context window defs

backend/agentpress/context_manager.py

@@ -11,6 +11,7 @@ from typing import List, Dict, Any, Optional, Union
 from litellm.utils import token_counter
 from services.supabase import DBConnection
 from utils.logger import logger
+from utils.constants import get_model_context_window
 
 DEFAULT_TOKEN_THRESHOLD = 120000
 
@@ -217,17 +218,22 @@ class ContextManager:
             token_threshold: Token threshold for individual message compression (must be a power of 2)
             max_iterations: Maximum number of compression iterations
         """
-        # Set model-specific token limits
-        if 'sonnet' in llm_model.lower():
-            max_tokens = 200 * 1000 - 64000 - 28000
-        elif 'gpt' in llm_model.lower():
-            max_tokens = 128 * 1000 - 28000
-        elif 'gemini' in llm_model.lower():
-            max_tokens = 1000 * 1000 - 300000
-        elif 'deepseek' in llm_model.lower():
-            max_tokens = 128 * 1000 - 28000
-        else:
-            max_tokens = 41 * 1000 - 10000
+        # Get model-specific token limits from constants
+        context_window = get_model_context_window(llm_model)
+        
+        # Reserve tokens for output generation and safety margin
+        if context_window >= 1_000_000:  # Very large context models (Gemini)
+            max_tokens = context_window - 300_000  # Large safety margin for huge contexts
+        elif context_window >= 400_000:  # Large context models (GPT-5)
+            max_tokens = context_window - 64_000  # Reserve for output + margin
+        elif context_window >= 200_000:  # Medium context models (Claude Sonnet)
+            max_tokens = context_window - 32_000  # Reserve for output + margin
+        elif context_window >= 100_000:  # Standard large context models
+            max_tokens = context_window - 16_000  # Reserve for output + margin
+        else:  # Smaller context models
+            max_tokens = context_window - 8_000   # Reserve for output + margin
+        
+        logger.debug(f"Model {llm_model}: context_window={context_window}, effective_limit={max_tokens}")
 
         result = messages
         result = self.remove_meta_messages(result)

backend/test_context_debug.py

@@ -0,0 +1,64 @@
+#!/usr/bin/env python3
+"""
+Test script to demonstrate the context manager debug functionality.
+"""
+
+import sys
+import os
+sys.path.append(os.path.dirname(os.path.abspath(__file__)))
+
+from agentpress.context_manager import ContextManager
+
+def test_context_compression():
+    """Test the context compression with debug output."""
+    
+    # Create sample messages that will trigger compression
+    sample_messages = [
+        {
+            "role": "system",
+            "content": "You are a helpful AI assistant.",
+            "message_id": "msg_001"
+        },
+        {
+            "role": "user", 
+            "content": "Hello, can you help me with a complex task that involves analyzing a large dataset and generating comprehensive reports? " * 100,  # Make it long
+            "message_id": "msg_002"
+        },
+        {
+            "role": "assistant",
+            "content": "Certainly! I'd be happy to help you with your dataset analysis and report generation. Let me break this down into manageable steps... " * 200,  # Make it very long
+            "message_id": "msg_003"
+        },
+        {
+            "role": "user",
+            "content": "Great! Here's my dataset: " + "x" * 10000,  # Large content
+            "message_id": "msg_004"
+        }
+    ]
+    
+    # Create context manager
+    cm = ContextManager()
+    
+    # Test compression with debug
+    print("Testing context compression with debug output...")
+    print(f"Original messages count: {len(sample_messages)}")
+    
+    compressed = cm.compress_messages(
+        messages=sample_messages,
+        llm_model="gpt-4",
+        thread_id="test_thread_123",
+        agent_run_id="test_run_456"
+    )
+    
+    print(f"Compressed messages count: {len(compressed)}")
+    print(f"Debug files saved to: {cm.debug_dir}")
+    
+    # List the debug files
+    import glob
+    debug_files = glob.glob(os.path.join(cm.debug_dir, "*.json"))
+    print(f"Debug files created: {len(debug_files)}")
+    for file in debug_files[-3:]:  # Show last 3 files
+        print(f"  - {os.path.basename(file)}")
+
+if __name__ == "__main__":
+    test_context_compression()

backend/test_model_limits.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python3
+"""
+Test script to verify model context window limits.
+"""
+
+import sys
+import os
+sys.path.append(os.path.dirname(os.path.abspath(__file__)))
+
+from utils.constants import get_model_context_window, MODEL_CONTEXT_WINDOWS
+
+def test_model_limits():
+    """Test the model context window limits."""
+    
+    print("=== All Model Context Windows ===")
+    for model, window in sorted(MODEL_CONTEXT_WINDOWS.items()):
+        print(f'{model}: {window:,} tokens')
+
+    print("\n=== Testing get_model_context_window function ===")
+    test_models = [
+        'gpt-5', 
+        'sonnet-3.5', 
+        'gemini-2.5-pro', 
+        'claude-sonnet-4', 
+        'grok-4', 
+        'unknown-model',
+        'anthropic/claude-sonnet-4-20250514',
+        'openai/gpt-5-mini'
+    ]
+    
+    for model in test_models:
+        window = get_model_context_window(model)
+        print(f'{model}: {window:,} tokens')
+
+    print("\n=== Context Manager Logic Simulation ===")
+    for model in ['gpt-5', 'anthropic/claude-sonnet-4', 'gemini/gemini-2.5-pro', 'unknown-model']:
+        context_window = get_model_context_window(model)
+        
+        # Simulate the logic from context manager
+        if context_window >= 1_000_000:  # Very large context models (Gemini)
+            max_tokens = context_window - 300_000
+        elif context_window >= 400_000:  # Large context models (GPT-5)
+            max_tokens = context_window - 64_000
+        elif context_window >= 200_000:  # Medium context models (Claude Sonnet)
+            max_tokens = context_window - 32_000
+        elif context_window >= 100_000:  # Standard large context models
+            max_tokens = context_window - 16_000
+        else:  # Smaller context models
+            max_tokens = context_window - 8_000
+        
+        print(f'{model}: context={context_window:,} → effective_limit={max_tokens:,} (reserved: {context_window-max_tokens:,})')
+
+if __name__ == "__main__":
+    test_model_limits()

backend/utils/constants.py

@@ -8,6 +8,7 @@ MODELS = {
             "input_cost_per_million_tokens": 3.00,
             "output_cost_per_million_tokens": 15.00
         },
+        "context_window": 200_000,  # 200k tokens
         "tier_availability": ["paid"]
     },
     # "openrouter/deepseek/deepseek-chat": {
@@ -48,6 +49,7 @@ MODELS = {
             "input_cost_per_million_tokens": 1.00,
             "output_cost_per_million_tokens": 3.00
         },
+        "context_window": 200_000,  # 200k tokens
         "tier_availability": ["free", "paid"]
     },
     "xai/grok-4": {
@@ -56,6 +58,7 @@ MODELS = {
             "input_cost_per_million_tokens": 5.00,
             "output_cost_per_million_tokens": 15.00
         },
+        "context_window": 128_000,  # 128k tokens
         "tier_availability": ["paid"]
     },
     
@@ -66,6 +69,7 @@ MODELS = {
             "input_cost_per_million_tokens": 1.25,
             "output_cost_per_million_tokens": 10.00
         },
+        "context_window": 2_000_000,  # 2M tokens
         "tier_availability": ["paid"]
     },
     # "openai/gpt-4o": {
@@ -90,6 +94,7 @@ MODELS = {
             "input_cost_per_million_tokens": 1.25,
             "output_cost_per_million_tokens": 10.00
         },
+        "context_window": 400_000,  # 400k tokens
         "tier_availability": ["paid"]
     },
     "openai/gpt-5-mini": {
@@ -98,6 +103,7 @@ MODELS = {
             "input_cost_per_million_tokens": 0.25,
             "output_cost_per_million_tokens": 2.00
         },
+        "context_window": 400_000,  # 400k tokens
         "tier_availability": ["paid"]
     },
     # "openai/gpt-4.1-mini": {
@@ -114,6 +120,7 @@ MODELS = {
             "input_cost_per_million_tokens": 3.00,
             "output_cost_per_million_tokens": 15.00
         },
+        "context_window": 200_000,  # 200k tokens
         "tier_availability": ["paid"]
     },
     "anthropic/claude-3-5-sonnet-latest": {
@@ -122,6 +129,7 @@ MODELS = {
             "input_cost_per_million_tokens": 3.00,
             "output_cost_per_million_tokens": 15.00
         },
+        "context_window": 200_000,  # 200k tokens
         "tier_availability": ["paid"]
     },   
 }
@@ -140,6 +148,9 @@ def _generate_model_structures():
     # Generate pricing
     pricing = {}
     
+    # Generate context window limits
+    context_windows = {}
+    
     for model_name, config in MODELS.items():
         # Add to tier lists
         if "free" in config["tier_availability"]:
@@ -154,29 +165,43 @@ def _generate_model_structures():
         # Add pricing
         pricing[model_name] = config["pricing"]
         
-        # Also add pricing for legacy model name variations
+        # Add context window limits
+        if "context_window" in config:
+            context_windows[model_name] = config["context_window"]
+        
+        # Also add pricing and context windows for legacy model name variations
         if model_name.startswith("openrouter/deepseek/"):
             legacy_name = model_name.replace("openrouter/", "")
             pricing[legacy_name] = config["pricing"]
+            if "context_window" in config:
+                context_windows[legacy_name] = config["context_window"]
         elif model_name.startswith("openrouter/qwen/"):
             legacy_name = model_name.replace("openrouter/", "")
             pricing[legacy_name] = config["pricing"]
+            if "context_window" in config:
+                context_windows[legacy_name] = config["context_window"]
         elif model_name.startswith("gemini/"):
             legacy_name = model_name.replace("gemini/", "")
             pricing[legacy_name] = config["pricing"]
+            if "context_window" in config:
+                context_windows[legacy_name] = config["context_window"]
         elif model_name.startswith("anthropic/"):
             # Add anthropic/claude-sonnet-4 alias for claude-sonnet-4-20250514
             if "claude-sonnet-4-20250514" in model_name:
                 pricing["anthropic/claude-sonnet-4"] = config["pricing"]
+                if "context_window" in config:
+                    context_windows["anthropic/claude-sonnet-4"] = config["context_window"]
         elif model_name.startswith("xai/"):
             # Add pricing for OpenRouter x-ai models
             openrouter_name = model_name.replace("xai/", "openrouter/x-ai/")
             pricing[openrouter_name] = config["pricing"]
+            if "context_window" in config:
+                context_windows[openrouter_name] = config["context_window"]
     
-    return free_models, paid_models, aliases, pricing
+    return free_models, paid_models, aliases, pricing, context_windows
 
 # Generate all structures
-FREE_TIER_MODELS, PAID_TIER_MODELS, MODEL_NAME_ALIASES, HARDCODED_MODEL_PRICES = _generate_model_structures()
+FREE_TIER_MODELS, PAID_TIER_MODELS, MODEL_NAME_ALIASES, HARDCODED_MODEL_PRICES, MODEL_CONTEXT_WINDOWS = _generate_model_structures()
 
 MODEL_ACCESS_TIERS = {
     "free": FREE_TIER_MODELS,
@@ -191,3 +216,40 @@ MODEL_ACCESS_TIERS = {
     "tier_6_42_yearly_commitment": PAID_TIER_MODELS,
     "tier_12_84_yearly_commitment": PAID_TIER_MODELS,
 }
+
+def get_model_context_window(model_name: str, default: int = 31_000) -> int:
+    """
+    Get the context window size for a given model.
+    
+    Args:
+        model_name: The model name or alias
+        default: Default context window if model not found
+        
+    Returns:
+        Context window size in tokens
+    """
+    # Check direct model name first
+    if model_name in MODEL_CONTEXT_WINDOWS:
+        return MODEL_CONTEXT_WINDOWS[model_name]
+    
+    # Check if it's an alias
+    if model_name in MODEL_NAME_ALIASES:
+        canonical_name = MODEL_NAME_ALIASES[model_name]
+        if canonical_name in MODEL_CONTEXT_WINDOWS:
+            return MODEL_CONTEXT_WINDOWS[canonical_name]
+    
+    # Fallback patterns for common model naming variations
+    if 'sonnet' in model_name.lower():
+        return 200_000  # Claude Sonnet models
+    elif 'gpt-5' in model_name.lower():
+        return 400_000  # GPT-5 models
+    elif 'gemini' in model_name.lower():
+        return 2_000_000  # Gemini models
+    elif 'grok' in model_name.lower():
+        return 128_000  # Grok models
+    elif 'gpt' in model_name.lower():
+        return 128_000  # GPT-4 and variants
+    elif 'deepseek' in model_name.lower():
+        return 128_000  # DeepSeek models
+    
+    return default