菜单
Choose the right model and effort per task phase and apply it when dispatching subagents — strong model + high effort for planning and review, a cheaper model + lower effort for mechanical execution. Use when a task splits into plan / execute / review phases or you're spawning subagents.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | model-router |
| description | Choose the right model and effort per task phase and apply it when dispatching subagents — strong model + high effort for planning and review, a cheaper model + lower effort for mechanical execution. Use when a task splits into plan / execute / review phases or you're spawning subagents. |
| model | inherit |
| effort | medium |
| allowed-tools | ["Read","Task"] |
| triggers | ["which model","route this","plan then implement","dispatch a subagent"] |
| tags | ["models","effort","routing","subagents","cost"] |
Match the model and effort to the work. Reasoning-heavy phases deserve a strong model at high effort; mechanical phases run cheaper and faster. This cuts cost and latency without hurting quality where it matters.
Defaults live in ${CLAUDE_PLUGIN_ROOT}/config.json under models.roles.
Read it first; treat it as the source of truth and let the user override.
| Phase / role | Model | Effort | Why |
|---|---|---|---|
| plan | strong (e.g. opus) | high | Design, trade-offs, architecture — errors here are expensive. |
| execute | mid (e.g. sonnet) | medium | Follow a plan, write code — throughput matters. |
| review | strong (e.g. opus) | high | Catch what execution missed; judge freshly. |
| cheap | small (e.g. haiku) | low | Formatting, renames, extraction, boilerplate. |
effort scales how much the model deliberates. Use inherit to keep the
session's current setting when a split isn't warranted.
inherit. Routing has overhead; only pay it when phases genuinely differ.cheap.When you dispatch a subagent for a phase, request that phase's model + effort. State it explicitly, e.g. "planner: strong model, high effort". If the runtime rejects a model or effort value, surface the exact error — do not silently fall back to a weaker setting, since that changes the quality contract.
plan and the
apply step to execute or cheap.decision via memory-keeper if it's a
standing convention for the project ("Review always runs on opus/high").Run a bounded change → verify → fix cycle until an objective check passes (tests green, lint clean, build/type-check ok) or a hard iteration cap is hit. Use for any task that needs repeated attempts against a verifiable signal instead of a single one-shot edit.
Review code for behavior-preserving simplification — collapse duplication, remove dead code and over-engineering, cut verbose patterns down to idiomatic ones. Use when code is longer or more complex than it needs to be, or the user asks to simplify, shorten, refactor, or clean up.
Implement a planned ticket on its own branch using test-first development, updating docs as part of the change, looping change→verify→fix until the suite is green. Use when a ticket has a plan and is ready to build, or the user says "implement X" / "build the ticket" / "make it pass".
Use to store durable context (decisions, facts, preferences, todos, summaries) and to recall or reply from it across sessions. Load when something worth remembering appears, or when the user asks what you were doing, what was decided, or to continue prior work.
Push a reviewed ticket branch and open a pull request with a structured title and body, after review-changes has passed. Use when a ticket is reviewed and ready to ship, or the user says "open a PR" / "create the pull request". Stops at opening the PR — never merges.
Turn a feature request into a concrete plan — scope, approach, affected files, a test plan mapped to acceptance criteria, and out-of-scope — before any code is written. Use when the user describes new work, asks to design or spec something, or says "plan this" / "how would you build X".