Alex314618-create/JevRev

JevRev – a decision‑layer for LLM‑driven coding agents

What it is JevRev is a Node‑JS library and CLI that sits next to a large language model (LLM) and decides which of the model’s proposals are worth pursuing. It does not replace the LLM; it filters, audits, and monitors the work the LLM generates, preventing wasted compute and token budget.

Core components

Component	Role
JevSift	Takes a JSON list of alternative approaches from the LLM, removes weak/duplicate/risky ones, and emits bounded work orders for the survivors.
JevLoop	After the LLM implements a work order, JevLoop receives the evidence (test results, benchmarks, etc.), asks the LLM only the narrow questions it cannot answer itself, and returns a next‑action token such as continue, fix_regression, verify, replan, waiting_human, or completed.
JevLong	Observes a long‑running session, reporting stalls, budget risk, tool‑call problems, and drift. It can output a one‑line status or a continuous JSON stream for external dashboards.

How it works

1. Propose – The host agent (e.g., Claude Code, Codex) writes proposals.json with several materially different solutions.
2. Sift – jevrev sift filters the list and returns a set of bounded work orders.
3. Implement & record – The agent runs the work order, producing evidence.json (test output, performance numbers, etc.).
4. Decide / Loop – jevrev decide or jevrev loop consumes the evidence, may ask the LLM a focused question, and returns the next action.
5. Long‑watch – For multi‑round projects, jevrev long ingests the session log (JSONL) and alerts a human if the run drifts or exceeds budget.

All communication is plain JSON, making JevRev easy to call from any LLM‑tool integration.

Demo

git clone https://github.com/Alex314618-create/JevRev.git
cd JevRev
npm install
npm run build
npm run demo:workflow   # runs a deterministic example with a regex shortcut vs. state‑machine

The demo shows a fast but incorrect regex solution being rejected after a correctness probe, while a slower but correct state‑machine wins.

Installation & usage

npm install   # pulls dependencies
npm run build # compiles TypeScript to dist/
# optional: install as a Codex skill
node scripts/install-skill.mjs --target codex

The CLI is invoked via node dist/cli.js <command> or, after installing the skill, simply jevrev <command>.

Providers JevRev can call a hosted “Jev” service, run a local LLM via llama.cpp (named SemIf), or replay pre‑recorded fixtures for offline testing. Provider selection is done with --provider <name>.

Why it matters LLMs are great at generating many ideas quickly, but they can waste resources on low‑value or incorrect paths. JevRev adds a cheap, deterministic audit step that:

• Cuts token cost by stopping bad proposals early.
• Guarantees that every accepted artifact has passed explicit correctness or performance checks.
• Gives humans a clear, logged decision trail.

Project health

• Node.js ≥ 20 required, MIT licensed.
• CI includes unit tests (npm test) and several demo scripts.
• Documentation is split into workflow guide, protocol spec, and design notes.

Typical use case

Give the host agent the instruction:
"Use JevRev for this task. Propose materially different approaches, ask JevRev to sift them, run only the bounded probes, record the evidence, and let JevRev audit the next round before continuing."

The agent then calls jevrev sift, jevrev loop, and optionally jevrev long as the project progresses.

---

Bottom line – JevRev is a lightweight, language‑agnostic decision layer that lets you keep the creative power of an LLM while adding a safety net that filters, audits, and monitors the work before you spend more compute or human time on it.