@miphamai/cli 0.27.0 → 0.28.0

package/src/core/auto-memory.ts

@@ -0,0 +1,613 @@
+/**
+ * CRSI Phase 2: Auto-Reflection Engine
+ *
+ * Structured API for post-turn conversation reflection, CRSI pipeline
+ * integration, and persistent memory management. Builds on top of the
+ * Phase 1 rule-level self-improvement (PatternAnalyzer, RuleEngine,
+ * EffectivenessTracker) to enable code-level and session-level learning.
+ *
+ * Architecture:
+ *   Turn completes → analyzeTurn() → persist() → CRSI pipeline feed
+ *   AI does the thinking; this module provides storage + integration.
+ */
+
+import { MemoryManager, type MemoryEntry } from './memory/memory-manager.js'
+import type { PatternAnalyzer, Pattern } from '../agent/pattern-analyzer.js'
+import type { ExperienceRuleEngine, ToolRule } from './rule-engine.js'
+import type { EffectivenessTracker } from '../agent/effectiveness-tracker.js'
+import { join } from 'node:path'
+import { homedir } from 'node:os'
+
+// ── Types ──
+
+export interface ToolCallRecord {
+  /** Tool name (Bash, Read, Write, Edit, Agent, etc.) */
+  name: string
+  /** Tool input parameters */
+  input: Record<string, unknown>
+  /** Whether the tool executed successfully */
+  success: boolean
+  /** Error message if the tool failed */
+  error?: string
+  /** Duration in ms */
+  durationMs?: number
+}
+
+export interface CrsiInsight {
+  /** Category: timeout | tool-params | search | import | semantic */
+  category: string
+  /** Human-readable description of the insight */
+  description: string
+  /** Severity: 'info' | 'warning' | 'critical' */
+  severity: 'info' | 'warning' | 'critical'
+  /** Suggested action (e.g. "增加 Bash timeout 到 300000ms") */
+  suggestion: string
+  /** Evidence: concrete examples supporting this insight */
+  evidence: string[]
+  /** Whether this insight can be auto-applied as a CRSI rule */
+  autoApplicable: boolean
+}
+
+export interface TurnReflection {
+  /** Unique ID for this reflection */
+  id: string
+  /** Session ID */
+  sessionId: string
+  /** Timestamp */
+  timestamp: string
+  /** Summary of the turn (1-3 sentences) */
+  summary: string
+  /** What went well */
+  successes: string[]
+  /** What failed or could be improved */
+  failures: string[]
+  /** CRSI-specific insights for rule generation */
+  crsiInsights: CrsiInsight[]
+  /** Key decisions made during this turn */
+  decisions: string[]
+  /** Action items for future sessions */
+  actionItems: string[]
+}
+
+// ── Constants ──
+
+const DEFAULT_MEMORY_DIR = join(homedir(), '.mipham', 'memory')
+
+/** Minimum number of similar failures before a CRSI rule is generated. */
+const CRSI_RULE_THRESHOLD = 2
+
+/** How many past reflections to recall for context. */
+const MAX_RECALL = 5
+
+// ── Engine ──
+
+export class AutoMemoryEngine {
+  private memoryManager: MemoryManager
+  private patternAnalyzer?: PatternAnalyzer
+  private ruleEngine?: ExperienceRuleEngine
+  private effectivenessTracker?: EffectivenessTracker
+
+  /** Accumulated reflections for the current session (not yet persisted to disk). */
+  private sessionReflections: TurnReflection[] = []
+
+  constructor(memoryDir: string = DEFAULT_MEMORY_DIR) {
+    this.memoryManager = new MemoryManager(memoryDir)
+    this.memoryManager.loadAll()
+  }
+
+  // ── CRSI Pipeline Wiring ──
+
+  /**
+   * Wire the CRSI Phase 1 pipeline into the auto-memory engine.
+   * Enables insights → rule generation and effectiveness tracking.
+   */
+  setCrsiPipeline(
+    patternAnalyzer: PatternAnalyzer,
+    ruleEngine: ExperienceRuleEngine,
+    effectivenessTracker: EffectivenessTracker,
+  ): void {
+    this.patternAnalyzer = patternAnalyzer
+    this.ruleEngine = ruleEngine
+    this.effectivenessTracker = effectivenessTracker
+  }
+
+  // ── Core API ──
+
+  /**
+   * Analyze a completed conversation turn and produce a structured reflection.
+   *
+   * The AI calls this after each response completes. The engine:
+   * 1. Analyzes tool call patterns (successes, failures, timeouts)
+   * 2. Cross-references with CRSI pipeline for rule suggestions
+   * 3. Produces a TurnReflection ready for persistence
+   */
+  analyzeTurn(params: {
+    sessionId: string
+    userMessage: string
+    assistantContent: string
+    toolCalls: ToolCallRecord[]
+    modelProvider: string
+    modelId: string
+    turnDurationMs: number
+  }): TurnReflection {
+    const { sessionId, userMessage, assistantContent, toolCalls, modelProvider, modelId } = params
+
+    // 1. Categorize tool calls
+    const successes: string[] = []
+    const failures: string[] = []
+    for (const tc of toolCalls) {
+      if (tc.success) {
+        successes.push(`${tc.name}: ${this.summarizeToolInput(tc)}`)
+      } else {
+        failures.push(`${tc.name}: ${tc.error || 'unknown error'} — ${this.summarizeToolInput(tc)}`)
+      }
+    }
+
+    // 2. Extract CRSI insights from tool failures
+    const crsiInsights = this.extractCrsiInsights(toolCalls)
+
+    // 3. Cross-reference with existing CRSI rules
+    this.crossReferenceWithRules(crsiInsights)
+
+    // 4. Analyze for decisions and action items
+    const decisions = this.extractDecisions(userMessage, assistantContent)
+    const actionItems = this.extractActionItems(assistantContent, toolCalls)
+
+    // 5. Build summary
+    const summary = this.buildSummary(successes, failures, crsiInsights, decisions)
+
+    const reflection: TurnReflection = {
+      id: `reflection-${sessionId}-${Date.now().toString(36)}`,
+      sessionId,
+      timestamp: new Date().toISOString(),
+      summary,
+      successes,
+      failures,
+      crsiInsights,
+      decisions,
+      actionItems,
+    }
+
+    this.sessionReflections.push(reflection)
+    return reflection
+  }
+
+  /**
+   * Persist a turn reflection to the memory store.
+   *
+   * Writes a structured .md file with frontmatter to ~/.mipham/memory/
+   * and updates the MEMORY.md index. CRSI insights that meet the threshold
+   * are fed into the rule engine for auto-rule generation.
+   */
+  persist(reflection: TurnReflection): void {
+    const content = this.formatReflectionContent(reflection)
+    const relevance = this.extractRelevance(reflection)
+
+    this.memoryManager.write(reflection.id, content, {
+      type: 'feedback',
+      relevance,
+      why: `CRSI Phase 2 自动复盘 — 会话 ${reflection.sessionId}`,
+      howToApply: '下次会话开始时回顾此文件，检查是否有可操作的改进项',
+    })
+
+    // Feed CRSI insights into the rule engine
+    this.feedCrsiPipeline(reflection)
+  }
+
+  /**
+   * Persist ALL accumulated session reflections at once (e.g. on session end).
+   */
+  persistAll(): void {
+    for (const reflection of this.sessionReflections) {
+      this.persist(reflection)
+    }
+
+    // Also write a session-level summary
+    if (this.sessionReflections.length > 0) {
+      this.writeSessionSummary()
+    }
+
+    // Flush CRSI state
+    if (this.effectivenessTracker) {
+      this.effectivenessTracker.persist()
+    }
+  }
+
+  /**
+   * Recall relevant past reflections for a given context.
+   */
+  recall(context: string, limit: number = MAX_RECALL): MemoryEntry[] {
+    return this.memoryManager.recall(context, limit)
+  }
+
+  /**
+   * Build a system reminder string from past reflections for injection
+   * into the next session's system prompt.
+   */
+  buildReminder(context: string, maxTokens?: number): string {
+    return this.memoryManager.buildSystemReminder(context, maxTokens)
+  }
+
+  /**
+   * Get the session reflection count (useful for stats).
+   */
+  get sessionReflectionCount(): number {
+    return this.sessionReflections.length
+  }
+
+  /**
+   * Get accumulated CRSI insights across all session reflections.
+   */
+  get accumulatedInsights(): CrsiInsight[] {
+    return this.sessionReflections.flatMap((r) => r.crsiInsights)
+  }
+
+  // ── Private: CRSI Integration ──
+
+  /**
+   * Extract CRSI insights from tool call records.
+   *
+   * Identifies patterns like:
+   * - timeout: heavy commands without explicit timeout
+   * - tool-params: git --force without sandbox disable
+   * - import: ESM imports missing .js extension (from Edit/Write errors)
+   * - search: full-repo Grep without directory scoping
+   */
+  private extractCrsiInsights(toolCalls: ToolCallRecord[]): CrsiInsight[] {
+    const insights: CrsiInsight[] = []
+    const seen = new Set<string>()
+
+    for (const tc of toolCalls) {
+      if (tc.success) continue
+
+      const category = this.categorizeFailure(tc)
+      const key = `${category}:${tc.name}`
+      if (seen.has(key)) continue
+      seen.add(key)
+
+      const evidence = this.gatherEvidence(tc, toolCalls)
+      const autoApplicable = ['timeout', 'tool-params'].includes(category)
+
+      const insight: CrsiInsight = {
+        category,
+        description: `${tc.name} 工具调用失败: ${tc.error || 'unknown'}`,
+        severity: tc.error?.includes('timeout') ? 'critical' : 'warning',
+        suggestion: this.suggestFix(category, tc),
+        evidence,
+        autoApplicable,
+      }
+
+      insights.push(insight)
+    }
+
+    return insights
+  }
+
+  /**
+   * Cross-reference new insights with existing CRSI rules.
+   * If a similar rule already exists, mark the insight as already-covered.
+   */
+  private crossReferenceWithRules(insights: CrsiInsight[]): void {
+    if (!this.ruleEngine) return
+
+    const activeRules = this.ruleEngine.getActiveRules()
+    for (const insight of insights) {
+      const covered = activeRules.some((r) => r.category === insight.category)
+      if (covered) {
+        insight.description += ' (已有对应 CRSI 规则覆盖)'
+        insight.autoApplicable = false // Don't create duplicate rules
+      }
+    }
+  }
+
+  /**
+   * Feed CRSI insights into the rule engine for automatic rule generation.
+   * Only insights that are auto-applicable and have sufficient evidence
+   * (≥ CRSI_RULE_THRESHOLD similar failures) are converted to rules.
+   */
+  private feedCrsiPipeline(reflection: TurnReflection): void {
+    if (!this.ruleEngine || !this.patternAnalyzer) return
+
+    // Count failures by category across ALL reflections (not just this one)
+    const categoryCounts = new Map<string, number>()
+    for (const r of this.sessionReflections) {
+      for (const insight of r.crsiInsights) {
+        if (!insight.autoApplicable) continue
+        const count = categoryCounts.get(insight.category) || 0
+        categoryCounts.set(insight.category, count + 1)
+      }
+    }
+
+    // Generate rules for categories that meet the threshold
+    for (const [category, count] of categoryCounts) {
+      if (count < CRSI_RULE_THRESHOLD) continue
+
+      // Use PatternAnalyzer to generate a proper ToolRule
+      const pattern: Pattern = {
+        id: `auto-${category}-${Date.now().toString(36)}`,
+        category: category as Pattern['category'],
+        agentName: 'auto-memory',
+        frequency: count,
+        confidence: count >= 5 ? 'high' : 'medium',
+        examples: this.sessionReflections
+          .flatMap((r) => r.crsiInsights)
+          .filter((i) => i.category === category)
+          .flatMap((i) => i.evidence)
+          .slice(0, 5),
+        firstSeen: this.sessionReflections[0]?.timestamp || '',
+        lastSeen: this.sessionReflections[this.sessionReflections.length - 1]?.timestamp || '',
+      }
+
+      const toolRule = this.patternAnalyzer.toToolRule(pattern)
+      this.ruleEngine.register(toolRule)
+
+      // Track the new rule's effectiveness
+      if (this.effectivenessTracker) {
+        this.effectivenessTracker.recordApplication(toolRule.id, true)
+      }
+    }
+  }
+
+  // ── Private: Analysis Helpers ──
+
+  private categorizeFailure(tc: ToolCallRecord): string {
+    const err = (tc.error || '').toLowerCase()
+    const cmd = String(tc.input?.command || tc.input?.description || '').toLowerCase()
+
+    if (err.includes('timeout') || err.includes('timed out')) return 'timeout'
+    if (cmd.includes('--force') || cmd.includes('rm -rf')) return 'tool-params'
+    if (err.includes('import') || err.includes('module') || err.includes('.js')) return 'import'
+    if (tc.name === 'Grep' && err.includes('no matches')) return 'search'
+    if (err.includes('permission') || err.includes('denied')) return 'tool-params'
+
+    return 'semantic'
+  }
+
+  private gatherEvidence(tc: ToolCallRecord, allCalls: ToolCallRecord[]): string[] {
+    const evidence: string[] = []
+    evidence.push(`${tc.name}: ${this.summarizeToolInput(tc)}`)
+
+    // Find similar failures in the same batch
+    const similar = allCalls.filter(
+      (c) => c !== tc && !c.success && this.categorizeFailure(c) === this.categorizeFailure(tc),
+    )
+    if (similar.length > 0) {
+      evidence.push(`同一轮中还有 ${similar.length} 个同类失败`)
+    }
+
+    return evidence
+  }
+
+  private suggestFix(category: string, tc: ToolCallRecord): string {
+    switch (category) {
+      case 'timeout':
+        return `为 ${tc.name} 增加 timeout 参数至 300000ms（5分钟）`
+      case 'tool-params':
+        return `检查 ${tc.name} 的参数安全性，考虑添加 dangerouslyDisableSandbox`
+      case 'import':
+        return '检查 ESM 模块导入是否缺少 .js 扩展名'
+      case 'search':
+        return '建议先用 Glob 缩小搜索范围，再使用 Grep'
+      default:
+        return '人工审查此工具调用的参数和上下文'
+    }
+  }
+
+  private extractDecisions(userMessage: string, assistantContent: string): string[] {
+    const decisions: string[] = []
+    const decisionPatterns = [
+      /决定使用\s*(.+)/g,
+      /选择\s*(.+?)(?:作为|方案)/g,
+      /采用\s*(.+?)(?:方案|架构|设计)/g,
+      /decided?\s*(?:to|on)\s*(.+)/gi,
+      /opted\s*(?:for|to)\s*(.+)/gi,
+    ]
+
+    for (const pattern of decisionPatterns) {
+      let match: RegExpExecArray | null
+      while ((match = pattern.exec(assistantContent)) !== null) {
+        decisions.push(match[1]!.trim())
+      }
+    }
+
+    // Also check for explicit decisions in user message
+    const userDecisionMatch = userMessage.match(/用\s*(.+?)(?:吧|方案|方法)/)
+    if (userDecisionMatch?.[1]) {
+      decisions.push(`用户决定: ${userDecisionMatch[1].trim()}`)
+    }
+
+    return decisions.slice(0, 5)
+  }
+
+  private extractActionItems(
+    assistantContent: string,
+    toolCalls: ToolCallRecord[],
+  ): string[] {
+    const items: string[] = []
+
+    // Check for TODO markers in assistant content
+    const todoMatches = assistantContent.matchAll(/[-*]\s*\[ \]\s*(.+)/g)
+    for (const match of todoMatches) {
+      items.push(match[1]!.trim())
+    }
+
+    // Failed tool calls become action items
+    for (const tc of toolCalls) {
+      if (!tc.success) {
+        items.push(`修复 ${tc.name} 调用失败: ${tc.error || 'unknown'}`)
+      }
+    }
+
+    return items.slice(0, 8)
+  }
+
+  private buildSummary(
+    successes: string[],
+    failures: string[],
+    insights: CrsiInsight[],
+    decisions: string[],
+  ): string {
+    const parts: string[] = []
+
+    if (successes.length > 0) {
+      parts.push(`${successes.length} 个工具调用成功`)
+    }
+    if (failures.length > 0) {
+      parts.push(`${failures.length} 个失败`)
+    }
+    if (insights.length > 0) {
+      const criticalCount = insights.filter((i) => i.severity === 'critical').length
+      if (criticalCount > 0) {
+        parts.push(`${criticalCount} 个关键 CRSI 洞察`)
+      } else {
+        parts.push(`${insights.length} 个 CRSI 洞察`)
+      }
+    }
+    if (decisions.length > 0) {
+      parts.push(`${decisions.length} 个关键决策`)
+    }
+
+    return parts.length > 0 ? parts.join('，') : '本轮无特殊事件'
+  }
+
+  private summarizeToolInput(tc: ToolCallRecord): string {
+    switch (tc.name) {
+      case 'Bash':
+        return String(tc.input?.command || '').slice(0, 80)
+      case 'Read':
+        return String(tc.input?.file_path || '').slice(0, 80)
+      case 'Write':
+        return String(tc.input?.file_path || '').slice(0, 80)
+      case 'Edit':
+        return `${String(tc.input?.file_path || '').slice(0, 60)}: ${String(tc.input?.old_string || '').slice(0, 20)}...`
+      case 'Grep':
+        return String(tc.input?.pattern || '').slice(0, 80)
+      case 'Agent':
+        return String(tc.input?.description || tc.input?.prompt || '').slice(0, 80)
+      default:
+        return JSON.stringify(tc.input).slice(0, 80)
+    }
+  }
+
+  // ── Private: Persistence ──
+
+  private formatReflectionContent(reflection: TurnReflection): string {
+    const lines: string[] = [
+      `# 会话复盘: ${reflection.id}`,
+      '',
+      `**会话**: ${reflection.sessionId}`,
+      `**时间**: ${reflection.timestamp}`,
+      `**摘要**: ${reflection.summary}`,
+      '',
+      '---',
+      '',
+      '## 成功项',
+      '',
+      ...(reflection.successes.length > 0
+        ? reflection.successes.map((s) => `- ✅ ${s}`)
+        : ['- _(本轮无记录的成功项)_']),
+      '',
+      '## 失败项',
+      '',
+      ...(reflection.failures.length > 0
+        ? reflection.failures.map((f) => `- ❌ ${f}`)
+        : ['- _(本轮无失败)_']),
+      '',
+      '## CRSI 洞察',
+      '',
+      ...(reflection.crsiInsights.length > 0
+        ? reflection.crsiInsights.map(
+            (i) =>
+              `- 🔧 [${i.severity === 'critical' ? '⚠️' : '📝'} ${i.category}] ${i.description}\n  → 建议: ${i.suggestion}${i.autoApplicable ? ' _(可自动应用)_' : ''}`,
+          )
+        : ['- _(本轮无 CRSI 洞察)_']),
+      '',
+      '## 关键决策',
+      '',
+      ...(reflection.decisions.length > 0
+        ? reflection.decisions.map((d) => `- 🎯 ${d}`)
+        : ['- _(本轮无关键决策)_']),
+      '',
+      '## 待办项',
+      '',
+      ...(reflection.actionItems.length > 0
+        ? reflection.actionItems.map((a) => `- [ ] ${a}`)
+        : ['- _(本轮无待办项)_']),
+      '',
+      '---',
+      '',
+      `_由 CRSI Phase 2 AutoMemoryEngine 自动生成_`,
+    ]
+
+    return lines.join('\n')
+  }
+
+  private extractRelevance(reflection: TurnReflection): string[] {
+    const keywords = new Set<string>()
+
+    // Extract from CRSI insights
+    for (const insight of reflection.crsiInsights) {
+      keywords.add(insight.category)
+    }
+
+    // Extract from decisions
+    for (const d of reflection.decisions) {
+      const words = d.split(/\s+/).filter((w) => w.length > 2)
+      for (const w of words.slice(0, 3)) {
+        keywords.add(w.toLowerCase())
+      }
+    }
+
+    // Always include core tags
+    keywords.add('crsi-phase-2')
+    keywords.add('turn-reflection')
+    keywords.add(reflection.sessionId)
+
+    return Array.from(keywords).slice(0, 10)
+  }
+
+  private writeSessionSummary(): void {
+    const totalSuccesses = this.sessionReflections.reduce((s, r) => s + r.successes.length, 0)
+    const totalFailures = this.sessionReflections.reduce((s, r) => s + r.failures.length, 0)
+    const totalInsights = this.sessionReflections.reduce((s, r) => s + r.crsiInsights.length, 0)
+    const criticalInsights = this.sessionReflections.flatMap((r) =>
+      r.crsiInsights.filter((i) => i.severity === 'critical'),
+    )
+
+    const content = [
+      `# 会话总结`,
+      '',
+      `**会话 ID**: ${this.sessionReflections[0]?.sessionId || 'unknown'}`,
+      `**复盘数**: ${this.sessionReflections.length}`,
+      `**时间**: ${new Date().toISOString()}`,
+      '',
+      '---',
+      '',
+      '## 统计',
+      '',
+      `| 指标 | 数值 |`,
+      `|------|------|`,
+      `| 成功工具调用 | ${totalSuccesses} |`,
+      `| 失败工具调用 | ${totalFailures} |`,
+      `| CRSI 洞察 | ${totalInsights} |`,
+      `| 关键洞察 | ${criticalInsights.length} |`,
+      `| 成功率 | ${totalSuccesses + totalFailures > 0 ? Math.round((totalSuccesses / (totalSuccesses + totalFailures)) * 100) : 100}% |`,
+      '',
+      ...(criticalInsights.length > 0
+        ? ['## ⚠️ 关键洞察', '', ...criticalInsights.map((i) => `- **${i.category}**: ${i.description}\n  → ${i.suggestion}`)]
+        : []),
+      '',
+      '---',
+      '',
+      `_由 CRSI Phase 2 AutoMemoryEngine 在会话结束时自动生成_`,
+    ].join('\n')
+
+    const name = `session-summary-${this.sessionReflections[0]?.sessionId || Date.now().toString(36)}`
+    this.memoryManager.write(name, content, {
+      type: 'feedback',
+      relevance: ['session-summary', 'crsi-phase-2', ...criticalInsights.map((i) => i.category)],
+      why: 'CRSI Phase 2 会话级自动总结',
+      howToApply: '在下次会话开始时回顾，重点关注关键洞察和失败模式',
+    })
+  }
+}