defer

{
  "name": "defer",
  "description": "Decide which model a piece of work goes to, at which thinking level, and produce the exact command that sends it there. One routing policy for every skill that hands work outside the session. Resolves user-preferred current model roles from the runtime first; a registry then maps nine task classes to calibrated lanes, each pinned to a model and an effort, with the CLI arguments, the environment GLM needs, and the check that proves the lane ran as routed. GPT-6 Astra is the OpenAI-family spine at three thinking levels \u2014 low is the workhorse tier alongside sol at high, fable at medium and grok at high, while medium and high are frontier capacity reserved for the hardest problems and reachable only by asking. Gemini 3.8 Flash is the fast implementation lane once a spec or plan exists, grok and GLM are trusted to orchestrate, and design work stays on Opus and Fable. Narrows lanes by what the work actually is, against a capability matrix measured over 106 benchmark tasks, and marks borrowed rows, placeholder prices and inert penalties as such rather than presenting them as measurements. Where several lanes remain it picks the one with the most plan headroom per remaining day. Use when a skill needs a second opinion, an out-of-family verdict, a completeness critic, an implementation lane, an orchestrator or a design review, and whenever someone asks which model should do something, why a lane was chosen, what a lane costs, whether a cheaper model could do it instead, or whether one is out of allowance.",
  "version": "1.5.0",
  "author": {
    "name": "Fledgeling",
    "email": "[email protected]"
  }
}