mapify-cli 1.3.2

MAP Framework installer - Modular Agentic Planner for Claude Code

MAP Framework for Claude Code

Implementation of Modular Agentic Planner (MAP) — a cognitive architecture for AI agents inspired by prefrontal cortex functions. Orchestrates 8 specialized agents for development with automatic quality validation.

Based on: Nature Communications research (2025) — 74% improvement in planning tasks Enhanced with: ACE (Agentic Context Engineering) — continuous learning from experience

📖 Documentation Structure

• README (this file) - Quick start and overview
• INSTALL.md - Complete installation guide with PATH setup and troubleshooting
• ARCHITECTURE.md - Technical deep dive, customization, and MCP integration
• USAGE.md - Practical examples, best practices, cost optimization, and dependency validation

🚀 Quick Start

Inside Claude Code (Recommended)

# Feature development (full quality - 8 agents)
/map-feature implement user authentication with JWT tokens

# ⭐ RECOMMENDED: Efficient workflow (30-40% token savings, preserves learning)
/map-efficient implement user profile page with avatar upload

# Debugging
/map-debug fix the API 500 error on login endpoint

# Refactoring
/map-refactor refactor UserService class with dependency injection

# Code review
/map-review review the recent changes in auth.py

# ⚠️ Fast workflow (40-50% savings, NO learning - throwaway code only)
/map-fast prototype a quick API endpoint mockup

Command Line Usage

MAP Framework works exclusively through slash commands in Claude Code:

# Start Claude Code in your project directory
cd your-project
claude

# Use slash commands inside Claude Code
/map-feature implement user authentication with JWT tokens

Note: Direct claude --agents syntax is not applicable to MAP Framework, as the orchestration logic is implemented in slash command prompts (.claude/commands/map-*.md), not as a separate agent file.

📦 Installation

Stable Release (Recommended)

# Using pip
pip install mapify-cli

# OR using UV (recommended for isolated tools)
uv tool install mapify-cli

# Verify installation
mapify --version

# Initialize in your project
cd your-project
mapify init

# Available commands
mapify --help                     # Show all available commands
mapify validate graph <file>      # Validate task dependency graphs
mapify playbook search <query>    # Search playbook patterns
mapify recitation get-context     # Get current task plan

Version Pinning:

# Install specific version
pip install mapify-cli==1.0.0

# Install with version constraints (semantic versioning: MAJOR.MINOR.PATCH)
pip install "mapify-cli>=1.0.0,<2.0.0"  # Allow 1.x versions, exclude 2.0.0+

Version Information:

• Check installed version: mapify --version
• PyPI releases - Available versions and package details
• GitHub releases - Changelog and release notes

Development Installation

For contributors or testing bleeding-edge features:

# Install from git repository
uv tool install --from git+https://github.com/azalio/map-framework.git mapify-cli

# OR clone and install locally
git clone https://github.com/azalio/map-framework.git
cd map-framework
pip install -e .

Other installation methods (manual copy, troubleshooting): See INSTALL.md

For maintainers: Release process documented in RELEASING.md

🔀 Workflow Variants

MAP Framework offers three workflow variants optimized for different scenarios:

Command	Token Usage	Learning	Quality Gates	Best For
/map-feature	100% (baseline)	✅ Full (per-subtask)	✅ All 8 agents	Critical features, first-time implementations, maximum QA
/map-efficient ⭐	60-70%	✅ Full (batched)	✅ Essential agents	RECOMMENDED: Most production tasks
/map-fast ⚠️	50-60%	❌ None	⚠️ Basic only	Throwaway prototypes, experiments (NOT production)

Which Workflow Should You Use?

Use /map-efficient (RECOMMENDED) when:

• ✅ Building production features where token costs matter
• ✅ Well-understood tasks with low to medium risk
• ✅ You want learning without excessive token usage
• ✅ Iterative development with frequent workflows

Use /map-feature (full workflow) when:

• 🔒 Implementing security-critical functionality
• 🔒 First-time complex features requiring maximum validation
• 🔒 High-risk changes affecting many files
• 🔒 You need complete quality assurance and impact analysis

Use /map-fast (minimal) ONLY when:

• 🗑️ Creating throwaway prototypes you'll discard
• 🗑️ Quick experiments where quality doesn't matter
• 🗑️ Learning/tutorial contexts where failure is acceptable
• ⚠️ NEVER for production code - no learning, quality risks

🎯 Auto-Activation System

Don't remember which workflow to use? MAP automatically suggests the right workflow based on your request!

Just describe your task naturally - no need to remember slash commands:

Your Request	MAP Suggests	Why
"Fix the failing tests"	/map-debug	Keywords: fix, failing test
"Implement user login"	/map-feature	Keywords: implement, feature
"Optimize database queries"	/map-efficient	Keywords: optimize
"Restructure auth module"	/map-refactor	Keywords: restructure
"Quick prototype for testing"	/map-fast	Keywords: quick, prototype

How it works:

1. Start typing your request normally
2. MAP analyzes keywords and intent patterns
3. Suggests the most appropriate workflow
4. You can accept the suggestion or proceed with your request

Customization: Edit .claude/workflow-rules.json to add project-specific trigger words and patterns.

Key Differences

/map-efficient optimizations:

• Conditional Predictor: Only called for high-risk tasks (security, breaking changes)
• Batched Learning: Reflector/Curator run once at end (vs per-subtask)
• Evaluator Skipped: Monitor provides sufficient validation for most tasks
• Result: 30-40% token savings while preserving learning and essential quality gates

/map-fast limitations:

• ❌ No impact analysis (Predictor skipped)
• ❌ No quality scoring (Evaluator skipped)
• ❌ No learning (Reflector/Curator skipped)
• ❌ Playbook never improves
• ❌ Cipher knowledge never accumulates
• ⚠️ Defeats MAP's core purpose of continuous improvement

See USAGE.md for detailed decision guide and real-world token usage examples.

📚 Skills System

MAP includes interactive skills that provide specialized guidance:

map-workflows-guide - Helps you choose the right workflow

Auto-suggested when you ask:

• "Which workflow should I use?"
• "What's the difference between workflows?"
• "When to use /map-efficient vs /map-feature?"

What you get:

• Quick decision tree (5 questions → recommended workflow)
• Comparison matrix (token cost, learning, agents, use cases)
• 8 detailed resource guides (progressive disclosure)

Skills vs Agents:

• Skills = Passive documentation (guidance)
• Agents = Active execution (code generation)

See docs/USAGE.md for full details.

Requirements

• Claude Code CLI — installed and configured
• Python 3.11+ — for mapify CLI (optional)
• Git — for cloning repository

🏗️ Architecture

MAP Framework orchestrates 8 specialized agents through slash commands:

• TaskDecomposer breaks goals into subtasks
• Actor generates code, Monitor validates quality
• Predictor analyzes impact, Evaluator scores solutions
• Reflector/Curator enable continuous learning via ACE playbook

The orchestration logic lives in .claude/commands/map-*.md prompts, coordinating agents via the Task tool.

See ARCHITECTURE.md for:

• Detailed agent specifications and responsibilities
• MCP integration architecture and tool usage patterns
• Agent coordination protocol and workflow stages
• Template customization guide with examples
• Hooks integration (automated validation, knowledge storage, context enrichment)
• Context engineering principles and optimizations

🔌 MCP Integration

MAP uses MCP (Model Context Protocol) servers for enhanced capabilities:

• cipher - Knowledge base for storing and retrieving successful patterns
• claude-reviewer - Professional code review with security analysis
• context7 - Up-to-date library documentation
• sequential-thinking - Chain-of-thought reasoning for complex problems
• codex-bridge - AI code generation (requires extended timeout)
• deepwiki - GitHub repository intelligence

Configuration files: .claude/mcp_config.json and mcp_config.json

See ARCHITECTURE.md for complete setup and usage patterns

📚 Usage Examples

# Feature development
/map-feature implement user profile page with avatar upload

# Bug fixing
/map-debug debug why payment processing fails for amounts over $1000

# Refactoring
/map-refactor refactor OrderService to use dependency injection

See USAGE.md for:

• Comprehensive usage examples with detailed scenarios
• Best practices for optimal results
• Cost optimization strategies (40-60% savings)
• Playbook management commands

🎓 ACE Playbook

Built-in learning system that improves with each task:

• Reflector extracts patterns from successes and failures
• Curator maintains structured knowledge base with quality tracking
• Semantic search (optional) finds patterns by meaning, not keywords
• Dependency validation ensures valid task graphs before execution
• Automatically grows high-quality pattern library

Playbook Commands

# View statistics
mapify playbook stats

# Search patterns
mapify playbook search "JWT authentication"

# View high-quality patterns
mapify playbook sync

# Validate task dependencies with visualization
python scripts/validate-dependencies.py decomposer-output.json --visualize

Optional semantic search: pip install -r requirements-semantic.txt for meaning-based matching. Details in SEMANTIC_SEARCH_SETUP.md and ARCHITECTURE.md.

Playbook configuration: See ARCHITECTURE.md for top_k settings and optimization.

Dependency validation: See USAGE.md for comprehensive guide on validating TaskDecomposer output, including cycle detection, visualization, and CI/CD integration.

💰 Cost Optimization

MAP Framework uses intelligent model selection per agent:

• Predictor & Evaluator use haiku (fast analysis) → ⬇️⬇️⬇️ cost
• Actor, Monitor, Reflector, Curator use sonnet (quality-critical) → balanced cost

Result: 40-60% cost reduction vs all-sonnet while maintaining code quality.

See USAGE.md for detailed cost breakdown and model override strategies

🔗 Hooks Integration

MAP integrates with Claude Code hooks for automated validation, knowledge storage, and context enrichment. Active hooks protect template variables, auto-store successful patterns, enrich prompts with relevant knowledge, and track performance metrics.

See ARCHITECTURE.md and .claude/hooks/README.md for configuration

🛠️ Troubleshooting

Command Not Found

Error: Slash command not recognized

Solution:

• Ensure you're in a directory with .claude/commands/ containing map-*.md files
• Use /map-feature, /map-debug, /map-refactor, or /map-review
• Run /help to see available commands

Agent Not Found

Error: Agent file not found

Solution: Ensure .claude/agents/ directory contains all 8 agent files (task-decomposer.md, actor.md, monitor.md, predictor.md, evaluator.md, reflector.md, curator.md, documentation-reviewer.md)

Semantic Search Warning

Warning: sentence-transformers not installed

Solution: pip install -r requirements-semantic.txt See SEMANTIC_SEARCH_SETUP.md for detailed troubleshooting

Infinite Loops

Actor-Monitor loop exceeding iterations

Solution: Orchestrator limits iterations to 3-5. Clarify requirements or add constraints.

More troubleshooting: See INSTALL.md for PATH issues, MCP configuration, and installation problems

🔧 Customization

Agent prompts in .claude/agents/*.md use Handlebars template syntax for dynamic context injection. You can safely modify instructions, examples, and validation criteria, but MUST NOT remove template variables like {{language}}, {{#if playbook_bullets}}, or {{feedback}} — these are critical for orchestration and ACE learning.

See ARCHITECTURE.md for:

• Safe vs unsafe modifications with examples
• Template variable reference
• Model selection per agent
• Adding custom agents
• Template validation and git hooks

📊 Success Metrics

• Monitor approval rate: >80% first try
• Evaluator scores: average >7.0/10
• Iteration count: <3 per subtask
• Playbook growth: increasing high-quality patterns

🤝 Contributing

Improvements welcome:

• Prompts for specific languages/frameworks
• New specialized agents
• CI/CD integrations
• Success story examples
• Plugin extensions for MAP Framework

📄 License

MIT License — see LICENSE file for details

🔗 References

• MAP Paper - Nature Communications
• ACE Paper - arXiv:2510.04618v1
• Claude Code Documentation

---

MAP is not just automation — it's systematic quality improvement through structured validation and iterative refinement.