@jungjaehoon/mama-core 1.9.0 → 2.1.0

package/dist/memory-inject.js

@@ -1,116 +0,0 @@
-"use strict";
-/**
- * MAMA (Memory-Augmented MCP Architecture) - Memory Injection Hook
- *
- * UserPromptSubmit hook that injects decision history into Claude's context
- * Tasks: 1.1-1.9 (Hook setup, timeout handling, context injection)
- * AC #1: Query intent → Graph query → Format → Inject (5s timeout for LLM latency)
- * AC #2: No history → null (graceful fallback)
- * AC #3: Timeout → graceful fallback
- *
- * @module memory-inject
- * @version 1.0
- * @date 2025-11-14
- */
-Object.defineProperty(exports, "__esModule", { value: true });
-exports.injectDecisionContext = injectDecisionContext;
-// Error handling policy:
-// - Timeout errors: thrown (caller handles retry/fallback)
-// - Vector search unavailable: returns empty array (recoverable, not critical)
-const debug_logger_js_1 = require("./debug-logger.js");
-const memory_store_js_1 = require("./memory-store.js");
-const decision_formatter_js_1 = require("./decision-formatter.js");
-// Configuration
-const TIMEOUT_MS = 5000; // LLM-based intent detection, user accepts longer thinking
-const TOKEN_BUDGET = 500; // AC #1: Max 500 tokens per injection
-/**
- * UserPromptSubmit Hook Handler
- *
- * Task 1.1-1.9: Main entry point for memory injection
- * AC #1, #2, #3: Intent analysis → Query → Format → Inject
- *
- * @param userMessage - User's message from prompt
- * @returns Injected context or null
- */
-async function injectDecisionContext(userMessage) {
-    const startTime = Date.now();
-    let timeoutId = null;
-    try {
-        // Task 1.3: Implement timeout wrapper (Promise.race with timeout)
-        const timeoutPromise = new Promise((_, reject) => {
-            timeoutId = setTimeout(() => {
-                reject(new Error(`Memory injection timeout (${TIMEOUT_MS}ms)`));
-            }, TIMEOUT_MS);
-        });
-        const context = await Promise.race([
-            performMemoryInjection(userMessage, startTime),
-            timeoutPromise,
-        ]);
-        // Clear timeout on success to prevent timer leak
-        if (timeoutId) {
-            clearTimeout(timeoutId);
-        }
-        return context;
-    }
-    catch (error) {
-        // Clear timeout on error to prevent timer leak
-        if (timeoutId) {
-            clearTimeout(timeoutId);
-        }
-        // CLAUDE.md Rule #1: NO FALLBACK
-        // Errors must be thrown for debugging (including timeout)
-        const message = error instanceof Error ? error.message : String(error);
-        (0, debug_logger_js_1.error)(`[MAMA] Memory injection FAILED: ${message}`);
-        throw error;
-    }
-}
-/**
- * Perform memory injection with all steps
- *
- * Simplified: Direct vector search without LLM intent analysis
- * Faster, more reliable, works with all query types
- *
- * @param userMessage - User's message
- * @param startTime - Start timestamp for latency tracking
- * @returns Formatted context or null
- */
-async function performMemoryInjection(userMessage, startTime) {
-    // 1. Generate query embedding
-    const { generateEmbedding } = await import('./embeddings.js');
-    const queryEmbedding = await generateEmbedding(userMessage, 'query');
-    const embeddingLatency = Date.now() - startTime;
-    (0, debug_logger_js_1.info)(`[MAMA] Embedding generation: ${embeddingLatency}ms`);
-    // 2. Adaptive threshold (shorter queries need higher confidence)
-    const wordCount = userMessage.split(/\s+/).length;
-    const adaptiveThreshold = wordCount < 3 ? 0.7 : 0.6;
-    // 3. Vector search (returns [] on error for graceful degradation)
-    let results;
-    try {
-        results = (await (0, memory_store_js_1.vectorSearch)(queryEmbedding, 10, 0.5)); // Get more candidates
-    }
-    catch (error) {
-        // Vector search unavailability should not block the main conversation flow
-        const message = error instanceof Error ? error.message : String(error);
-        (0, debug_logger_js_1.error)(`[MAMA] Vector search failed: ${message}`);
-        return null;
-    }
-    // 4. Filter by adaptive threshold
-    results = results.filter((r) => r.similarity >= adaptiveThreshold);
-    const searchLatency = Date.now() - startTime;
-    (0, debug_logger_js_1.info)(`[MAMA] Vector search: ${searchLatency - embeddingLatency}ms (${results.length} results, threshold: ${adaptiveThreshold})`);
-    // 5. Check if we have any decisions
-    if (results.length === 0) {
-        // Redact user content for privacy - only log length, not content
-        (0, debug_logger_js_1.info)(`[MAMA] No relevant decisions found (query length: ${userMessage.length} chars)`);
-        return null;
-    }
-    // 6. Format context summary
-    const formattedContext = (0, decision_formatter_js_1.formatContext)(results, {
-        maxTokens: TOKEN_BUDGET,
-    });
-    const formatLatency = Date.now() - startTime;
-    (0, debug_logger_js_1.info)(`[MAMA] Format context: ${formatLatency - searchLatency}ms (total: ${formatLatency}ms)`);
-    // 7. Return formatted context (Claude Code will inject it)
-    return formattedContext;
-}
-//# sourceMappingURL=memory-inject.js.map