@dommaker/harness 0.16.6 → 0.16.8

package/src/cli/commands/update-user-model.ts

@@ -12,12 +12,20 @@
  *
  * 模型状态: ~/.claude/user-model-state.json
  * 画像输出: ~/.claude/projects/-root-projects/memory/user_profile.md
+ *
+ * 工单 19-C：transcript 解析/纠正模式/相似度收敛至 cli/session-mining/。
  */
 
 import chalk from 'chalk';
 import * as fs from 'fs';
 import * as path from 'path';
 import * as os from 'os';
+import {
+  readTranscriptSessions,
+  extractCorrectionMatches,
+  cleanCorrectionConcept,
+  jaccardChinese,
+} from '../session-mining';
 
 export interface UpdateUserModelOptions {
   json?: boolean;
@@ -130,56 +138,17 @@ interface SimpleSession {
 }
 
 function findNewSessions(dir: string, processed: string[]): SimpleSession[] {
-  const sessions: SimpleSession[] = [];
-  try {
-    if (!fs.existsSync(dir)) return [];
-    for (const file of fs.readdirSync(dir)) {
-      if (!file.endsWith('.jsonl')) continue;
-      const sessionId = file.replace('.jsonl', '');
-      if (processed.includes(sessionId)) continue;
-
-      const filePath = path.join(dir, file);
-      const stat = fs.statSync(filePath);
-      const turns: Array<{ role: string; content: string }> = [];
-      const toolCalls: string[] = [];
-
-      const content = fs.readFileSync(filePath, 'utf-8');
-      for (const line of content.split('\n')) {
-        if (!line.trim()) continue;
-        try {
-          const entry = JSON.parse(line);
-          const msg = entry.message;
-          if (msg?.role && msg.content) {
-            const text = typeof msg.content === 'string'
-              ? msg.content
-              : Array.isArray(msg.content)
-                ? msg.content.filter((b: any) => b.type === 'text').map((b: any) => b.text).join(' ')
-                : '';
-            turns.push({ role: msg.role, content: text });
-          }
-          // Track tool calls from assistant messages
-          if (msg?.role === 'assistant' && Array.isArray(msg.content)) {
-            for (const block of msg.content) {
-              if (block.type === 'tool_use' && block.name) {
-                toolCalls.push(block.name);
-              }
-            }
-          }
-        } catch {}
-      }
-
-      if (turns.length > 0) {
-        sessions.push({
-          id: sessionId.slice(0, 40),
-          date: stat.mtime.toISOString().slice(0, 10),
-          userText: turns.filter(t => t.role === 'user').map(t => t.content).join('\n'),
-          assistantText: turns.filter(t => t.role === 'assistant').map(t => t.content).join('\n'),
-          toolCalls: [...new Set(toolCalls)],
-        });
-      }
-    }
-  } catch {}
-  return sessions.sort((a, b) => a.date.localeCompare(b.date));
+  const sessions = readTranscriptSessions(dir)
+    .filter(s => !processed.includes(s.id))
+    .sort((a, b) => a.date.localeCompare(b.date));
+
+  return sessions.map(s => ({
+    id: s.id,
+    date: s.date,
+    userText: s.turns.filter(t => t.role === 'user').map(t => t.content).join('\n'),
+    assistantText: s.turns.filter(t => t.role === 'assistant').map(t => t.content).join('\n'),
+    toolCalls: s.toolCalls,
+  }));
 }
 
 // ── Signal extraction ──
@@ -196,33 +165,11 @@ interface SessionSignals {
   knowledgeCaptured: boolean;         // Write to knowledge-docs
 }
 
-// ── Correction clustering (from analyze-sessions) ──
-
 function extractCorrectionConcepts(userText: string): string[] {
-  const correctionPatterns = [
-    /你(?:又|还是)(?:在)?(?:犯|忘|没|不).{0,30}?[了]?/g,
-    /我不是(?:说|让|让做|讲)[了过]?.{0,50}?[了吗?？]?/g,
-    /怎么(?:又|还|老)(?:是|在)?.{0,30}?[了]?/g,
-    /我(?:一直|反复|总是)(?:在|说|强调)?.{0,30}?[了]?/g,
-    /老(?:是|在|犯|忘)(?:了)?.{0,20}?[了]?/g,
-    /这(?:个|种)(?:问题|模式|错误).{0,20}?/g,
-    /(?:补上|加上|记上|修复).{0,10}?[了吗?？]?/g,
-    /(?:沉淀|监控|日志|记录).{0,5}?[了吗?？]?/g,
-  ];
   const concepts: string[] = [];
-  for (const pat of correctionPatterns) {
-    pat.lastIndex = 0;
-    let m;
-    while ((m = pat.exec(userText)) !== null) {
-      const cleaned = m[0]
-        .replace(/你(?:又|还是)(?:在)?(?:犯|忘|没|不)/g, '')
-        .replace(/我不是(?:说|让|让做|讲)[了过]?/g, '')
-        .replace(/怎么(?:又|还|老)(?:是|在)?/g, '')
-        .replace(/我(?:一直|反复|总是)(?:在|说|强调)?/g, '')
-        .replace(/[，。！？、；：""''（）\s]+/g, '')
-        .trim();
-      if (cleaned.length >= 4 && cleaned.length <= 60) concepts.push(cleaned);
-    }
+  for (const sentence of extractCorrectionMatches(userText)) {
+    const cleaned = cleanCorrectionConcept(sentence);
+    if (cleaned.length >= 4 && cleaned.length <= 60) concepts.push(cleaned);
   }
   return [...new Set(concepts)];
 }
@@ -265,17 +212,6 @@ function clusterConcepts(
   return merged;
 }
 
-function jaccardChinese(a: string, b: string): number {
-  const charsA = new Set(a.split('').filter(c => /[\u4e00-\u9fff]/.test(c)));
-  const charsB = new Set(b.split('').filter(c => /[\u4e00-\u9fff]/.test(c)));
-  if (charsA.size === 0 || charsB.size === 0) return 0;
-  const intersection = new Set([...charsA].filter(x => charsB.has(x)));
-  const union = new Set([...charsA, ...charsB]);
-  return intersection.size / union.size;
-}
-
-// ── Old ──
-
 function extractSignals(sessions: SimpleSession[]): SessionSignals[] {
   return sessions.map(s => {
     const correctionPhrases = extractCorrectionConcepts(s.userText);
@@ -430,7 +366,7 @@ function applySignals(state: ModelState, signals: SessionSignals[], mergedConcep
   }
 
   // Detect falling/stable shifts
-  for (const [key, p] of Object.entries(state.patterns)) {
+  for (const [, p] of Object.entries(state.patterns)) {
     if (p.trend === 'stable' && p.lastSeen < new Date(Date.now() - 7 * 86400000).toISOString().slice(0, 10)) {
       p.trend = 'falling';
     }

package/src/cli/commands/validate.ts

@@ -10,7 +10,6 @@ import * as path from 'path';
 import * as yaml from 'js-yaml';
 import { CheckpointValidator } from '../../core/validators/checkpoint';
 import type { Checkpoint, CheckpointContext } from '../../types/checkpoint';
-import { detectSourceRoots } from '../../utils/detect-source-roots';
 
 export interface ValidateOptions {
   /** 检查点文件路径 */
@@ -96,9 +95,8 @@ export async function validate(options: ValidateOptions): Promise<void> {
 
   if (failed > 0) {
     console.log(chalk.red(`\n🛑 ${failed} 个检查点未通过`));
-    if (options.strict) {
-      process.exit(1);
-    }
+    // 工单 23：门控语义——检查点失败一律非零退出（此前仅 --strict 退出，钩子/CI 形同虚设）
+    process.exit(1);
   } else {
     console.log(chalk.green('\n✅ 所有检查点验证通过'));
   }
@@ -106,6 +104,9 @@ export async function validate(options: ValidateOptions): Promise<void> {
 
 /**
  * 默认检查点列表（供 init 和 upgrade 共用）
+ *
+ * 注：no-console 检查点已移除（工单 23/24）——CLI 产品 src/ 必然有合法 console 输出，
+ * 且旧配置 expected:'' 语义恒错；output_* 族修复为真正执行 config.command 后该检查会恒失败。
  */
 export const DEFAULT_CHECKPOINTS: Checkpoint[] = [
   {
@@ -132,21 +133,6 @@ export const DEFAULT_CHECKPOINTS: Checkpoint[] = [
       },
     ],
   },
-  {
-    id: 'no-console',
-    name: '无 console.log',
-    checks: [
-      {
-        id: 'check-console',
-        type: 'output_not_contains',
-        config: {
-          command: 'grep -r "console.log" src/ || true',
-          expected: '',
-        },
-        message: '源代码中不应包含 console.log',
-      },
-    ],
-  },
 ];
 
 /**
@@ -167,27 +153,7 @@ export async function createExampleCheckpoint(projectPath: string): Promise<void
 
   await fs.mkdir(dir, { recursive: true });
 
-  // 动态替换 no-console 检查点的源目录
-  const roots = detectSourceRoots(projectPath);
-  let checkpoints = DEFAULT_CHECKPOINTS;
-  if (roots.length > 0) {
-    checkpoints = DEFAULT_CHECKPOINTS.map(cp => {
-      if (cp.id === 'no-console') {
-        return {
-          ...cp,
-          checks: cp.checks.map(c => {
-            if (c.id === 'check-console') {
-              return { ...c, config: { command: `grep -r "console.log" ${roots.join(' ')} || true` } };
-            }
-            return c;
-          }),
-        };
-      }
-      return cp;
-    });
-  }
-
-  const content = yaml.dump({ checkpoints }, { indent: 2 });
+  const content = yaml.dump({ checkpoints: DEFAULT_CHECKPOINTS }, { indent: 2 });
   await fs.writeFile(filePath, content, 'utf-8');
 
   console.log(chalk.green(`✅ 已创建示例检查点文件: ${filePath}`));

package/src/cli/session-mining/corrections.ts

@@ -0,0 +1,56 @@
+/**
+ * 纠正信号模式表与概念清洗（工单 19-C）
+ *
+ * 合并 analyze-sessions 与 update-user-model 两份几乎相同的模式表，
+ * 采用较宽窗口（update-user-model 版本）作为唯一实现。
+ */
+
+/** 用户纠正/重复强调语句的匹配模式（两命令合并后的超集，取较宽窗口） */
+export const CORRECTION_PATTERNS: RegExp[] = [
+  /你(?:又|还是)(?:在)?(?:犯|忘|没|不).{0,30}?[了]?/g,
+  /我不是(?:说|让|让做|讲)[了过]?.{0,50}?[了吗?？]?/g,
+  /怎么(?:又|还|老)(?:是|在)?.{0,30}?[了]?/g,
+  /我(?:一直|反复|总是)(?:在|说|强调)?.{0,30}?[了]?/g,
+  /老(?:是|在|犯|忘)(?:了)?.{0,20}?[了]?/g,
+  /这(?:个|种)(?:问题|模式|错误).{0,20}?[.。！]?/g,
+  /(?:第[一二三四五六七八九十\d]+次|反复|重复)(?:说|提醒|强调).{0,20}?/g,
+  /不(?:要|能|想|愿意).{0,20}?老.{0,10}?[了]?/g,
+  /(?:补上|加上|记上|修复).{0,10}?[了吗?？]?/g,
+  /(?:沉淀|监控|日志|记录).{0,5}?[了吗?？]?/g,
+];
+
+/** 纠正语句前缀（提取概念时剥离） */
+const CORRECTION_PREFIX_STRIP: RegExp[] = [
+  /你(?:又|还是)(?:在)?(?:犯|忘|没|不)/g,
+  /我不是(?:说|让|让做|讲)[了过]?/g,
+  /怎么(?:又|还|老)(?:是|在)?/g,
+  /我(?:一直|反复|总是)(?:在|说|强调)?/g,
+  /老(?:是|在|犯|忘)(?:了)?/g,
+];
+
+/**
+ * 从纠正语句中提取概念：剥离纠正前缀与标点
+ */
+export function cleanCorrectionConcept(sentence: string): string {
+  let cleaned = sentence;
+  for (const re of CORRECTION_PREFIX_STRIP) {
+    re.lastIndex = 0;
+    cleaned = cleaned.replace(re, '');
+  }
+  return cleaned.replace(/[，。！？、；：""''（）\s]+/g, '').trim();
+}
+
+/**
+ * 用模式表在文本中抽取纠正语句（去重）
+ */
+export function extractCorrectionMatches(text: string): string[] {
+  const matches: string[] = [];
+  for (const pattern of CORRECTION_PATTERNS) {
+    pattern.lastIndex = 0;
+    let match: RegExpExecArray | null;
+    while ((match = pattern.exec(text)) !== null) {
+      matches.push(match[0].trim());
+    }
+  }
+  return matches;
+}

package/src/cli/session-mining/index.ts

@@ -0,0 +1,13 @@
+/**
+ * 会话挖掘共享层（工单 19-C）
+ *
+ * update-user-model 与 analyze-sessions 两个命令此前各自维护一份
+ * transcript 解析 / 纠正模式 / 分词与相似度实现，收敛到本目录：
+ *   - transcript.ts  Claude Code transcript（.jsonl）读取
+ *   - corrections.ts 纠正信号模式表与概念清洗
+ *   - text.ts        分词 / 停用词 / Jaccard / 文本清洗
+ */
+
+export * from './transcript';
+export * from './corrections';
+export * from './text';

package/src/cli/session-mining/text.ts

@@ -0,0 +1,117 @@
+/**
+ * 分词 / 停用词 / 相似度 / 文本清洗（工单 19-C）
+ *
+ * 合并 analyze-sessions 与 update-user-model 的重复实现。
+ */
+
+/** N-gram 停用词表（中英） */
+export const STOP_WORDS = new Set([
+  '的', '了', '在', '是', '我', '有', '和', '就', '不', '人', '都', '一',
+  '一个', '上', '也', '很', '到', '说', '要', '去', '你', '会', '着',
+  '没有', '看', '好', '自己', '这', '他', '她', '它', '们', '那',
+  'the', 'a', 'an', 'is', 'are', 'was', 'were', 'be', 'been', 'being',
+  'have', 'has', 'had', 'do', 'does', 'did', 'will', 'would', 'could',
+  'should', 'may', 'might', 'can', 'shall', 'to', 'of', 'in', 'for',
+  'on', 'with', 'at', 'by', 'from', 'as', 'into', 'through', 'during',
+  'before', 'after', 'above', 'below', 'between', 'under', 'again',
+  'further', 'then', 'once', 'here', 'there', 'when', 'where', 'why',
+  'how', 'all', 'both', 'each', 'few', 'more', 'most', 'other', 'some',
+  'such', 'only', 'own', 'same', 'so', 'than', 'too', 'very', 'just',
+  '因为', '所以', '但是', '而且', '如果', '虽然', '可以', '这个', '那个',
+  '什么', '怎么', '哪', '吗', '吧', '呢', '啊', '嗯', '哦',
+]);
+
+const CN_CHAR = /[\u4e00-\u9fff\u3400-\u4dbf]/;
+
+/**
+ * 中英文混合分词：中文产出 bigram/trigram（短序列保留整体），
+ * 连续的拉丁/数字/路径字符作为整体 token。
+ */
+export function tokenize(text: string): string[] {
+  const tokens: string[] = [];
+  let current = '';
+
+  for (const char of text) {
+    if (CN_CHAR.test(char)) {
+      current += char;
+    } else if (/[a-zA-Z0-9_\-/.]/.test(char)) {
+      current += char;
+    } else {
+      pushToken(current, tokens);
+      current = '';
+    }
+  }
+  pushToken(current, tokens);
+
+  return tokens;
+}
+
+function pushToken(current: string, tokens: string[]): void {
+  if (current.length < 2) return;
+  if (CN_CHAR.test(current)) {
+    for (let i = 0; i <= current.length - 2; i++) {
+      tokens.push(current.slice(i, i + 2));
+    }
+    for (let i = 0; i <= current.length - 3; i++) {
+      tokens.push(current.slice(i, i + 3));
+    }
+    if (current.length <= 6) tokens.push(current);
+  } else {
+    tokens.push(current);
+  }
+}
+
+/**
+ * 基于 tokenize 的 Jaccard 相似度
+ */
+export function jaccardSimilarity(a: string, b: string): number {
+  const setA = new Set(tokenize(a));
+  const setB = new Set(tokenize(b));
+  const intersection = new Set([...setA].filter(x => setB.has(x)));
+  const union = new Set([...setA, ...setB]);
+  return union.size === 0 ? 0 : intersection.size / union.size;
+}
+
+/**
+ * 仅取汉字字符集的 Jaccard 相似度（概念聚类用）
+ */
+export function jaccardChinese(a: string, b: string): number {
+  const charsA = new Set(a.split('').filter(c => /[\u4e00-\u9fff]/.test(c)));
+  const charsB = new Set(b.split('').filter(c => /[\u4e00-\u9fff]/.test(c)));
+  if (charsA.size === 0 || charsB.size === 0) return 0;
+  const intersection = new Set([...charsA].filter(x => charsB.has(x)));
+  const union = new Set([...charsA, ...charsB]);
+  return intersection.size / union.size;
+}
+
+/**
+ * 剥离代码块 / 行内代码 / JSON 片段，保留自然语言
+ */
+export function stripCodeBlocks(text: string): string {
+  return text
+    .replace(/```[\s\S]*?```/g, '')
+    .replace(/`[^`]+`/g, '')
+    .replace(/\{[^{}]*"[^"]+"\s*:\s*"[^"]*"[^}]*\}/g, '')
+    .replace(/\[[\s\S]*?\]/g, '')
+    .replace(/\\n/g, ' ')
+    .replace(/[{}[\]"':,\\]+/g, ' ');
+}
+
+/** 是否纯标点/空白 */
+export function isPunctuation(str: string): boolean {
+  return /^[，。！？、；：""''（）\[\]【】「」『』《》〈〉\s]+$/.test(str);
+}
+
+/** 是否含语义内容（汉字，或 ≥7 字符的合法英文标识词） */
+export function hasSemanticContent(phrase: string): boolean {
+  if (/[\u4e00-\u9fff]/.test(phrase)) return true;
+  if (phrase.length >= 7 && /^[a-zA-Z][a-zA-Z0-9_]*[a-zA-Z]$/.test(phrase)) return true;
+  return false;
+}
+
+/** 是否代码噪声（短拉丁片段 / 常见 JSON 字段名碎片） */
+export function isCodeNoise(phrase: string): boolean {
+  if (!/[\u4e00-\u9fff]/.test(phrase) && phrase.length <= 6) return true;
+  if (/^(?:clla|cood|odde|oppe|laaw|peen|enne|ennc|ddec|deco|ecod|code|type|role|name|text|file|path|tool|user|session)$/i.test(phrase)) return true;
+  return false;
+}

package/src/cli/session-mining/transcript.ts

@@ -0,0 +1,107 @@
+/**
+ * Claude Code transcript（.jsonl）读取（工单 19-C）
+ *
+ * 原 analyze-sessions.findSessions 与 update-user-model.findNewSessions
+ * 各持一份解析逻辑，收敛于此。
+ */
+
+import * as fs from 'fs';
+import * as path from 'path';
+
+export interface MinedTurn {
+  role: string;
+  content: string;
+}
+
+export interface MinedSession {
+  /** 会话 ID（文件名去掉 .jsonl，截断至 40 字符） */
+  id: string;
+  /** 文件修改日期（YYYY-MM-DD） */
+  date: string;
+  /** 修改时间戳（毫秒） */
+  mtimeMs: number;
+  turns: MinedTurn[];
+  /** 出现过的工具调用名（去重，保持出现顺序） */
+  toolCalls: string[];
+}
+
+function extractText(content: unknown): string {
+  if (typeof content === 'string') return content;
+  if (Array.isArray(content)) {
+    return content
+      .filter((b: any) => b.type === 'text')
+      .map((b: any) => b.text)
+      .join(' ');
+  }
+  return '';
+}
+
+/**
+ * 扫描目录下的 .jsonl transcript 文件
+ *
+ * 不可读/损坏的行静默跳过（与原实现一致）。
+ */
+export function readTranscriptSessions(dir: string): MinedSession[] {
+  const sessions: MinedSession[] = [];
+
+  let files: string[];
+  try {
+    files = fs.readdirSync(dir);
+  } catch {
+    return sessions;
+  }
+
+  for (const file of files) {
+    if (!file.endsWith('.jsonl')) continue;
+    const filePath = path.join(dir, file);
+
+    let stat: fs.Stats;
+    try {
+      stat = fs.statSync(filePath);
+    } catch {
+      continue;
+    }
+
+    const turns: MinedTurn[] = [];
+    const toolCalls: string[] = [];
+
+    let content: string;
+    try {
+      content = fs.readFileSync(filePath, 'utf-8');
+    } catch {
+      continue;
+    }
+
+    for (const line of content.split('\n')) {
+      if (!line.trim()) continue;
+      try {
+        const entry = JSON.parse(line);
+        const msg = entry.message;
+        if (msg?.role && (msg.content || entry.type === 'user')) {
+          turns.push({ role: msg.role, content: extractText(msg.content) });
+        }
+        if (msg?.role === 'assistant' && Array.isArray(msg.content)) {
+          for (const block of msg.content) {
+            if (block.type === 'tool_use' && block.name) {
+              toolCalls.push(block.name);
+            }
+          }
+        }
+      } catch {
+        // 损坏行跳过
+      }
+    }
+
+    if (turns.length > 0) {
+      sessions.push({
+        id: file.replace('.jsonl', '').slice(0, 40),
+        date: stat.mtime.toISOString().slice(0, 10),
+        mtimeMs: stat.mtimeMs,
+        turns,
+        toolCalls: [...new Set(toolCalls)],
+      });
+    }
+  }
+
+  return sessions;
+}

package/src/constraints/__tests__/lifecycle-runner.test.ts

@@ -5,7 +5,7 @@
 import { describe, it, expect, beforeEach } from '@jest/globals';
 import { ConstraintLifecycleRunner, type ExecutionResult } from '../lifecycle-runner';
 import { ConstraintRegistry } from '../registry';
-import type { ConstraintProposal } from '../../monitoring/constraint-evolver';
+import type { ConstraintProposal } from '../../types/monitoring-types';
 
 function makeProposal(overrides: Partial<ConstraintProposal> = {}): ConstraintProposal {
   return {

package/src/constraints/lifecycle-runner.ts

@@ -6,7 +6,7 @@
  */
 
 import { ConstraintRegistry } from './registry';
-import type { ConstraintProposal } from '../monitoring/constraint-evolver';
+import type { ConstraintProposal } from '../types/monitoring-types';
 import type { ConstraintLevel } from '../types/constraint';
 import type { DeprecationSchedule } from './types';
 

package/src/context/compaction.ts

@@ -81,7 +81,7 @@ export class SessionCompaction {
   /**
    * 中等压缩：结构化总结 + 保留最近 N 轮
    */
-  private summaryCompact(messages: SessionMessage[], budget: number, originalTokens: number): CompactionResult {
+  private summaryCompact(messages: SessionMessage[], _budget: number, originalTokens: number): CompactionResult {
     const recentCount = Math.min(6, messages.length); // 保留最近 6 条
     const recent = messages.slice(-recentCount);
     const older = messages.slice(0, -recentCount);
@@ -111,7 +111,7 @@ export class SessionCompaction {
   /**
    * 重型压缩：checkpoint + 清空历史
    */
-  private checkpointCompact(messages: SessionMessage[], budget: number, originalTokens: number): CompactionResult {
+  private checkpointCompact(messages: SessionMessage[], _budget: number, originalTokens: number): CompactionResult {
     const summary = this.generateStructuredSummary(messages);
 
     const checkpointSessionMessage: SessionMessage = {
@@ -169,7 +169,6 @@ export class SessionCompaction {
    * 如果 tool 结果被保留，对应的 tool 调用也要保留
    */
   private preserveToolCallPairs(all: SessionMessage[], compacted: SessionMessage[]): SessionMessage[] {
-    const compactedIds = new Set(compacted.map(m => m.toolCallId).filter(Boolean));
     const result: SessionMessage[] = [];
 
     for (const msg of compacted) {

package/src/context/knowledge-injector.ts

@@ -7,7 +7,7 @@
 
 import { KnowledgeQuery } from '../knowledge/query';
 import { TokenEstimator } from './token-budget';
-import type { KnowledgeEntry, QueryResult, KnowledgeType } from '../knowledge/types';
+import type { KnowledgeEntry, KnowledgeType } from '../knowledge/types';
 import type { ContextSource } from './types';
 
 export interface InjectionConfig {

package/src/context/tool-output-budget.ts

@@ -77,7 +77,7 @@ export class ToolOutputBudget {
   /**
    * 应用预算策略
    */
-  applyBudget(output: string, strategy: OutputStrategy, totalBudget?: number): OutputApplyResult {
+  applyBudget(output: string, strategy: OutputStrategy, _totalBudget?: number): OutputApplyResult {
     const originalTokens = TokenEstimator.estimateText(output);
 
     switch (strategy) {

package/src/core/constraints/capabilities-parser.ts

@@ -0,0 +1,70 @@
+/**
+ * CAPABILITIES.md 表格解析（工单 19-B）
+ *
+ * 此前 checker.parseCapabilitiesFiles 与 sync-docs.parseCapabilitiesFiles
+ * 各持一份正则，收敛到此处。
+ */
+
+import * as fs from 'fs';
+
+/** 匹配表格单元格中的源码文件条目（如 `src/foo.ts`、`foo.tsx`） */
+const FILE_CELL_REGEX = /\|\s*([^|]+?\.(?:ts|tsx|js|jsx))\s*\|/g;
+
+/** 仅匹配表格第二列的文件条目（checker 的历史语义：第一列为模块名，第二列为文件） */
+const FILE_SECOND_COLUMN_REGEX = /^\|[^|]+\|\s*([^|]+?\.(?:ts|tsx|js|jsx))\s*\|/gm;
+
+/**
+ * 匹配表格单元格中的目录条目（如 `agents/`）。
+ *
+ * 仅匹配纯路径样式的单元格（字母/数字/._/-/@ 与 / 组成、以 / 结尾），
+ * 否则以 / 结尾的散文描述（如 JSDoc 首行在 "gitRepo /" 处换行）会被误判为
+ * 目录条目，导致 --check 永远报「包含已删除的模块」不收敛（2026-08-04 studio PR #44）
+ */
+const DIR_CELL_REGEX = /\|\s*([\w][\w@./-]*\/)\s*\|/g;
+
+export interface CapabilitiesParseOptions {
+  /** 是否同时收集目录条目（sync-docs 需要；checker 只做文件覆盖比对，不需要） */
+  includeDirs?: boolean;
+  /** 只取表格第二列的文件条目（checker 的历史严格语义），默认 false */
+  secondColumnOnly?: boolean;
+}
+
+/**
+ * 从 CAPABILITIES.md 内容中提取表格列出的条目
+ *
+ * @param content markdown 内容
+ * @returns 文件路径原样 + （可选）目录条目，按出现顺序去重
+ */
+export function parseCapabilitiesEntries(content: string, options: CapabilitiesParseOptions = {}): string[] {
+  const { includeDirs = false, secondColumnOnly = false } = options;
+  const entries: string[] = [];
+
+  const fileRegex = secondColumnOnly ? FILE_SECOND_COLUMN_REGEX : FILE_CELL_REGEX;
+  let match: RegExpExecArray | null;
+  fileRegex.lastIndex = 0;
+  while ((match = fileRegex.exec(content)) !== null) {
+    const raw = match[1].trim();
+    if (!entries.includes(raw)) entries.push(raw);
+  }
+
+  if (includeDirs) {
+    DIR_CELL_REGEX.lastIndex = 0;
+    while ((match = DIR_CELL_REGEX.exec(content)) !== null) {
+      const dir = match[1].trim();
+      if (!entries.includes(dir)) entries.push(dir);
+    }
+  }
+
+  return entries;
+}
+
+/**
+ * 读取并解析 CAPABILITIES.md；文件缺失/不可读时返回 []
+ */
+export function readCapabilitiesEntries(capabilitiesPath: string, options: CapabilitiesParseOptions = {}): string[] {
+  try {
+    return parseCapabilitiesEntries(fs.readFileSync(capabilitiesPath, 'utf-8'), options);
+  } catch {
+    return [];
+  }
+}