feat: dual model support for cost optimization

[aryasaatvik]

Summary

This PR implements dual model support in open, allowing users to configure separate main and lightweight models for different tasks. The lightweight model is intended for quick, low-cost operations like window title generation and status verb generation.

Key Features

• Dual Model Configuration: Added lightweight_model configuration field that accepts any provider/model combination
• Smart Cost-Based Selection: Automatically selects models with cost.output <= 4 as lightweight options when not explicitly configured
• Cross-Provider Support: Users can mix providers (e.g., Anthropic for main model, OpenAI for lightweight tasks)
• Two-Pane Model Selector: Completely redesigned model selection UI with separate panes for main and lightweight models
• Visual Indicators: Lightning emoji (⚡) marks lightweight models in the UI
• State Persistence: Both model selections are saved and restored between sessions

Implementation Details

Configuration Layer

• Added lightweight_model field to config schema as optional string
• Supports format: provider/model (e.g., openai/gpt-4o-mini)

Provider Layer

• New getLightweightModel() function with intelligent fallback logic:
  1. First checks user configuration
  2. Falls back to cheapest model meeting cost threshold (cost.output <= 4)
  3. Returns null if no suitable model found

TUI Layer

• two pane model selection dialog:
  • Two independent panes for main and lightweight model selection
  • Tab key switches between panes
  • Arrow keys navigate within each pane
  • Separate scroll state for each pane
  • Shows provider name in header for clarity
• updated state management to track four values:
  • MainProvider / MainModel
  • LightProvider / LightModel
• model info display below textarea shows both models

Impact on Related PRs

This PR provides the foundation for cost-optimized model selection that will benefit:

• PR feat: add dynamic status verbs #269 (Dynamic Status Verbs): Can use getLightweightModel() instead of hardcoded models
• PR feat: add dynamic window titles based on user messages #275 (Dynamic Window Titles): Can use getLightweightModel() instead of hardcoded models

Both PRs currently hardcode specific models (haiku for Anthropic, 4o-mini for OpenAI). With this change, they can dynamically use the configured lightweight model, providing better flexibility and cost optimization.

Breaking Changes

None. The feature is fully backward compatible - if no lightweight model is configured, the system continues to use the main model for all operations.

Future Enhancements

• Provider-specific smart defaults (planned but not implemented)
• Make cost threshold configurable (currently hardcoded to 4)
• Add metrics to track lightweight vs main model usage

Demo

CleanShot.2025-06-21.at.15.36.20.mp4