v1.0.3 - Enhanced Reliability & User Experience
🎯 Major Improvements
Auto-Launch Claude Code
- claude-* aliases now automatically launch Claude Code after starting the backend, matching the codex-* workflow
- No more manual claudel command needed - just run claude-local-glm-9b and you're ready to code!
Robust Model Download Handling
- Added comprehensive model validation for sharded models (checks all parts are present)
- Detects incomplete downloads and automatically cleans up corrupted model caches
- Waits for models to be fully loaded into memory before proceeding
- Prevents launching CLI tools with unready backends
Better Process Management
- Ctrl-C now properly cleans up background processes (no more orphaned downloads)
- PID files saved immediately to enable proper cleanup on interruption
- Improved child process termination (kills download subprocesses)
📚 Documentation
- Added ASCII architecture diagram to README for quick understanding
- Created comprehensive CLAUDE.md with setup instructions, architecture overview, and troubleshooting
- Improved Python version compatibility checks in setup
🐛 Bug Fixes
- Fixed default Codex command to avoid unintended profile overrides
- Reduced verbose output in status.sh for cleaner display
- Better alignment between codex-models and claude-models output
🔧 Technical Details
Validation Improvements:
- Checks model.safetensors.index.json for sharded models
- Verifies all weight files are present and non-empty
- Validates tokenizer files exist
- Tests actual model functionality before proceeding
Signal Handling:
- Trap SIGINT/SIGTERM for graceful cleanup
- Kills both parent and child processes
- Removes PID files on exit
Full Changelog: 1.0.2...1.0.3