Intelligent Context Condensing (Experimental)

The Intelligent Context Condensing feature helps manage long conversations by summarizing earlier parts of the dialogue. This prevents important information from being lost when the context window nears its limit. This is an experimental feature and is disabled by default.

How It Works

As your conversation with Roo Code grows, it might approach the context window limit of the underlying AI model. When this happens, older messages would typically be removed to make space. Intelligent Context Condensing aims to prevent this abrupt loss by:

1. Summarizing: Using an AI model, it condenses earlier parts of the conversation.
2. Retaining Essentials: The goal is to reduce the overall token count while keeping the key information from the summarized messages.
3. Maintaining Flow: This allows the AI to have a more coherent understanding of the entire conversation, even very long ones.

Important Considerations:

• Summarization Impact: While original messages are preserved if you use Checkpoints to rewind, the summarized version is what's used in ongoing LLM calls to keep the context manageable.
• Cost: The AI call to perform the summarization incurs a cost. This cost is included in the context condensing metrics displayed in the UI.

Enabling This Feature

As an experimental feature, Intelligent Context Condensing is disabled by default.

1. Open Roo Code settings ( icon in the top right corner of the Roo Code panel).
2. Navigate to the "Experimental" section.
3. Toggle the "Automatically trigger intelligent context condensing" (autoCondenseContext) option to enable it.
4. Optionally, adjust the "Threshold to trigger intelligent context condensing" (autoCondenseContextPercent) slider to control the trigger point for automatic context condensing.
5. Save your changes.

The image above shows settings for Intelligent Context Condensing: the toggle to "Automatically trigger intelligent context condensing" and the "Threshold to trigger intelligent context condensing" slider.

Controlling and Understanding Context Condensing

Roo Code provides several ways to control and understand the Intelligent Context Condensing feature:

Controlling Context Condensing

• Automatic Threshold: In Roo Code Settings () > "Experimental," the autoCondenseContextPercent setting allows you to define a percentage (e.g., 80%). Roo Code will attempt to condense the context automatically when the conversation reaches this level of the context window's capacity.

• Manual Trigger: A Condense Context button ( icon) is available when a task is expanded, typically located at the bottom of the task view, next to other task action icons like the trash can. This allows you to initiate the context condensing process at any time.

  

  The Manual Condense Context button (highlighted with a yellow arrow) appears in the expanded task view.

Understanding Context Condensing Activity

• Context Condensing Metrics: When context condensing occurs, Roo Code displays:
  • The context token counts before and after context condensing.
  • The cost associated with the context condensing AI call.
  • An expandable summary detailing what was condensed (this information is part of the ContextCondenseRow component visible in the chat history).

After context condensing, a message indicates the context has been condensed, showing token changes and cost.

• Visual Indicators:
  • A progress indicator ("Condensing context...") is shown in the chat interface while context condensing is active.

The "Condensing context..." indicator appears in the chat during the process.

• The task header also displays the current context condensing status.
• The ContextWindowProgress bar offers a visual representation of token distribution, including current usage, space reserved for the AI's output, available space, and raw token numbers.
• Interface Clarity: The "Condense Context" button includes a tooltip explaining its function, available in all supported languages. The icon for context condensing-related actions is codicon-compress.

Accurate Token Information

• Roo Code employs accurate token counting methods, with some AI providers utilizing their native token counting endpoints. This ensures that context size and associated costs are calculated reliably.

Internationalization

• All user interface elements for this feature, such as button labels, tooltips, status messages, and settings descriptions, are available in multiple supported languages.

Technical Implementation

Token Counting

Roo Code uses a sophisticated token counting system that:

• Employs native token counting endpoints when available (e.g., Anthropic's API)
• Falls back to tiktoken estimation if API calls fail
• Provides accurate counting for different content types:
  • Text content: Uses word-based estimation with punctuation and newline overhead
  • Image content: Uses a conservative estimate of 300 tokens per image
  • System prompts: Includes additional overhead for structural elements

Context Window Management

• By default, 30% of the context window is reserved (20% for model output and 10% as a safety buffer), leaving 70% available for conversation history.
• This reservation can be overridden by model-specific settings
• The system automatically calculates available space while maintaining this reservation

Performance Considerations

Optimization

• The system optimizes token counting to minimize performance impact
• Token calculations are cached where possible
• Background processing prevents UI blocking during context condensing

Resource Usage

• Context condensing operations are performed asynchronously
• The UI remains responsive during the process
• System resources are managed to prevent excessive memory usage

Feedback

Your experience with experimental features is valuable. When reporting issues, please include:

• The current threshold setting
• The token counts before and after context condensing
• Any error messages displayed
• Steps to reproduce the issue

Please report any issues or suggestions regarding Intelligent Context Condensing on the Roo Code GitHub Issues page.