model-escalation

Advisory model escalation tracks task outcomes and provides recommendations for escalating or de-escalating models based on failure patterns. It operates in an advisory-only capacity, meaning the orchestrator retains the final decision on whether to act on the escalation advice.

You track task failure patterns by running the task-outcome-recorder.sh and model-escalation-advisor.sh scripts. The recorder persists task outcomes as JSON lines to /tmp, which the advisor script then reads to generate structured escalation advisories.

The model escalation advisor never blocks input or interrupts orchestration. It operates asynchronously by reading recorded outcomes and returning advisories, ensuring the primary agent workflow continues without latency or forced interruptions.

Yes, advisory model escalation applies across multiple agent types and models. It evaluates failure patterns against configured thresholds to determine when to escalate or de-escalate, supporting diverse AI architectures without requiring specific model dependencies.

The best way to handle de-escalation is by persisting task outcomes to support stability analysis. The advisor script reads this historical data to recommend de-escalation when failure patterns subside and outcomes stabilize below configured thresholds.

Advisory-only escalation is needed because it provides structured next-model recommendations and cost impact analysis without forcing a switch. This allows the orchestrator to evaluate context and make the final decision, preventing unnecessary or costly model changes.