Agent long-session context health check

Catch context drift before long conversations go off the rails

Problem

Users report flaws in long conversations, system messages leaking into user turns, and behavior changes after compaction—but they only have a feeling. There’s no way to visualize “context health,” so they restart sessions by trial and error.

Solution

A context health inspector that samples the structure of long sessions, detects system/user role drift, key information lost after compaction, instruction conflicts and token distribution anomalies, and returns a readable health score with fixes like splitting sessions, re-pasting key instructions, or disabling a compactor.

Users

Developers and automation teams running long multi-turn agent sessions

MVP

Ingest local agent logs/session files; parse roles and token distribution; flag role drift and compaction traces; output health score and fixes; export reports

Revenue

Subscription (team session upload, scheduled health scans) or open-source core with paid cloud analysis

Why now

Long agent sessions are mainstream but context management is still a black box; log formats have de-facto standards, keeping scope manageable