Context Management Environment Variables

Status: 🟢 In-use Category: Configuration & Prompts Date: November 5, 2025

Overview

Global context management configuration that supports provider-specific overrides via environment variables.

Environment Variables

Provider Selection

Variable	Default	Description
LLM_PROVIDER	azure-openai	LLM provider (azure-openai, aws-bedrock, google-gemini, etc.)

Provider-Specific Context Limits

Override context limits for specific providers:

Variable	Provider	Default	Description
AZURE_OPENAI_MAX_CONTEXT_TOKENS	azure-openai	100,000	Azure OpenAI (GPT-4o) context limit
OPENAI_MAX_CONTEXT_TOKENS	openai	100,000	OpenAI context limit
AWS_BEDROCK_MAX_CONTEXT_TOKENS	aws-bedrock	150,000	AWS Bedrock (Claude) context limit
ANTHROPIC_MAX_CONTEXT_TOKENS	anthropic-claude	150,000	Anthropic Claude context limit
GOOGLE_GEMINI_MAX_CONTEXT_TOKENS	google-gemini	800,000	Google Gemini context limit
GCP_VERTEXAI_MAX_CONTEXT_TOKENS	gcp-vertexai	150,000	GCP Vertex AI context limit

Global Overrides

Variable	Default	Description
MAX_CONTEXT_TOKENS	provider-specific	Override for ALL providers
MIN_MESSAGES_TO_KEEP	10	Minimum recent messages to preserve
ENABLE_AUTO_COMPRESSION	true	Enable/disable auto-trimming (LangGraph only)

Priority Order

When determining context limits, the system uses this priority:

1. Provider-specific ENV var (e.g., AWS_BEDROCK_MAX_CONTEXT_TOKENS)
2. Global override (MAX_CONTEXT_TOKENS)
3. Default for provider (from DEFAULT_PROVIDER_CONTEXT_LIMITS)
4. Fallback (100,000)

Examples

Use Provider Defaults

# docker-compose.yml
environment:
  - LLM_PROVIDER=aws-bedrock  # Uses 150K default

Override Specific Provider

environment:
  - LLM_PROVIDER=aws-bedrock
  - AWS_BEDROCK_MAX_CONTEXT_TOKENS=180000  # Increase Claude limit

Override All Providers Globally

environment:
  - MAX_CONTEXT_TOKENS=120000  # Applies to all providers

Per-Provider Configuration (Multi-Provider Setup)

# Platform Engineer (Azure OpenAI)
platform-engineer-p2p:
  environment:
    - LLM_PROVIDER=azure-openai
    - AZURE_OPENAI_MAX_CONTEXT_TOKENS=110000

# AWS Agent (Bedrock Claude)
agent-aws-p2p:
  environment:
    - LLM_PROVIDER=aws-bedrock
    - AWS_BEDROCK_MAX_CONTEXT_TOKENS=180000

# GitHub Agent (Google Gemini)
agent-github-p2p:
  environment:
    - LLM_PROVIDER=google-gemini
    - GOOGLE_GEMINI_MAX_CONTEXT_TOKENS=900000

Testing

Test configuration loading:

# Set provider-specific override
export LLM_PROVIDER=aws-bedrock
export AWS_BEDROCK_MAX_CONTEXT_TOKENS=180000

# Start agent and check logs
docker compose up agent-aws-p2p

# Expected log output:
# INFO: Using provider-specific context limit from AWS_BEDROCK_MAX_CONTEXT_TOKENS: 180,000 tokens
# INFO: Context management initialized for provider=aws-bedrock: max_tokens=180,000, ...

Code Usage

In Python Code

from ai_platform_engineering.utils.a2a_common.context_config import (
    get_context_limit_for_provider,
    get_min_messages_to_keep,
    is_auto_compression_enabled,
    get_context_config,
)

# Get limit for specific provider
limit = get_context_limit_for_provider("aws-bedrock")
print(f"AWS Bedrock limit: {limit:,} tokens")

# Get limit for current provider (from LLM_PROVIDER env)
limit = get_context_limit_for_provider()
print(f"Current provider limit: {limit:,} tokens")

# Get complete config
config = get_context_config()
print(config)
# {
#   "provider": "aws-bedrock",
#   "max_context_tokens": 150000,
#   "min_messages_to_keep": 10,
#   "auto_compression_enabled": True
# }

Benefits

1. Centralized Configuration

One source of truth for context limits across all agents (LangGraph and Strands).

2. Flexible Overrides

• Development: Lower limits for faster testing
• Production: Provider-optimized defaults
• Cost-sensitive: Custom limits per environment

3. Multi-Provider Support

Different agents can use different providers with appropriate limits:

• Platform Engineer (Azure GPT-4o): 100K
• AWS Agent (Bedrock Claude): 150K
• GitHub Agent (Gemini): 800K

4. Easy Debugging

Clear logging shows which configuration is being used:

INFO: Using provider-specific context limit from AWS_BEDROCK_MAX_CONTEXT_TOKENS: 180,000 tokens
INFO: Context management initialized for provider=aws-bedrock: max_tokens=180,000, ...

Migration

From Hardcoded Values

Before:

max_context_tokens = 100000  # Hardcoded

After:

from .context_config import get_context_limit_for_provider
max_context_tokens = get_context_limit_for_provider()

From ENV-Only Configuration

Before:

max_context_tokens = int(os.getenv("MAX_CONTEXT_TOKENS", "100000"))

After:

from .context_config import get_context_limit_for_provider
max_context_tokens = get_context_limit_for_provider()
# Now supports provider-specific overrides!

Files Updated

• context_config.py (new): Global configuration module
• base_langgraph_agent.py: Uses get_context_limit_for_provider()
• base_strands_agent.py: Uses get_context_limit_for_provider()

See Also

• Context Management - Full documentation on context management implementation