aadhar-build

Claude Code Jev Router — AI skill for Claude Code

AI community

Picks the model tier for a delegated Claude Code task before it spawns.

How to install Claude Code Jev Router

This entry records only its repository, not the path inside it, so there is no exact command to give. Open aadhar-build/claude-code-jev-router and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Claude Code Jev Router does

Picks the model tier for a delegated Claude Code task before it spawns. Opt-in per repo, reversible in one command, fails safe. v0.1 is a lookup table on purpose.

Alternatives in AI

Codex Plugin — 's /codex:adversarial-review command 9.6k ★
Botmux — Bridge Feishu/Lark to AI coding CLIs — Claude Code, Codex, Gemini, OpenCode… every DM, group or topic spawns i 1.2k ★
StereOS — A Linux based operating system hardened and purpose built for AI agents 489 ★

README

claude-code-jev-router

**Picks the model tier for a delegated Claude Code task before it spawns.** Cheap work goes to a cheap model. Hard work doesn't.

Opt-in per repository. Reversible in one command. Fails safe.

./jev install /path/to/your/repo --yes

Should you use this?

**Yes, if** you delegate work to subagents in Claude Code and you'd rather not pay frontier rates for every `Explore` task.

**No, if** you want it to make Claude Code *feel* faster. It won't — see below. That's measured, not modesty.

**Not yet, if** you want a proven savings number. **There isn't one.** This went live only just now, and no result has been collected. A README claiming a percentage would be making it up.

What it actually does

Claude Code fires a `PreToolUse` hook before it spawns a subagent. This registers a hook there, looks at the `subagent_type`, and rewrites `model` according to a table you control.

  you: "go explore the codebase"
        |
        v
  Claude Code prepares a subagent  ──►  [ jev hook, ~25ms ]
        |                                      |
        |                          reads config/tiers.json
        |                          Explore -> haiku
        v                                      |
  subagent runs on haiku  ◄────────────────────┘

That's it. **v0.1 contains no AI.** It is a lookup table.

That's deliberate, and it's the most interesting decision here.

Why a lookup table, and not a classifier

Five independent studies find learned routers frequently fail to beat a trivial static baseline — LLMRouterBench (400K instances, 21 datasets), RouterArena's "routing plateau", kNN outperforming MLP and GNN routers, RouteLLM near-random on MMLU. Meanwhile published static heuristics deliver real savings.

So the dumb version ships first, and anything cleverer has to beat it on measured cost at equal task success. **The baseline is the product until something outperforms it.**