Jev in production › Model and agent routing

claude-subagent-router

claude-subagent-router uses a Jev classifier to pick Sonnet or Opus per Claude Code sub-task; the author measured Sonnet at high effort passing the same hidden tests as Opus at max effort (stas4000).

Open on GitHub ↗

The source publishes no measured number.
Use
Model and agent routing
Industry
Developer tools
Form
Open-source tool
Stage
Beta
Plugs into
Claude Code
Listed
2026-09-30
Found via
github
Repository
stas4000/claude-subagent-router
Stars
10
Forks
0
Last push
2026-09-29
Language
Python
License
MIT

The README opens with

A Claude Code sub-agent setup where an Opus 5.5 lead hands sub-tasks to Sonnet 5.5 or Opus 5.5 builders by effort level. A PreToolUse hook picks the builder for each sub-task. A Jev classifier makes the pick, and an offline rule-based picker answers when Jev is not available. A hard guard makes sure Sonnet never runs at max effort.

It is the setup we run ourselves, packaged so you can copy it. Python 3.10+, standard library only.

We ran every sub-agent on Opus 5.5, because we never compromise on quality. Then we measured Sonnet 5.5 at each effort level on real coding tasks, and added Sonnet where it is just as good: Sonnet 5.5 on high and extra high passed the same hidden tests as max, on a fraction of the output tokens (numbers in Measured)....

Badge

For the project's own README, linking back here:

Listed in Jev in production

Also used for model and agent routing