dsh-hybrid-coder

A local small model (say, Qwen running through Ollama) is lightweight and free for day-to-day coding, but it struggles the moment a task demands multi-step planning or a streak of failed tool calls. dsh-hybrid-coder is built for exactly that gap: it sits as a routing-policy layer that lets the local model handle routine steps while delegating planning and rescue to a premium model (DeepSeek, GLM, etc.), switching automatically between the two based on explicit escalation and de-escalation rules.

The core mechanism operates on three routing priorities: plan mode always routes to premium, routine steps go to local, and an escalation latch keeps everything on premium until enough clean premium steps accumulate. When the local model's consecutive tool failures hit a configurable threshold, the plugin escalates, injects a bounded guidance message carrying the recent failure trajectory, and lets the premium model take over. Transport-level failures such as Ollama not running or a refused connection trigger an immediate request-level fallback to premium. All routing state is persisted as session events, so fork, resume, and process restarts recover identical decisions with no in-process state to lose.

Ideal for DeepSeek Harness users who run a local inference engine like Ollama and want to keep daily coding costs low without giving up the safety net of a frontier model. The plugin is experimental; public contracts (event names, config fields, guidance text) may change before the first tagged release.