setup_routing

Configure per-project model & effort routing — spend Opus where it changes output quality, run mechanical work on Sonnet.

What it does

setup_routing independently analyzes the current repository, classifies it into a sensitivity profile, and proposes a tag-driven routing policy — then, only after you confirm, writes two coordinated artifacts:

1. a concise (≤10-line), sentinel-delimited routing block in the project's CLAUDE.md (the policy layer — which executor a tagged task is delegated to), and
2. the pinned-model executor subagents in .claude/agents/ that the policy references (the enforcement layer — plan-task-executor on Sonnet, plan-task-executor-opus on Opus).

The policy is only trustworthy because the pinned executors exist to back it, so the skill writes the executors first and never references one that does not exist. It is non-destructive and idempotent: re-running updates the managed block in place and creates executors only if absent.

When to use it

• You want a project to reserve Opus + high/xhigh thinking for the work that needs it and let Sonnet handle high-volume mechanical tasks.
• You are setting up a compiler/DSL/web project and want routing wired into make_plan / exec_plan for that repo.
• Not for global configuration — it only ever touches the current project's CLAUDE.md and .claude/agents/, never ~/.claude/.

Trigger phrases

"setup_routing", "set up model routing", "configure model routing", "route tasks by model/cost", "make this project use Opus/Sonnet per task". Explicit: /codeops:setup_routing <description> or the alias /setup_routing.

Sensitivity profiles

Profile	Fits	Default tag → model
A — Opus-dominant	compilers, interpreters, type systems, codegen	untagged → complex (Opus); only mechanical work de-escalates to Sonnet
B — Mixed core/scaffold	DSL/query languages that lower to a target (e.g. SQL)	lowering/semantics → sensitive (Opus); scaffolding → Sonnet
C — Sonnet-default	React/Vue/Node, REST/GraphQL, CRUD	untagged → standard (Sonnet); escalate only security/concurrency/perf
Balanced (fallback)	ambiguous detection	Sonnet default, Opus on complex/sensitive — flagged as a fallback to correct

Routing is expressed over a fixed tag vocabulary — trivial · standard · complex · sensitive — with the constant rule trivial/standard → Sonnet, complex/sensitive → Opus. Each profile only sets the default tag and what it escalates or de-escalates.

Worked example

/codeops:setup_routing a Rust compiler with a parser, IR, and codegen backend

The skill reads the repo, sees parser/, ir/, and codegen.rs, classifies it Opus-dominant, and presents the routing block plus the two executors to be created. After you confirm, it writes the <!-- CODEOPS-ROUTING --> block into CLAUDE.md and plan-task-executor{,-opus}.md into .claude/agents/, then tells you to run /agents to verify and to watch your rework rate before trusting Sonnet on borderline tasks.

Related skills

• make_plan — the routing block asks it to tag each generated task.
• exec_plan — consumes the routing block, delegating tasks to executors by tag.
• preflight — always Opus under every profile.
• See the Commands page for the /setup_routing alias.