Skip to content

IMPROVEMENT_ROADMAP.md

Latest commit

 

History

History
306 lines (213 loc) · 6.62 KB

IMPROVEMENT_ROADMAP.md

File metadata and controls

306 lines (213 loc) · 6.62 KB
Raw

Improvement Roadmap

Strategic plan for Steve AI development and enhancements.

Last Updated: 2026-03-03

Executive Summary

Steve AI is 85% production-ready with comprehensive core systems. This roadmap outlines the path to 100% production readiness and beyond.

Current State:

  • 234 source files, 85,752 lines of production code
  • 91 test files, 33,349 lines of test code
  • 425 documentation files
  • All P0 tests complete (ActionExecutor: 909 lines, AgentStateMachine: 1048 lines)

Phase 1: Production Readiness (Current)

1.1 Skill Composition System ✅ IN PROGRESS

Status: Implementation started

Components:

  • SkillComposer.java - Compose skills into complex behaviors
  • ComposedSkill.java - Represents a multi-step skill
  • CompositionStep.java - Individual step in composition
  • Tests for SkillComposer
  • Integration with ActionExecutor

Impact: Enables Voyager-style skill composition (3.3x performance improvement)

1.2 Multi-Agent Coordination

Status: 50% Complete

Remaining Work:

  • Implement Contract Net Protocol bidding
  • Add capability matching
  • Add workload tracking
  • Test with multiple mock agents

Impact: Emergent intelligence from agent cooperation

1.3 Quality Tools

Status: Configured but disabled

Tasks:

  • Run ./gradlew checkstyleMain and fix warnings
  • Run ./gradlew spotbugsMain and fix bugs
  • Change ignoreFailures = false in build.gradle

Phase 2: Feature Completion

2.1 Script DSL System

Purpose: Declarative automation scripts to reduce LLM token usage

Components:

  • Script DSL grammar definition
  • ScriptDSLParser.java
  • ScriptDSLExecutor.java
  • LLM prompt templates for DSL generation

Example DSL:

SCRIPT "mine_iron"
  TRIGGER "need_iron"
  PRECONDITION "has_pickaxe"
  SEQUENCE {
    PATHFIND nearest("iron_ore")
    MINE "iron_ore" quantity(10)
    RETURN_TO player
  }
  ON_FAIL { SAY "Couldn't find iron!" }
END

2.2 LLM→Script Generation Pipeline

Purpose: Automate script creation from natural language

Components:

  • ScriptGenerator.java enhancement
  • Script validation logic
  • Refinement loop based on execution feedback

2.3 Small Model Specialization

Purpose: Reduce API costs by 40-60%

Tasks:

  • Train/fine-tune small models for specific tasks
  • Integrate with CascadeRouter
  • Add fallback to large models

Phase 3: Research & Development

3.1 MUD Automation Learning

Purpose: Extract principles from 1990s MUD automation for LLM learning

Research Questions:

  • How did TinTin++/ZMud solve complex problems without LLMs?
  • What patterns can LLMs learn from trigger/alias systems?
  • How to translate procedural scripts into learnable principles?

3.2 DEPS Planning Integration

Purpose: Add Detective-style verification to planning

Components:

  • Study DEPS paper patterns
  • Implement verification step
  • Add self-correction loop

3.3 DreamerV3 World Model

Purpose: Enable imagination-based planning

Research:

  • Study DreamerV3 architecture
  • Evaluate applicability to Minecraft
  • Design integration approach

Phase 4: Performance Optimization

4.1 Memory Optimization

Issues Identified:

  • Unbounded collections in CompanionMemory
  • CopyOnWriteArrayList overhead in emotional memories

Solutions:

  • Add size limits with LRU eviction
  • Replace with ReentrantReadWriteLock
  • Profile and optimize hot paths

4.2 Pathfinding Optimization

Current State: Good performance (<50ms for 100-block paths)

Improvements:

  • Parallel path exploration
  • Better chunk caching
  • Dynamic heuristic adjustment

4.3 LLM Caching

Current: 40-60% hit rate with semantic caching

Improvements:

  • Increase cache hit rate to 70%+
  • Add context-aware caching
  • Implement cache warming

Phase 5: Dissertation Completion

5.1 Chapter 3: Emotional AI

Status: 60% Complete

Remaining:

  • Complete emotional AI section
  • Add 2024-2025 citations
  • Integration with code examples

5.2 Chapter 6: Architecture

Status: In Progress

Tasks:

  • Add more implementation details
  • Discuss limitations
  • Future work section

5.3 Publication Preparation

Target Venues: ICLR, NeurIPS, AAAI

Requirements:

  • Benchmark results
  • Comparison with Voyager, DreamerV3, DEPS
  • Novel contribution statement

Metrics & Success Criteria

Production Readiness

Metric Current Target
Test Coverage 40% 60%
Checkstyle Disabled Clean
SpotBugs Disabled Clean
Documentation 425 files Complete
Build Time 30s <30s

Performance

Metric Current Target
LLM Latency 1-5s <3s avg
Cache Hit Rate 40-60% 70%+
Pathfinding <50ms <30ms
Memory/Agent ~5MB <3MB

Features

Feature Current Target
Skill Composition 80% 100%
Multi-Agent 50% 100%
Script DSL 0% 100%
Small Models 0% 50%

Risk Assessment

Risk Probability Impact Mitigation
API rate limits Medium High Local LLM fallback
Memory leaks Low High Profiling, limits
LLM hallucination Medium Medium Validation layer
Test flakiness Low Medium Better mocking

Timeline

Q1 2026:

  • Complete skill composition system
  • Re-enable quality tools
  • Improve test coverage to 50%

Q2 2026:

  • Complete multi-agent coordination
  • Implement Script DSL
  • Begin dissertation finalization

Q3 2026:

  • Performance optimization
  • Small model integration
  • Publication preparation

Q4 2026:

  • Release candidate
  • Documentation finalization
  • Community feedback integration

Resource Requirements

Development

  • Java 17+ JDK
  • Minecraft Forge 1.20.1
  • LLM API keys (OpenAI, Groq, Gemini)
  • Optional: Ollama for local inference

Testing

  • JUnit 5
  • Mockito
  • Minecraft test framework

Documentation

  • Markdown editor
  • Diagram tools (Mermaid, PlantUML)

Next Actions

Immediate (This Session):

  1. ✅ Create SkillComposer, ComposedSkill, CompositionStep
  2. Add tests for SkillComposer
  3. Push improvements to repo

Short-term (Next Session):

  1. Complete multi-agent coordination bidding
  2. Run and fix Checkstyle warnings
  3. Run and fix SpotBugs issues

Medium-term (This Week):

  1. Implement Script DSL grammar
  2. Create ScriptDSLParser
  3. Integrate with LLM generation

Last Updated: 2026-03-03 Next Review: After Phase 1 completion