dsh-model-router

In the real world, model pools are the norm: a cheap fast one, a strong one, a spare. The problem is that hardcoding a single model ID means any single provider outage becomes your outage. dsh-model-router pulls routing out of the harness core: declare a virtual model ID that looks identical to a real model in the picker, backed by a candidate pool and a routing algorithm that decides where each call actually lands.

The headline experience is transparent failover. When a candidate times out or errors, the plugin re-dispatches the same request to the next candidate with no user-visible error. Meanwhile the response provenance honestly records which real model produced the answer and which candidates are still cooling down in their backoff window. Two built-in algorithms ship out of the box: priority always tries the first candidate and applies exponential backoff on failure, giving a struggling model time to recover; round-robin distributes calls evenly across the pool. For custom strategies, the RoutingAlgorithm factory exposes select, onFailure, onDispatch, and onSuccess hooks — implement one, register it, and the rest of the harness never needs to know.

It fits a very concrete scenario: you span two or three providers, your agent workflow needs at least one model always reachable, and you do not want to hand-roll retry logic for every failure. One route declaration is all it takes; the algorithm handles the rest.