Context Overflow

Detects when agent context windows are approaching or exceeding capacity, causing information loss and degraded performance.

Examples

• Agent conversation has consumed 95% of the 128K token context window
• Per-turn token usage averaging 8K tokens with only 12K remaining
• System prompt + tool definitions consume 40% of available context
• Token usage trending upward with estimated overflow in 3 turns

Detection methods

Token Counting

Precise token counting using tiktoken per model

Usage Tracking

Monitors safe (<70%), warning (70-85%), critical (85-95%), overflow (>95%)

Overflow Prediction

Estimates turns until overflow based on per-turn averages

Token Breakdown

Separates system, message, and tool token usage

Calibration accuracy

From the Pisama calibration set. See detector scoreboard for the full table.