dsh-model-controller

When juggling v4-pro and v4-flash in DeepSeek Harness, the most tedious part is rarely the code itself—it is deciding, turn by turn, which model to switch to and what reasoning level to set. dsh-model-controller removes that overhead entirely: at the start of every user turn it automatically picks a route, so you can stay focused on the problem.

The routing strategy works in two layers. First, keyword regex rules scan the input text in priority order; if a rule matches, the corresponding model and reasoning tier are locked in instantly. If nothing matches, a semantic classifier steps in as a fallback. You can enable either layer alone or combine them in hybrid mode (regex first, then semantic), all configured from the plugin settings panel without touching any code. A built-in default table already covers common scenarios—complex refactoring, code review, small scripts, small methods, and Plan Mode—and can be extended, reordered, or reset at any time. Once a route is selected it stays effective for the entire turn; if no rule fires, the Agent's original model is preserved. Optionally, route metadata prepends the provider, model name, reasoning tier, and route id to the top of every reasoning block for quick audit.

This plugin suits developers who already mix multiple DeepSeek tiers in Harness and want model selection to become a declarative rule rather than a manual toggle each turn. It also fits tech leads who want to establish a team-wide convention for mapping specific tasks to the right models.