Skip to main content

Context Window Manager

The ContextWindowManager v0.3.3.3.post1 handles advanced token budgeting and content trimming for conversation context with multimodal content handling and dynamic budget allocation.

Overview​

The ContextWindowManager is responsible for:

  1. Advanced Token Budgeting: Dynamic allocation based on content type and priority
  2. Image Handling: Intelligent trimming of multimodal content to preserve tokens
  3. Category-Based Management: Fine-grained control over different message types
  4. Real-time Usage Tracking: Live monitoring of token consumption
  5. Smart Trimming Strategies: Context-aware content removal with recency preservation
  6. Multi-Provider Token Counting: Automatic fallback tokenization strategies
  7. Rich Reporting: Human-readable and programmatic token usage statistics

Message Categories and Priorities​

Messages are categorized into five priority levels for intelligent trimming:

  1. SYSTEM (highest priority): System instructions and prompts - never trimmed (10% budget)
  2. CONTEXT: Important reference information, documentation, and context files (35% budget)
  3. DIALOG: Main conversation between user and assistant (50% budget)
  4. SYSTEM_OUTPUT: Tool outputs, code execution results, and action responses (5% budget)
  5. ERROR: Error messages and debugging information (fallback 5% budget)

Priority System​

  • SYSTEM messages: Guaranteed preservation with strict 10% minimum allocation
  • Chronological trimming: Within each category, oldest messages are trimmed first to preserve recency
  • Dynamic budget management: Automatic budget redistribution when categories are under-utilized
  • Image-aware processing: Special handling for multimodal content with token-intensive elements
  • Fallback categories: Automatic handling of new message types with default budgets

Trimming Strategy​

Special Content Handling​

Initialization​

def __init__(
    self,
    model_config=None,
    token_counter: Optional[Callable[[Any], int]] = None,
    api_client=None,
    config_obj: Optional[Any] = None,
)

Smart Configuration Resolution​

The ContextWindowManager automatically resolves configuration from multiple sources in priority order:

  1. Model Configuration: model_config.max_tokens (if provided)
  2. Live Config Object: config_obj.model_config.max_history_tokens (preferred)
  3. Global Config: config.yml model settings
  4. Fallback: Default 150,000 tokens

Intelligent Token Counter Selection​

The system automatically selects the best available token counter:

  1. API Client Counter: api_client.count_tokens() (provider-specific)
  2. Diagnostics Counter: diagnostics.count_tokens() (tiktoken-based)
  3. Fallback Counter: Rough estimation (4 chars per token)

Features​

  • Dynamic Max Tokens: Adapts to model capabilities automatically
  • Multi-source Configuration: Supports live config updates
  • Provider-aware Tokenization: Uses provider-specific token counting when available
  • Robust Fallbacks: Graceful degradation when components are unavailable

Clamp Notices (Sub-Agents)​

When a sub-agent is created with an isolated context window and a shared_cw_max_tokens value, the child's ContextWindowManager.max_tokens is set to the lower of the parent’s max and the provided clamp. A system note with metadata type=cw_clamp_notice is recorded on both the parent and child conversations to make this visible in transcripts and dashboards. The metadata includes:

  • sub_agent: child agent id
  • child_max: effective child max tokens
  • parent_max: parent’s max tokens at time of spawn
  • clamped: true if the child was reduced relative to parent

Key Methods​

Token Management​

@property
def total_budget(self) -> int

Gets the total token budget.

@property
def available_tokens(self) -> int

Gets the number of available tokens.

def get_budget(self, category: MessageCategory) -> TokenBudget

Gets the token budget for a specific category.

def update_usage(self, category: MessageCategory, tokens: int) -> None

Updates token usage for a category.

def reset_usage(self, category: Optional[MessageCategory] = None) -> None

Resets token usage for a category or all categories.

def is_over_budget(self, category: Optional[MessageCategory] = None) -> bool

Checks if a category or the entire context is over budget.

Session Processing​

def analyze_session(self, session: Session) -> Dict[str, Any]

Analyzes a session for token usage statistics and multimodal content.

def trim_session(self, session: Session, preserve_recency: bool = True) -> Session

Trims session messages to fit within token budget.

def process_session(self, session: Session) -> Session

Processes a session through token budgeting and trimming - main entry point for integration.

Advanced Reporting and Analytics​

def get_token_usage(self) -> Dict[str, int]

Returns comprehensive token usage statistics with structure:

{
    "total": 45000,
    "available": 105000,
    "max": 150000,
    "usage_percentage": 30.0,
    "MessageCategory.SYSTEM": 5000,
    "MessageCategory.CONTEXT": 15000,
    "MessageCategory.DIALOG": 20000,
    "MessageCategory.SYSTEM_OUTPUT": 5000
}
def format_token_usage(self) -> str

Returns human-readable token usage summary with category breakdowns.

def format_token_usage_rich(self) -> str

Returns rich-formatted token usage with progress bars and color coding.

def get_current_allocations(self) -> Dict[MessageCategory, float]

Returns current token allocations as percentages for monitoring.

def get_usage(self, category: MessageCategory) -> int

Returns current token usage for a specific category.

Advanced Trimming Strategy​

The trimming strategy v0.3.3.3.post1 follows a sophisticated multi-pass approach:

Phase 1: Image Optimization​

  1. Multi-image Detection: Identify sessions with multiple images
  2. Recency Preservation: Keep the most recent image intact
  3. Token-Efficient Replacement: Replace older images with placeholders
  4. Metadata Preservation: Store original image references for context

Phase 2: Budget Analysis​

  1. Total Budget Check: Compare session tokens against model limits
  2. Category Budget Analysis: Check individual category limits
  3. Priority Queue Creation: Order categories by trim priority

Phase 3: Intelligent Trimming​

  1. Chronological Trimming: Remove oldest messages first within each category
  2. Minimum Budget Enforcement: Respect minimum token allocations
  3. SYSTEM Protection: Never trim SYSTEM messages
  4. Iterative Reduction: Continue trimming until budget is met

Phase 4: Session Reconstruction​

  1. Order Preservation: Maintain original message ordering
  2. Category Integrity: Keep messages in their logical categories
  3. Metadata Updates: Update session statistics and metadata

Image Handling​

The system includes specialized handling for multimodal content:

Image Token Estimation​

  • Claude Models: ~4,000 tokens per image (higher for safety)
  • Vision Models: Provider-specific token calculations
  • Fallback Estimation: Character-based approximation

Smart Image Trimming​

def _handle_image_trimming(self, session: Session) -> Session:
    """
    Replace all but the most recent image with placeholders.
    Preserves conversation context while saving significant tokens.
    """

Placeholder Generation​

  • Context Preservation: Maintains conversation flow
  • Reference Tracking: Stores original image metadata
  • User-Friendly Messages: Clear indication of image removal

Token Counter Fallback Chain​

The system implements a robust fallback strategy for token counting:

  1. Provider-Specific Counter: Uses native model tokenizer (optimal accuracy)
  2. LiteLLM Generic Counter: Fallback to model-specific tiktoken
  3. Diagnostics Counter: Uses tiktoken with model mapping
  4. Character Estimation: Rough fallback (4 chars = 1 token)

This ensures accurate token counting across all scenarios while maintaining performance.

Advanced Usage Examples​

Basic Setup with Configuration​

from penguin.system.context_window import ContextWindowManager
from penguin.llm.model_config import ModelConfig

# Initialize with configuration
context_window = ContextWindowManager(
    model_config=model_config,
    api_client=api_client,
    config_obj=live_config  # Supports live config updates
)

# Automatic configuration resolution from multiple sources
print(f"Max tokens: {context_window.max_tokens}")
print(f"Available tokens: {context_window.available_tokens}")

Comprehensive Session Analysis​

# Analyze session with multimodal content detection
stats = context_window.analyze_session(session)
print(f"Total tokens: {stats['total_tokens']}")
print(f"Images detected: {stats['image_count']}")
print(f"Tokens by category: {stats['per_category']}")
print(f"Over budget: {stats['over_budget']}")

# Detailed breakdown
for category, tokens in stats['per_category'].items():
    percentage = (tokens / stats['total_tokens']) * 100 if stats['total_tokens'] > 0 else 0
    print(f"  {category.name}: {tokens} tokens ({percentage:.1f}%)")

Intelligent Session Processing​

# Process session with automatic trimming and optimization
original_token_count = context_window.analyze_session(session)['total_tokens']
trimmed_session = context_window.process_session(session)
final_token_count = context_window.analyze_session(trimmed_session)['total_tokens']

print(f"Original tokens: {original_token_count}")
print(f"Final tokens: {final_token_count}")
print(f"Reduction: {original_token_count - final_token_count} tokens")

# Session integrity preserved
print(f"Messages removed: {len(session.messages) - len(trimmed_session.messages)}")
print(f"SYSTEM messages preserved: {len([m for m in trimmed_session.messages if m.category.name == 'SYSTEM'])}")

Real-time Token Monitoring​

# Monitor token usage in real-time
usage = context_window.get_token_usage()
print(context_window.format_token_usage())

# Rich formatting for CLI
if rich_available:
    print(context_window.format_token_usage_rich())

# Programmatic access
for category in MessageCategory:
    budget = context_window.get_budget(category)
    usage = context_window.get_usage(category)
    percentage = (usage / budget.max_tokens) * 100 if budget.max_tokens > 0 else 0
    print(f"{category.name}: {usage}/{budget.max_tokens} ({percentage:.1f}%)")

Advanced Category Management​

# Update usage tracking manually
context_window.update_usage(MessageCategory.CONTEXT, 5000)

# Check if any category is over budget
for category in MessageCategory:
    if context_window.is_over_budget(category):
        print(f"Warning: {category.name} is over budget!")

# Reset usage tracking
context_window.reset_usage()  # Reset all
context_window.reset_usage(MessageCategory.DIALOG)  # Reset specific category

Token Budgeting System​

The dynamic token budget allocation system v0.3.3.3.post1 includes:

Intelligent Budget Distribution​

  • SYSTEM: 10% (guaranteed preservation, strict minimum enforcement)
  • CONTEXT: 35% (documentation and reference materials)
  • DIALOG: 50% (main conversation, flexible allocation)
  • SYSTEM_OUTPUT: 5% (tool results and action outputs)
  • ERROR: 5% (fallback for error messages and debugging)

Dynamic Budget Management​

  • Automatic Redistribution: Unused budget from one category can benefit others
  • Minimum Guarantees: Each category has a minimum token allocation
  • Maximum Limits: Prevents any category from consuming excessive tokens
  • Runtime Adjustments: Budgets can be modified during session processing

Token Efficiency Features​

  • Image Optimization: Intelligent replacement of redundant images
  • Smart Trimming: Chronological removal preserving conversation coherence
  • Category-Aware Processing: Different strategies for different content types
  • Metadata Preservation: Maintains conversation context during optimization

This system ensures optimal token utilization while preserving conversation quality and system functionality.