feat: Add dynamic multi-model support for OpenAI-compatible APIs

[benzntech]

TLDR

This PR adds dynamic multi-model support to Qwen Code, allowing users to fetch and switch between models from OpenAI-compatible API endpoints at runtime without restarting. It removes hardcoded model configuration and adds a new /model command that dynamically discovers available models from any OpenAI-compatible service (LocalAI, Ollama, LM Studio, OpenRouter, Azure OpenAI, etc.). Selected models are now persisted to settings.json so users' latest model choice is remembered across sessions.

Dive Deeper

Problem:
Previously, users could only use a single hardcoded model via environment variable (OPENAI_MODEL). This prevented:

• Runtime model switching without restarting
• Discovery of available models from the API
• Leveraging model aggregators (like OpenRouter) that provide 100+ models
• Seamless switching between different models for experimentation
• Remembering the selected model across sessions

Solution:
Implemented a new ModelsService that fetches models from OpenAI-compatible endpoints and integrated it with the CLI's model command system. Users can now:

1. Configure API endpoint via /auth command (no model selection needed)
2. Use /model command to dynamically fetch available models
3. Interactively select a model
4. Model selection is persisted to settings.json
5. Switch models at runtime without restarting
6. Selected model is remembered across CLI sessions

Implementation Details:

New Service: ModelsService (packages/cli/src/services/ModelsService.ts)

• Fetches models from {OPENAI_BASE_URL}/v1/models standard endpoint
• Timeout protection: 5 seconds default (prevents hanging on unresponsive endpoints)
• Comprehensive error handling for network failures
• Returns models in standard OpenAI API format
• Caches results to avoid repeated API calls

Updated Components:

1. modelCommand.ts - Enhanced to fetch and display available models dynamically
2. ModelDialog.tsx - Added loading states, error handling, and model persistence to settings.json
3. OpenAIKeyPrompt.tsx - Simplified to remove hardcoded model field
4. useAuth.ts - Updated auth flow to work without model selection

Model Persistence:
When users select a model via /model command:

• Model is updated in-memory via config.setModel(model)
• Selected model is persisted to ~/.qwen/settings.json via settings.setValue('model.name', model)
• On next CLI startup, the last selected model is automatically loaded from settings
• Consistent with how other settings (theme, approval mode, editor) are persisted

Workflow Example:

# Setup (no model needed, just endpoint and key)
export OPENAI_BASE_URL='http://localhost:8317'
export OPENAI_API_KEY='your-key'

# Run qwen
qwen

# At runtime, switch models
/model
# Shows: Loading models...
# Then displays: [1] gpt-4, [2] gpt-3.5-turbo, [3] claude-3-opus, etc.
# User selects one, model is now active for the session AND saved to settings.json

# Close and restart qwen
qwen
# Model automatically loads the previously selected model from settings.json

Supported APIs:

• ✅ LocalAI (local server)
• ✅ Ollama (local model runner)
• ✅ LM Studio (local model server)
• ✅ OpenRouter (cloud model aggregator)
• ✅ Azure OpenAI (Microsoft's implementation)
• ✅ Any OpenAI-compatible service

Reviewer Test Plan

1. Build and test:

   npm run build
   npm run test --workspace=packages/cli
   npm run test --workspace=packages/core

   Expected: All tests pass ✅

2. Test with local OpenAI-compatible service:

   # Start a local model server (e.g., Ollama, LocalAI, or LM Studio)
   export OPENAI_BASE_URL='http://localhost:8000'
   export OPENAI_API_KEY='dummy-key'
   
   qwen
   # Type: /model
   # Verify models are fetched and displayed
   # Select a model and verify it's now active

3. Test model persistence across sessions:

   # Session 1: Select a model
   qwen
   /model
   # Select 'gpt-4' (or any model)
   # Quit with /quit or Ctrl+C
   
   # Session 2: Verify persisted model
   qwen
   # Check status bar - should show the previously selected 'gpt-4'
   # Verify /model still shows the last selected model as active

4. Test error handling:

   • Set OPENAI_BASE_URL to invalid endpoint
   • Run /model
   • Expected: Clear error message instead of hanging

5. Test timeout protection:

   • Set OPENAI_BASE_URL to slow/unresponsive endpoint
   • Run /model
   • Expected: Times out after 5 seconds with helpful message

6. Test backward compatibility:

   • Users with existing OPENAI_MODEL environment variable should still work
   • Test with OpenRouter, Azure OpenAI, standard OpenAI API
   • Expected: All continue to work without issues

7. Test UI/UX:

   • Launch CLI with /auth command
   • Verify model field is no longer requested
   • Verify /model command appears in help (/help)
   • Verify loading indicator appears while fetching models
   • Verify model persistence for the current session and across restarts

Testing Matrix

	🍏	🪟	🐧
npm run	✅	❓	❓
npx	❓	❓	❓
Docker	❓	❓	❓
Podman	❓	-	-
Seatbelt	❓	-	-