Which coding runtimes does it support?

It targets Claude Code, Codex, Opencode, Pi Agent, Devin, Cursor, and Hermes CLI through plugins, hooks, and extensions rather than hand wiring. Seven CLI bridges also let a runtime dispatch other AI CLIs as sub-tools with bounded lineage.

What happens when the skill daemon is down?

A resident daemon scores prompts across five signal lanes against a live skill graph, while a local scorer answers and marks itself stale when the daemon is down. No skill names need memorizing.

Does memory need a database or service?

Session context is written into the spec folder it belongs to and found through a committed trigger index plus ripgrep recipes, with no database or daemon on that path. An unwritten phrase is a clean no-hit.

How do the gates stop sloppy changes?

Five mandatory gates run around file changes, including a hard block requiring a spec folder and repo-rules loading. Reviews rank findings P0/P1/P2, re-challenge critical findings, and force another pass on open blockers.

How do I install and verify it?

Clone the repo, then npm --prefix .skilled/skills/system-spec-kit install, install and build the advisor runtime, and run node .skilled/bin/skill-advisor.cjs list-tools --format json. It needs Node 20.11+, with 22.12+ for the spec-kit runtime CLI, plus npm, git, and a POSIX shell; no global TypeScript is required.

What do the deep loops do?

Research, review, and improvement loops run unattended and stop only when their own evidence says done. Progress is written to disk, so a loop survives crashes, new sessions, and long runs, and one shared runtime underlies every loop. You can run fully hands-off or pause at each step.

How does the agent library dispatch work?

Twelve specialized agents own focused roles and seven CLI bridges let a runtime dispatch other AI CLIs as sub-tools. A review can fan out across several models with the safest verdict winning, while the conducting AI stays in charge. Every bridge but cli-pi refuses to dispatch the runtime it runs in, and fan-out lineages and dispatch stacks are bounded.

CommunityJev Agents & MCP

Skilled Agent Harness Spec Driven Loops

Assistant framework that stores specs, memory, and specialist agents as plain markdown files in your repo for Claude Code, Codex, Opencode, and similar runtimes.

See highlights

Open github.com

Stars: 37
Language: TypeScript
License: MIT
Published by: Community
Author: MichelKerkmeester
Last verified: Oct 2026

MichelKerkmeester/skilled-agent-harness_spec-driven-loops

User repository on GitHub

View on GitHub

AI-assisted coding setup that helps you spend less time re-explaining context, and more time shipping with better output. Includes a custom spec kit, memory, agent and skill framework.

Forks: 3
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

An assistant framework that gives your AI coding agent a memory, a paper trail and a team of specialists. All of it lives inside your own repository, as files you can read and diff. Built for Claude Code, Codex, Opencode, Pi Agent, Devin, Cursor and Hermes CLI. Spec Kit Framework - structured plans, task tracking, validation gates and handover docs. Spec Memory & Search - session context saved into each spec folder and recovered through instant keyword search, surviving resets and compactions.

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

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Routing Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Records every file change in a spec folder with scaled docs, validation gates, P0/P1/P2 reviews, and handover notes.
Recovers session context through /speckit:resume plus a committed trigger index and ripgrep search with no database.
Suggests skills at prompt time with a five-lane fusion daemon and live dependency graph, plus a stale-marked local fallback.
Ships 12 specialized agents, 14 on-demand skills, unattended research loops with on-disk progress, and a Code Mode MCP server.

Watch out

Needs Node 20.11+, git, and a POSIX shell, with a newer Node for the spec-kit CLI; behavior is per-runtime plugin wiring, not a hosted service. 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

duet-agent

Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.