@waterwx/dsh-novel-forge 1.3.2-alpha → 1.3.3-alpha

package/src/client/sidebar-entry.ts

@@ -8,7 +8,7 @@ import { tt } from './panel/helpers.ts'
 import css from './panel/panel.module.css'
 
 /** Stable data attribute identifying the injected entry row. */
-export const ENTRY_SELECTOR = '[data-dsh-novelforge-entry]'
+const ENTRY_SELECTOR = '[data-dsh-novelforge-entry]'
 
 /** Inline icon: an open book / writing glyph. */
 const ICON = '<svg viewBox="0 0 24 24" width="16" height="16" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M12.67 19a2 2 0 0 0 1.416-.588l6.154-6.172a6 6 0 0 0-8.49-8.49L5.586 9.914A2 2 0 0 0 5 11.328V18a1 1 0 0 0 1 1z"/><path d="M16 8 2 22"/><path d="M17.5 15H9"/></svg>'
@@ -35,7 +35,7 @@ function newSessionButton(root: HTMLElement): HTMLButtonElement | undefined {
 function createEntry(controller: PanelController): HTMLButtonElement {
   const entry = document.createElement('button')
   entry.type = 'button'
-  entry.dataset.dshNovelforgeEntry = 'true'
+  entry.setAttribute(ENTRY_SELECTOR.slice(1, -1), 'true')
   entry.className = css.entry
   entry.setAttribute('aria-label', tt('entry.label'))
   entry.setAttribute('title', tt('entry.tooltip'))
@@ -52,7 +52,7 @@ function placeEntry(root: HTMLElement, entry: HTMLButtonElement): boolean {
     const row = button.closest('[class*="logoRow"]')
     const base = (row !== null && row.parentElement === root) ? row : button
     const family = Array.from(root.children).filter(
-      (el): el is HTMLElement => el instanceof HTMLElement && el.matches('[data-dsh-taskboard-entry], [data-dsh-ssh-entry], [data-dsh-novelforge-entry]'),
+      (el): el is HTMLElement => el instanceof HTMLElement && el.matches(`[data-dsh-taskboard-entry], [data-dsh-ssh-entry], ${ENTRY_SELECTOR}`),
     )
     const last = family.length > 0 ? family[family.length - 1] : undefined
     const anchor = last !== undefined ? last.nextElementSibling : base.nextElementSibling

package/src/engine.ts

@@ -23,7 +23,7 @@ export const COMPLIANCE_REDLINES: ReadonlyArray<string> = [
   '9. 不得出现法律法规禁止的其他内容。',
 ]
 
-/** 审稿维度取值（与 review-policy.ts 的 REVIEW_DIMENSIONS 对齐，用于归一化模型输出的 dimension 字段）。 */
+/** 审稿维度取值：归一化模型输出的 dimension 字段。维度清单以 reviewSystemPrompt 的九条为准。 */
 const REVIEW_DIMENSION_IDS = new Set(['character', 'setting', 'redline', 'writing', 'pacing', 'logic', 'anti-ai', 'presentation', 'compliance'])
 
 import { mkdirSync, readFileSync, writeFileSync, readdirSync, existsSync, rmSync, renameSync } from 'node:fs'
@@ -35,6 +35,8 @@ import { BUILTIN_GENRE_LIBRARY, BUILTIN_PROGRESSION_MODES, emptyProjectAssets, r
 import { scanAiFlavor } from './ai-scan.ts'
 import { beginLiveCall, endLiveCall, markFirstToken } from './llm-live.ts'
 import { shouldSwitchModel, withModelFallback } from './llm-retry.ts'
+import { convergeRequestParams } from './model-capability.ts'
+import { renderLengthContract, resolveOutputBudget, STRUCTURED_FIELD_LIMITS } from './output-budget.ts'
 import { renderTensionBlock } from './tension.ts'
 import { renderPromptSlots } from './prompt-slots.ts'
 import { normalizeTimelineEvent, renderTimelineBlock, sortTimeline } from './timeline.ts'
@@ -347,11 +349,64 @@ export function createProject(outline: string, outlinePath?: string): ProjectSta
 
 // ------------------------------------------------------------------- llm
 
+/** 一次非流式 LLM 调用的入参。 */
+interface CompleteOptions {
+  system: string
+  user: string
+  temperature?: number
+  maxTokens?: number
+  reasoning?: 'off' | 'low' | 'high' | 'max'
+  model?: string
+  liveLabel?: string
+  /**
+   * 该调用是否要求严格结构化输出（JSON）。为 true 时，模型参数能力层可以在
+   * 结构化稳定性优先的前提下覆盖用户的思考档位。
+   */
+  structured?: boolean
+}
+
+/**
+ * 把已经拼好的请求过一遍模型参数能力层。
+ *
+ * **凡是自己构造 `GenerateOptions`、不走 {@link complete} 的调用点都必须经过它**，
+ * 否则那条路径会绕过厂商参数约束（固定温度、不支持的 reasoning 等），对 Kimi
+ * 这类模型直接报参数错误 —— 而"测试连通"恰恰是用户用来判断配置对不对的地方。
+ *
+ * 抽成函数而不是在每个调用点内联：流式调用点有 5 处，还要各自处理备用模型切换，
+ * 内联五遍就是同一件事写五份，必然漂移。
+ * @param base - 已拼好的请求（调用方已确定 temperature / reasoningEffort）。
+ * @param model - **本次实际要调用的模型**（切换备用模型后必须重算，故在循环内调用）。
+ * @param options - `structured` 表示严格结构化输出任务。
+ * @returns 收敛后的请求（固定温度的模型会省略 temperature / 不支持的参数会被移除）。
+ */
+function applyModelCapability(
+  base: GenerateOptions,
+  model: string,
+  options: { structured?: boolean } = {},
+): GenerateOptions {
+  const converged = convergeRequestParams(
+    {
+      temperature: base.temperature ?? 0.7,
+      reasoningEffort: base.reasoningEffort === undefined
+        ? undefined
+        : String(base.reasoningEffort) as 'off' | 'low' | 'high' | 'max',
+    },
+    model,
+    { structured: options.structured === true, omitFixedTemperature: true },
+  )
+  return {
+    ...base,
+    model,
+    ...(converged.temperature === undefined ? { temperature: undefined } : { temperature: converged.temperature }),
+    ...(converged.reasoningEffort === undefined ? { reasoningEffort: undefined } : { reasoningEffort: ReasoningEffortId(converged.reasoningEffort) }),
+  }
+}
+
 /** One complete non-streaming LLM call. */
 async function complete(
   ctx: Context,
   config: NovelConfig,
-  options: { system: string; user: string; temperature?: number; maxTokens?: number; reasoning?: 'off' | 'low' | 'high' | 'max'; model?: string; liveLabel?: string },
+  options: CompleteOptions,
 ): Promise<string> {
   const liveLabel = options.liveLabel ?? 'LLM 调用'
   const messages: Message[] = [createUserMessage({
@@ -361,14 +416,31 @@ async function complete(
   /** 用指定模型跑一次完整调用（含实况打点、token 记账与诊断日志）。 */
   const runOnce = async (model: string): Promise<string> => {
     const live = beginLiveCall({ label: liveLabel, model, system: options.system, user: options.user })
+    // 用户偏好先成型，再经能力层按**实际模型**收敛：同一模型走不同 provider/中转
+    // 时约束不变，换备用模型后约束随新模型重算。
+    const converged = convergeRequestParams(
+      {
+        temperature: options.temperature ?? 0.7,
+        reasoningEffort: options.reasoning ?? config.reasoningEffort ?? 'off',
+      },
+      model,
+      {
+        structured: options.structured === true,
+        // 厂商建议固定温度的模型不要显式传入，优先省略，省得多余参数被拒。
+        omitFixedTemperature: true,
+      },
+    )
+    if (converged.adjustments.length > 0) {
+      ctx.logger('novel-forge').info('模型参数收敛（%s @ %s）：%s', liveLabel, model, converged.adjustments.join('；'))
+    }
     const request: GenerateOptions = {
       provider: config.provider,
       model,
       messages,
       system: options.system,
       maxTokens: options.maxTokens ?? config.maxTokens,
-      temperature: options.temperature ?? 0.7,
-      reasoningEffort: ReasoningEffortId(options.reasoning ?? config.reasoningEffort ?? 'off'),
+      ...(converged.temperature === undefined ? {} : { temperature: converged.temperature }),
+      ...(converged.reasoningEffort === undefined ? {} : { reasoningEffort: ReasoningEffortId(converged.reasoningEffort) }),
     }
     const assembler = new BlockAssembler()
     for await (const chunk of ctx.llm.stream(request)) {
@@ -421,7 +493,7 @@ function textFromAssembler(assembler: BlockAssembler): string {
 async function completeJsonArray(
   ctx: Context,
   config: NovelConfig,
-  options: { system: string; user: string; temperature?: number; maxTokens?: number; reasoning?: 'off' | 'low' | 'high' | 'max'; model?: string; liveLabel?: string },
+  options: CompleteOptions,
   parse: (text: string) => unknown[],
 ): Promise<unknown[]> {
   let parsed = parse(await complete(ctx, config, options))
@@ -671,13 +743,21 @@ function volumeSystemPrompt(): string {
     '[{"no": 1, "title": "卷名", "summary": "卷定位与剧情重心", "chapterStart": 1, "chapterEnd": 80}]',
     '重要：所有字符串值内部不得包含换行符，JSON 必须在一段内完整结束。',
     '重要：直接输出 JSON 结果本身，不要把思考过程或推理内容写在输出里。',
+    // 长度合同与 resolve 端的截断口径共用一份声明，避免 prompt 与解析漂移。
+    renderLengthContract([
+      ['卷名', STRUCTURED_FIELD_LIMITS.chapterTitle],
+      ['每卷定位', STRUCTURED_FIELD_LIMITS.chapterSummary],
+    ]),
   ].join('\n')
 }
 
 /** Plan volumes from an outline. */
 export async function planVolumes(ctx: Context, config: NovelConfig, outline: string): Promise<Volume[]> {
   const user = `请为下面这部小说划分卷：\n\n${outline}`
-  const text = await complete(ctx, config, { system: volumeSystemPrompt(), user, temperature: 0.4, maxTokens: Math.max(config.maxTokens, 12000) })
+  // 卷数由大纲决定、此处未知，按「大纲暗示的规模」估算：约每 1200 字大纲对应一卷，
+  // 夹在 3-24 卷之间（与卷规划自身的 3-8 卷建议不冲突，只是给预算一个量级）。
+  const expectedVolumes = Math.min(24, Math.max(3, Math.round(outline.length / 1_200)))
+  const text = await complete(ctx, config, { system: volumeSystemPrompt(), user, temperature: 0.4, maxTokens: resolveOutputBudget('volume-strategy', { volumes: expectedVolumes }) })
   const parsed = parseJsonArray<Record<string, unknown>>(text)
   const volumes: Volume[] = []
   for (let i = 0; i < parsed.length; i++) {
@@ -690,8 +770,8 @@ export async function planVolumes(ctx: Context, config: NovelConfig, outline: st
     const end = typeof entry.chapterEnd === 'number' ? entry.chapterEnd : undefined
     volumes.push({
       no,
-      title: title.slice(0, 40),
-      summary: summary.slice(0, 300),
+      title: title.slice(0, STRUCTURED_FIELD_LIMITS.chapterTitle),
+      summary: summary.slice(0, STRUCTURED_FIELD_LIMITS.chapterSummary),
       chapterStart: start ?? 1,
       chapterEnd: end ?? 1,
     })
@@ -981,7 +1061,7 @@ export async function planChapters(
     : '')
   const parsed = (await completeJsonArray(
     ctx, config,
-    { system, user, temperature: 0.7, maxTokens: Math.max(config.maxTokens, 40000), liveLabel: '章节规划' },
+    { system, user, temperature: 0.7, maxTokens: resolveOutputBudget('chapter-list', { chapters: chapterCount }), liveLabel: '章节规划' },
     t => parseJsonArray<Record<string, unknown>>(t),
   )) as Record<string, unknown>[]
   const chapters: ChapterPlan[] = []
@@ -1561,7 +1641,8 @@ export async function extractRoles(
       : '',
     '只输出 JSON 数组。',
   ].join('\n\n')
-  let text = await complete(ctx, config, { system, user, temperature: 0.3, maxTokens: Math.max(config.maxTokens, 16000), reasoning: config.analysisReasoning ?? 'low' })
+  // 角色阵容：一次要产出 8-20 个角色，预算按预期角色数推算。
+  let text = await complete(ctx, config, { system, user, temperature: 0.3, maxTokens: resolveOutputBudget('character-cast', { characters: 20 }), structured: true, reasoning: config.analysisReasoning ?? 'low' })
   let raw = parseJsonArray<Record<string, unknown>>(text)
   const hasProtagonist = raw.some(e => typeof e === 'object' && e !== null && e.roleLabel === 'protagonist')
   const tooFew = raw.length > 0 && raw.length < 6
@@ -1572,7 +1653,7 @@ export async function extractRoles(
       : tooFew
         ? '\n上一次输出角色过少（不足 6 个）。这是一部完整故事，请重新核对正文摘录：主角、主要反派、重要配角与所有有名有姓的角色都要收录（宁多勿漏），输出 8-20 个。'
         : '\n上一次输出中缺少主角（roleLabel 为 protagonist 的角色）。请重新输出完整 JSON 数组，务必包含正文中的主角。'
-    text = await complete(ctx, config, { system: system + hint, user, temperature: 0.3, maxTokens: Math.max(config.maxTokens, 12000), reasoning: config.analysisReasoning ?? 'low' })
+    text = await complete(ctx, config, { system: system + hint, user, temperature: 0.3, maxTokens: resolveOutputBudget('character-cast', { characters: 24 }), structured: true, reasoning: config.analysisReasoning ?? 'low' })
     raw = parseJsonArray<Record<string, unknown>>(text)
   }
   const labels = new Set(['protagonist', 'female_lead', 'female_support', 'support', 'antagonist', 'extra'])
@@ -2029,13 +2110,16 @@ export async function testLlmModel(ctx: Context, provider: string, model: string
     content: [{ type: 'text', text: '只回复两个字：OK' }],
     source: { kind: 'plugin', plugin: 'dsh-novel-forge' },
   })]
-  const request: GenerateOptions = {
+  // 这里自己构造请求（为省 token 拿到首块即停），所以**必须显式过一遍能力层**：
+  // 否则「测试连通」会原样下发参数，对固定温度的模型必然报错 —— 而测试连通
+  // 正是用户用来判断配置对不对的地方。
+  const request = applyModelCapability({
     provider,
     model,
     messages,
     maxTokens: 16,
     temperature: 0,
-  }
+  }, model)
   // 真实流式调用 + 30 秒超时（GenerateOptions.signal 由适配器响应并中止）。
   const controller = new AbortController()
   const timer = setTimeout(() => controller.abort(), 30_000)
@@ -2467,7 +2551,7 @@ export async function* rewriteChapterStream(
     })
     const assembler = new BlockAssembler()
     let produced = 0
-    for await (const chunk of ctx.llm.stream({ ...request, model })) {
+    for await (const chunk of ctx.llm.stream(applyModelCapability(request, model))) {
       assembler.push(chunk)
       if (chunk.type === 'text-delta') { produced += chunk.text.length; markFirstToken(live); yield { frame: 'delta', text: chunk.text } }
     }
@@ -2578,7 +2662,7 @@ export async function* polishChapterStream(
     })
     const assembler = new BlockAssembler()
     let produced = 0
-    for await (const chunk of ctx.llm.stream({ ...request, model })) {
+    for await (const chunk of ctx.llm.stream(applyModelCapability(request, model))) {
       assembler.push(chunk)
       if (chunk.type === 'text-delta') { produced += chunk.text.length; markFirstToken(live); yield { frame: 'delta', text: chunk.text } }
     }
@@ -2754,7 +2838,7 @@ export async function* generateChapterStream(
     })
     const assembler = new BlockAssembler()
     let produced = 0
-    for await (const chunk of ctx.llm.stream({ ...request, model })) {
+    for await (const chunk of ctx.llm.stream(applyModelCapability(request, model))) {
       assembler.push(chunk)
       if (chunk.type === 'text-delta') {
         produced += chunk.text.length
@@ -4152,6 +4236,9 @@ export async function runMarketRadar(
   return {
     signals: signalAs(raw.signals),
     productionFoundation,
+    // 报告 id 让前端能区分「同一份报告的重渲染」与「换了新报告」：只有后者才重置
+    // 作者手选的信号。用短随机 id 即可，无需持久化。
+    reportId: `mr-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 8)}`,
   }
 }
 

package/src/global-assets.ts

@@ -2,17 +2,16 @@
  * 全局资源库（跨书可复用的自定义题材/推进模式）。
  * 雷达「同步底座到全局资源库」写入此处；/assets 返回时与内置库合并展示。
  */
-import { homedir } from 'node:os'
-import { join } from 'node:path'
 import { mkdirSync, readFileSync, writeFileSync } from 'node:fs'
 import type { GenreNode, ProgressionMode } from './protocol.ts'
+import { dshHome, dshHomePath } from './home.ts'
 
 interface GlobalAssets {
   genres: GenreNode[]
   modes: ProgressionMode[]
 }
 
-const FILE = join(homedir(), '.dsh', 'novel-forge-global-assets.json')
+const FILE = dshHomePath('novel-forge-global-assets.json')
 let cache: GlobalAssets | null = null
 
 function load(): GlobalAssets {
@@ -29,7 +28,7 @@ function load(): GlobalAssets {
 
 function persist(): void {
   if (cache === null) return
-  mkdirSync(join(homedir(), '.dsh'), { recursive: true })
+  mkdirSync(dshHome(), { recursive: true })
   writeFileSync(FILE, JSON.stringify(cache, null, 2), 'utf8')
 }
 

package/src/home.ts

@@ -0,0 +1,33 @@
+/**
+ * DeepSeek Harness home resolution for this plugin.
+ *
+ * Precedence mirrors `resolveDshHome` from `@deepseek-ai/dsh-home-paths`:
+ * `$DSH_HOME` when set and non-blank, else `~/.dsh`. That package's remaining
+ * input — the harness's own explicit override — is not reachable from a
+ * plugin, so this reproduces the part a plugin can observe instead of taking
+ * a dependency whose published range tracks a different harness line.
+ *
+ * State files must resolve through here rather than `~/.dsh` directly: the
+ * desktop build points `DSH_HOME` at its own root, and a hardcoded path made
+ * both builds share one store and lose it whenever that directory was cleaned.
+ */
+import { homedir } from 'node:os'
+import { join, resolve } from 'node:path'
+
+/**
+ * Absolute Harness home for this process.
+ * @returns the normalized home directory.
+ */
+export function dshHome(): string {
+  const configured = process.env.DSH_HOME?.trim()
+  return resolve(configured !== undefined && configured.length > 0 ? configured : join(homedir(), '.dsh'))
+}
+
+/**
+ * Join path segments onto the Harness home.
+ * @param segments - path segments appended to the home.
+ * @returns the normalized absolute joined path.
+ */
+export function dshHomePath(...segments: string[]): string {
+  return join(dshHome(), ...segments)
+}

package/src/index.ts

@@ -7,18 +7,18 @@
  * changes.
  */
 
-import { homedir } from 'node:os'
-import { join } from 'node:path'
 import type { Context } from '@deepseek-ai/cordis'
 import type {} from '@deepseek-ai/dsh-settings'
 import z from 'schemastery'
 import type {} from '@deepseek-ai/dsh-host-webserver'
 import type {} from '@deepseek-ai/dsh-llm'
 import type {} from '@deepseek-ai/dsh-system-prompt'
-import type { ConfigPatch, NovelConfig } from './protocol.ts'
+import type { ConfigPatch, NovelConfig, ConfigResponse } from './protocol.ts'
 import { makeRoutes } from './routes.ts'
 import { activeBookOutputDir } from './bookshelf.ts'
 import { loadProject } from './engine.ts'
+import { dshHomePath } from './home.ts'
+import { novelSettingsNamespace, persistConfigPatch, type SettingsLike } from './settings-persist.ts'
 
 /** Stable cordis plugin name. */
 export const name = 'novel-forge'
@@ -28,11 +28,18 @@ export const inject = ['webServer', 'llm', 'systemPrompt', 'settings']
 
 /**
  * Settings namespace of the novel-forge capability — the section the web
- * settings surface edits. Spelled here rather than imported: the browser half
- * spells the same value and must not depend on a Host package.
+ * settings surface edits.
+ *
+ * Legacy fallback only: from Harness 0.2 the Loader derives the section from this
+ * plugin's exported `Config`, and addresses it by the **entry id** it assigned in
+ * the active profile. That id is not guaranteed to be this name (bundle rows may
+ * be given a generated id), so the live id is read from the context instead —
+ * see {@link novelSettingsNamespace}. The constant stays for the browser half and
+ * for installations whose entry id really is this spelling.
  */
 export const NOVEL_SETTINGS_NAMESPACE = 'dsh-novel-forge' as const
 
+
 /** Plugin config, validated by the same-named schemastery schema. */
 export interface Config {
   /** When true (default), a system-prompt section announces the plugin to every agent. */
@@ -83,36 +90,70 @@ export interface Config {
   enableAdaptMode?: boolean
 }
 
+/**
+ * Mark a config field as live-editable.
+ *
+ * From Harness 0.2 the settings surface builds its editable form from this
+ * plugin's exported `Config` and only exposes fields whose nearest ancestor
+ * declares volatility (`SettingsForms.schema` + `volatileForm`). Without the
+ * marker the host refuses a write with `Plugin entry "…" has no volatile
+ * fields`, so the plugin's own 保存设置 route could never persist anything.
+ *
+ * The canonical form is schemastery's `.volatile()` builder, but that builder
+ * ships only in the Harness fork (`@deepseek-ai/schemastery`); the npm
+ * `schemastery` this package depends on exposes the same switch as metadata
+ * (`extra('volatile', true)`), which is what the fork's builder sets anyway.
+ * Prefer the builder when the loaded copy has it, fall back to the metadata so
+ * the schema is correct on either copy.
+ * @param schema - the field schema.
+ * @returns the same field marked live-editable.
+ */
+function live<T>(schema: T): T {
+  const candidate = schema as unknown as {
+    volatile?: () => unknown
+    extra?: (key: string, value: boolean) => unknown
+  }
+  if (typeof candidate.volatile === 'function') return candidate.volatile() as T
+  if (typeof candidate.extra === 'function') return candidate.extra('volatile', true) as T
+  return schema
+}
+
+/**
+ * Plugin config schema.
+ *
+ * Every field is live-editable (see {@link live}): the plugin re-reads the
+ * settings section on every request, so a change must apply without remounting.
+ */
 export const Config: z<Config> = z.object({
-  announceToAgent: z.boolean().default(true),
-  enabled: z.boolean().default(true),
-  outlinePath: z.string().default(''),
-  outputDir: z.string().default(join(homedir(), '.dsh', 'novels')),
-  provider: z.string().default('deepseek-official'),
-  model: z.string().default('deepseek-flash'),
-  generateModel: z.string().default(''),
-  reviewModel: z.string().default(''),
-  auditModel: z.string().default(''),
-  fallbackModel: z.string().default(''),
-  autoTimeline: z.boolean().default(true),
-  reasoningEffort: z.union(['off', 'low', 'high', 'max']).default('off'),
-  analysisReasoning: z.union(['off', 'low', 'high', 'max']).default('low'),
-  chapterChars: z.number().default(3500),
-  maxTokens: z.number().default(12000),
-  reviewPassScore: z.number().default(70),
-  autoReview: z.boolean().default(true),
-  autoAuthorReview: z.boolean().default(true),
-  autoReviewAfterRevise: z.boolean().default(true),
-  themeBackground: z.string().default(''),
-  themeBackgroundBlur: z.number().default(0),
-  themeOpacity: z.number().default(100),
-  enableAdaptMode: z.boolean().default(true),
+  announceToAgent: live(z.boolean().default(true)),
+  enabled: live(z.boolean().default(true)),
+  outlinePath: live(z.string().default('')),
+  outputDir: live(z.string().default(dshHomePath('novels'))),
+  provider: live(z.string().default('deepseek-official')),
+  model: live(z.string().default('deepseek-flash')),
+  generateModel: live(z.string().default('')),
+  reviewModel: live(z.string().default('')),
+  auditModel: live(z.string().default('')),
+  fallbackModel: live(z.string().default('')),
+  autoTimeline: live(z.boolean().default(true)),
+  reasoningEffort: live(z.union(['off', 'low', 'high', 'max']).default('off')),
+  analysisReasoning: live(z.union(['off', 'low', 'high', 'max']).default('low')),
+  chapterChars: live(z.number().default(3500)),
+  maxTokens: live(z.number().default(12000)),
+  reviewPassScore: live(z.number().default(70)),
+  autoReview: live(z.boolean().default(true)),
+  autoAuthorReview: live(z.boolean().default(true)),
+  autoReviewAfterRevise: live(z.boolean().default(true)),
+  themeBackground: live(z.string().default('')),
+  themeBackgroundBlur: live(z.number().default(0)),
+  themeOpacity: live(z.number().default(100)),
+  enableAdaptMode: live(z.boolean().default(true)),
 })
 
 /** Schema defaults, re-read for hand-built test contexts. */
 const DEFAULT_ANNOUNCE = true
 const DEFAULT_OUTLINE_PATH = ''
-const DEFAULT_OUTPUT_DIR = join(homedir(), '.dsh', 'novels')
+const DEFAULT_OUTPUT_DIR = dshHomePath('novels')
 const DEFAULT_PROVIDER = 'deepseek-official'
 const DEFAULT_MODEL = 'deepseek-flash'
 const DEFAULT_REASONING_EFFORT = 'off' as const
@@ -130,7 +171,7 @@ const DEFAULT_AUTO_REVIEW_AFTER_REVISE = true
 const SECTION_ORDER = 160
 
 /** Model-facing announcement: plugin presence, capabilities, and limits. */
-export const NOVEL_GUIDANCE = '本机已安装 dsh-novel-forge 插件（AI 编译小说工作台）：侧边栏「小说工坊」入口。能力：读取 docx 大纲、粘贴大纲文本或把 txt 全本拆章导入；用 LLM 提炼道藏（人设/世界观/金手指规则/写作红线）与大世界（境界/区域/势力）；生成卷计划与结构化章节计划（本章目标/剧情要点/必达项/义务合约/章末钩子）；逐章生成 3000-4000 字正文并保存为 Markdown（默认输出到用户主目录 ~/.dsh/novels）；每章自动生成摘要与编年录事实、抽取故事时间线（故事内时间/地点/在场角色/事件）、自动 AI 审稿（人设/设定/红线/文笔/爽点/逻辑/反 AI/呈现/合规九维）与作者复盘；支持按审稿意见合并修订（审稿 high + 时间线矛盾 + 张力偏差合成一轮，每章最多两轮后转人工）、去 AI 味润色、章节历史版本回滚、暗线（伏笔）管理、剧情线与角色库、张力曲线、全书一致性质检与敏感词扫描、分范围导出与项目备份、批量连写（生产单，可暂停/续跑）。限制：所有 AI 操作消耗 LLM API 额度；输出目录与模型可在插件设置中修改；章节正文质量取决于大纲完整度；张力/时间线问题默认只是建议（写日志与待办），不会自动改正文。用户提到「小说 / 大纲 / 写小说 / 章节 / 审稿 / 润色」时即指本插件，请据此协作。'
+export const NOVEL_GUIDANCE = '本机已安装 dsh-novel-forge 插件（AI 编译小说工作台）：侧边栏「小说工坊」入口。能力：读取 docx 大纲、粘贴大纲文本或把 txt 全本拆章导入；用 LLM 提炼道藏（人设/世界观/金手指规则/写作红线）与大世界（境界/区域/势力）；生成卷计划与结构化章节计划（本章目标/剧情要点/必达项/义务合约/章末钩子）；逐章生成 3000-4000 字正文并保存为 Markdown（默认输出到插件设置里「输出目录」指定的位置，未配置时落到 Harness 数据目录下的 novels/书名）；每章自动生成摘要与编年录事实、抽取故事时间线（故事内时间/地点/在场角色/事件）、自动 AI 审稿（人设/设定/红线/文笔/爽点/逻辑/反 AI/呈现/合规九维）与作者复盘；支持按审稿意见合并修订（审稿 high + 时间线矛盾 + 张力偏差合成一轮，每章最多两轮后转人工）、去 AI 味润色、章节历史版本回滚、暗线（伏笔）管理、剧情线与角色库、张力曲线、全书一致性质检与敏感词扫描、分范围导出与项目备份、批量连写（生产单，可暂停/续跑）。限制：所有 AI 操作消耗 LLM API 额度；输出目录与模型可在插件设置中修改；章节正文质量取决于大纲完整度；张力/时间线问题默认只是建议（写日志与待办），不会自动改正文。用户提到「小说 / 大纲 / 写小说 / 章节 / 审稿 / 润色」时即指本插件，请据此协作。'
 
 /** Resolve a config-like value into the full runtime config. */
 export function resolveConfig(value: Partial<Config> | undefined): NovelConfig {
@@ -165,9 +206,13 @@ export function resolveConfig(value: Partial<Config> | undefined): NovelConfig {
  * @param config - resolved plugin config (schema defaults applied by the loader).
  */
 export function apply(ctx: Context, config?: Config): void {
-  // The live source the routes read: the settings section once the web
-  // settings surface is served, the composition entry otherwise.
-  let current: () => Config = () => config ?? {}
+  // The live source the routes read. Held as an explicit state object, not a
+  // self-referential closure: `current = () => ({ ...current(), ...patch })`
+  // re-reads `current` inside the function it is assigning, so the first save
+  // recurses until the stack blows (`Maximum call stack size exceeded`).
+  let currentState: Config = config ?? {}
+  const current = (): Config => currentState
+  const setCurrent = (next: Config): void => { currentState = next }
   const resolve = (): NovelConfig => {
     const resolved = resolveConfig(current())
     // 书架激活的书优先决定输出目录（settings 仍可改默认值）。
@@ -190,7 +235,25 @@ export function apply(ctx: Context, config?: Config): void {
     return book
   }
 
-  const patchConfig = async (patch: ConfigPatch): Promise<NovelConfig> => {
+  /**
+   * The settings section this instance is addressed by, resolved once at mount.
+   * Undefined when the host exposes neither an entry id nor a registered section,
+   * in which case config edits stay in memory (see {@link patchConfig}).
+   */
+  const settingsNs = novelSettingsNamespace(ctx)
+
+  /**
+   * Persist one config patch.
+   *
+   * Applies to the live config either way. Persistence goes through the host
+   * settings section when that section is addressable; when it is not (unknown
+   * entry id, or the section never registered), the patch is kept in memory and
+   * the caller is told it will not survive a restart — a config save must never
+   * fail the request just because the host does not offer persistence.
+   * @param patch - the subset to merge.
+   * @returns the config the host will read next, plus a persistence caveat.
+   */
+  const patchConfig = async (patch: ConfigPatch): Promise<ConfigResponse> => {
     const next: ConfigPatch = {}
     if (patch.outlinePath !== undefined) next.outlinePath = patch.outlinePath
     if (patch.outputDir !== undefined) next.outputDir = patch.outputDir
@@ -213,16 +276,24 @@ export function apply(ctx: Context, config?: Config): void {
     if (patch.themeBackgroundBlur !== undefined) next.themeBackgroundBlur = patch.themeBackgroundBlur
     if (patch.themeOpacity !== undefined) next.themeOpacity = patch.themeOpacity
     if (patch.enableAdaptMode !== undefined) next.enableAdaptMode = patch.enableAdaptMode
-    // Persist through the settings seam when available; otherwise keep in memory.
-    // (ctx.get is the non-strict service access — no inject requirement, same
-    // pattern installSettingsSection itself uses.)
-    const settings = ctx.get('settings')
-    if (settings !== undefined) {
-      await settings.update(NOVEL_SETTINGS_NAMESPACE, next as Record<string, unknown>)
-    } else {
-      current = () => ({ ...current(), ...next })
+    // Apply + persist via the shared helper: the host section is addressed by the
+    // entry id the Loader assigned, never by a guessed constant. (ctx.get is the
+    // non-strict service access — same pattern the 0.1.x wiring used.)
+    const settings = ctx.get('settings') as SettingsLike | undefined
+    const persisted = await persistConfigPatch(
+      current() as unknown as Record<string, unknown>,
+      next as Record<string, unknown>,
+      settings,
+      settingsNs,
+      NOVEL_SETTINGS_NAMESPACE,
+    )
+    // Keep the live source in sync even when persistence went through the host:
+    // until the host reloads the plugin, the state object is still the old snapshot.
+    setCurrent(persisted.config as Config)
+    if (persisted.settingsWarning !== undefined) {
+      ctx.logger('novel-forge').warn('%s', persisted.settingsWarning)
     }
-    return resolve()
+    return { config: resolve(), ...(persisted.settingsWarning === undefined ? {} : { settingsWarning: persisted.settingsWarning }) }
   }
 
   let disposeSection: (() => void) | undefined
@@ -246,7 +317,7 @@ export function apply(ctx: Context, config?: Config): void {
         text: NOVEL_GUIDANCE,
       })
     }
-    const routes = makeRoutes({ ctx, getConfig: resolve, patchConfig, rawConfig: () => current() })
+    const routes = makeRoutes({ ctx, getConfig: resolve, patchConfig, rawConfig: () => current(), settingsNs })
     disposeRoutes = ctx.effect(
       () => {
         const disposers = routes.map(route => ctx.webServer.register(route))
@@ -257,13 +328,27 @@ export function apply(ctx: Context, config?: Config): void {
     void value
   }
 
-  ctx.settings.installSection(ctx, NOVEL_SETTINGS_NAMESPACE, Config, config ?? {}, {
-    setSource: (source) => {
-      current = source
-      sync()
-    },
-    onChange: sync,
-  })
+  // settings 接线做跨版本兼容（2026-09-30 适配 dsh 0.2.0-rc.2）：
+  // 0.1.x 提供 installSection（注册配置段并回调 setSource/onChange），段 id 由
+  // 调用方指定，因此用声明的常量；
+  // 0.2.x 移除了它，改由 Loader 从本插件导出的 Config schema 推导配置段（段 id
+  // = 该 entry 的 id），settings 服务只剩 configure/describe/update/replace/mutate。
+  // 缺失时必须静默跳过，不能让整棵插件挂载失败（此前 novel-forge 因此整包不加载）。
+  const settingsService = ctx.get('settings') as unknown as
+    | { installSection?: (c: unknown, ns: string, schema: unknown, cfg: unknown, hooks: unknown) => void }
+    | undefined
+  if (settingsService !== undefined && typeof settingsService.installSection === 'function') {
+    settingsService.installSection(ctx, settingsNs ?? NOVEL_SETTINGS_NAMESPACE, Config, config ?? {}, {
+      setSource: (source: () => Config) => {
+        // Snapshot the host-provided source. The live source is one state object
+        // now, so a source that recomputed per call would have to be polled;
+        // 0.1.x wiring is legacy and the composition entry is authoritative here.
+        setCurrent(source())
+        sync()
+      },
+      onChange: sync,
+    })
+  }
 
   // 运行时 skill：章节批量生产与值守处理（只在本插件环境可用）。
   // 通过 ctx.get('skills') 获取注册表（服务存在时注册，缺失则跳过，不影响其他能力）。

package/src/llm-live.ts

@@ -19,8 +19,6 @@ import type {
 
 export type { LlmLiveFrame }
 
-export type LlmLivePhase = 'requesting' | 'streaming' | 'completed' | 'failed'
-
 /** 帧缓冲上限（新连接重放用）。 */
 const MAX_BUFFER = 400
 /** Prompt 记录上限（环形，避免长跑吃内存）。 */
@@ -68,14 +66,6 @@ export function emitLive(frame: LlmLiveFrame): void {
   for (const l of listeners) l(frame)
 }
 
-export function liveLatest(): LlmLiveFrame[] {
-  return [...buffer]
-}
-
-export function clearLiveFeed(): void {
-  buffer = []
-}
-
 // ------------------------------------------------------------------ 用量账本
 
 /** 把 dsh-llm 的 usage 转成线上形状。 */
@@ -183,11 +173,6 @@ export function livePrompt(sessionId: string): LlmPromptRecord | undefined {
   return prompts.get(sessionId)
 }
 
-/** 清空 Prompt 记录。 */
-export function clearLivePrompts(): void {
-  prompts.clear()
-}
-
 // --------------------------------------------------------------- 调用打点
 
 /** 一次调用的打点句柄。 */