@mrxkun/mcfast-mcp 2.2.1 → 2.2.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mrxkun/mcfast-mcp",
-  "version": "2.2.1",
+  "version": "2.2.2",
   "description": "Ultra-fast code editing with fuzzy patching, auto-rollback, and 5 unified tools.",
   "type": "module",
   "bin": {

package/src/strategies/fuzzy-patch.js

@@ -1,8 +1,16 @@
 /**
- * Fuzzy Patching Engine for mcfast v2.1
- * Applies code changes with tolerance for whitespace and minor formatting differences
+ * Fuzzy Patch Strategy for mcfast v2.1+
+ * Applies unified diffs with whitespace tolerance and semantic similarity
  */
 
+import { parsePatch } from './patch-parser.js';
+import {
+    tokenSimilarity,
+    calculateConfidence,
+    contextAwareMatch,
+    isSemanticMatchingEnabled
+} from './semantic-similarity.js';
+
 /**
  * Calculate Levenshtein distance between two strings
  * Used for fuzzy matching to find best location for patch
@@ -122,59 +130,102 @@ export function parseDiff(diffText) {
 
 /**
  * Find best match location for a pattern in target text
- * Returns { index, score, lineNumber } or null if no good match
+ * Returns { index, distance, confidence } or null if no good match
  */
-export function findBestMatch(targetLines, patternLines, threshold = 0.8) {
-    // Input validation
-    if (!targetLines || !patternLines || patternLines.length === 0) {
-        return null;
-    }
+export function findBestMatch(targetLines, fileLines, startHint = 0) {
+    let bestMatch = null;
+    let bestScore = Infinity;
+    const maxIterations = 10000;
+    let iterations = 0;
 
-    if (targetLines.length < patternLines.length) {
-        return null;
+    const useSemanticMatching = isSemanticMatchingEnabled();
+
+    if (useSemanticMatching) {
+        console.error('[FUZZY] Semantic matching enabled');
     }
 
-    // Performance limit: skip if pattern is too large (>500 lines)
-    if (patternLines.length > 500) {
-        return null;
+    // Try exact match first at hint location
+    if (startHint >= 0 && startHint + targetLines.length <= fileLines.length) {
+        const exactMatch = targetLines.every((line, i) =>
+            fileLines[startHint + i] === line
+        );
+        if (exactMatch) {
+            return { index: startHint, distance: 0, confidence: 1.0 };
+        }
     }
 
-    let bestMatch = null;
-    let bestScore = 0;
-    const maxIterations = Math.min(targetLines.length - patternLines.length + 1, 10000);
+    // Fuzzy search with semantic similarity
+    for (let i = 0; i <= fileLines.length - targetLines.length; i++) {
+        iterations++;
+        if (iterations > maxIterations) {
+            console.error(`[FUZZY] Max iterations (${maxIterations}) reached`);
+            break;
+        }
 
-    // Sliding window search with iteration limit
-    for (let i = 0; i < maxIterations; i++) {
-        const window = targetLines.slice(i, i + patternLines.length);
-        const windowText = window.join('\n');
-        const patternText = patternLines.join('\n');
+        let totalDistance = 0;
+        let tokenSimilaritySum = 0;
+        let contextMatchSum = 0;
 
-        // Calculate max acceptable distance based on threshold
-        const maxLen = Math.max(windowText.length, patternText.length);
-        const maxDistance = Math.ceil(maxLen * (1 - threshold));
+        for (let j = 0; j < targetLines.length; j++) {
+            const targetLine = targetLines[j];
+            const fileLine = fileLines[i + j];
 
-        const distance = levenshteinDistance(windowText, patternText, maxDistance);
+            // Levenshtein distance
+            const distance = levenshteinDistance(targetLine, fileLine);
+            totalDistance += distance;
 
-        // Skip if distance exceeds threshold
-        if (distance > maxDistance) continue;
+            // Token similarity (always available)
+            const tokSim = tokenSimilarity(targetLine, fileLine);
+            tokenSimilaritySum += tokSim;
 
-        const score = 1.0 - (distance / maxLen);
+            // Context-aware matching (use surrounding lines)
+            const surroundingLines = [
+                fileLines[i + j - 1],
+                fileLines[i + j + 1]
+            ].filter(Boolean);
 
-        if (score > bestScore && score >= threshold) {
+            const contextScore = contextAwareMatch(targetLine, fileLine, surroundingLines);
+            contextMatchSum += contextScore;
+        }
+
+        const avgTokenSim = tokenSimilaritySum / targetLines.length;
+        const avgContextMatch = contextMatchSum / targetLines.length;
+
+        // Calculate confidence using enhanced scoring
+        const confidence = calculateConfidence(null, {
+            levenshteinDistance: totalDistance,
+            maxDistance: targetLines.length * 100,
+            tokenSimilarity: avgTokenSim,
+            structuralSimilarity: 0, // Would need AST parsing
+            lineNumberMatch: i === startHint,
+            surroundingContextMatch: avgContextMatch
+        });
+
+        // Prefer matches with higher confidence
+        const score = totalDistance * (1 - confidence * 0.5); // Confidence reduces effective distance
+
+        if (score < bestScore) {
             bestScore = score;
             bestMatch = {
                 index: i,
-                score: score,
-                lineNumber: i + 1
+                distance: totalDistance,
+                confidence,
+                tokenSimilarity: avgTokenSim,
+                contextMatch: avgContextMatch
             };
 
-            // Early exit if perfect match found
-            if (score >= 0.99) {
+            // Early termination if we find a very good match
+            if (confidence > 0.99) {
+                console.error(`[FUZZY] Early termination at ${confidence.toFixed(2)} confidence`);
                 break;
             }
         }
     }
 
+    if (bestMatch) {
+        console.error(`[FUZZY] Best match: line ${bestMatch.index}, distance ${bestMatch.distance}, confidence ${bestMatch.confidence.toFixed(2)}, token_sim ${bestMatch.tokenSimilarity.toFixed(2)}`);
+    }
+
     return bestMatch;
 }
 

package/src/strategies/semantic-similarity.js

@@ -0,0 +1,292 @@
+/**
+ * Semantic Similarity for mcfast v2.2
+ * Optional embedding-based code similarity (controlled by MCFAST_SEMANTIC_MATCHING env var)
+ */
+
+import crypto from 'crypto';
+
+/**
+ * Check if semantic matching is enabled
+ */
+export function isSemanticMatchingEnabled() {
+    return process.env.MCFAST_SEMANTIC_MATCHING === 'true' ||
+        process.env.MCFAST_SEMANTIC_MATCHING === '1';
+}
+
+/**
+ * Simple hash-based similarity (always available, no dependencies)
+ * Uses token-level comparison for code similarity
+ */
+export function tokenSimilarity(code1, code2) {
+    // Normalize whitespace and tokenize
+    const normalize = (code) => {
+        return code
+            .replace(/\s+/g, ' ')           // Normalize whitespace
+            .replace(/[{}();,]/g, ' $& ')   // Separate punctuation
+            .trim()
+            .toLowerCase()
+            .split(/\s+/)
+            .filter(t => t.length > 0);
+    };
+
+    const tokens1 = normalize(code1);
+    const tokens2 = normalize(code2);
+
+    // Jaccard similarity
+    const set1 = new Set(tokens1);
+    const set2 = new Set(tokens2);
+
+    const intersection = new Set([...set1].filter(x => set2.has(x)));
+    const union = new Set([...set1, ...set2]);
+
+    return intersection.size / union.size;
+}
+
+/**
+ * Structural similarity based on AST depth and node types
+ * Lightweight alternative to full embeddings
+ */
+export function structuralSimilarity(ast1, ast2) {
+    const getStructure = (ast) => {
+        const structure = {
+            nodeTypes: new Map(),
+            depth: 0,
+            totalNodes: 0
+        };
+
+        const traverse = (node, depth = 0) => {
+            if (!node || typeof node !== 'object') return;
+
+            structure.totalNodes++;
+            structure.depth = Math.max(structure.depth, depth);
+
+            if (node.type) {
+                structure.nodeTypes.set(
+                    node.type,
+                    (structure.nodeTypes.get(node.type) || 0) + 1
+                );
+            }
+
+            for (const key in node) {
+                if (key === 'loc' || key === 'range') continue;
+                const value = node[key];
+                if (Array.isArray(value)) {
+                    value.forEach(child => traverse(child, depth + 1));
+                } else if (value && typeof value === 'object') {
+                    traverse(value, depth + 1);
+                }
+            }
+        };
+
+        traverse(ast);
+        return structure;
+    };
+
+    const s1 = getStructure(ast1);
+    const s2 = getStructure(ast2);
+
+    // Compare depth similarity
+    const depthSim = 1 - Math.abs(s1.depth - s2.depth) / Math.max(s1.depth, s2.depth, 1);
+
+    // Compare node count similarity
+    const countSim = 1 - Math.abs(s1.totalNodes - s2.totalNodes) / Math.max(s1.totalNodes, s2.totalNodes, 1);
+
+    // Compare node type distribution
+    const allTypes = new Set([...s1.nodeTypes.keys(), ...s2.nodeTypes.keys()]);
+    let typeMatchScore = 0;
+    let typeTotal = 0;
+
+    for (const type of allTypes) {
+        const count1 = s1.nodeTypes.get(type) || 0;
+        const count2 = s2.nodeTypes.get(type) || 0;
+        const maxCount = Math.max(count1, count2);
+        const minCount = Math.min(count1, count2);
+        typeMatchScore += minCount;
+        typeTotal += maxCount;
+    }
+
+    const typeSim = typeTotal > 0 ? typeMatchScore / typeTotal : 0;
+
+    // Weighted average
+    return (depthSim * 0.2) + (countSim * 0.3) + (typeSim * 0.5);
+}
+
+/**
+ * Semantic code similarity (optional, requires OpenAI API)
+ * Only used when MCFAST_SEMANTIC_MATCHING=true
+ */
+export async function semanticSimilarity(code1, code2, options = {}) {
+    if (!isSemanticMatchingEnabled()) {
+        // Fallback to token similarity
+        return tokenSimilarity(code1, code2);
+    }
+
+    const { apiKey = process.env.OPENAI_API_KEY, model = 'text-embedding-3-small' } = options;
+
+    if (!apiKey) {
+        console.warn('[SEMANTIC] OpenAI API key not found, falling back to token similarity');
+        return tokenSimilarity(code1, code2);
+    }
+
+    try {
+        // Get embeddings for both code snippets
+        const [embedding1, embedding2] = await Promise.all([
+            getEmbedding(code1, apiKey, model),
+            getEmbedding(code2, apiKey, model)
+        ]);
+
+        // Cosine similarity
+        return cosineSimilarity(embedding1, embedding2);
+    } catch (error) {
+        console.warn(`[SEMANTIC] Embedding failed: ${error.message}, falling back to token similarity`);
+        return tokenSimilarity(code1, code2);
+    }
+}
+
+/**
+ * Get embedding from OpenAI API
+ */
+async function getEmbedding(text, apiKey, model) {
+    const response = await fetch('https://api.openai.com/v1/embeddings', {
+        method: 'POST',
+        headers: {
+            'Content-Type': 'application/json',
+            'Authorization': `Bearer ${apiKey}`
+        },
+        body: JSON.stringify({
+            input: text,
+            model: model
+        })
+    });
+
+    if (!response.ok) {
+        throw new Error(`OpenAI API error: ${response.status} ${response.statusText}`);
+    }
+
+    const data = await response.json();
+    return data.data[0].embedding;
+}
+
+/**
+ * Cosine similarity between two vectors
+ */
+function cosineSimilarity(vec1, vec2) {
+    if (vec1.length !== vec2.length) {
+        throw new Error('Vectors must have same length');
+    }
+
+    let dotProduct = 0;
+    let norm1 = 0;
+    let norm2 = 0;
+
+    for (let i = 0; i < vec1.length; i++) {
+        dotProduct += vec1[i] * vec2[i];
+        norm1 += vec1[i] * vec1[i];
+        norm2 += vec2[i] * vec2[i];
+    }
+
+    return dotProduct / (Math.sqrt(norm1) * Math.sqrt(norm2));
+}
+
+/**
+ * Enhanced fuzzy matching with semantic similarity
+ * Combines Levenshtein distance with semantic understanding
+ */
+export function enhancedFuzzyMatch(expected, actual, options = {}) {
+    const {
+        useSemanticSimilarity = isSemanticMatchingEnabled(),
+        threshold = 0.7
+    } = options;
+
+    // 1. Exact match
+    if (expected === actual) {
+        return { score: 1.0, method: 'exact' };
+    }
+
+    // 2. Normalized match (whitespace-insensitive)
+    const normalizedExpected = expected.replace(/\s+/g, ' ').trim();
+    const normalizedActual = actual.replace(/\s+/g, ' ').trim();
+
+    if (normalizedExpected === normalizedActual) {
+        return { score: 0.95, method: 'normalized' };
+    }
+
+    // 3. Token similarity (fast, no API calls)
+    const tokenScore = tokenSimilarity(expected, actual);
+
+    if (tokenScore >= threshold) {
+        return { score: tokenScore, method: 'token' };
+    }
+
+    // 4. Semantic similarity (optional, requires API)
+    if (useSemanticSimilarity) {
+        // Note: This is async, caller must await
+        return semanticSimilarity(expected, actual).then(score => ({
+            score,
+            method: 'semantic'
+        }));
+    }
+
+    return { score: tokenScore, method: 'token' };
+}
+
+/**
+ * Improved confidence scoring for fuzzy patches
+ * Combines multiple signals for better accuracy
+ */
+export function calculateConfidence(match, context = {}) {
+    const {
+        levenshteinDistance = 0,
+        maxDistance = 100,
+        tokenSimilarity = 0,
+        structuralSimilarity = 0,
+        lineNumberMatch = false,
+        surroundingContextMatch = 0
+    } = context;
+
+    // Base score from Levenshtein distance
+    const distanceScore = 1 - (levenshteinDistance / maxDistance);
+
+    // Weighted combination
+    let confidence = 0;
+    confidence += distanceScore * 0.3;           // 30% from edit distance
+    confidence += tokenSimilarity * 0.25;        // 25% from token similarity
+    confidence += structuralSimilarity * 0.2;    // 20% from structure
+    confidence += surroundingContextMatch * 0.15; // 15% from context
+    confidence += (lineNumberMatch ? 0.1 : 0);   // 10% bonus for line match
+
+    // Clamp to [0, 1]
+    return Math.max(0, Math.min(1, confidence));
+}
+
+/**
+ * Context-aware code matching
+ * Uses surrounding lines to improve match accuracy
+ */
+export function contextAwareMatch(targetLine, candidateLine, surroundingLines = []) {
+    // Match the target line
+    const lineScore = tokenSimilarity(targetLine, candidateLine);
+
+    if (surroundingLines.length === 0) {
+        return lineScore;
+    }
+
+    // Check if surrounding context also matches
+    let contextScore = 0;
+    let contextCount = 0;
+
+    for (const contextLine of surroundingLines) {
+        if (contextLine && contextLine.trim().length > 0) {
+            // Simple check: does context appear nearby?
+            contextScore += 0.5; // Placeholder - would need actual implementation
+            contextCount++;
+        }
+    }
+
+    if (contextCount > 0) {
+        contextScore /= contextCount;
+    }
+
+    // Combine line score with context score
+    return (lineScore * 0.7) + (contextScore * 0.3);
+}