efficient-dispatch — Claude Code skill: model routing for subagents

What efficient-dispatch does

Two splits matter here. The main model orchestrates and reviews; subagents take the mechanical load. The one that moves your bill is the tier of each dispatch. Well-specified implementation, bulk edits and search go out with an explicit model: sonnet, pure locate work with haiku. Anything that needs judgement — review, design, audit — stays on the strong model, because a cheap reviewer misses real bugs.

Forget the model: field on an implement dispatch and it silently inherits the expensive one. The route-dispatch hook only advises; Claude hooks cannot force a model onto a subagent dispatch, they just inject context. So explicit model: on every call is the actual enforcement. The routing table lives at ~/.ax/hooks/routing-table.json and is rebuilt with ax dispatches compile-routing.

Context isolation is the other reason to dispatch. A screenshot read into the main thread stays in the window and gets re-sent on every later turn. When you want a verdict rather than the bytes, a subagent returns text and the image dies with its context. Tight visual iteration — look, tweak, look again — is the case where this loses: the round-trips cost more than they save. The skill targets subagent-heavy work and stays out of single-shot questions and small tasks with no dispatching.