feat(v13): add file-layer PolisherAgent to main pipeline after reviewer cycle [phase 6/8 part 4]

Polisher is a new agent file (packages/core/src/agents/polisher.ts) that runs
AFTER the reviewer+reviser cycle accepts a chapter. It only touches prose
surface — sentence craft, paragraph shape, wording, punctuation, sensory
detail, dialogue naturalness — and is EXPLICITLY FORBIDDEN in its system
prompt from adding/removing plot beats, changing character setup, or
touching the mainline. Plot/structure findings are emitted as
'[polisher-note] ...' lines for the next reviewer pass, not patched in-place.

Prompt encodes the 6 prose-level 雷点: ineffective description, over-purple
prose, weak prose, bad formatting, AI-tell residue, crowd-face reactions.
Plus the file-layer hard rules from new.txt: 3-5 line paragraphs, verbs
over adjectives, five-sense immersion in moderation, character-specific
dialogue voice, ±15% length guard.

Pipeline order is now: planner → writer → reviewer (structural audit) →
reviser (local patch OR full rewrite by failure type) → polisher (file
layer) → persist → (offline) consolidator. Polisher only runs when the
reviewer cycle passed; a failing draft is left visible for the next cycle.

No schema fields added.

packages/core/src/agents/polisher.ts

@@ -0,0 +1,153 @@
+import { BaseAgent } from "./base.js";
+import type { ChapterMemo } from "../models/input-governance.js";
+
+export interface PolishChapterInput {
+  readonly chapterContent: string;
+  readonly chapterNumber: number;
+  readonly chapterMemo?: ChapterMemo;
+  readonly language?: "zh" | "en";
+  readonly temperature?: number;
+}
+
+export interface PolishChapterOutput {
+  readonly polishedContent: string;
+  readonly changed: boolean;
+  readonly tokenUsage?: {
+    readonly promptTokens: number;
+    readonly completionTokens: number;
+    readonly totalTokens: number;
+  };
+}
+
+/**
+ * File-layer polisher — runs AFTER the reviewer+reviser cycle accepts the
+ * chapter's structure. Polisher ONLY touches prose surface: sentence craft,
+ * paragraph shape, wording, punctuation, five-sense immersion, dialogue
+ * naturalness. It is forbidden from changing plot, character, or mainline.
+ *
+ * If a structural/plot issue is found, the polisher marks it in a comment
+ * line (`[polisher-note] ...`) for the next reviewer iteration and leaves
+ * the prose untouched — it does NOT attempt to rewrite across that boundary.
+ */
+export class PolisherAgent extends BaseAgent {
+  get name(): string {
+    return "polisher";
+  }
+
+  async polishChapter(input: PolishChapterInput): Promise<PolishChapterOutput> {
+    const language = input.language ?? "zh";
+    const isEnglish = language === "en";
+
+    const memoBlock = input.chapterMemo
+      ? isEnglish
+        ? `\n\n## Chapter Memo (do NOT let polish drift from this goal)\nGoal: ${input.chapterMemo.goal}\n\n${input.chapterMemo.body}`
+        : `\n\n## 章节备忘（润色不得偏离此目标）\ngoal：${input.chapterMemo.goal}\n\n${input.chapterMemo.body}`
+      : "";
+
+    const systemPrompt = isEnglish
+      ? buildEnglishSystemPrompt()
+      : buildChineseSystemPrompt();
+
+    const userPrompt = isEnglish
+      ? `Polish chapter ${input.chapterNumber}. Return the polished chapter in full, nothing else — no JSON, no headers, no commentary.${memoBlock}\n\n## Chapter Under Polish\n${input.chapterContent}`
+      : `请润色第${input.chapterNumber}章。只返回完整的润色后正文，不要 JSON、不要标题、不要解释。${memoBlock}\n\n## 待润色章节\n${input.chapterContent}`;
+
+    const response = await this.chat(
+      [
+        { role: "system", content: systemPrompt },
+        { role: "user", content: userPrompt },
+      ],
+      { temperature: input.temperature ?? 0.4, maxTokens: 16384 },
+    );
+
+    const raw = response.content.trim();
+    // Strip any leading fenced code block wrapper if the model wraps the
+    // chapter body defensively.
+    const stripped = stripWrappingFence(raw);
+    const polishedContent = stripped.length > 0 ? stripped : input.chapterContent;
+    return {
+      polishedContent,
+      changed: polishedContent !== input.chapterContent,
+      tokenUsage: response.usage,
+    };
+  }
+}
+
+function stripWrappingFence(text: string): string {
+  const fence = text.match(/^```[a-zA-Z]*\n([\s\S]*?)\n```\s*$/);
+  return fence?.[1]?.trim() ?? text;
+}
+
+function buildChineseSystemPrompt(): string {
+  return `你是一位专业中文网文文字层润色编辑。
+
+## 润色边界（硬约束）
+
+你只改文字层——句式 / 段落 / 排版 / 用词 / 五感 / 对话自然度。你禁止增删情节、改变人设、调整主线。发现情节/结构问题只能以 [polisher-note] 形式附在章末供下一轮 reviewer 参考，不能动正文。
+
+结构的事归 Reviewer，不归你。如果读到人设崩、主线偏、冲突缺、memo 未兑现之类的问题，保留原意，不要替作者补情节。
+
+## 6 条文笔类雷点（你要消灭的）
+
+- 描写无效：冗长的环境描写、与主线无关的对话塞满页面。把无效描写删到"一笔带过"。
+- 文笔华丽过度：为辞藻堆辞藻，情感失真，形容词地毯轰炸。让文字服从情绪，不要炫技。
+- 文笔欠佳：句意含混、指代不清、逻辑跳跃、语言干瘪。重写成通顺、有画面感的句子。
+- 排版不规范：段落过长、格式不统一、对话无换行。统一为手机阅读友好格式。
+- （延伸）AI 味痕迹：转折词泛滥、"了"字堆砌、"仿佛/宛如/竟然"等情绪中介词、编剧旁白、分析报告式语言。替换成口语化表达或具体动作。
+- （延伸）群像脸谱化：不写"众人齐声惊呼"，而是挑 1-2 个角色写具体反应。
+
+## 文字层硬规约
+
+- 段落：3-5 行/段（手机阅读），连续 7 行以上必须拆段，但不可把动作+反应拆碎到失去节奏。
+- 句式：多样化，禁止连续 3 句以上同结构/同主语开头；长短交替。
+- 动词 > 形容词：名词+动词驱动画面，一句话最多 1-2 个精准形容词。
+- 五感代入：场景里至少 1-2 种感官细节（视/听/嗅/触/味），但不机械叠加，适度即可。
+- 对话自然度：
+  - 不同角色说话方式有辨识度（用词、句子长短、口头禅、方言痕迹）。
+  - 对话符合说话人当前身份、情绪、信息掌握。
+  - 不写"……"式敷衍应答替代实质交锋。
+- 情绪外化：把"他感到愤怒"改为"他捏碎了茶杯，滚烫的茶水流过指缝"。
+- 删除无意义的叙述者结论（"这一刻他终于明白了力量"—删）和"显然/不禁/仿佛"这类 AI 标记词。
+- 禁止破折号 "——"，禁止"不是……而是……"句式（存量出现一律改写）。
+
+## 输出契约
+
+直接返回润色后的完整章节正文——不要 JSON、不要章节标题行、不要任何解释或进度说明。如果发现必须交给 reviewer 的情节/结构问题，在正文末尾另起一行以 "[polisher-note] " 开头写明，每条一行。没有问题就不加。
+
+保留原文绝大多数句子。只改真正有问题的句子，不要整段重写。修改后章节总长变化不得超过原文字数 ±15%。`;
+}
+
+function buildEnglishSystemPrompt(): string {
+  return `You are a professional English web-fiction prose polisher.
+
+## Polisher Scope (hard constraints)
+
+You touch the prose surface only — sentence craft, paragraph shape, wording, punctuation, sensory detail, dialogue naturalness. You are FORBIDDEN from adding or removing plot beats, changing character setup, or altering the mainline. If you notice plot/structure problems, append a "[polisher-note] ..." line at the very end of the chapter for the next reviewer pass — do NOT attempt to fix them in the prose.
+
+Structure is the Reviewer's job. Do not invent beats to patch a weak chapter.
+
+## 6 prose-level reader-pain patterns you must eliminate
+
+- Ineffective description: long-winded environment setup or off-topic dialogue filler. Compress to a single telling stroke.
+- Over-purple prose: adjective carpet-bombing, words chosen for flourish instead of emotion. Let language serve feeling, not performance.
+- Weak prose: muddy meaning, unclear referents, illogical jumps, flat language. Rewrite into clear, image-carrying sentences.
+- Bad formatting: walls of text, inconsistent layout, un-broken dialogue. Standardise to mobile-reader-friendly shape.
+- (extension) AI-tell residue: excessive transitions, rhetorical hedges, stage-direction voiceover, analytical-report phrasing. Replace with colloquial idiom or concrete action.
+- (extension) Crowd-face reactions: do not write "everyone gasped in unison" — pick one or two characters and write specific reactions.
+
+## Prose-layer hard rules
+
+- Paragraphs: 3-5 lines each for mobile reading; break anything over 7 lines, but do not shatter an action+reaction beat into loss of rhythm.
+- Sentence variety: forbid 3+ consecutive sentences with the same structure or subject; alternate long and short.
+- Verbs > adjectives: noun+verb drives the image; at most 1-2 precise adjectives per sentence.
+- Five senses: at least 1-2 sensory details per scene (sight / sound / smell / touch / taste), but avoid mechanical stacking.
+- Dialogue naturalness: each character has distinct voice (vocabulary, sentence length, verbal tics); dialogue must fit current identity, emotion, information scope; no "..." filler in place of real exchange.
+- Externalise emotion: replace "he felt angry" with "he crushed the teacup, scalding tea running through his fingers".
+- Delete narrator conclusions ("At this moment he finally understood power" — cut) and AI hedge words ("obviously", "as if", "couldn't help but").
+
+## Output contract
+
+Return the polished chapter in full — no JSON, no section headers, no commentary or progress notes. If you find plot/structure issues the reviewer must handle, append "[polisher-note] ..." lines at the very end, one per line. Omit the block if there are no notes.
+
+Preserve the vast majority of sentences. Only rewrite those that truly need it — do not rewrite whole paragraphs. Total length change must stay within ±15% of the original.`;
+}

packages/core/src/index.ts

@@ -88,6 +88,7 @@ export { WriterAgent, type WriteChapterInput, type WriteChapterOutput, type Toke
 export { LengthNormalizerAgent, type NormalizeLengthInput, type NormalizeLengthOutput } from "./agents/length-normalizer.js";
 export { ContinuityAuditor, type AuditResult, type AuditIssue } from "./agents/continuity.js";
 export { ReviserAgent, DEFAULT_REVISE_MODE, type ReviseOutput, type ReviseMode } from "./agents/reviser.js";
+export { PolisherAgent, type PolishChapterInput, type PolishChapterOutput } from "./agents/polisher.js";
 export { RadarAgent, type RadarResult, type RadarRecommendation } from "./agents/radar.js";
 export { FanqieRadarSource, QidianRadarSource, TextRadarSource, type RadarSource, type PlatformRankings, type RankingEntry } from "./agents/radar-source.js";
 export { readGenreProfile, readBookRules, listAvailableGenres, getBuiltinGenresDir } from "./agents/rules-reader.js";

packages/core/src/pipeline/runner.ts

@@ -1248,6 +1248,36 @@ export class PipelineRunner {
     const postReviseCount = reviewResult.postReviseCount;
     const normalizeApplied = reviewResult.normalizeApplied;
 
+    // 3b. File-layer polish pass — runs AFTER structural audit accepts the
+    // chapter. Polisher only touches sentence craft / paragraph shape /
+    // wording / sensory detail / dialogue naturalness. Plot / character /
+    // mainline stay frozen. Skipped when audit did not produce a passing
+    // chapter (leave the failing draft visible for the next cycle) or when
+    // length is dangerously off-range (normalize should handle it, not polish).
+    if (auditResult.passed) {
+      try {
+        const { PolisherAgent } = await import("../agents/polisher.js");
+        const polisher = new PolisherAgent(this.agentCtxFor("polisher", bookId));
+        this.logStage(stageLanguage, { zh: "文字层润色", en: "polishing prose" });
+        const polishOutput = await polisher.polishChapter({
+          chapterContent: finalContent,
+          chapterNumber,
+          chapterMemo: reducedControlInput?.chapterMemo,
+          language: pipelineLang === "en" ? "en" : "zh",
+        });
+        totalUsage = PipelineRunner.addUsage(totalUsage, polishOutput.tokenUsage);
+        if (polishOutput.changed && polishOutput.polishedContent.trim().length > 0) {
+          finalContent = polishOutput.polishedContent;
+          finalWordCount = countChapterLength(finalContent, lengthSpec.countingMode);
+        }
+      } catch (error) {
+        this.logWarn(pipelineLang, {
+          zh: `润色阶段失败：${error instanceof Error ? error.message : String(error)}`,
+          en: `polish stage failed: ${error instanceof Error ? error.message : String(error)}`,
+        });
+      }
+    }
+
     // 4. Save the final chapter and truth files from a single persistence source
     this.logStage(stageLanguage, { zh: "落盘最终章节", en: "persisting final chapter" });
     this.logStage(stageLanguage, { zh: "生成最终真相文件", en: "rebuilding final truth files" });