|
|
@@ -13,6 +13,7 @@ export interface WriteChapterInput {
|
|
|
readonly chapterNumber: number;
|
|
|
readonly externalContext?: string;
|
|
|
readonly wordCountOverride?: number;
|
|
|
+ readonly temperatureOverride?: number;
|
|
|
}
|
|
|
|
|
|
export interface WriteChapterOutput {
|
|
|
@@ -71,6 +72,9 @@ export class WriterAgent extends BaseAgent {
|
|
|
book, genreProfile, bookRules, bookRulesBody, genreBody, styleGuide, styleFingerprint,
|
|
|
);
|
|
|
|
|
|
+ const dialogueFingerprints = this.extractDialogueFingerprints(recentChapters, storyBible);
|
|
|
+ const relevantSummaries = this.findRelevantSummaries(chapterSummaries, volumeOutline, chapterNumber);
|
|
|
+
|
|
|
const userPrompt = this.buildUserPrompt({
|
|
|
chapterNumber,
|
|
|
storyBible,
|
|
|
@@ -85,17 +89,31 @@ export class WriterAgent extends BaseAgent {
|
|
|
subplotBoard,
|
|
|
emotionalArcs,
|
|
|
characterMatrix,
|
|
|
+ dialogueFingerprints,
|
|
|
+ relevantSummaries,
|
|
|
});
|
|
|
|
|
|
+ const temperature = input.temperatureOverride ?? 0.7;
|
|
|
+
|
|
|
const response = await this.chat(
|
|
|
[
|
|
|
{ role: "system", content: systemPrompt },
|
|
|
{ role: "user", content: userPrompt },
|
|
|
],
|
|
|
- { maxTokens: 16384, temperature: 0.7 },
|
|
|
+ { maxTokens: 16384, temperature },
|
|
|
);
|
|
|
|
|
|
- return this.parseOutput(chapterNumber, response.content, genreProfile);
|
|
|
+ const output = this.parseOutput(chapterNumber, response.content, genreProfile);
|
|
|
+
|
|
|
+ // #4: Post-write constraint validation (regex-based, zero cost)
|
|
|
+ const violations = this.validateConstraints(output.content, bookRules);
|
|
|
+ if (violations.length > 0) {
|
|
|
+ process.stderr.write(
|
|
|
+ `[writer] Constraint violations in chapter ${chapterNumber}: ${violations.join("; ")}\n`,
|
|
|
+ );
|
|
|
+ }
|
|
|
+
|
|
|
+ return output;
|
|
|
}
|
|
|
|
|
|
async saveChapter(
|
|
|
@@ -145,6 +163,8 @@ export class WriterAgent extends BaseAgent {
|
|
|
readonly subplotBoard: string;
|
|
|
readonly emotionalArcs: string;
|
|
|
readonly characterMatrix: string;
|
|
|
+ readonly dialogueFingerprints?: string;
|
|
|
+ readonly relevantSummaries?: string;
|
|
|
}): string {
|
|
|
const contextBlock = params.externalContext
|
|
|
? `\n## 外部指令\n以下是来自外部系统的创作指令，请在本章中融入：\n\n${params.externalContext}\n`
|
|
|
@@ -170,6 +190,14 @@ export class WriterAgent extends BaseAgent {
|
|
|
? `\n## 角色交互矩阵\n${params.characterMatrix}\n`
|
|
|
: "";
|
|
|
|
|
|
+ const fingerprintBlock = params.dialogueFingerprints
|
|
|
+ ? `\n## 角色对话指纹\n${params.dialogueFingerprints}\n`
|
|
|
+ : "";
|
|
|
+
|
|
|
+ const relevantBlock = params.relevantSummaries
|
|
|
+ ? `\n## 相关历史章节摘要\n${params.relevantSummaries}\n`
|
|
|
+ : "";
|
|
|
+
|
|
|
return `请续写第${params.chapterNumber}章。
|
|
|
${contextBlock}
|
|
|
## 当前状态卡
|
|
|
@@ -177,7 +205,7 @@ ${params.currentState}
|
|
|
${ledgerBlock}
|
|
|
## 伏笔池
|
|
|
${params.hooks}
|
|
|
-${summariesBlock}${subplotBlock}${emotionalBlock}${matrixBlock}
|
|
|
+${summariesBlock}${subplotBlock}${emotionalBlock}${matrixBlock}${fingerprintBlock}${relevantBlock}
|
|
|
## 最近章节
|
|
|
${params.recentChapters || "(这是第一章，无前文)"}
|
|
|
|
|
|
@@ -329,6 +357,151 @@ ${params.volumeOutline}
|
|
|
}
|
|
|
}
|
|
|
|
|
|
+ /** Validate hard constraints against content. Returns list of violation descriptions. */
|
|
|
+ private validateConstraints(content: string, bookRules: BookRules | null): ReadonlyArray<string> {
|
|
|
+ const violations: string[] = [];
|
|
|
+
|
|
|
+ // Built-in hard constraints from writer-prompts
|
|
|
+ if (/不是[^，。！？\n]{0,30}[，,]?\s*而是/.test(content)) {
|
|
|
+ violations.push("硬性禁令：出现了「不是……而是……」句式");
|
|
|
+ }
|
|
|
+ if (content.includes("——")) {
|
|
|
+ violations.push("硬性禁令：出现了破折号「——」");
|
|
|
+ }
|
|
|
+
|
|
|
+ // Book-level prohibitions (check as substring)
|
|
|
+ if (bookRules?.prohibitions) {
|
|
|
+ for (const prohibition of bookRules.prohibitions) {
|
|
|
+ // Only check short prohibitions as literal patterns (long ones are descriptive)
|
|
|
+ if (prohibition.length <= 15 && content.includes(prohibition)) {
|
|
|
+ violations.push(`本书禁忌：出现了"${prohibition}"`);
|
|
|
+ }
|
|
|
+ }
|
|
|
+ }
|
|
|
+
|
|
|
+ return violations;
|
|
|
+ }
|
|
|
+
|
|
|
+ /**
|
|
|
+ * Extract dialogue fingerprints from recent chapters.
|
|
|
+ * For each character with multiple dialogue lines, compute speaking style markers.
|
|
|
+ */
|
|
|
+ private extractDialogueFingerprints(recentChapters: string, _storyBible: string): string {
|
|
|
+ if (!recentChapters) return "";
|
|
|
+
|
|
|
+ // Match dialogue patterns: "speaker said" or dialogue in quotes
|
|
|
+ // Chinese dialogue typically uses "" or 「」
|
|
|
+ const dialogueRegex = /(?:(.{1,6})(?:说道|道|喝道|冷声道|笑道|怒道|低声道|大声道|喝骂道|冷笑道|沉声道|喊道|叫道|问道|答道)\s*[：:]\s*["""「]([^"""」]+)["""」])|["""「]([^"""」]{2,})["""」]/g;
|
|
|
+
|
|
|
+ const characterDialogues = new Map<string, string[]>();
|
|
|
+ let match: RegExpExecArray | null;
|
|
|
+
|
|
|
+ while ((match = dialogueRegex.exec(recentChapters)) !== null) {
|
|
|
+ const speaker = match[1]?.trim();
|
|
|
+ const line = match[2] ?? match[3] ?? "";
|
|
|
+ if (speaker && line.length > 1) {
|
|
|
+ const existing = characterDialogues.get(speaker) ?? [];
|
|
|
+ characterDialogues.set(speaker, [...existing, line]);
|
|
|
+ }
|
|
|
+ }
|
|
|
+
|
|
|
+ // Only include characters with >=2 dialogue lines
|
|
|
+ const fingerprints: string[] = [];
|
|
|
+ for (const [character, lines] of characterDialogues) {
|
|
|
+ if (lines.length < 2) continue;
|
|
|
+
|
|
|
+ const avgLen = Math.round(lines.reduce((sum, l) => sum + l.length, 0) / lines.length);
|
|
|
+ const isShort = avgLen < 15;
|
|
|
+
|
|
|
+ // Find frequent words/phrases (2+ occurrences)
|
|
|
+ const wordCounts = new Map<string, number>();
|
|
|
+ for (const line of lines) {
|
|
|
+ // Extract 2-3 char segments as "words"
|
|
|
+ for (let i = 0; i < line.length - 1; i++) {
|
|
|
+ const bigram = line.slice(i, i + 2);
|
|
|
+ wordCounts.set(bigram, (wordCounts.get(bigram) ?? 0) + 1);
|
|
|
+ }
|
|
|
+ }
|
|
|
+ const frequentWords = [...wordCounts.entries()]
|
|
|
+ .filter(([, count]) => count >= 2)
|
|
|
+ .sort((a, b) => b[1] - a[1])
|
|
|
+ .slice(0, 3)
|
|
|
+ .map(([w]) => `「${w}」`);
|
|
|
+
|
|
|
+ // Detect style markers
|
|
|
+ const markers: string[] = [];
|
|
|
+ if (isShort) markers.push("短句为主");
|
|
|
+ else markers.push("长句为主");
|
|
|
+
|
|
|
+ const questionCount = lines.filter((l) => l.includes("？") || l.includes("?")).length;
|
|
|
+ if (questionCount > lines.length * 0.3) markers.push("反问多");
|
|
|
+
|
|
|
+ if (frequentWords.length > 0) markers.push(`常用${frequentWords.join("")}`);
|
|
|
+
|
|
|
+ fingerprints.push(`${character}：${markers.join("，")}`);
|
|
|
+ }
|
|
|
+
|
|
|
+ return fingerprints.length > 0 ? fingerprints.join("；") : "";
|
|
|
+ }
|
|
|
+
|
|
|
+ /**
|
|
|
+ * Find relevant chapter summaries based on volume outline context.
|
|
|
+ * Extracts character names and hook IDs from the current volume's outline,
|
|
|
+ * then searches chapter summaries for matching entries.
|
|
|
+ */
|
|
|
+ private findRelevantSummaries(
|
|
|
+ chapterSummaries: string,
|
|
|
+ volumeOutline: string,
|
|
|
+ chapterNumber: number,
|
|
|
+ ): string {
|
|
|
+ if (!chapterSummaries || chapterSummaries === "(文件尚未创建)") return "";
|
|
|
+ if (!volumeOutline || volumeOutline === "(文件尚未创建)") return "";
|
|
|
+
|
|
|
+ // Extract character names from volume outline (Chinese name patterns)
|
|
|
+ const nameRegex = /[\u4e00-\u9fff]{2,4}(?=[，、。：]|$)/g;
|
|
|
+ const outlineNames = new Set<string>();
|
|
|
+ let nameMatch: RegExpExecArray | null;
|
|
|
+ while ((nameMatch = nameRegex.exec(volumeOutline)) !== null) {
|
|
|
+ outlineNames.add(nameMatch[0]);
|
|
|
+ }
|
|
|
+
|
|
|
+ // Extract hook IDs from volume outline
|
|
|
+ const hookRegex = /H\d{2,}/g;
|
|
|
+ const hookIds = new Set<string>();
|
|
|
+ let hookMatch: RegExpExecArray | null;
|
|
|
+ while ((hookMatch = hookRegex.exec(volumeOutline)) !== null) {
|
|
|
+ hookIds.add(hookMatch[0]);
|
|
|
+ }
|
|
|
+
|
|
|
+ if (outlineNames.size === 0 && hookIds.size === 0) return "";
|
|
|
+
|
|
|
+ // Search chapter summaries for matching rows
|
|
|
+ const rows = chapterSummaries.split("\n").filter((line) =>
|
|
|
+ line.startsWith("|") && !line.startsWith("| 章节") && !line.startsWith("|--") && !line.startsWith("| -"),
|
|
|
+ );
|
|
|
+
|
|
|
+ const matchedRows = rows.filter((row) => {
|
|
|
+ for (const name of outlineNames) {
|
|
|
+ if (row.includes(name)) return true;
|
|
|
+ }
|
|
|
+ for (const hookId of hookIds) {
|
|
|
+ if (row.includes(hookId)) return true;
|
|
|
+ }
|
|
|
+ return false;
|
|
|
+ });
|
|
|
+
|
|
|
+ // Skip rows for the current chapter and recent chapters (they're already in context)
|
|
|
+ const recentCutoff = Math.max(1, chapterNumber - 3);
|
|
|
+ const filteredRows = matchedRows.filter((row) => {
|
|
|
+ const chNumMatch = row.match(/\|\s*(\d+)\s*\|/);
|
|
|
+ if (!chNumMatch) return true;
|
|
|
+ const num = parseInt(chNumMatch[1]!, 10);
|
|
|
+ return num < recentCutoff;
|
|
|
+ });
|
|
|
+
|
|
|
+ return filteredRows.length > 0 ? filteredRows.join("\n") : "";
|
|
|
+ }
|
|
|
+
|
|
|
private sanitizeFilename(title: string): string {
|
|
|
return title
|
|
|
.replace(/[/\\?%*:|"<>]/g, "")
|
majx_mac