Does this actually run the skill it picks, or only recommend one?

Only recommend one. Classification, skill matching and tier selection are implemented; the handlers return example responses. Doing the work requires wiring in your own rule functions, skill runner or model client, so treat this as a router to build on rather than a finished skill executor.

What does the Verified & Ready badge actually verify?

Each profile runs one broad presence check, such as finding a Git repository, a manifest or a Node executable. Passing that check is the whole meaning of the badge. Skill installation, tool availability, credentials and output quality all need validation in the execution layer you add yourself.

Why does it call Jev even for arithmetic it is about to route to a local rule?

In live mode the classifier calls Jev to get a task type and a local-suitability score before choosing any tier, then applies JEV_RULE_THRESHOLD for deterministic-rule tasks and JEV_THRESHOLD for structured decisions. If you want zero calls, leave TYPESAFE_API_KEY unset and it uses keyword matching on profile text instead.

Can I point the skill registry at my own directories?

Yes, by constructing SkillRegistry with your own search_paths. The default scan looks one level deep for */SKILL.md under .agent/skills/, ../cjbs-backend/.agent/skills/ and the Gemini antigravity skills folder, so a fresh clone of this repo alone will not discover anything in your project.

What are the three routing tiers?

Tier 1 for local rules, Tier 2 for a structured decision or a specialist skill, and Tier 3 for foundation-model work, which also covers vision, complex reasoning, or unknown tasks.

How is the routing score computed?

In live mode Jev returns a task choice and a local-suitability score, and the classifier multiplies task_type_confidence by can_handle_locally; it is a routing aid and has not been calibrated as a probability of success.

What do the offline routing examples show?

With no API key, 15 + 27 routes to Tier 1 local rule, matching answer-sheet entries to question numbers goes to Tier 2 structured decision, and interpreting a shape in a photo goes to Tier 3 foundation model.

CommunityJev Agents & MCP

Todo Jev

Python CLI that classifies a request into one of three tiers — local rule, structured decision or skill, or foundation model — using TypeSafe Jev, 20 bundled skill profiles with exclusion conditions, and local environment checks.

See highlights

Open github.com

Stars: 3
Language: Python
License: MIT
Published by: Community
Author: maker-KK
Last verified: Oct 2026

maker-KK/todo-jev

User repository on GitHub

View on GitHub

⚡ Ultra-fast, low-cost intelligent task classifier and 3-tier routing engine powered by TypeSafe Jev (System One)

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Choose the next step before your AI agent takes it. A task-routing experiment powered by TypeSafe Jev, structured skill profiles, and basic environment checks. Quick start · How it works · Skill profiles · Evaluation. todo-jev classifies a request and recommends a handling path: a local rule, a structured decision or specialist skill, or a foundation model. It combines when a skill applies and when to exclude it with basic environment checks performed locally.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Routing

Used inJev for Software & DevTools

Languages: Python

Highlights

Twenty bundled skill profiles each carry application conditions, exclusions, an expected input, a preflight contract and source metadata.
Preflight checks are broad presence checks such as a Git repository, a dependency manifest or a Node executable.
Offline keyword fallback runs with no API key; a Jev network or HTTP failure falls back to the same local heuristic.
CLI covers route, catalog, skills and demo, with live thresholds for Tier 1 and Tier 2 and a --no-auto-sync switch.

Watch out

The execution handlers return example responses, so nothing is actually performed until you connect your own rule functions, skill runner or model client. The CLI badge Verified & Ready only means the presence check passed; the recorded SHA256 fingerprints are metadata and identity is not verified during routing. matching_rate is a routing score, not a calibrated probability of success, and offline mode uses fixed heuristic scores. Auto-sync only reads up to 15 discovered skills from three hardcoded default directories. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

claude-jev

Coding-agent plugin that routes prompt classification, subagent tiering, edit rule checks, and fast Jev-backed compaction through small typed judgments.