Why not just fine-tune Jev on our past reviews?

Jev is not fine-tuned with customer data, so reviewer rulings live in the Meko datapack where promotion makes them visible to the next review without retraining.

Will one ruling suppress the same violation elsewhere?

No. The re-review shows the fixture exempted while production code keeps the finding, the other rule is unchanged, and shared helpers go to ask-a-person.

What breaks if guidelines or IDs are misconfigured?

Empty output means guidelines were not found or the title mismatches, and a UUID error means MEKO_DATAPACK_ID is not the datapack UUID from its page.

Do I need human approval before promoting a ruling?

Yes. Promotion moves a private memory to Shared Knowledge for everyone and cannot be undone, and promote.py refuses without --yes.

What probability thresholds drive a finding, exemption, or human review?

A rule probability of 0.3 or more triggers a search for covering rulings. A covering answer of 0.7 or more exempts the finding, 0.3 to 0.7 asks a person, and below 0.3 the finding stands.

What format must a ruling follow, and why?

Rule name, scope, then because and the reason. The README stresses the reason matters: without it, Jev reads the ruling's scope more broadly than the reviewer meant.

How do the scripts reach Meko, and who decides what is stored?

Over MCP at https://mcp.mekodata.ai/mcp with a Bearer key. No LLM decides what is stored; the code does, and promotion requires the --yes flag from a human.

CommunityJev Agents & MCP

Meko Jev Code Review Agent

Code review demo pairing Jev per-rule yes/no checks with a Meko datapack that stores guidelines, rulings, and decision traces, exempting only the ruled case on re-review.

See highlights

Open github.com

Stars: 0
Language: Python
License: APACHE-2.0
Published by: Community
Author: YugabyteDB-Samples
Last verified: Oct 2026

YugabyteDB-Samples/meko-jev-code-review-agent

Organization repository on GitHub

View on GitHub

A code review check that learns from reviewer rulings without fine-tuning. Jev (TypeSafe AI) judges each diff hunk against your team's guidelines; a Meko datapack (powered by YugabyteDB) holds the guidelines, the rulings, and a trace of every decision. Companion repo for the blog post "Jev and Meko Make Code Review Feedback Stick".

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

A code review check that learns from reviewer rulings without fine-tuning. Jev (TypeSafe AI) judges each diff hunk against your team's written guidelines and returns probabilities, not text. Jev is not fine-tuned with customer data, so when a reviewer overrules a finding, the ruling has to live somewhere else. A Meko datapack holds the guidelines, the rulings, and a trace of every decision. After a ruling is promoted, the next review finds it and exempts the case it covers, and only that case.

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

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Asks Jev one yes/no question per rule per hunk; rulings are never included in that first request.
Re-asks Jev per candidate ruling with 0.7 exempt, 0.3-0.7 ask-a-person, below 0.3 finding stands.
Requires ruling format Rule name plus scope plus because reason, or Jev reads scope too broadly.
Stores every hunk decision as a conversation viewable in the Meko portal Observe view.

Watch out

Needs TypeSafe and Meko keys plus a portal-uploaded datapack; promotion is visible to all and cannot be undone. 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

wald-4b

Open-weight 4B decision model on Qwen3.5-4B-Base serving a Jev-compatible /v1/systemone API, returning per-option probabilities with adjustable thinking.