Skip to content

Codex Skill Router v0.1.0

Latest

Choose a tag to compare

@zcccc-mxc zcccc-mxc released this 12 Jul 03:25

Release Notes: 0.1.0

Date: 2026-07-02

Codex Skill Router v0.1.0 is the first stable release of the local Skill routing quality checker for Codex.

What It Does

Codex Skill Router helps users inspect, audit, route, evaluate, and estimate local Codex Skills before relying on them in real work.

It answers practical questions:

  • Which Skills are discoverable?
  • Are Skill descriptions clear enough?
  • Which Skills look suitable for this task?
  • Did a routing change improve or regress behavior?
  • How large is the local Skill metadata context?

Install

After npm publication:

npm install -g codex-skill-router
csr --help

Repository install:

git clone https://github.com/zcccc-mxc/codex-skill-router.git
cd codex-skill-router
npm install
npm link
csr --help

Main Commands

csr scan
csr audit
csr route "check login authorization bypass"
csr eval ./examples/eval.yml --path ./examples/skills
csr budget

Local-First Guarantees

  • No API key required.
  • No external AI service required.
  • No cloud database.
  • No Skill upload.
  • No modification of user Skill files.
  • Paths are hidden by default.

RC1 Validation Result

v0.1.0-rc.1 completed real validation before this stable release decision:

30 real tasks reviewed
28 complete
2 failed
Required Skill Recall: 100.0%
No-Match Accuracy: 100.0%
Exclusion Accuracy: 91.7%
npm test: 21 tests passed

Known P2 issue:

docs-authoring may be over-recommended in a small number of web/mobile tasks.
The correct primary Skills are still selected.
This does not block v0.1.0.

Current Limitations

  • Local prediction is not Codex actual Skill invocation.
  • Routing is rule-based and has limited semantic understanding.
  • budget is a rough estimate, not real Codex token accounting.
  • The tool does not modify user Skills.
  • The tool does not call external AI services.

Reporting Routing Errors

When reporting a routing problem, include:

  • command used;
  • sanitized task prompt;
  • expected include, optional, and exclude;
  • actual recommendations;
  • whether no-match was expected.

Remove private paths, secrets, customer names, and private Skill contents before posting.