dsh-peak-avoidance

DeepSeek peak-hour pricing adds up fast if you run long agent workflows. dsh-peak-avoidance takes a straightforward approach: it automatically switches to a lower-cost alternate model before a scheduled peak window and switches back to the official model once that window ends, with zero manual intervention.

Switching covers both the default route and live sessions. You can define multiple peak windows, including cross-midnight ranges such as 22:00 to 02:00 the next day, with an adjustable lead time (default 30 minutes). Provider, model, and reasoning-effort dropdowns are fetched live from the DSH llm service, so nothing is hardcoded against a specific deployment. A frame-wide banner announces entry and exit (toggleable), and you can always step in manually via card buttons or the peak_avoidance_status / peak_avoidance_control model tools. Sessions you have already reverted yourself are never force-switched back, and the plugin restores your original model when you stop it, keeping the experience safe.

If you run heavy agent workloads through DSH, are cost-conscious about inference pricing, and would rather not juggle model switches by hand every single peak window, this plugin is built for you. All you need to configure are peakModel and normalModel; every other setting ships with a sensible default and takes effect immediately without restarting DSH.