python-sdk/docs/GPT5_ORCHESTRATION.md

GPT-5/Codex Orchestration Guide

Overview

Hanzo Dev supports using GPT-5, GPT-4, Codex, and other advanced models as orchestrators for AI-powered code review, development, and system improvements.

Available Orchestrator Models

Model	Best For	Context Window	Cost
gpt-5	Complex reasoning, architecture	128K	High
gpt-4o	Code review, optimization	128K	Medium
gpt-4-turbo	Fast iterations	128K	Medium
gpt-4	General development	32K	Medium
codex	Code generation, completion	8K	Low
o3	Advanced reasoning	200K	High
claude-3-5-sonnet	Long context, analysis	200K	Medium
local:llama3.2	Simple tasks (free!)	8K	Free

Quick Start

1. Basic GPT-5 Orchestration

# Use GPT-5 as the main orchestrator
hanzo dev --orchestrator gpt-5

2. GPT-4o for Code Review

# Optimized GPT-4 for code review
hanzo dev --orchestrator gpt-4o --instances 3 --critic-instances 2

3. Cost-Optimized Hybrid Mode

# GPT-5 orchestrator + local workers (90% cost reduction)
hanzo dev --orchestrator gpt-5 --use-hanzo-net

This configuration:

• Uses GPT-5 only for high-level orchestration
• Deploys local models for simple tasks
• Routes complex tasks to API models when needed
• Reduces costs by 90%

Advanced Usage

Comprehensive Code Review

hanzo dev --orchestrator gpt-4o \
          --instances 3 \
          --critic-instances 2 \
          --enable-guardrails \
          --workspace /path/to/project

Then provide review instructions:

Please review the codebase for:
1. Security vulnerabilities
2. Performance bottlenecks
3. Architecture issues
4. Code quality problems
5. Best practice violations

Generate a detailed report with fixes.

Multi-Model Orchestration

You can combine different models for specialized tasks:

# In your Python code
from hanzo.dev import HanzoDevOrchestrator

orchestrator = HanzoDevOrchestrator(
    orchestrator_model="gpt-5",      # High-level planning
    worker_models=[
        "gpt-4o",                     # Code generation
        "codex",                      # Code completion
        "local:llama3.2"              # Simple tasks
    ],
    critic_models=[
        "claude-3-5-sonnet",          # Deep analysis
        "gpt-4-turbo"                 # Fast validation
    ]
)

Cost Optimization Strategies

1. Tiered Routing

The system automatically routes tasks based on complexity:

Task Type	Routed To	Examples
Simple	Local models	Formatting, linting, simple checks
Medium	GPT-4-turbo	Refactoring, documentation
Complex	GPT-4o/GPT-5	Architecture, security analysis
Critical	GPT-5/O3	System design, complex debugging

2. Batch Processing

Group related tasks for efficient processing:

# Process multiple files in parallel
hanzo dev --orchestrator gpt-4o \
          --batch-mode \
          --files "src/**/*.py"

3. Caching and Reuse

The system caches:

• Common code patterns
• Previous review results
• Frequent queries

Model-Specific Features

GPT-5 Features

• Advanced reasoning chains
• Multi-step planning
• Cross-codebase analysis
• Architectural recommendations

GPT-4o Features

• Optimized for code tasks
• Fast response times
• Excellent at refactoring
• Strong type inference

Codex Features

• Code completion
• Docstring generation
• Test generation
• Code translation

O3 Features

• Complex problem solving
• Mathematical proofs
• Algorithm optimization
• Security analysis

Integration with MCP Tools

All orchestrators can use MCP tools:

# Enable all MCP tools for the orchestrator
hanzo dev --orchestrator gpt-5 --mcp-tools

Available tools:

• File operations
• Code search
• Git operations
• Browser automation
• Database queries
• API calls

Monitoring and Debugging

View Orchestrator Decisions

# Enable verbose logging
hanzo dev --orchestrator gpt-4o --verbose

# Monitor in real-time
hanzo dev --orchestrator gpt-5 --monitor

Performance Metrics

The system tracks:

• Token usage per model
• Response times
• Task success rates
• Cost per operation

View metrics:

hanzo metrics --orchestrator gpt-5

Best Practices

1. Start with GPT-4o: Good balance of performance and cost
2. Use local models for simple tasks: Free and fast
3. Reserve GPT-5 for complex problems: Maximum capability when needed
4. Enable critic agents: Catch issues early
5. Use guardrails: Prevent code degradation
6. Monitor costs: Track usage with hanzo metrics

Example Workflows

Security Audit

hanzo dev --orchestrator gpt-5 \
          --focus security \
          --workspace . \
          --output security-report.md

Performance Optimization

hanzo dev --orchestrator gpt-4o \
          --focus performance \
          --profile \
          --suggest-optimizations

Architecture Review

hanzo dev --orchestrator o3 \
          --focus architecture \
          --generate-diagrams \
          --suggest-patterns

Environment Variables

# Required for OpenAI models
export OPENAI_API_KEY=sk-...

# Optional: Custom endpoints
export OPENAI_API_BASE=https://api.openai.com/v1
export GPT5_ENDPOINT=https://api.openai.com/v1  # When available

# Cost limits
export HANZO_DAILY_LIMIT=100  # USD
export HANZO_HOURLY_LIMIT=10  # USD

Troubleshooting

Model Not Available

If GPT-5 is not yet available:

# Fallback to GPT-4o
hanzo dev --orchestrator gpt-4o

Rate Limits

Handle rate limits with:

# Add retry logic and backoff
hanzo dev --orchestrator gpt-4o \
          --retry-on-rate-limit \
          --max-retries 3

High Costs

Reduce costs with:

# Use local models for most tasks
hanzo dev --orchestrator gpt-4o \
          --use-hanzo-net \
          --prefer-local

Future Features

• GPT-5 Vision: Code review from screenshots
• Voice Control: Natural language commands
• Auto-fix: Automatic issue resolution
• Continuous Monitoring: 24/7 code quality checks
• Team Collaboration: Multiple orchestrators working together

Support

For issues or questions:

• GitHub: https://github.com/hanzoai/python-sdk
• Discord: https://discord.gg/CJCyAsm9Vr
• Email: [email protected]