- Fix Claude model IDs to use dots instead of dashes (4.5 not 4-5)
- Remove non-existent GPT models (5.2-thinking, 5.2-instant)
- Add GPT-4o as reliable fallback model
- Fix DeepSeek model ID to use correct free tier variant
- Update Claude Sonnet 4.5 context limit to 1M (actual value)
- All models now use verified OpenRouter API model IDs

- Remove all tool passing to AI SDK since built-in tools from JSON don't have proper Zod validation
- Tools are now processed server-side only, avoiding zod-to-json-schema conversion errors
- This prevents 'Cannot read properties of undefined (reading "typeName")' errors
- Functionality is maintained through server-side tool processing (MCP + built-in tools)

…ndows

- Add useLocalProviders hook that detects running Ollama and LMStudio instances
- Automatically enable detected local providers in settings
- Improve Ollama provider with proper context window calculation based on model size
- Fix updateProviderSettings function signature for correct type handling
- Local models now appear in provider/model selectors when available

Closes issues with local model detection and context window limits.

Fix: #69 local model server failure