context-health

Context drift in long-running tasks is detected using heuristic checks that measure divergence between user intent and assistant output across multiple turns. Threshold-based monitoring flags degradation so recovery strategies can be applied.

Context window pressure is handled by monitoring token budget thresholds and applying recovery strategies like truncation or checkpoint restart. These mechanisms prevent context degradation from breaking long conversation progress.

Recover from tool failure context degradation by applying re-anchoring or checkpoint restart strategies. The system detects high tool failure rates and triggers safe reset mechanisms with cooldowns to restore task progress.

Re-anchoring or truncation is needed when heuristic drift detection identifies a significant gap between the original user intent and current assistant output. Applying these recovery strategies realigns the conversation safely.

Yes, you can monitor token budget thresholds to prevent context degradation in long horizons. The system tracks token usage and triggers user warnings or truncation when thresholds are exceeded during multi-step tasks.

Checkpoint restarts for long-running conversation recovery require re-establishing context from the saved state, which may lose intermediate tool call results. Cooldowns and safe reset mechanisms mitigate but do not eliminate data loss risks.