MCP Standards

# Package Manager Intelligence System - Executive Summary ## The Problem Users repeatedly correct Claude about package manager preferences: ``` "Use uv not pip" ← Correction #1 "Use uv not pip!" ← Correction #2 "USE UV NOT PIP!!!" ← Correction #3 ``` **Current System**: Requires 3-5 identical corrections before learning ## The Solution **Package Manager Intelligence System (PMIS)** reduces corrections from 3-5 to 1-2 through five integrated components: ### 1. Intelligent Detection (0ms overhead) Automatically detects preferred package manager from project files: - `uv.lock` → uv (confidence: 0.9) - `poetry.lock` → poetry (confidence: 0.9) - `package-lock.json` → npm (confidence: 0.7) **Impact**: Immediate context without user input ### 2. Semantic Clustering (AgentDB) Merges similar corrections into single pattern: - "use uv not pip" + "prefer uv" + "always use uv" → **1 pattern** **Technology**: 384-dim embeddings, HNSW graph, <1ms search **Impact**: 67% fewer corrections needed (3 → 1) ### 3. Bayesian Confidence Scoring Updates confidence based on outcomes: - SUCCESS: confidence × 1.2 (capped at 0.95) - FAILURE: confidence × 0.6 (floor at 0.1) - DECAY: -0.05/week if unused **Impact**: Self-adjusting accuracy >90% ### 4. Cross-Project Learning Shares patterns across similar projects: - Django Project #1: "use pytest" → learned - Django Project #2: "use pytest" → **auto-applied** (0 corrections!) **Impact**: Zero-shot learning for new projects ### 5. Proactive Application Predicts and applies before user asks: ```python User: "install pytest" [System detects uv.lock + finds pattern "use uv" (0.85 confidence)] Claude: "uv pip install pytest" # Auto-applied! ``` **Impact**: Eliminates correction loop entirely ## Architecture ``` ┌─────────────────────────────────────────────────────────┐ │ Package Manager Intelligence System │ │ │ │ ProjectFileDetector → SemanticClusterer (AgentDB) │ │ ↓ ↓ │ │ BayesianScorer ← CrossProjectLearner → ProactiveApp │ │ ↓ ↓ │ │ SQLite (audit) + AgentDB (vectors) → CLAUDE.md │ └─────────────────────────────────────────────────────────┘ ``` **Hybrid Storage**: - **AgentDB**: Fast semantic search (<1ms) - **SQLite**: Audit trail, compliance, metadata ## Performance Targets | Metric | Current | Target | Improvement | |--------|---------|--------|-------------| | Corrections to learn | 3-5 | 1-2 | **60-70% reduction** | | Learning time | 2-3 days | Same session | **10x faster** | | Detection latency | N/A | <10ms | **Instant** | | Search latency | 50ms+ | <1ms | **50x faster** | | Prediction accuracy | N/A | >85% | **New capability** | | Context token reduction | 0 | -2,000 | **Less noise** | ## Key Algorithms ### Semantic Clustering ```python # "use uv" + "prefer uv" → same pattern embedding1 = embed("use uv not pip") # [0.234, -0.567, ...] embedding2 = embed("prefer uv over pip") # [0.241, -0.562, ...] similarity = cosine(embedding1, embedding2) # 0.93 → MERGE ``` ### Bayesian Confidence ```python # First correction confidence = 0.4 (prior) + 0.3 (uv.lock detected) = 0.7 # Success → boost confidence = 0.7 × 1.2 = 0.84 # Auto-apply threshold reached (>0.7) ``` ### Cross-Project ```python # Project 1: Django app, learned "use pytest" # Project 2: New Django app similar_projects = find_similar(project2) # Finds project1 patterns = get_patterns(similar_projects) # Gets "use pytest" apply_pattern(project2, "use pytest") # 0 corrections! ``` ## Implementation Roadmap **Phase 1 (Week 1)**: Foundation - Project file detection - Database schema - AgentDB setup **Phase 2 (Week 2)**: Semantic Clustering - Embedding generation - Pattern merging - Confidence scoring **Phase 3 (Week 3)**: Cross-Project & Proactive - Project profiles - Command prediction - Auto-application **Phase 4 (Week 4)**: Polish & Deploy - Performance optimization - Error handling - A/B testing ## Success Criteria **Primary**: Reduce corrections from 3-5 to 1-2 (60-70% reduction) **Secondary**: - Prediction accuracy >85% - False positive rate <5% - End-to-end latency <50ms - User satisfaction >4.0/5 ## Innovation Highlights 1. **Hybrid Architecture**: AgentDB (speed) + SQLite (compliance) 2. **Zero-Configuration**: Auto-detects from project files 3. **Semantic Understanding**: Natural language clustering 4. **Self-Improving**: Bayesian updates from outcomes 5. **Cross-Project Intelligence**: Learn once, apply everywhere ## Business Value **For Users**: - Fewer repetitive corrections - Faster learning - Better cross-project experience - Less context pollution **For mcp-standards**: - Differentiated feature (competitors lack this) - Demonstrates AgentDB/ReasoningBank integration - Measurable UX improvement - Foundation for other preference learning ## Next Steps 1. Review design document: `/docs/architecture/package-manager-intelligence-system.md` 2. Approve architecture and algorithms 3. Begin Phase 1 implementation 4. Set up A/B test framework --- **Contact**: See full specification in `package-manager-intelligence-system.md`