Problem:
- Multiple instances of the Roo Code application were launching separate MCP server instances
- This led to unnecessary resource consumption and potential conflicts between instances
Solution:
1. Created new McpServerManager singleton class to manage MCP server instances:
- Static getInstance() method ensures only one McpHub instance exists
- Tracks registered ClineProvider instances
- Handles cleanup on extension deactivation
2. Modified ClineProvider class:
- Changed mcpHub from private to protected
- Added getMcpHub() public getter method
- Updated initialization to use McpServerManager
- Added unregister logic in dispose()
3. Updated extension.ts to handle cleanup:
- Added McpServerManager cleanup in deactivate()
Technical Implementation:
- Uses WeakRef for provider tracking to allow proper garbage collection
- Maintains global state to track instance IDs
- Implements proper cleanup of resources on disposal
- Ensures backward compatibility with existing code
This change significantly improves resource usage and prevents potential conflicts
between multiple instances of the application while maintaining all existing
functionality.
- Add enableMcpServerCreation setting to control whether MCP server creation is allowed
- Add UI toggle in settings view for this feature
- Update system prompt to conditionally include MCP server creation documentation
- Add tests for new functionality
Display the percentage of total context window used alongside token count in the task header. This helps users better understand their context usage relative to the model's capacity.
Implementation details:
- Retrieve model's context window size from normalized API configuration
- Calculate percentage of context window used based on current context tokens
- Display percentage in parentheses next to token count (e.g. "40,000 (20%)")
- Default to context window size of 1 if model info is unavailable
- Format numbers with commas for better readability
Adds the current conversation's context token count to the environment details section by:
- Using existing getApiMetrics() to retrieve contextTokens from the last API request
- Adding a new "Context Tokens" section between "Current Time" and "Current Mode"
- Formatting the token count with toLocaleString() for better readability
- Showing "(Not available)" when no context tokens are present
This provides visibility into the current conversation's token usage directly in the environment details, helping track context window utilization.
Change ExperimentId type to be value-based rather than key-based
Make experiment record types more strict with proper typing
Pass full experiment config object instead of single boolean flag
Update type definitions and usages across codebase
- Add experiments.ts to manage experimental features
- Refactor experimental diff strategy into experiments system
- Add UI components for managing experimental features
- Add tests for experimental tools
- Update system prompts to handle experiments
- Rename customPrompts to customModePrompts for mode-specific prompts
- Add new customSupportPrompts type for support action prompts
- Update types to be more specific (CustomModePrompts and CustomSupportPrompts)
- Fix all related tests and component implementations
- Added `updateDiffStrategy` method to dynamically adjust the diff strategy based on the current state and experimental settings.
- Updated ClineProvider to call `updateDiffStrategy` when the experimental diff strategy is modified, ensuring real-time updates in the Cline instance.
