Summary

This PR fixes issue #6290 where the VS Code Language Model API was unable to correctly calculate context size when using models like "copilot - gpt-4.1". The issue was caused by VS Code updating their countTokens API to expect LanguageModelChatMessage2 instead of LanguageModelChatMessage.

Problem

• When using VS Code LM API with models like "copilot - gpt-4.1", the context size calculation was incorrect
• System prompt was accurately accounted for, but conversation content had minimal impact on context size
• The root cause was that VS Code's countTokens method now expects a different message format

Solution

1. Updated internalCountTokens method to handle both old and new API versions:

   • First attempts to use the message object directly (new API)
   • Falls back to extracting text content and counting tokens as a string if the new API fails
   • Properly handles both LanguageModelChatMessage instances and string inputs

2. Enhanced calculateTotalInputTokens method:

   • Attempts batch token counting first for better performance
   • Falls back to individual message counting if batch counting fails
   • Ensures compatibility with both API versions

3. Added comprehensive test coverage:

   • Tests for successful token counting with text content
   • Tests for handling image content with placeholders
   • Tests for API error handling and graceful fallback
   • Tests for the specific scenario where the new API format causes errors

Testing

• All existing tests pass ✅
• Added new tests specifically for the fallback behavior ✅
• Ran full test suite for all API providers to ensure no regressions ✅
• All linting and type checks pass ✅

Impact

This fix ensures that Roo Code continues to work correctly with VS Code's Language Model API regardless of which version of the API is being used, providing a seamless experience for users.

Fixes #6290