guard

To prevent auto-compaction from discarding critical agent team state, you can run a background sentinel that monitors session size and proactively prunes context bloat. This preserves TeamCreate, SendMessage, and task state during long-running interactive sessions.

Session pruning removes noisy context and metadata bloat to manage token budgets. Threshold-based pruning applies soft trims when a soft limit is reached and full prescriptions plus optional auto-reload at a hard limit to protect critical state.

Recover from rapid context overflow using reactive file-watcher overflow recovery. It detects inbox floods and applies escalating pruning prescriptions with a circuit breaker to prevent recovery loops while preserving team state.

Yes, you can checkpoint session state to disk at configurable intervals for safe reloads. Continuous checkpointing saves team state periodically, enabling recovery without losing critical agent team context.

Yes, the guard sentinel works with Claude Code interactive sessions and long-running agent teams. It monitors token budgets, repeated tool outputs, and metadata bloat that accumulate in these environments to prevent auto-compaction issues.

A limitation of auto-pruning is that aggressive pruning prescriptions can trigger recovery loops if rapid inbox floods persist. A circuit breaker mitigates this, but pruning is designed for threshold-based overflow rather than continuous high-volume context generation.