Testing ADK Agent Context Window and History Limits

ShopMax India customer support conversations can span dozens of turns as customers ask about multiple orders, check return eligibility, and inquire about product stock. Without a history truncation strategy, the agent context window fills up and either throws an error or drops important earlier context. Testing the history management logic verifies that the agent keeps only the most recent turns and stays within the token budget.

A sliding window strategy keeps the N most recent turns in history, where each turn contributes two messages - one user and one agent. Token budget enforcement estimates total tokens from word counts using an average tokens-per-word ratio and discards additional turns until the budget is met. These two constraints can be independently tested.

The example below simulates a 50-turn ShopMax India conversation and tests that the sliding window keeps exactly 20 messages and that the estimated token count stays well under 8000.

It gives the following output,

Truncated history: 20 messages
Estimated tokens: 205
.. (2 passed in 0.01s)

In production, call truncate_history before every agent invocation so the context window never overflows. Combine the turn limit with a token budget check using actual tokenizer counts from tiktoken or the model SDK for more precise enforcement. Run these tests as part of your pre-release regression suite to catch any history management logic that is accidentally removed during refactoring.