hook-context-window-monitor

To monitor context window usage, configure threshold-based warnings that trigger when token consumption exceeds defined limits. This approach injects non-invasive reminders during long-running LLM sessions to prevent hidden token overflows without automatically triggering context compaction.

Context window monitoring only observes usage and emits warnings when thresholds are exceeded, whereas preemptive compaction actively summarizes and discards previous conversation turns. Monitoring suggests running compaction but leaves the actual execution to the user.

Set up token budget warnings by configuring the warnThreshold and criticalThreshold parameters. When context usage crosses these configured limits during multi-turn AI conversations, the system injects reminders to improve dialogue efficiency and avoid overflows.

Yes, you can adjust the warning interval for context window alerts. Configurable parameters allow you to control the warningInterval alongside critical and warning thresholds to manage alert frequency during sessions with tight token budgets.

Use a non-invasive context monitor when you need to preserve the full conversation history for multi-turn planning but want alerts before hitting token limits. It prompts the LLM to be more concise and suggests manual compaction only when necessary.