LiPu-jpg/Openwrite/issues/74)），以及 `novel_*` 工具的对象/数组参数在模型输出 JSON 字符串时预览失败、`data` 字段丢失（[#75](https://github.com/LiPu-jpg/Openwrite/issues/75)）。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，再刷新浏览器页面。
+
+`0.2.18` 候选包修正 `auto:cheapest`（Core 5.8.9）：零价模型不再参与比价，自动落到付费低价模型上。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，再刷新浏览器页面。
+
 `0.2.17` 候选包新增自动选模（Core 5.8.8）：模型档案的模型名写成 `auto:cheapest` 会自动选用服务商价格目录中总价最低且满足上下文/输出要求的模型（目录缓存一小时，价格变动自动跟随），写成 `auto:popular`（OpenRouter）则由官方 `openrouter/auto` 按社区用量自动选模型。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，再刷新浏览器页面。
 
 `0.2.16` 候选包补漏 [#64](https://github.com/LiPu-jpg/Openwrite/issues/64)/[#67](https://github.com/LiPu-jpg/Openwrite/issues/67)：字数精简重试耗尽后保稿时，`length_out_of_range`/`length_warning` 标记与校验问题现在随草稿产物持久化（Core 5.8.7），导出/审阅侧可追溯。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，再刷新浏览器页面。

package/docs/RELEASE_ACCEPTANCE.md

@@ -6,9 +6,26 @@
 
 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。
 
+## 0.2.19
+
+<!-- TODO: 发布后回填 CI、Release、npm 链接与 SHA -->
+
+修复 [#74](https://github.com/LiPu-jpg/Openwrite/issues/74) 与 [#75](https://github.com/LiPu-jpg/Openwrite/issues/75)，均为插件侧改动，Core 保持 5.8.9（native-core `d0a75e30047c28c99af909e1b85b07190abb1fb7`），wheel 与 requirements.lock 未变。
+
+- #74：vendor/dsh-dog 运行时不再裸导入 `@deepseek-ai/dsh-home-paths` 与 `@deepseek-ai/dsh-tools` 两个 peer 宿主包（pnpm 隔离布局下从 vendor 目录不可解析 → openwrite-dog 条目导入失败 → openwriteDog 未注册 → 预设从列表消失）。`dshHomePath` 与 `defineTool` 由新增 `src/host-shims.ts` 自包含实现，行为对齐宿主；构建产物 `lib/index.js` 只导入 node 内置模块与 dependencies 内的 `@deepseek-ai/schemastery`，独立 `import()` 成功。新增 host-shims 单元测试 10 例，dsh-dog 共 60/60 通过。
+- #75：bridge 的 90 个 `novel_*` 工具统一经本地 `defineTool`（复用宿主 `parameterSchemaSpecToJsonSchema`/`valueSchemaSpecToJsonSchema`/`validateJsonSchemaValue`/`ToolArgsError`），在 schema 校验前对 `type: 'json'` 参数（含 array 的 json items）做字符串 `JSON.parse` 归一化；报错带参数名与实际类型。8 项归一化冒烟用例通过；四个宿主符号已核对在 0.1.2-rc.1 / 0.1.5-alpha.1 / 0.2.0-rc.2 三个受支持宿主中均导出且签名一致。
+
+## 0.2.18
+
+[GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.18) 已发布，直接使用[完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/38059156641)验收的同一份包。来源提交 `6e84c2a`，npm 包 SHA-256 为 `edd510c6d4de58bfa09eb30f8a3b74e5a0e6cf7cf2425bc6fe3681b732c1a196`。六组原生平台/Node、源码安装、alpha 兼容及三代宿主客户端共 12 份报告全部通过；Release 四份附件的校验值已与 CI `validated-release` 产物逐一核对。合并来源 [PR #73](https://github.com/LiPu-jpg/Openwrite/pull/73)。npm `dsh-openwrite@0.2.18` 已发布到官方 registry（registry.npmjs.org），`latest` 已更新到 0.2.18；从官方 registry 无认证重新下载后，SHA-256 与上述 CI 包一致。main CI 首次运行 source-install 因 pnpm `--find-links` 目录瞬时不存在失败，重跑同一提交通过，判定为运行器 flake。
+
+修正 0.2.17 的 `auto:cheapest`：真实 Key 实测发现 OpenRouter 上部分免费模型没有 `:free` 后缀（`inclusionai/ling-3.1-flash` 定价 0/0），会被低价策略选中；免费模型可用性不稳定，Core 5.8.9（native-core `d0a75e30047c28c99af909e1b85b07190abb1fb7`）把零价模型一并排除，`auto:cheapest` 现在落在付费低价模型上。回归测试补充零价排除断言，9/9 通过。
+
+Core 5.8.9 由 `native-core` 分支干净树 `pip wheel` 构建，manifest 三处已回填。
+
 ## 0.2.17
 
-<!-- TODO: 发布后回填 PR、CI、Release 链接与 SHA -->
+[GitHub Release](https://github.com/LiPu-jpg/Openwrite/releases/tag/v0.2.17) 已发布，直接使用[完整 CI](https://github.com/LiPu-jpg/Openwrite/actions/runs/38057362486)验收的同一份包。来源提交 `de2cacb`，npm 包 SHA-256 为 `da2a0b2ae264392f7d0f8c5d6c98de08cd0497f8a761bf0c96f9a43ce95771c5`。六组原生平台/Node、源码安装、alpha 兼容及三代宿主客户端共 12 份报告全部通过；Release 四份附件的校验值已与 CI `validated-release` 产物逐一核对。合并来源 [PR #72](https://github.com/LiPu-jpg/Openwrite/pull/72)。npm `dsh-openwrite@0.2.17` 已发布到官方 registry（registry.npmjs.org），`latest` 已更新到 0.2.17；从官方 registry 无认证重新下载后，SHA-256 与上述 CI 包一致。
 
 新增自动选模（Core 5.8.8，native-core `a7f4d8769de7496cb18b9d6dc38872c60fa5517a`）：档案模型名支持 `auto:cheapest`（实时拉取 `/models` 价格目录，过滤后取总价最低，缓存一小时、失败回退过期缓存）与 `auto:popular`（OpenRouter `openrouter/auto`）。解析在 `resolve`/`resolve_profile` 统一发生，结果写入 `auto_select` 字段；静态模型档案不受影响。新增 9 项回归测试全部通过（比价规则、缓存命中、过期缓存兜底、目录不可用报错、popular 非 OpenRouter 拒绝、未知策略拒绝、静态档案零网络）；档案/连接测试/工作室相关面 227 项通过（2 项失败为 native-core 预存在的 `test_real_litellm_backend_accepts_openwrite_request_shape_without_network` 与 `test_studio_outline_task_uses_planning_snapshot_and_validated_artifact` 漂移，与本版无关）。真实 OpenRouter Key 实测：`auto:popular` 路由到 `qwen/qwen3.8-flash` 返回 200。
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-openwrite",
-  "version": "0.2.17",
+  "version": "0.2.19",
   "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": {

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

@@ -8,7 +8,7 @@ import { createHash } from 'node:crypto';
 import { realpathSync } from 'node:fs';
 import { mkdir, writeFile } from 'node:fs/promises';
 import { isAbsolute, join } from 'node:path';
-import { defineTool } from '@deepseek-ai/dsh-tools';
+import { ToolArgsError, parameterSchemaSpecToJsonSchema, validateJsonSchemaValue, valueSchemaSpecToJsonSchema, } from '@deepseek-ai/dsh-tools';
 import { StudioError } from './client.js';
 import { materializeChapterDelivery } from './dog-delivery.js';
 import { materializeCompletedDogTaskReview, materializeDogReview, reviewChapterId } from './dog-review.js';
@@ -23,6 +23,135 @@ const JSON_OUTPUT_SCHEMA = {
 };
 /** Manuscript chapter ids accepted by OpenWrite (`ch_<digits>`). */
 const CHAPTER_ID_PATTERN = /^ch_\d+$/;
+/** Short, model-facing description of a received argument value's actual shape. */
+function describeValue(value) {
+    if (value === undefined)
+        return 'undefined';
+    if (value === null)
+        return 'null';
+    if (Array.isArray(value))
+        return `an array of ${value.length} item(s)`;
+    if (typeof value === 'string')
+        return 'a string';
+    return `a value of type ${typeof value}`;
+}
+/**
+ * Parse a JSON string argument once, naming the parameter and the received
+ * type when the text is not valid JSON.
+ */
+function parseJsonArgument(name, value) {
+    try {
+        return JSON.parse(value);
+    }
+    catch (error) {
+        const detail = error instanceof Error ? error.message : String(error);
+        throw new Error(`parameter "${name}" expects a JSON value; received a string that is not valid JSON (${detail})`);
+    }
+}
+/**
+ * Normalize arguments for parameters declared `type: 'json'` (and
+ * `type: 'array'` with json items). DSH keeps `json` an annotation-only
+ * schema, so the value reaches the plugin in whatever shape the model
+ * emitted; models that serialize the object/array as a JSON *string* would
+ * otherwise fail the plugin-side object/array checks (GitHub issue #75).
+ */
+function normalizeJsonArguments(parameters, args) {
+    let normalized;
+    for (const [name, spec] of Object.entries(parameters)) {
+        const value = args[name];
+        if (value === undefined || spec === null || typeof spec !== 'object')
+            continue;
+        const type = spec.type;
+        if (type === 'json' && typeof value === 'string') {
+            normalized ??= { ...args };
+            normalized[name] = parseJsonArgument(name, value);
+        }
+        else if (type === 'array') {
+            const items = spec.items;
+            const itemType = items !== null && typeof items === 'object' ? items.type : undefined;
+            if (itemType !== 'json')
+                continue;
+            if (typeof value === 'string') {
+                normalized ??= { ...args };
+                normalized[name] = parseJsonArgument(name, value);
+            }
+            else if (Array.isArray(value) && value.some((item) => typeof item === 'string')) {
+                normalized ??= { ...args };
+                normalized[name] = value.map((item) => (typeof item === 'string' ? parseJsonArgument(`${name}[]`, item) : item));
+            }
+        }
+    }
+    return normalized ?? args;
+}
+/**
+ * Local `defineTool`: mirrors the host helper (`@deepseek-ai/dsh-tools`
+ * `defineTool`) using the host's own schema compiler and validator, except
+ * every argument flow (execute and the replay presenters) first normalizes
+ * string-typed JSON arguments — see {@link normalizeJsonArguments}.
+ * Normalization must run *before* validation, otherwise an `array` parameter
+ * receiving a JSON string would be rejected by the schema check before the
+ * plugin ever sees it (GitHub issue #75). Centralizing here keeps all
+ * `novel_*` call sites free of per-parameter parsing.
+ */
+function defineTool(options) {
+    const userExecute = options.execute;
+    const userFinalizeContent = options.finalizeContent;
+    const userRender = options.output.render;
+    const userPresentationMeta = options.output.presentationMeta;
+    const userPresentCall = options.presentCall;
+    const userPresentResult = options.presentResult;
+    const userIsConcurrencySafe = options.isConcurrencySafe;
+    if (options.timeoutMs !== undefined && (!Number.isFinite(options.timeoutMs) || options.timeoutMs <= 0))
+        throw new Error(`defineTool(${options.name}): timeoutMs must be a positive finite number`);
+    const parameters = parameterSchemaSpecToJsonSchema(options.parameters);
+    const outputSchema = valueSchemaSpecToJsonSchema(options.output.schema);
+    const normalize = (args) => normalizeJsonArguments(options.parameters, objectValue(args));
+    const validate = (args) => validateJsonSchemaValue(parameters, args, '');
+    const tool = {
+        name: options.name,
+        description: options.description,
+        parameters,
+        output: {
+            schema: outputSchema,
+            render: (args, value) => userRender(args, value),
+            ...(userPresentationMeta !== undefined
+                ? { presentationMeta: (args, value) => userPresentationMeta(args, value) }
+                : {}),
+        },
+        ...(options.timeoutMs !== undefined ? { timeoutMs: options.timeoutMs } : {}),
+        async execute(args, exec) {
+            const normalized = normalize(args);
+            const violations = validate(normalized);
+            if (violations.length > 0)
+                throw new ToolArgsError(violations);
+            return userExecute(normalized, exec);
+        },
+    };
+    if (userFinalizeContent)
+        tool.finalizeContent = (exec, result) => userFinalizeContent(exec, result);
+    if (userPresentCall)
+        tool.presentCall = (args) => {
+            const normalized = normalize(args);
+            if (validate(normalized).length > 0)
+                return undefined;
+            return userPresentCall(normalized);
+        };
+    if (userPresentResult)
+        tool.presentResult = (args, result) => {
+            const normalized = normalize(args);
+            if (validate(normalized).length > 0)
+                return undefined;
+            return userPresentResult(normalized, result);
+        };
+    if (userIsConcurrencySafe)
+        tool.isConcurrencySafe = (args) => {
+            const normalized = normalize(args);
+            if (validate(normalized).length > 0)
+                return false;
+            return userIsConcurrencySafe(normalized);
+        };
+    return tool;
+}
 /** Resolve a review target: an explicit manuscript path, or a chapter id mapped to `data/manuscript/<id>.md`. */
 function resolveReviewPath(args) {
     const path = args.path?.trim();
@@ -734,7 +863,8 @@ export function registerNovelTools(ctx, clientFactory, options) {
         timeoutMs,
         async execute(args, exec) {
             if (args.action === 'preview' && (!Array.isArray(args.edits) || args.edits.length !== 1)) {
-                throw new Error('preview requires exactly one edit so the author can accept or reject each change independently');
+                throw new Error('preview requires exactly one edit so the author can accept or reject each change independently ' +
+                    `(received ${describeValue(args.edits)})`);
             }
             if (args.action !== 'preview' && !args.preview_token?.trim()) {
                 throw new Error(`${args.action} requires preview_token`);
@@ -771,7 +901,7 @@ export function registerNovelTools(ctx, clientFactory, options) {
                 if (!args.change_kind)
                     throw new Error('preview requires change_kind');
                 if (args.change === null || typeof args.change !== 'object' || Array.isArray(args.change)) {
-                    throw new Error('preview requires one change object');
+                    throw new Error(`preview requires one change object (received ${describeValue(args.change)})`);
                 }
             }
             else if (!args.preview_token?.trim()) {