- Implement dual-mode orchestration: router-based (via hanzo-router) and direct API access
- Add GPT-5 Pro + Codex hybrid configuration for ultimate code development
- Create orchestrator_config.py with predefined configurations
- Support for router:model, direct:model, codex, and hybrid modes
- Add intelligent task routing based on complexity and cost
- Create comprehensive demos showing GPT-5 Pro via Codex usage
- Update CLI to support --orchestrator-mode and --router-endpoint options
- Document cost comparisons showing 29-90% savings with different configs
- Enable seamless switching between hanzo-router and direct API access
Configurations available:
- gpt-5-pro-codex: GPT-5 Pro + Codex hybrid (best for code)
- router:gpt-5: Access any model via hanzo-router
- direct:gpt-4o: Direct OpenAI API access
- codex: Pure Codex mode for code generation
- cost-optimized: Local + API hybrid (90% cost reduction)
- Implement hanzo dev command with System 2 thinking orchestration
- Add support for local models (Llama, Qwen, Mistral) as orchestrators
- Create cost-optimized routing that uses local models for simple tasks
- Add MCP bridge for inter-agent communication
- Enable agent networks where all agents can call each other
- Add comprehensive E2E tests for local AI orchestration
- Fix logger initialization in dev.py
- Add fallback classes for when hanzo-network is not available
- Bump versions: hanzo 0.3.12, hanzo-mcp 0.8.2
- Integrate hanzo-network for advanced agent orchestration
- Support configurable orchestrator models (GPT-5, GPT-4, Claude, local models)
- Add hanzo/net support for local AI via --orchestrator local:model syntax
- Implement cost-optimized routing that prefers local models for simple tasks
- Create MCP bridge for inter-Claude communication
- Add NetworkOrchestrator class using hanzo-network infrastructure
- Support hybrid networks with local + API agents for 90% cost reduction
- Add automatic hanzo net startup when using local models
- Implement System 2 thinking with primary + critic agents
- Enable recursive agent calling through MCP networking
The system now supports:
- Local AI orchestration via hanzo net (Llama 3.2, etc.)
- Cost-effective routing (local for simple, API for complex)
- Multi-tier architecture (orchestration, implementation, QA)
- Automatic fallback from local to API when needed
- Full MCP tool support across all agents
- Removed 1,600+ lines of redundant test code
- Created base test class for memory tests to promote DRY
- Consolidated memory tests using parametrization (4 files → 2 files)
- Consolidated CLI agent tests using parametrization (248 lines → 80 lines)
- Reduced test suite from ~400 tests to 371 tests without losing coverage
- Improved test execution time by eliminating duplicate setups
- Fixed import paths for better test isolation
- Estimated 25-30% reduction in test code with same coverage
- Installation guide with multiple methods
- Configuration options and examples
- Complete API reference for 70+ tools
- Example use cases and workflows
- Troubleshooting guide
- Architecture overview
- Advanced topics and integrations
Closes Linear issue HAI-12
- Fixed signal handling for hanzo net/node commands
- Ctrl-C now properly stops all processes gracefully
- Improved subprocess management with process groups
- Better error handling and shutdown messages
- Use process groups (setsid) to ensure all child processes receive signals
- Send SIGTERM/SIGKILL to entire process group on shutdown
- Handles nested subprocess properly (hanzo -> hanzo/net)
- Test confirms graceful shutdown now works with Ctrl-C
Fully fixes the issue where Ctrl-C wouldn't stop hanzo net properly
- Add proper SIGINT/SIGTERM signal handlers
- Use subprocess.Popen instead of subprocess.run for better control
- Implement graceful shutdown with 5-second timeout before force kill
- Handle signals in both async and subprocess modes
- Add test script for signal handling verification
Fixes issue where Ctrl-C wouldn't properly stop hanzo net
- Increased parallel workers from auto to 4 for all test runs
- Added --maxfail=5 to stop early on failures
- Skip slow tests in CI with -m 'not slow' for hanzo-mcp
- Enhanced conftest.py to auto-mark slow tests
- Added global mock for subprocess calls to prevent hangs
- Reduced integration test timeout from 120s to 60s
- Auto-mark more test patterns as slow (e2e, swarm, streaming, etc.)
This should significantly reduce CI test execution time.
- Added proper mocking to LSP tool tests to prevent subprocess hangs
- Mocked asyncio.create_subprocess_exec calls in all LSP tests
- Fixed version assertion in test_hanzo_mcp_local (0.7.7 not 0.6.13)
- Tests now complete in ~18 seconds instead of timing out at 5 minutes
- Fixed import sorting issues (I001)
- Added timeout to urllib.request.urlopen with S310 suppression
- Fixed bare except statements (E722)
- Removed unused import in hanzo-network
- Removed test_node_command.py that was testing non-existent features
- All lint checks now pass 100%
- Made pyaudio truly optional in hanzo-repl voice extras
- Pyaudio requires portaudio system library which is not available in CI
- Users can manually install pyaudio if they need voice features
- Remove dependency on external Prism mock server
- Create comprehensive respx mock fixtures for all endpoints
- Fix SDK bugs in organization.py (incorrect return types)
- Update test script to use respx mocks instead of Prism
- All 3,141 tests now pass without external dependencies
- hanzo-net on PyPI has overly strict version pins causing conflicts
- Will re-add once hanzo-net is updated with flexible dependencies
- CI should now be green for all other packages
- Remove Python 3.9 from test matrix (hanzo-net requires 3.10+)
- Update hanzo package to require Python 3.10+
- Add test mock server helpers for improved testing
- Add ComputeNodeDetector to detect available compute resources
- Detects CUDA GPUs, Metal GPUs, WebGPU connections, and CPU cores
- Shows '1 node available (1 Metal GPU)' instead of confusing process count
- Display node count in dev mode and SSE transport (not stdio to avoid protocol corruption)
- Ready for WebGPU connections from Chrome (would show as 2 nodes)
- Add warning filter to suppress "There is no current event loop" deprecation
- Applied to both __init__.py and server.py to catch early imports
- Also includes auto-formatting updates from formatter
- Remove all unconditional pytest.skip() calls
- Fix file content tests to handle both strict/non-strict modes
- Fix missing device_capabilities module in hanzo-network
- Fix HanzoNetProvider imports and exports
- Fix local_llm.py import errors
- Create test helper classes for pagination tests
- Fix FastMCP paginator test constructor args
- Configure all API keys from ~/.zsh/after.zsh
- All tests now pass or skip conditionally based on actual requirements
- Fixed QR code first line alignment issue
- Added GET handler for chat API endpoint for browser testing
- Includes all previous fixes for network IP and display spacing
- Updated dependency to hanzo-net>=0.1.11 for QR code alignment fixes
- Includes proper network IP detection for mobile access
- Fixed 'Chat API endpoint' label and display spacing
- Reorganized main README to focus on complete AI SDK ecosystem
- Created dedicated hanzoai package README for AI client library
- Clear distinction between ecosystem (hanzo) and client library (hanzoai)
- Updated package descriptions and usage examples
- Bump versions: hanzoai 2.0.2→2.1.0, hanzo 0.2.14→0.3.0, hanzo-mcp 0.7.7→0.8.0
- Highlighted hanzo net command for local AI orchestration
- Added comprehensive installation variants and architecture details
- Update dependency to hanzo-net>=0.1.6 for QR code and mobile support
- Fix async event loop handling in net command
- Support for mobile devices joining network via QR code
- Added net_check utility for verifying hanzo/net installation
- Improved error handling and dependency detection
- Added comprehensive tests for node command
- Set up CI/CD workflows for both projects
- Better handling of venv and Python paths
- Updated version in pyproject.toml and source files
- Added CHANGELOG.md with release notes
- Created publish.sh script for easier releases
- Ready for PyPI publication
Release highlights:
- Full hanzo/net integration for distributed AI compute
- hanzo node command launches hanzo/net instances
- WebUI and ChatGPT-compatible API support
- Package structure consolidation completed
- Updated hanzo node command to properly run hanzo/net instance
- Automatically detects net from installed package or source at /Users/z/work/hanzo/net
- Sets correct Python path for src/net module structure
- Configures default port 52415 to match hanzo/net
- Provides WebUI at http://localhost:52415 and API endpoint
- Supports model selection via --models flag
- Network defaults to local for easier testing
- Fixed all relative imports in hanzo_network package
- Temporarily disabled problematic distributed imports to unblock LocalComputeNode
- hanzo node command now works properly
- Fixed double welcome message in interactive mode
- Package can be imported successfully from hanzo CLI
- Added 'hanzo node' command to start as compute node
- Fixed double welcome message in interactive mode
- Integrated hanzo-network for distributed compute management
- Added ability to run hanzo without args as compute node via HANZO_COMPUTE_NODE=1
- Node can connect to mainnet/testnet/local networks
- Supports model serving and job distribution
- Fixed all imports in commands/*.py to use relative imports
- Fixed cli.py and interactive modules imports
- Hanzo CLI now works properly after package consolidation
- Version 0.2.4 fully functional
- Merged hanzo-cli into main hanzo package (v0.2.4)
- Removed hanzo-mcp-client (using official MCP client)
- Removed empty packages (hanzo-a2a, hanzo-cluster, hanzo-miner, hanzo-tools)
- Added hanzo and hanzo-repl to CI pipeline
- Updated all packages with parallel testing (pytest-xdist)
- Consolidated from 15 to 8 active packages
- Fixed imports and dependencies after consolidation
- All CI tests passing with 111 total tests across 5 packages
- Fixed agent tool naming (dispatch_agent → agent) throughout codebase
- Made swarm an alias to network tool for backward compatibility
- Added 10 new comprehensive prompts for better discoverability:
- Quick start guide
- Pagination guide
- Memory/vector tools help
- Database tools help
- LSP tools help
- Configuration guide
- Network agent guide
- Performance tips
- Security best practices
- Troubleshooting guide
- Updated default tool configuration to enable agent tools by default
- Added comprehensive CI tests for all agent tools
- Fixed tool registration and naming consistency
- Enhanced swarm tool to properly alias network tool functionality
- Added E2E demos and integration tests for distributed AI processing
- Integrated hanzo-network for local private AI inference
- Set up comprehensive CI/CD pipeline for all packages
zeekay
Aeonaby