opc-agent 4.1.0 → 4.1.1

package/src/memory/context-compressor.ts

@@ -1,189 +1,189 @@
-import type { Message } from '../core/types';
-
-export interface CompressorConfig {
-  maxTokens: number;
-  compressThreshold: number;
-  preserveRecent: number;
-  brain?: any;
-}
-
-export interface CompressResult {
-  messages: Message[];
-  learnedCount: number;
-  savedTokens: number;
-  summary: string;
-}
-
-const DEFAULT_CONFIG: CompressorConfig = {
-  maxTokens: 8000,
-  compressThreshold: 0.8,
-  preserveRecent: 10,
-};
-
-/**
- * Context compression with optional DeepBrain memory offloading.
- */
-export class ContextCompressor {
-  private config: CompressorConfig;
-
-  constructor(config: Partial<CompressorConfig> = {}) {
-    this.config = { ...DEFAULT_CONFIG, ...config };
-  }
-
-  /**
-   * Estimate token count using language-aware heuristic.
-   * English: ~1 token per 4 chars. Chinese: ~1 token per 2 chars.
-   */
-  estimateTokens(text: string): number {
-    let tokens = 0;
-    for (const char of text) {
-      // CJK Unicode range detection
-      const code = char.codePointAt(0) ?? 0;
-      if (
-        (code >= 0x4e00 && code <= 0x9fff) ||  // CJK Unified
-        (code >= 0x3400 && code <= 0x4dbf) ||  // CJK Extension A
-        (code >= 0x3000 && code <= 0x303f)     // CJK Punctuation
-      ) {
-        tokens += 0.5; // 1 token per 2 chars
-      } else {
-        tokens += 0.25; // 1 token per 4 chars
-      }
-    }
-    return Math.ceil(tokens);
-  }
-
-  private estimateMessagesTokens(messages: Message[]): number {
-    return messages.reduce((sum, m) => sum + this.estimateTokens(m.content), 0);
-  }
-
-  /**
-   * Extract key insights from messages for brain storage.
-   */
-  private extractInsights(messages: Message[]): Array<{ content: string; type: string }> {
-    const insights: Array<{ content: string; type: string }> = [];
-    for (const msg of messages) {
-      const c = msg.content;
-      // Decisions
-      if (/\b(decided|decision|choose|chose|will use|going with|let's go|确定|决定)\b/i.test(c)) {
-        insights.push({ content: c.slice(0, 500), type: 'decision' });
-      }
-      // Facts / definitions
-      else if (/\b(is defined as|means|equals|refers to|是指|定义)\b/i.test(c)) {
-        insights.push({ content: c.slice(0, 500), type: 'fact' });
-      }
-      // Preferences
-      else if (/\b(prefer|like|want|don't like|不喜欢|喜欢|偏好)\b/i.test(c)) {
-        insights.push({ content: c.slice(0, 500), type: 'preference' });
-      }
-      // Code snippets
-      else if (/```[\s\S]{20,}```/.test(c)) {
-        insights.push({ content: c.slice(0, 800), type: 'code' });
-      }
-      // Long assistant messages likely contain useful info
-      else if (msg.role === 'assistant' && c.length > 200) {
-        insights.push({ content: c.slice(0, 500), type: 'knowledge' });
-      }
-    }
-    return insights;
-  }
-
-  /**
-   * Generate a simple summary from messages (no-brain fallback).
-   */
-  private summarize(messages: Message[]): string {
-    const topics = new Set<string>();
-    const keyLines: string[] = [];
-
-    for (const msg of messages) {
-      // Extract first meaningful sentence
-      const firstLine = msg.content.split(/[.\n!?。！？]/)[0]?.trim();
-      if (firstLine && firstLine.length > 10 && firstLine.length < 200) {
-        if (keyLines.length < 5) keyLines.push(`[${msg.role}] ${firstLine}`);
-      }
-      // Extract topic words (capitalized words, Chinese phrases)
-      const words = msg.content.match(/[A-Z][a-z]{2,}/g) ?? [];
-      words.forEach(w => topics.add(w));
-    }
-
-    const topicStr = [...topics].slice(0, 10).join(', ');
-    const linesStr = keyLines.join('; ');
-    return `Topics: ${topicStr || 'general discussion'}. Key points: ${linesStr || 'varied conversation'}`;
-  }
-
-  /**
-   * Compress messages when token count exceeds threshold.
-   */
-  async compress(messages: Message[], config?: Partial<CompressorConfig>): Promise<CompressResult> {
-    const cfg = { ...this.config, ...config };
-    const totalTokens = this.estimateMessagesTokens(messages);
-    const threshold = cfg.maxTokens * cfg.compressThreshold;
-
-    // Under threshold — return as-is
-    if (totalTokens <= threshold) {
-      return {
-        messages: [...messages],
-        learnedCount: 0,
-        savedTokens: 0,
-        summary: '',
-      };
-    }
-
-    const recentCount = Math.min(cfg.preserveRecent, messages.length);
-    const splitIdx = messages.length - recentCount;
-    const oldMessages = messages.slice(0, splitIdx);
-    const recentMessages = messages.slice(splitIdx);
-
-    if (oldMessages.length === 0) {
-      return { messages: [...messages], learnedCount: 0, savedTokens: 0, summary: '' };
-    }
-
-    const oldTokens = this.estimateMessagesTokens(oldMessages);
-    let learnedCount = 0;
-    let summary: string;
-
-    if (cfg.brain) {
-      // Extract and learn insights
-      const insights = this.extractInsights(oldMessages);
-      for (const insight of insights) {
-        try {
-          await cfg.brain.learn(insight.content, { insight_type: insight.type });
-          learnedCount++;
-        } catch { /* non-critical */ }
-      }
-      summary = `${oldMessages.length} messages compressed. Extracted ${learnedCount} insights (${insights.map(i => i.type).filter((v, i, a) => a.indexOf(v) === i).join(', ')}).`;
-    } else {
-      summary = this.summarize(oldMessages);
-    }
-
-    const compressionMessage: Message = {
-      id: `compressed-${Date.now()}`,
-      role: 'system',
-      content: `[Context compressed: ${oldMessages.length} messages → ${summary}${cfg.brain ? ' Details stored in Brain, use recall() to retrieve.' : ''}]`,
-      timestamp: Date.now(),
-    };
-
-    return {
-      messages: [compressionMessage, ...recentMessages],
-      learnedCount,
-      savedTokens: oldTokens - this.estimateTokens(compressionMessage.content),
-      summary,
-    };
-  }
-
-  /**
-   * Restore context from brain for a given query.
-   */
-  async restore(query: string, brain: any): Promise<string[]> {
-    if (!brain?.recall) return [];
-    try {
-      const results = await brain.recall(query);
-      if (Array.isArray(results)) {
-        return results.map((r: any) => typeof r === 'string' ? r : r.content ?? JSON.stringify(r));
-      }
-      if (typeof results === 'string') return [results];
-      return [];
-    } catch {
-      return [];
-    }
-  }
-}
+import type { Message } from '../core/types';
+
+export interface CompressorConfig {
+  maxTokens: number;
+  compressThreshold: number;
+  preserveRecent: number;
+  brain?: any;
+}
+
+export interface CompressResult {
+  messages: Message[];
+  learnedCount: number;
+  savedTokens: number;
+  summary: string;
+}
+
+const DEFAULT_CONFIG: CompressorConfig = {
+  maxTokens: 8000,
+  compressThreshold: 0.8,
+  preserveRecent: 10,
+};
+
+/**
+ * Context compression with optional DeepBrain memory offloading.
+ */
+export class ContextCompressor {
+  private config: CompressorConfig;
+
+  constructor(config: Partial<CompressorConfig> = {}) {
+    this.config = { ...DEFAULT_CONFIG, ...config };
+  }
+
+  /**
+   * Estimate token count using language-aware heuristic.
+   * English: ~1 token per 4 chars. Chinese: ~1 token per 2 chars.
+   */
+  estimateTokens(text: string): number {
+    let tokens = 0;
+    for (const char of text) {
+      // CJK Unicode range detection
+      const code = char.codePointAt(0) ?? 0;
+      if (
+        (code >= 0x4e00 && code <= 0x9fff) ||  // CJK Unified
+        (code >= 0x3400 && code <= 0x4dbf) ||  // CJK Extension A
+        (code >= 0x3000 && code <= 0x303f)     // CJK Punctuation
+      ) {
+        tokens += 0.5; // 1 token per 2 chars
+      } else {
+        tokens += 0.25; // 1 token per 4 chars
+      }
+    }
+    return Math.ceil(tokens);
+  }
+
+  private estimateMessagesTokens(messages: Message[]): number {
+    return messages.reduce((sum, m) => sum + this.estimateTokens(m.content), 0);
+  }
+
+  /**
+   * Extract key insights from messages for brain storage.
+   */
+  private extractInsights(messages: Message[]): Array<{ content: string; type: string }> {
+    const insights: Array<{ content: string; type: string }> = [];
+    for (const msg of messages) {
+      const c = msg.content;
+      // Decisions
+      if (/\b(decided|decision|choose|chose|will use|going with|let's go|确定|决定)\b/i.test(c)) {
+        insights.push({ content: c.slice(0, 500), type: 'decision' });
+      }
+      // Facts / definitions
+      else if (/\b(is defined as|means|equals|refers to|是指|定义)\b/i.test(c)) {
+        insights.push({ content: c.slice(0, 500), type: 'fact' });
+      }
+      // Preferences
+      else if (/\b(prefer|like|want|don't like|不喜欢|喜欢|偏好)\b/i.test(c)) {
+        insights.push({ content: c.slice(0, 500), type: 'preference' });
+      }
+      // Code snippets
+      else if (/```[\s\S]{20,}```/.test(c)) {
+        insights.push({ content: c.slice(0, 800), type: 'code' });
+      }
+      // Long assistant messages likely contain useful info
+      else if (msg.role === 'assistant' && c.length > 200) {
+        insights.push({ content: c.slice(0, 500), type: 'knowledge' });
+      }
+    }
+    return insights;
+  }
+
+  /**
+   * Generate a simple summary from messages (no-brain fallback).
+   */
+  private summarize(messages: Message[]): string {
+    const topics = new Set<string>();
+    const keyLines: string[] = [];
+
+    for (const msg of messages) {
+      // Extract first meaningful sentence
+      const firstLine = msg.content.split(/[.\n!?。！？]/)[0]?.trim();
+      if (firstLine && firstLine.length > 10 && firstLine.length < 200) {
+        if (keyLines.length < 5) keyLines.push(`[${msg.role}] ${firstLine}`);
+      }
+      // Extract topic words (capitalized words, Chinese phrases)
+      const words = msg.content.match(/[A-Z][a-z]{2,}/g) ?? [];
+      words.forEach(w => topics.add(w));
+    }
+
+    const topicStr = [...topics].slice(0, 10).join(', ');
+    const linesStr = keyLines.join('; ');
+    return `Topics: ${topicStr || 'general discussion'}. Key points: ${linesStr || 'varied conversation'}`;
+  }
+
+  /**
+   * Compress messages when token count exceeds threshold.
+   */
+  async compress(messages: Message[], config?: Partial<CompressorConfig>): Promise<CompressResult> {
+    const cfg = { ...this.config, ...config };
+    const totalTokens = this.estimateMessagesTokens(messages);
+    const threshold = cfg.maxTokens * cfg.compressThreshold;
+
+    // Under threshold — return as-is
+    if (totalTokens <= threshold) {
+      return {
+        messages: [...messages],
+        learnedCount: 0,
+        savedTokens: 0,
+        summary: '',
+      };
+    }
+
+    const recentCount = Math.min(cfg.preserveRecent, messages.length);
+    const splitIdx = messages.length - recentCount;
+    const oldMessages = messages.slice(0, splitIdx);
+    const recentMessages = messages.slice(splitIdx);
+
+    if (oldMessages.length === 0) {
+      return { messages: [...messages], learnedCount: 0, savedTokens: 0, summary: '' };
+    }
+
+    const oldTokens = this.estimateMessagesTokens(oldMessages);
+    let learnedCount = 0;
+    let summary: string;
+
+    if (cfg.brain) {
+      // Extract and learn insights
+      const insights = this.extractInsights(oldMessages);
+      for (const insight of insights) {
+        try {
+          await cfg.brain.learn(insight.content, { insight_type: insight.type });
+          learnedCount++;
+        } catch { /* non-critical */ }
+      }
+      summary = `${oldMessages.length} messages compressed. Extracted ${learnedCount} insights (${insights.map(i => i.type).filter((v, i, a) => a.indexOf(v) === i).join(', ')}).`;
+    } else {
+      summary = this.summarize(oldMessages);
+    }
+
+    const compressionMessage: Message = {
+      id: `compressed-${Date.now()}`,
+      role: 'system',
+      content: `[Context compressed: ${oldMessages.length} messages → ${summary}${cfg.brain ? ' Details stored in Brain, use recall() to retrieve.' : ''}]`,
+      timestamp: Date.now(),
+    };
+
+    return {
+      messages: [compressionMessage, ...recentMessages],
+      learnedCount,
+      savedTokens: oldTokens - this.estimateTokens(compressionMessage.content),
+      summary,
+    };
+  }
+
+  /**
+   * Restore context from brain for a given query.
+   */
+  async restore(query: string, brain: any): Promise<string[]> {
+    if (!brain?.recall) return [];
+    try {
+      const results = await brain.recall(query);
+      if (Array.isArray(results)) {
+        return results.map((r: any) => typeof r === 'string' ? r : r.content ?? JSON.stringify(r));
+      }
+      if (typeof results === 'string') return [results];
+      return [];
+    } catch {
+      return [];
+    }
+  }
+}