@zwbigi/ink-xy 0.1.2 → 0.1.4

package/inkos/packages/core/src/agent/context-transform.ts

@@ -0,0 +1,85 @@
+import type { AgentMessage } from "@mariozechner/pi-agent-core";
+import type { UserMessage } from "@mariozechner/pi-ai";
+import { readdir, readFile } from "node:fs/promises";
+import { join } from "node:path";
+import { isNewLayoutBook } from "../utils/outline-paths.js";
+
+/** Files read in this order; anything else in story/ comes after, sorted alphabetically. */
+const PRIORITY_FILES = [
+  "story_bible.md",
+  "volume_outline.md",
+  "book_rules.md",
+  "current_focus.md",
+];
+
+const UPGRADE_HINT =
+  "[提示] 当前这本书的架构稿是旧的条目式格式（story_bible.md / volume_outline.md / character_matrix.md）。" +
+  "如果作者有意愿升级成段落式架构稿 + 一人一卡的角色目录（outline/story_frame.md + outline/volume_map.md + roles/），" +
+  "可以调用 `sub_agent(architect, { revise: true, bookId, feedback: \"把架构稿从条目式升级成段落式架构稿，并把角色矩阵拆成 roles 目录一人一卡\" })`。" +
+  "升级只改架构稿，不动已写的章节。在作者没明确同意前不要主动触发。";
+
+export function createBookContextTransform(
+  bookId: string | null,
+  projectRoot: string,
+): (messages: AgentMessage[], signal?: AbortSignal) => Promise<AgentMessage[]> {
+  if (bookId === null) {
+    return async (messages) => messages;
+  }
+
+  const bookDir = join(projectRoot, "books", bookId);
+  const storyDir = join(bookDir, "story");
+
+  return async (messages) => {
+    const sections = await readTruthFiles(storyDir);
+    if (sections.length === 0) return messages;
+
+    const isNew = await isNewLayoutBook(bookDir);
+    const hintBlock = isNew ? "" : `\n\n${UPGRADE_HINT}`;
+
+    const body =
+      "[以下是当前书籍的真相文件，每次对话时自动从磁盘读取注入。请基于这些内容进行创作和判断。]" +
+      hintBlock + "\n\n" +
+      sections.map((s) => `=== ${s.name} ===\n${s.content}`).join("\n\n");
+
+    const injected: UserMessage = {
+      role: "user",
+      content: body,
+      timestamp: Date.now(),
+    };
+
+    return [injected, ...messages];
+  };
+}
+
+interface TruthFileSection {
+  name: string;
+  content: string;
+}
+
+async function readTruthFiles(storyDir: string): Promise<TruthFileSection[]> {
+  let entries: string[];
+  try {
+    entries = await readdir(storyDir);
+  } catch {
+    return [];
+  }
+
+  const mdFiles = entries.filter((f) => f.endsWith(".md"));
+  if (mdFiles.length === 0) return [];
+
+  const prioritySet = new Set(PRIORITY_FILES);
+  const prioritized = PRIORITY_FILES.filter((f) => mdFiles.includes(f));
+  const rest = mdFiles.filter((f) => !prioritySet.has(f)).sort();
+  const ordered = [...prioritized, ...rest];
+
+  const sections: TruthFileSection[] = [];
+  for (const fileName of ordered) {
+    try {
+      const content = await readFile(join(storyDir, fileName), "utf-8");
+      sections.push({ name: fileName, content });
+    } catch {
+      // skip unreadable files
+    }
+  }
+  return sections;
+}

package/inkos/packages/core/src/agent/index.ts

@@ -0,0 +1,14 @@
+export { buildAgentSystemPrompt } from "./agent-system-prompt.js";
+export {
+  createSubAgentTool,
+  createReadTool,
+  createWriteTruthFileTool,
+  createRenameEntityTool,
+  createPatchChapterTextTool,
+  createEditTool,
+  createWriteFileTool,
+  createGrepTool,
+  createLsTool,
+} from "./agent-tools.js";
+export { runAgentSession, evictAgentCache, type AgentSessionConfig, type AgentSessionResult } from "./agent-session.js";
+export { createBookContextTransform } from "./context-transform.js";

package/inkos/packages/core/src/agents/ai-tells.ts

@@ -0,0 +1,161 @@
+/**
+ * Structural AI-tell detection — pure rule-based analysis (no LLM).
+ *
+ * Detects patterns common in AI-generated Chinese text:
+ * - dim 20: Paragraph length uniformity (low variance)
+ * - dim 21: Filler/hedge word density
+ * - dim 22: Formulaic transition patterns
+ * - dim 23: List-like structure (consecutive same-prefix sentences)
+ */
+
+export interface AITellIssue {
+  readonly severity: "warning" | "info";
+  readonly category: string;
+  readonly description: string;
+  readonly suggestion: string;
+}
+
+export interface AITellResult {
+  readonly issues: ReadonlyArray<AITellIssue>;
+}
+
+type AITellLanguage = "zh" | "en";
+
+const HEDGE_WORDS: Record<AITellLanguage, ReadonlyArray<string>> = {
+  zh: ["似乎", "可能", "或许", "大概", "某种程度上", "一定程度上", "在某种意义上"],
+  en: ["seems", "seemed", "perhaps", "maybe", "apparently", "in some ways", "to some extent"],
+};
+
+const TRANSITION_WORDS: Record<AITellLanguage, ReadonlyArray<string>> = {
+  zh: ["然而", "不过", "与此同时", "另一方面", "尽管如此", "话虽如此", "但值得注意的是"],
+  en: ["however", "meanwhile", "on the other hand", "nevertheless", "even so", "still"],
+};
+
+/**
+ * Analyze text content for structural AI-tell patterns.
+ * Returns issues that can be merged into audit results.
+ */
+export function analyzeAITells(content: string, language: AITellLanguage = "zh"): AITellResult {
+  const issues: AITellIssue[] = [];
+  const isEnglish = language === "en";
+  const joiner = isEnglish ? ", " : "、";
+
+  const paragraphs = content
+    .split(/\n\s*\n/)
+    .map((p) => p.trim())
+    .filter((p) => p.length > 0);
+
+  // dim 20: Paragraph length uniformity (needs ≥3 paragraphs)
+  if (paragraphs.length >= 3) {
+    const paragraphLengths = paragraphs.map((p) => p.length);
+    const mean = paragraphLengths.reduce((a, b) => a + b, 0) / paragraphLengths.length;
+    if (mean > 0) {
+      const variance = paragraphLengths.reduce((sum, l) => sum + (l - mean) ** 2, 0) / paragraphLengths.length;
+      const stdDev = Math.sqrt(variance);
+      const cv = stdDev / mean;
+      if (cv < 0.15) {
+        issues.push({
+          severity: "warning",
+          category: isEnglish ? "Paragraph uniformity" : "段落等长",
+          description: isEnglish
+            ? `Paragraph-length coefficient of variation is only ${cv.toFixed(3)} (threshold <0.15), which suggests unnaturally uniform paragraph sizing`
+            : `段落长度变异系数仅${cv.toFixed(3)}（阈值<0.15），段落长度过于均匀，呈现AI生成特征`,
+          suggestion: isEnglish
+            ? "Increase paragraph-length contrast: use shorter beats for impact and longer blocks for immersive detail"
+            : "增加段落长度差异：短段落用于节奏加速或冲击，长段落用于沉浸描写",
+        });
+      }
+    }
+  }
+
+  // dim 21: Hedge word density
+  const totalChars = content.length;
+  if (totalChars > 0) {
+    let hedgeCount = 0;
+    for (const word of HEDGE_WORDS[language]) {
+      const regex = new RegExp(word, isEnglish ? "gi" : "g");
+      const matches = content.match(regex);
+      hedgeCount += matches?.length ?? 0;
+    }
+    const hedgeDensity = hedgeCount / (totalChars / 1000);
+    if (hedgeDensity > 3) {
+      issues.push({
+        severity: "warning",
+        category: isEnglish ? "Hedge density" : "套话密度",
+        description: isEnglish
+          ? `Hedge-word density is ${hedgeDensity.toFixed(1)} per 1k characters (threshold >3), making the prose sound overly tentative`
+          : `套话词（似乎/可能/或许等）密度为${hedgeDensity.toFixed(1)}次/千字（阈值>3），语气过于模糊犹豫`,
+        suggestion: isEnglish
+          ? "Replace hedges with firmer narration: remove vague qualifiers and use concrete detail instead"
+          : "用确定性叙述替代模糊表达：去掉「似乎」直接描述状态，用具体细节替代「可能」",
+      });
+    }
+  }
+
+  // dim 22: Formulaic transition repetition
+  const transitionCounts: Record<string, number> = {};
+  for (const word of TRANSITION_WORDS[language]) {
+    const regex = new RegExp(word, isEnglish ? "gi" : "g");
+    const matches = content.match(regex);
+    const count = matches?.length ?? 0;
+    if (count > 0) {
+      transitionCounts[isEnglish ? word.toLowerCase() : word] = count;
+    }
+  }
+  const repeatedTransitions = Object.entries(transitionCounts)
+    .filter(([, count]) => count >= 3);
+  if (repeatedTransitions.length > 0) {
+    const detail = repeatedTransitions
+      .map(([word, count]) => `"${word}"×${count}`)
+      .join(joiner);
+    issues.push({
+      severity: "warning",
+      category: isEnglish ? "Formulaic transitions" : "公式化转折",
+      description: isEnglish
+        ? `Transition words repeat too often: ${detail}. Reusing the same transition pattern 3+ times creates a formulaic AI texture`
+        : `转折词重复使用：${detail}。同一转折模式≥3次暴露AI生成痕迹`,
+      suggestion: isEnglish
+        ? "Let scenes pivot through action, timing, or viewpoint shifts instead of repeating the same transitions"
+        : "用情节自然转折替代转折词，或换用不同的过渡手法（动作切入、时间跳跃、视角切换）",
+    });
+  }
+
+  // dim 23: List-like structure (consecutive sentences with same prefix pattern)
+  const sentences = content
+    .split(isEnglish ? /[.!?\n]/ : /[。！？\n]/)
+    .map((s) => s.trim())
+    .filter((s) => s.length > 2);
+
+  if (sentences.length >= 3) {
+    let consecutiveSamePrefix = 1;
+    let maxConsecutive = 1;
+    for (let i = 1; i < sentences.length; i++) {
+      const prevPrefix = isEnglish
+        ? sentences[i - 1]!.split(/\s+/)[0]?.toLowerCase() ?? ""
+        : sentences[i - 1]!.slice(0, 2);
+      const currPrefix = isEnglish
+        ? sentences[i]!.split(/\s+/)[0]?.toLowerCase() ?? ""
+        : sentences[i]!.slice(0, 2);
+      if (prevPrefix === currPrefix) {
+        consecutiveSamePrefix++;
+        maxConsecutive = Math.max(maxConsecutive, consecutiveSamePrefix);
+      } else {
+        consecutiveSamePrefix = 1;
+      }
+    }
+    if (maxConsecutive >= 3) {
+      issues.push({
+        severity: "info",
+        category: isEnglish ? "List-like structure" : "列表式结构",
+        description: isEnglish
+          ? `Detected ${maxConsecutive} consecutive sentences with the same opening pattern, creating a list-like generated cadence`
+          : `检测到${maxConsecutive}句连续以相同开头的句子，呈现列表式AI生成结构`,
+        suggestion: isEnglish
+          ? "Vary how sentences open: change subject, timing, or action entry to break the list effect"
+          : "变换句式开头：用不同主语、时间词、动作词开头，打破列表感",
+      });
+    }
+  }
+
+  return { issues };
+}