@waterwx/dsh-novel-forge 1.2.1-alpha → 1.3.0-alpha

package/src/assistant.ts

@@ -37,6 +37,9 @@ import {
 } from './engine.ts'
 import { rewriteChapterStream, countHanzi } from './engine.ts'
 import { computeBookStage, isGenericContinue, renderStageContract, splitByStage, stageAllows, type BookStage } from './stage-contract.ts'
+import { beginLiveCall, endLiveCall, markFirstToken } from './llm-live.ts'
+import { withModelFallback } from './llm-retry.ts'
+import { createSnapshot } from './snapshots.ts'
 import { NovelActionError, classifyActionError, FailureLedger } from './action-guard.ts'
 
 /** History file name inside the output dir. */
@@ -481,6 +484,8 @@ export async function* executeAction(
       if (chapter === undefined || draft === undefined || draft === '') {
         throw new NovelActionError('contract', `章节 ${no} 修订后没有产出草稿`)
       }
+      // 覆盖前留存历史版本（助手直接修订会就地覆盖正文）。
+      if (chapter.file !== undefined && chapter.file !== '') createSnapshot(outputDir, chapter, '助手修订前')
       const fileName = chapterFileName(chapter)
       mkdirSync(outputDir, { recursive: true })
       writeFileSync(join(outputDir, fileName), `# 第${chapter.no}章 ${chapter.title}\n\n${draft}\n`, 'utf8')
@@ -806,37 +811,48 @@ async function chatOnce(
       messages[messages.length - 1] = { ...last, content: newBlocks }
     }
   }
-  const request: GenerateOptions = {
-    provider: config.provider,
-    model: config.model,
-    messages,
-    system,
-    // v4-flash 推理模型：reasoning channel 占预算，给足避免回复被截断。
-    maxTokens: Math.max(config.maxTokens, 16000),
-    temperature: 0.7,
-  }
-  const assembler = new BlockAssembler()
-  for await (const chunk of ctx.llm.stream(request)) {
-    assembler.push(chunk)
-  }
-  const finish = assembler.finish
-  if (finish.kind === 'error' || finish.kind === 'aborted') {
-    throw new Error(`助手调用失败（${finish.kind}）: ${finish.failure.message}`)
-  }
-  const blocks = assembler.blocks()
-  const textBlocks = blocks
-    .filter((block): block is Extract<StreamChunk, { type: 'block-end' }>['block'] & { type: 'text' } => block.type === 'text')
-    .map(block => block.text)
-  let text = textBlocks.join('\n').trim()
-  if (text === '') {
-    const reasoning = blocks
-      .filter((block): block is { type: 'reasoning'; text: string } => block.type === 'reasoning')
+  /** 用指定模型跑一次（含实况打点与 token 记账）。 */
+  const runOnce = async (model: string): Promise<string> => {
+    const request: GenerateOptions = {
+      provider: config.provider,
+      model,
+      messages,
+      system,
+      // 推理模型的 reasoning channel 占预算，给足避免回复被截断。
+      maxTokens: Math.max(config.maxTokens, 16000),
+      temperature: 0.7,
+    }
+    // 实况打点：助手对话同样计入用量账本，并能查看实际发送的 Prompt。
+    const live = beginLiveCall({ label: 'AI 编辑对话', model, system, user: messages.map(m => JSON.stringify(m.content)).join(' / ').slice(0, 8000) })
+    const assembler = new BlockAssembler()
+    for await (const chunk of ctx.llm.stream(request)) {
+      if (chunk.type === 'text-delta') markFirstToken(live)
+      assembler.push(chunk)
+    }
+    const finish = assembler.finish
+    if (finish.kind === 'error' || finish.kind === 'aborted') {
+      const message = `助手调用失败（${finish.kind}）: ${finish.failure.message}`
+      endLiveCall(live, assembler, { phase: 'failed', error: message })
+      throw new Error(message)
+    }
+    const blocks = assembler.blocks()
+    const textBlocks = blocks
+      .filter((block): block is Extract<StreamChunk, { type: 'block-end' }>['block'] & { type: 'text' } => block.type === 'text')
       .map(block => block.text)
-      .join('\n')
-      .trim()
-    if (reasoning !== '') text = reasoning
+    let text = textBlocks.join('\n').trim()
+    if (text === '') {
+      const reasoning = blocks
+        .filter((block): block is { type: 'reasoning'; text: string } => block.type === 'reasoning')
+        .map(block => block.text)
+        .join('\n')
+        .trim()
+      if (reasoning !== '') text = reasoning
+    }
+    endLiveCall(live, assembler, { chars: text.length, preview: text.slice(0, 320), phase: text === '' ? 'failed' : 'completed' })
+    return text
   }
-  return text
+  // 主模型失败（网络/限流/额度类）且配置了备用模型时，自动换模型重试一次。
+  return withModelFallback({ model: config.model, fallbackModel: config.fallbackModel }, runOnce)
 }
 
 /** Run one user turn. Yields stream frames; persists history. */

package/src/client/api.ts

@@ -152,8 +152,8 @@ export class NovelApi {
     return postJson<ForeshadowResponse>(NOVEL_API.foreshadow, req)
   }
 
-  async exportBook(format: 'txt' | 'md'): Promise<ExportResponse> {
-    return postJson<ExportResponse>(NOVEL_API.exportBook, { format })
+  async exportBook(format: 'txt' | 'md' | 'json', scope: import('../protocol.ts').ExportScope = 'book'): Promise<ExportResponse> {
+    return postJson<ExportResponse>(NOVEL_API.exportBook, { format, scope })
   }
 
   async chapter(no: number): Promise<ChapterResponse> {
@@ -579,6 +579,80 @@ export class NovelApi {
     }
   }
 
+  /** 张力曲线：数据 + 规则核对。 */
+  async tension(): Promise<import('../protocol.ts').TensionResponse> {
+    const response = await fetch(withBookId(NOVEL_API.tension))
+    return readJson<import('../protocol.ts').TensionResponse>(response)
+  }
+
+  /** 设置张力曲线预设。 */
+  async tensionPreset(preset: import('../protocol.ts').TensionCurvePreset): Promise<void> {
+    await postJson(NOVEL_API.tension, { op: 'preset', preset })
+  }
+
+  /** 设置/清除某一章的目标张力。 */
+  async tensionSet(chapterNo: number, tension?: number): Promise<void> {
+    await postJson(NOVEL_API.tension, { op: 'set', chapterNo, tension })
+  }
+
+  /** 提示词槽位：列表。 */
+  async promptSlots(): Promise<import('../protocol.ts').PromptSlotsResponse> {
+    const response = await fetch(withBookId(NOVEL_API.promptSlots))
+    return readJson<import('../protocol.ts').PromptSlotsResponse>(response)
+  }
+
+  /** 写入一个提示词槽位（空串 = 清空）。 */
+  async promptSlotSet(id: import('../protocol.ts').PromptSlotId, value: string): Promise<void> {
+    await postJson(NOVEL_API.promptSlots, { id, value })
+  }
+
+  /** 故事时间线：列表（含规则初筛问题）。 */
+  async timeline(): Promise<import('../protocol.ts').TimelineResponse> {
+    const response = await fetch(withBookId(NOVEL_API.timeline))
+    return readJson<import('../protocol.ts').TimelineResponse>(response)
+  }
+
+  /** 抽取某一章的时间线事件。 */
+  async timelineExtract(chapterNo: number): Promise<{ events: import('../protocol.ts').TimelineEvent[] }> {
+    return postJson<{ events: import('../protocol.ts').TimelineEvent[] }>(NOVEL_API.timeline, { op: 'extract', chapterNo })
+  }
+
+  /** 检查时间线矛盾（规则初筛 + AI 复核）。 */
+  async timelineCheck(): Promise<import('../protocol.ts').TimelineResponse> {
+    return postJson<import('../protocol.ts').TimelineResponse>(NOVEL_API.timeline, { op: 'check' })
+  }
+
+  /** 手工修正一条时间线事件。 */
+  async timelineUpdate(event: import('../protocol.ts').TimelineEvent): Promise<void> {
+    await postJson(NOVEL_API.timeline, { op: 'update', event })
+  }
+
+  /** 删除一条时间线事件。 */
+  async timelineRemove(id: string): Promise<void> {
+    await postJson(NOVEL_API.timeline, { op: 'remove', id })
+  }
+
+  /** 章节历史版本列表。 */
+  async snapshots(): Promise<import('../protocol.ts').SnapshotResponse> {
+    const response = await fetch(withBookId(NOVEL_API.snapshot))
+    return readJson<import('../protocol.ts').SnapshotResponse>(response)
+  }
+
+  /** 回滚到某个历史版本。 */
+  async snapshotRestore(id: string): Promise<void> {
+    await postJson(NOVEL_API.snapshot, { op: 'restore', id })
+  }
+
+  /** 给某章手工存档。 */
+  async snapshotCreate(chapterNo: number): Promise<void> {
+    await postJson(NOVEL_API.snapshot, { op: 'create', chapterNo })
+  }
+
+  /** 删除一个历史版本。 */
+  async snapshotRemove(id: string): Promise<void> {
+    await postJson(NOVEL_API.snapshot, { op: 'remove', id })
+  }
+
   /** Load the persisted assistant conversation. */
   async assistantHistory(): Promise<import('../protocol.ts').AssistantMessage[]> {
     const response = await fetch(withBookId(NOVEL_API.assistantHistory))

package/src/client/llmLive.ts

@@ -4,22 +4,10 @@
  */
 
 import { useEffect, useMemo, useState } from 'react'
-import { NOVEL_API } from '../protocol.ts'
+import { NOVEL_API, type LlmLiveFrame, type LlmTokenUsage } from '../protocol.ts'
 
-export interface LlmLiveFrame {
-  type: 'session_started' | 'output_delta' | 'reasoning_delta' | 'phase_changed' | 'session_completed'
-  sessionId: string
-  label?: string
-  model?: string
-  at: string
-  content?: string
-  totalChars?: number
-  totalReasoningChars?: number
-  phase?: 'requesting' | 'streaming' | 'completed' | 'failed'
-  phaseMessage?: string
-  preview?: string
-  error?: string
-}
+// 帧结构定义在 protocol.ts（宿主与浏览器共用同一份线上形状），此处只再导出。
+export type { LlmLiveFrame }
 
 export interface LlmLiveSession {
   sessionId: string
@@ -31,6 +19,16 @@ export interface LlmLiveSession {
   totalChars: number
   startedAt: string
   updatedAt: string
+  /** 本次调用的 token 用量（供应商未上报时缺省）。 */
+  usage?: LlmTokenUsage
+  /** 本次调用耗时（毫秒）。 */
+  elapsedMs?: number
+  /** 首字耗时（毫秒）。 */
+  firstTokenMs?: number
+  /** 是否保存了可查看的 Prompt。 */
+  hasPrompt?: boolean
+  /** Prompt 字符数（system + user）。 */
+  promptChars?: number
 }
 
 const MAX_PREVIEW_CHARS = 20_000
@@ -50,6 +48,8 @@ function applyFrame(current: Record<string, LlmLiveSession>, frame: LlmLiveFrame
       totalChars: 0,
       startedAt: frame.at,
       updatedAt: frame.at,
+      hasPrompt: frame.hasPrompt,
+      promptChars: frame.promptChars,
     }
     next[id] = session
     return next
@@ -66,7 +66,17 @@ function applyFrame(current: Record<string, LlmLiveSession>, frame: LlmLiveFrame
     session = { ...session, phase: frame.phase ?? session.phase, phaseMessage: frame.phaseMessage ?? session.phaseMessage, updatedAt: frame.at }
     next[id] = session
   } else if (frame.type === 'session_completed') {
-    session = { ...session, phase: frame.phase ?? 'completed', phaseMessage: frame.phase === 'failed' ? (frame.error ?? '调用失败') : '模型结果已准备完成', totalChars: frame.totalChars ?? session.totalChars, preview: session.preview || (frame.preview ?? ''), updatedAt: frame.at }
+    session = {
+      ...session,
+      phase: frame.phase ?? 'completed',
+      phaseMessage: frame.phase === 'failed' ? (frame.error ?? '调用失败') : '模型结果已准备完成',
+      totalChars: frame.totalChars ?? session.totalChars,
+      preview: session.preview || (frame.preview ?? ''),
+      updatedAt: frame.at,
+      usage: frame.usage ?? session.usage,
+      elapsedMs: frame.elapsedMs ?? session.elapsedMs,
+      firstTokenMs: frame.firstTokenMs ?? session.firstTokenMs,
+    }
     next[id] = session
   }
   return next

package/src/client/panel/AssetsTab.tsx

@@ -1,6 +1,6 @@
 /**
  * 写作资产页签：题材基底库 / 推进模式库 / 反 AI 规则 / 写法引擎。
- * 学习自 AI-Novel-Writing-Assistant 的四大资产模块，注入到生成与审稿提示词中。
+ * 四大资产模块（题材 / 推进模式 / 反 AI 规则 / 叙事风格），注入到生成与审稿提示词中。
  */
 import { useCallback, useEffect, useState } from 'react'
 import type { NovelApi } from '../api.ts'
@@ -294,7 +294,7 @@ export function AssetsTab({ api, initialTab = 'genre' }: AssetsTabProps) {
       {assetTab === 'templates' && (
         <div className={css.card} style={{ flex: 1, minHeight: 0 }}>
           <span className={css.cardTitle}>笔法帖</span>
-          <span className={css.meta}>从内置 8 套叙事风格模板中一键选用（来自 AI-Novel-Writing-Assistant 写法引擎），无需样本文本；绑定后生成与润色都遵循该风格。</span>
+          <span className={css.meta}>从内置 8 套叙事风格模板中一键选用，无需样本文本；绑定后生成与润色都遵循该风格。</span>
           <div style={{ display: 'flex', flexDirection: 'column', gap: 'var(--nf-space-8)', flex: 1, minHeight: 0, overflowY: 'auto' }}>
             {data.styleTemplates.map(template => {
               const bound = assets.styleAssets.some(s => s.name === template.name)

package/src/client/panel/LiveFeedLog.tsx

@@ -1,67 +1,204 @@
-/**
- * 「实时调用 / LIVE」—— 内嵌到 AI 进度窗的下半区（详情层），不再独立浮窗。
- * 连接 /llm-live/stream，按 LLM 调用分组展示模型实时输出。
- *
- * 状态色全部走 panel.module.css 的 .lfPill / .lfDot（主题令牌 + --nf-streaming），
- * 不再内联硬编码色；布局性 inline style 保留（均为令牌引用或动态值）。
- */
-
-import { useMemo, useState } from 'react'
-import { useLlmLiveFeed, type LlmLiveSession } from '../llmLive.ts'
-import css from './panel.module.css'
-
-function PhasePill({ s }: { s: LlmLiveSession }): JSX.Element {
-  return (
-    <span className={css.lfPill} data-phase={s.phase}>
-      {s.phase === 'requesting' ? '请求中' : s.phase === 'streaming' ? '生成中' : s.phase === 'failed' ? '失败' : '完成'}
-    </span>
-  )
-}
-
-function SessionCard({ s, brief }: { s: LlmLiveSession; brief: boolean }): JSX.Element {
-  const active = s.phase === 'requesting' || s.phase === 'streaming'
-  const dotState = active ? 'streaming' : s.phase
-  return (
-    <div style={{ borderBottom: '1px solid var(--nf-border)', padding: '7px 6px' }}>
-      <div style={{ display: 'flex', alignItems: 'center', gap: 8, fontSize: 12 }}>
-        <span className={css.lfDot} data-state={dotState} />
-        <b style={{ flex: 1, overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{s.label}</b>
-        {s.model !== undefined && <span style={{ color: 'var(--nf-text-2)', fontSize: 10, maxWidth: 130, overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{s.model}</span>}
-        <PhasePill s={s} />
-      </div>
-      <div style={{ fontSize: 10, color: 'var(--nf-text-2)', marginTop: 2, display: 'flex', gap: 8 }}>
-        <span>{s.totalChars} 字符</span>
-        {s.phaseMessage !== '' && <span style={{ overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{s.phaseMessage}</span>}
-      </div>
-      {!brief && s.preview !== '' && (
-        <pre style={{ whiteSpace: 'pre-wrap', wordBreak: 'break-word', fontSize: 11, lineHeight: 1.5, color: 'var(--nf-text)', backgroundColor: 'var(--nf-bg-inset)', borderRadius: 8, padding: 6, marginTop: 6, marginBottom: 0 }}>{s.preview}</pre>
-      )}
-    </div>
-  )
-}
-
-export default function LiveFeedLog(): JSX.Element {
-  const [brief, setBrief] = useState(true)
-  const [cleared, setCleared] = useState(false)
-  const { connected, sessions } = useLlmLiveFeed(true)
-  const active = sessions.filter(s => s.phase === 'requesting' || s.phase === 'streaming').length
-  const shown = useMemo(() => (cleared ? sessions.slice(-8) : sessions), [sessions, cleared])
-  return (
-    <div style={{ borderTop: '1px solid var(--nf-border)', marginTop: 8, paddingTop: 6, flex: '1 1 auto', minHeight: 140, display: 'flex', flexDirection: 'column' }}>
-      <div style={{ display: 'flex', alignItems: 'center', gap: 8, fontSize: 12, padding: '2px 4px' }}>
-        <span className={css.lfDot} data-size="lg" data-state={connected ? 'ok' : 'idle'} />
-        <b style={{ flex: 1 }}>实时调用 / LIVE</b>
-        <span className={css.lfPill} data-phase={active > 0 ? 'streaming' : 'requesting'}>
-          {active > 0 ? `${active} 项进行中` : connected ? '等待生成' : '正在连接'}
-        </span>
-        <button type="button" title={brief ? '详细' : '简略'} onClick={() => setBrief(b => !b)} style={{ border: '1px solid var(--nf-border)', background: 'var(--nf-bg)', color: 'var(--nf-text)', borderRadius: 6, fontSize: 10, padding: '1px 6px', cursor: 'pointer' }}>{brief ? '详细' : '简略'}</button>
-        <button type="button" title="清空当前窗口" onClick={() => setCleared(true)} style={{ border: '1px solid var(--nf-border)', background: 'var(--nf-bg)', color: 'var(--nf-text)', borderRadius: 6, fontSize: 10, padding: '1px 6px', cursor: 'pointer' }}>清空</button>
-      </div>
-      <div style={{ flex: 1, minHeight: 0, overflowY: 'auto', display: 'flex', flexDirection: 'column' }}>
-        {shown.length === 0 ? (
-          <div style={{ padding: 10, fontSize: 12, color: 'var(--nf-text-2)' }}>{connected ? '暂无实时调用。开始一次生成/审稿/计划即可看到。' : '正在连接 AI 实况服务…'}</div>
-        ) : shown.map(s => <SessionCard key={s.sessionId} s={s} brief={brief} />)}
-      </div>
-    </div>
-  )
-}
+/**
+ * 「实时调用 / LIVE」—— 内嵌到 AI 进度窗的下半区（详情层），不再独立浮窗。
+ * 连接 /llm-live/stream，按 LLM 调用分组展示模型实时输出。
+ *
+ * 除实时输出外，这里还承担两件与「花了多少」有关的事：
+ *   1) 每次调用显示 token 用量（输入↑ / 输出↓ / 思考）与耗时、首字耗时；
+ *   2) 「查看 Prompt」按 sessionId 拉取这次调用**实际发送**的 system / user。
+ * 顶部另有一条「本次运行」累计条（进程内，重启归零）。
+ *
+ * 状态色全部走 panel.module.css 的 .lfPill / .lfDot（主题令牌 + --nf-streaming），
+ * 不再内联硬编码色；布局性 inline style 保留（均为令牌引用或动态值）。
+ */
+
+import { useCallback, useEffect, useMemo, useState } from 'react'
+import { useLlmLiveFeed, type LlmLiveSession } from '../llmLive.ts'
+import { NOVEL_API, type LlmPromptRecord, type LlmUsageSummary } from '../../protocol.ts'
+import css from './panel.module.css'
+
+/** 毫秒 → 紧凑可读（840ms / 1.2s）。 */
+function fmtMs(ms: number | undefined): string {
+  if (ms === undefined || ms <= 0) return ''
+  return ms < 1000 ? Math.round(ms) + 'ms' : (ms / 1000).toFixed(1) + 's'
+}
+
+/** token 数 → 紧凑（12.3k）。 */
+function fmtTokens(n: number | undefined): string {
+  if (n === undefined || n <= 0) return '0'
+  return n < 1000 ? String(n) : (n / 1000).toFixed(1) + 'k'
+}
+
+/** 一次调用的用量摘要（输入↑ / 输出↓ / 思考 / 缓存）。 */
+function usageText(s: LlmLiveSession): string {
+  const u = s.usage
+  if (u === undefined) return '用量未上报'
+  const parts = ['↑' + fmtTokens(u.inputTokens), '↓' + fmtTokens(u.outputTokens)]
+  if ((u.reasoningTokens ?? 0) > 0) parts.push('思考 ' + fmtTokens(u.reasoningTokens))
+  if ((u.cacheReadTokens ?? 0) > 0) parts.push('缓存 ' + fmtTokens(u.cacheReadTokens))
+  return parts.join(' · ')
+}
+
+function PhasePill({ s }: { s: LlmLiveSession }): JSX.Element {
+  return (
+    <span className={css.lfPill} data-phase={s.phase}>
+      {s.phase === 'requesting' ? '请求中' : s.phase === 'streaming' ? '生成中' : s.phase === 'failed' ? '失败' : '完成'}
+    </span>
+  )
+}
+
+/** 「查看 Prompt」弹层：拉取这次调用实际发送的 system + user。 */
+function PromptDialog({ sessionId, onClose }: { sessionId: string; onClose: () => void }): JSX.Element {
+  const [record, setRecord] = useState<LlmPromptRecord | null>(null)
+  const [error, setError] = useState('')
+  const [tab, setTab] = useState<'user' | 'system'>('user')
+  useEffect(() => {
+    let cancelled = false
+    void (async () => {
+      try {
+        const res = await fetch(NOVEL_API.llmPrompt + '?sessionId=' + encodeURIComponent(sessionId))
+        if (!res.ok) {
+          const body = await res.json().catch(() => ({})) as { error?: string }
+          throw new Error(body.error ?? 'HTTP ' + res.status)
+        }
+        const data = await res.json() as LlmPromptRecord
+        if (!cancelled) setRecord(data)
+      } catch (err) {
+        if (!cancelled) setError((err as Error).message)
+      }
+    })()
+    return () => { cancelled = true }
+  }, [sessionId])
+  const text = record === null ? '' : (tab === 'user' ? (record.user ?? '') : (record.system ?? ''))
+  return (
+    <div
+      style={{ position: 'fixed', inset: 0, background: 'rgba(0,0,0,.45)', zIndex: 60, display: 'flex', alignItems: 'center', justifyContent: 'center', padding: 20 }}
+      onClick={onClose}
+    >
+      <div
+        onClick={e => e.stopPropagation()}
+        style={{ background: 'var(--nf-bg)', border: '1px solid var(--nf-border)', borderRadius: 10, width: 'min(880px, 96vw)', maxHeight: '86vh', display: 'flex', flexDirection: 'column', padding: 12, gap: 8 }}
+      >
+        <div style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
+          <b style={{ flex: 1, fontSize: 13 }}>实际发送的 Prompt{record?.label !== undefined ? ' · ' + record.label : ''}</b>
+          {record !== null && <span className={css.meta}>共 {record.chars.toLocaleString()} 字符{record.truncated === true ? '（记录已截断）' : ''}</span>}
+          <button type="button" className={css.iconButton} title="关闭" onClick={onClose}>×</button>
+        </div>
+        <div style={{ display: 'flex', gap: 6 }}>
+          <button type="button" className={css.button + (tab === 'user' ? ' ' + css.buttonPrimary : '')} onClick={() => setTab('user')}>User / 输入</button>
+          <button type="button" className={css.button + (tab === 'system' ? ' ' + css.buttonPrimary : '')} onClick={() => setTab('system')}>System / 系统提示</button>
+        </div>
+        {error !== '' && <span style={{ color: 'var(--nf-error)', fontSize: 12 }}>{error}</span>}
+        {record === null && error === '' && <span className={css.meta}>加载中…</span>}
+        {record !== null && (
+          <pre style={{ flex: 1, minHeight: 0, overflow: 'auto', whiteSpace: 'pre-wrap', wordBreak: 'break-word', fontSize: 11, lineHeight: 1.55, background: 'var(--nf-bg-inset)', borderRadius: 8, padding: 8, margin: 0 }}>
+            {text === '' ? '（空）' : text}
+          </pre>
+        )}
+      </div>
+    </div>
+  )
+}
+
+function SessionCard({ s, brief, onShowPrompt }: { s: LlmLiveSession; brief: boolean; onShowPrompt: (sessionId: string) => void }): JSX.Element {
+  const active = s.phase === 'requesting' || s.phase === 'streaming'
+  const dotState = active ? 'streaming' : s.phase
+  return (
+    <div style={{ borderBottom: '1px solid var(--nf-border)', padding: '7px 6px' }}>
+      <div style={{ display: 'flex', alignItems: 'center', gap: 8, fontSize: 12 }}>
+        <span className={css.lfDot} data-state={dotState} />
+        <b style={{ flex: 1, overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{s.label}</b>
+        {s.model !== undefined && <span style={{ color: 'var(--nf-text-2)', fontSize: 10, maxWidth: 130, overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{s.model}</span>}
+        <PhasePill s={s} />
+      </div>
+      <div style={{ fontSize: 10, color: 'var(--nf-text-2)', marginTop: 2, display: 'flex', gap: 8, flexWrap: 'wrap', alignItems: 'center' }}>
+        <span>{s.totalChars} 字符</span>
+        <span title="输入 / 输出 token">{usageText(s)}</span>
+        {s.elapsedMs !== undefined && <span title="总耗时">用时 {fmtMs(s.elapsedMs)}</span>}
+        {s.firstTokenMs !== undefined && <span title="首个字返回耗时">首字 {fmtMs(s.firstTokenMs)}</span>}
+        {s.phaseMessage !== '' && <span style={{ overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap' }}>{s.phaseMessage}</span>}
+        {s.hasPrompt === true && (
+          <button
+            type="button"
+            title="查看这次调用实际发送的 Prompt"
+            onClick={() => onShowPrompt(s.sessionId)}
+            style={{ border: '1px solid var(--nf-border)', background: 'var(--nf-bg)', color: 'var(--nf-text)', borderRadius: 6, fontSize: 10, padding: '0 5px', cursor: 'pointer' }}
+          >查看 Prompt</button>
+        )}
+      </div>
+      {!brief && s.preview !== '' && (
+        <pre style={{ whiteSpace: 'pre-wrap', wordBreak: 'break-word', fontSize: 11, lineHeight: 1.5, color: 'var(--nf-text)', backgroundColor: 'var(--nf-bg-inset)', borderRadius: 8, padding: 6, marginTop: 6, marginBottom: 0 }}>{s.preview}</pre>
+      )}
+    </div>
+  )
+}
+
+export default function LiveFeedLog(): JSX.Element {
+  const [brief, setBrief] = useState(true)
+  const [cleared, setCleared] = useState(false)
+  const [promptId, setPromptId] = useState<string | null>(null)
+  const { connected, sessions } = useLlmLiveFeed(true)
+  const active = sessions.filter(s => s.phase === 'requesting' || s.phase === 'streaming').length
+  const shown = useMemo(() => (cleared ? sessions.slice(-8) : sessions), [sessions, cleared])
+
+  // 本次运行用量（进程内累计，来自 /status.usage）。
+  const [usage, setUsage] = useState<LlmUsageSummary | null>(null)
+  const loadUsage = useCallback(async () => {
+    try {
+      const res = await fetch(NOVEL_API.status + '?slim=1')
+      if (!res.ok) return
+      const data = await res.json() as { usage?: LlmUsageSummary }
+      setUsage(data.usage ?? null)
+    } catch { /* 用量是尽力而为，不影响实况 */ }
+  }, [])
+  useEffect(() => {
+    void loadUsage()
+    const timer = window.setInterval(() => { void loadUsage() }, 15_000)
+    return () => { window.clearInterval(timer) }
+  }, [loadUsage])
+
+  const resetUsage = useCallback(async (): Promise<void> => {
+    try {
+      await fetch(NOVEL_API.usageReset, { method: 'POST', headers: { 'content-type': 'application/json' }, body: '{}' })
+      await loadUsage()
+    } catch { /* ignore */ }
+  }, [loadUsage])
+
+  return (
+    <div style={{ borderTop: '1px solid var(--nf-border)', marginTop: 8, paddingTop: 6, flex: '1 1 auto', minHeight: 140, display: 'flex', flexDirection: 'column' }}>
+      <div style={{ display: 'flex', alignItems: 'center', gap: 8, fontSize: 12, padding: '2px 4px' }}>
+        <span className={css.lfDot} data-size="lg" data-state={connected ? 'ok' : 'idle'} />
+        <b style={{ flex: 1 }}>实时调用 / LIVE</b>
+        <span className={css.lfPill} data-phase={active > 0 ? 'streaming' : 'requesting'}>
+          {active > 0 ? `${active} 项进行中` : connected ? '等待生成' : '正在连接'}
+        </span>
+        <button type="button" title={brief ? '详细' : '简略'} onClick={() => setBrief(b => !b)} style={{ border: '1px solid var(--nf-border)', background: 'var(--nf-bg)', color: 'var(--nf-text)', borderRadius: 6, fontSize: 10, padding: '1px 6px', cursor: 'pointer' }}>{brief ? '详细' : '简略'}</button>
+        <button type="button" title="清空当前窗口" onClick={() => setCleared(true)} style={{ border: '1px solid var(--nf-border)', background: 'var(--nf-bg)', color: 'var(--nf-text)', borderRadius: 6, fontSize: 10, padding: '1px 6px', cursor: 'pointer' }}>清空</button>
+      </div>
+      {usage !== null && usage.calls > 0 && (
+        <div
+          title="本次运行累计（进程内，重启归零）"
+          style={{ display: 'flex', alignItems: 'center', gap: 10, flexWrap: 'wrap', fontSize: 10, color: 'var(--nf-text-2)', padding: '3px 4px', borderTop: '1px dashed var(--nf-border)' }}
+        >
+          <span>本次运行：{usage.calls} 次调用{usage.failed > 0 ? `（失败 ${usage.failed}）` : ''}</span>
+          <span>↑{fmtTokens(usage.inputTokens)}</span>
+          <span>↓{fmtTokens(usage.outputTokens)}</span>
+          {usage.reasoningTokens > 0 && <span>思考 {fmtTokens(usage.reasoningTokens)}</span>}
+          <span>用时 {fmtMs(usage.elapsedMs)}</span>
+          {usage.callsWithoutUsage > 0 && <span>（{usage.callsWithoutUsage} 次未上报用量）</span>}
+          <button
+            type="button"
+            title="清零本次运行用量"
+            onClick={() => { void resetUsage() }}
+            style={{ marginLeft: 'auto', border: '1px solid var(--nf-border)', background: 'var(--nf-bg)', color: 'var(--nf-text)', borderRadius: 6, fontSize: 10, padding: '1px 6px', cursor: 'pointer' }}
+          >重新计数</button>
+        </div>
+      )}
+      <div style={{ flex: 1, minHeight: 0, overflowY: 'auto', display: 'flex', flexDirection: 'column' }}>
+        {shown.length === 0 ? (
+          <div style={{ padding: 10, fontSize: 12, color: 'var(--nf-text-2)' }}>{connected ? '暂无实时调用。开始一次生成/审稿/计划即可看到。' : '正在连接 AI 实况服务…'}</div>
+        ) : shown.map(s => <SessionCard key={s.sessionId} s={s} brief={brief} onShowPrompt={setPromptId} />)}
+      </div>
+      {promptId !== null && <PromptDialog sessionId={promptId} onClose={() => setPromptId(null)} />}
+    </div>
+  )
+}

package/src/client/panel/ModelManager.tsx

@@ -11,8 +11,8 @@ import { LLM_VENDORS } from '../../protocol.ts'
 import { tt } from './helpers.ts'
 import css from './panel.module.css'
 
-/** 目录不可用时回退的内置预设（历史行为）。 */
-export const FALLBACK_MODEL_PRESETS = ['deepseek-v4-flash', 'deepseek-v4-pro', 'deepseek-chat', 'deepseek-reasoner'] as const
+/** 目录不可用时回退的内置预设（对齐当前模型目录）。 */
+export const FALLBACK_MODEL_PRESETS = ['deepseek-flash', 'deepseek-v4-pro', 'deepseek-chat', 'deepseek-reasoner'] as const
 
 const CUSTOM = '__custom__'
 interface TestState { testing: boolean; ok?: boolean; ms?: number; message?: string }