cost-optimize

Analyze recent usage to find agents running simple tasks on expensive models, then downgrade them to cheaper tiers like Haiku. Also check cache hit rates per agent, since enabling prompt caching can cut costs significantly on repeated context.

Query the cost-tracking memory namespace for the last 7 days of usage data, then break down consumption per agent and model. Compare each agent's task complexity against its model tier to find mismatches.

This Skill flags agents below a 60% cache hit rate and recommends enabling or improving prompt caching. Cache reads can reduce costs by up to 90% compared to uncached input tokens.

Avoid downgrading agents handling complex tasks like architecture decisions or security analysis, since cheaper models create quality risk there. The Skill flags these cases instead of recommending a downgrade.

Each downgrade recommendation emits a model-outcome event via the outcome script, reporting success or escalation. Without this feedback signal, the routing layer cannot learn and the booster bypass rate does not improve.