lynkr 7.2.0 → 7.2.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "lynkr",
-  "version": "7.2.0",
+  "version": "7.2.2",
   "description": "Self-hosted Claude Code & Cursor proxy with Databricks,AWS BedRock,Azure  adapters, openrouter, Ollama,llamacpp,LM Studio, workspace tooling, and MCP integration.",
   "main": "index.js",
   "bin": {
@@ -47,7 +47,6 @@
     "@azure/openai": "^2.0.0",
     "@babel/parser": "^7.29.0",
     "@babel/traverse": "^7.29.0",
-    "better-sqlite3": "^12.6.2",
     "compression": "^1.7.4",
     "diff": "^5.2.0",
     "dockerode": "^4.0.2",
@@ -62,6 +61,7 @@
     "undici": "^6.22.0"
   },
   "optionalDependencies": {
+    "better-sqlite3": "^12.6.2",
     "tree-sitter": "^0.21.1",
     "tree-sitter-javascript": "^0.21.0",
     "tree-sitter-python": "^0.21.0",

package/src/agents/store.js

@@ -1,24 +1,39 @@
-const Database = require("better-sqlite3");
+let Database;
+try {
+  Database = require("better-sqlite3");
+} catch {
+  Database = null;
+}
 const path = require("path");
 const fs = require("fs");
 const logger = require("../logger");
 
 class AgentStore {
   constructor() {
-    // Use same database location as main app
-    const dbDir = path.join(process.cwd(), "data");
-    if (!fs.existsSync(dbDir)) {
-      fs.mkdirSync(dbDir, { recursive: true });
+    if (!Database) {
+      this.db = null;
+      return;
     }
 
-    const dbPath = path.join(dbDir, "lynkr.db");
-    this.db = new Database(dbPath, {
-      verbose: process.env.DEBUG_SQL ? console.log : null,
-      fileMustExist: false
-    });
-
-    this.initTables();
-    this.prepareStatements();
+    try {
+      // Use same database location as main app
+      const dbDir = path.join(process.cwd(), "data");
+      if (!fs.existsSync(dbDir)) {
+        fs.mkdirSync(dbDir, { recursive: true });
+      }
+
+      const dbPath = path.join(dbDir, "lynkr.db");
+      this.db = new Database(dbPath, {
+        verbose: process.env.DEBUG_SQL ? console.log : null,
+        fileMustExist: false
+      });
+
+      this.initTables();
+      this.prepareStatements();
+    } catch (err) {
+      logger.warn({ err: err.message }, "AgentStore: better-sqlite3 not available");
+      this.db = null;
+    }
   }
 
   initTables() {

package/src/budget/index.js

@@ -16,15 +16,21 @@ class BudgetManager {
       return;
     }
 
-    const dbPath = path.join(process.cwd(), 'data', 'budgets.db');
-    const dbDir = path.dirname(dbPath);
+    try {
+      const dbPath = path.join(process.cwd(), 'data', 'budgets.db');
+      const dbDir = path.dirname(dbPath);
 
-    if (!fs.existsSync(dbDir)) {
-      fs.mkdirSync(dbDir, { recursive: true });
-    }
+      if (!fs.existsSync(dbDir)) {
+        fs.mkdirSync(dbDir, { recursive: true });
+      }
 
-    this.db = new Database(dbPath);
-    this.initDatabase();
+      this.db = new Database(dbPath);
+      this.initDatabase();
+    } catch (err) {
+      logger.warn({ err: err.message }, "BudgetManager: better-sqlite3 not available");
+      this.enabled = false;
+      return;
+    }
 
     logger.info({ dbPath }, 'Budget manager initialized');
   }

package/src/cache/prompt.js

@@ -1,5 +1,10 @@
 const crypto = require("crypto");
-const Database = require("better-sqlite3");
+let Database;
+try {
+  Database = require("better-sqlite3");
+} catch {
+  Database = null;
+}
 const path = require("path");
 const fs = require("fs");
 const config = require("../config");
@@ -49,9 +54,11 @@ class PromptCache {
     this.isClosed = false; // Track if database has been closed
 
     // Initialize persistent cache database
-    if (this.enabled) {
+    if (this.enabled && Database) {
       this.initDatabase();
       this.startPruning();
+    } else if (!Database) {
+      this.enabled = false;
     }
   }
 

package/src/cache/semantic.js

@@ -142,9 +142,9 @@ class SemanticCache {
 
   /**
    * Extract cacheable text from messages
-   * IMPORTANT: Only extracts user message content, NOT system prompt.
-   * System prompt is handled separately via hash for exact matching.
-   * This prevents false cache hits when system prompts are large and similar.
+   * IMPORTANT: Only extracts the ACTUAL user query, NOT system-like content.
+   * When messages are merged (e.g., Codex sends AGENTS.md as user role),
+   * we need to extract only the real user query from the end.
    *
    * @param {Array} messages - Chat messages
    * @returns {string|null} - Extracted user prompt or null
@@ -169,12 +169,121 @@ class SemanticCache {
         .join('\n');
     }
 
-    // REMOVED: System prompt inclusion was causing false cache hits
-    // The system prompt is now handled via hash matching instead
+    // Extract ONLY the actual user query when content contains merged system-like prefixes
+    // Codex and other clients send AGENTS.md, environment_context, etc. as user role messages
+    // which get merged into one large user message. We need to extract just the real query.
+    const originalLength = content.length;
+    content = this._extractActualUserQuery(content);
+
+    // Log extraction for debugging
+    if (originalLength !== content.length) {
+      logger.info({
+        originalLength,
+        extractedLength: content.length,
+        extracted: content.substring(0, 100),
+      }, '[SemanticCache] Extracted user query from merged content');
+    }
 
     return content.trim() || null;
   }
 
+  /**
+   * Extract the actual user query from potentially merged content.
+   * Codex and other clients merge system instructions with user queries.
+   * We need to find the ACTUAL user query, which is usually short and at the end.
+   *
+   * @param {string} content - Potentially merged user content
+   * @returns {string} - The actual user query
+   */
+  _extractActualUserQuery(content) {
+    if (!content) return content;
+
+    // Short content is likely the actual query - no extraction needed
+    if (content.length < 100) {
+      return content;
+    }
+
+    // Patterns that indicate SYSTEM/INSTRUCTION content (NOT user queries)
+    const systemPatterns = [
+      /^#\s*(AGENTS|CLAUDE|README)/i,           // Markdown doc headers
+      /^<[a-z_-]+[\s>]/i,                       // XML-like tags
+      /^```/,                                   // Code blocks
+      /^---\s*$/m,                              // YAML/markdown separators
+      /^IMPORTANT:/i,                           // Instruction markers
+      /^(permissions|environment|collaboration|context|instructions|Focus on)/i,
+      /sandboxing|workspace|cwd|shell/i,        // Environment info
+      /Do not summarize|respond ONLY/i,         // Instruction text
+    ];
+
+    // Split content by double newlines or single newlines
+    const segments = content.split(/\n\n+|\n(?=[A-Z#<])/);
+
+    // Strategy 1: Find the LAST SHORT segment that looks like a real query
+    // Real user queries are usually short (< 200 chars) and don't match system patterns
+    for (let i = segments.length - 1; i >= 0; i--) {
+      const segment = segments[i].trim();
+
+      // Skip empty or very short segments (< 2 chars)
+      if (!segment || segment.length < 2) continue;
+
+      // Skip if too long (system content tends to be verbose)
+      if (segment.length > 300) continue;
+
+      // Check if this looks like system content
+      const isSystemContent = systemPatterns.some(pattern => pattern.test(segment));
+
+      if (!isSystemContent) {
+        // Found a non-system segment - likely the real query
+        logger.debug({
+          originalLength: content.length,
+          extractedLength: segment.length,
+          extracted: segment.substring(0, 100),
+        }, '[SemanticCache] Extracted actual user query');
+        return segment;
+      }
+    }
+
+    // Strategy 2: Look for content after the last XML closing tag
+    const afterXmlMatch = content.match(/<\/[^>]+>\s*\n*([^<\n]{2,200})$/);
+    if (afterXmlMatch) {
+      const extracted = afterXmlMatch[1].trim();
+      if (extracted.length >= 2) {
+        logger.debug({
+          extractedLength: extracted.length,
+          extracted: extracted.substring(0, 100),
+        }, '[SemanticCache] Extracted query after XML tag');
+        return extracted;
+      }
+    }
+
+    // Strategy 3: Take the very last line if it's short
+    const lines = content.split('\n').filter(l => l.trim());
+    const lastLine = lines[lines.length - 1]?.trim();
+    if (lastLine && lastLine.length >= 2 && lastLine.length <= 200) {
+      const isSystem = systemPatterns.some(p => p.test(lastLine));
+      if (!isSystem) {
+        logger.debug({
+          extractedLength: lastLine.length,
+          extracted: lastLine.substring(0, 100),
+        }, '[SemanticCache] Extracted last line as query');
+        return lastLine;
+      }
+    }
+
+    // Strategy 4: If all else fails, return last 150 chars
+    // This ensures we don't cache based on system prompt prefix
+    if (content.length > 500) {
+      const tail = content.slice(-150).trim();
+      logger.debug({
+        originalLength: content.length,
+        extractedLength: tail.length,
+      }, '[SemanticCache] Using tail extraction fallback');
+      return tail;
+    }
+
+    return content;
+  }
+
   /**
    * Find the most similar cached response
    * IMPORTANT: Only matches entries with the same system prompt hash.