fix: flush LiteLLM cache when credentials change on refresh

[daniel-lxs]

Problem

When changing the LiteLLM API key or base URL in settings and clicking 'Refresh Models', the model list wasn't updating. Users had to manually delete cache directories or restart VS Code to see the updated models.

Root Cause

The caching system uses only the provider name as the cache key. When credentials change, cached models for the old credentials remain and are served instead of fetching new models with the updated credentials.

Solution

This PR modifies the requestRouterModels handler to detect when explicit LiteLLM credentials are provided via the 'Refresh Models' button and flushes the cache before fetching.

The fix distinguishes between two scenarios:

1. Explicit 'Refresh Models' button click - Includes credentials in message.values, cache is flushed
2. Automatic/debounced refresh - No credentials, uses cached models for better performance

This approach is consistent with how Ollama and LMStudio handle explicit model refreshes.

Changes

• Modified webviewMessageHandler.ts to flush LiteLLM cache when credentials are provided in the message
• Added comprehensive tests to verify cache flushing behavior

Testing

• ✅ All 5 router models tests pass
• ✅ All 20 existing LiteLLM tests pass
• ✅ All 17 webview message handler tests pass
• ✅ Manual testing: changing API key and clicking 'Refresh Models' now shows updated model list

Impact

• Users can now change their LiteLLM API key or base URL and immediately see the updated model list
• No need to manually delete cache directories or restart VS Code
• Automatic refreshes still benefit from caching for faster UI loading

---

Important

Flush LiteLLM cache in webviewMessageHandler.ts when credentials change, ensuring updated models are fetched immediately.

• Behavior:
  • In webviewMessageHandler.ts, flush LiteLLM cache when new credentials are provided in requestRouterModels message.
  • Distinguishes between explicit refresh (flush cache) and automatic refresh (use cache).
• Testing:
  • Added tests in webviewMessageHandler.routerModels.spec.ts to verify cache flushing when credentials change.
  • Ensures cache is not flushed when using stored credentials.
• Impact:
  • Users see updated LiteLLM models immediately after changing API key or base URL.
  • No need to manually clear cache or restart VS Code.

^{This description was created by for f0b1d13. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task on Roo Cloud

Review completed - no issues found.

The implementation correctly solves the LiteLLM cache refresh problem by detecting when explicit credentials are provided via the 'Refresh Models' button and flushing the cache before fetching. The approach is consistent with how Ollama and LMStudio handle explicit refreshes, and the tests comprehensively verify both the explicit refresh and stored credentials scenarios.

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}