Does Jev itself find the bugs in my diff?

No. The digest states Jev never judges code and that asking it whether a hunk has a bug failed; it recognizes text such as what a hunk touches or what a finding claims, while code and verifier agents check evidence.

What happens when Jev does not answer mid-run?

The run fails closed: triage assumes high risk, nothing is published inline, everything goes to the human queue, and the check reflects that the automated review could not complete.

Can a malicious pull request rewrite its own review rules?

No. The config is strict and read from the PR base, so a PR cannot choose the rules that judge it. The digest calls this out as a deliberate safety property.

Will this flood my pull requests with noisy comments?

Agentic mode publishes inline comments only for high-confidence findings with fingerprints against duplicates, and format-only hunks never reach the LLM. Per-run and monthly spend caps bound the rest.

How do I install it in a repository?

Add .github/workflows/jevest.yml using tincke10/Jevest@v1, then add TYPESAFE_API_KEY and, for the default reviewer, CLAUDE_CODE_OAUTH_TOKEN from claude setup-token as repository secrets. Optionally add a .jevest.yml at the root, which is a strict partial override read from the PR base.

What changed in version 1.0?

Agentic review became the default: one read-only claude -p agent per PR (Opus 5.5, effort xhigh) over a checkout of the head, evidence checked against the code and a refuting verifier (Sonnet 5). On 10 held-out PRs it measured 58.2% weighted recall, 82.4% precision, about $1.60 and 4.5 minutes per PR, against 7.7% and 25% for the 0.1 per-hunk reviewer.

How do the confidence bands choose an action?

A Jev answer's confidence lands in one of three bands configured per stage and risk level: at or above autoMin it acts automatically (publish, skip or green check); at or above confirmMin it asks a human; below that it escalates and fails closed with a red check. Thresholds rise with risk, so a critical merge gate needs 0.999 to go green and a low one 0.95.

CommunityJev Tools & CLIs

Jevest

GitHub Action that reviews pull requests with a read-only finder agent plus a refuting verifier, using Jev only to triage, route, label, and gate merging.

See highlights

Open github.com

Stars: 2
Language: TypeScript
License: MIT
Published by: Community
Author: tincke10
Last verified: Oct 2026

tincke10/Jevest

User repository on GitHub

View on GitHub

Automated PR review pipeline using Jev (TypeSafe AI) as a millisecond decision layer over an LLM reviewer

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Automated pull-request review with Jev as the decision layer around an LLM reviewer. A read-only agent finds the issues, a verifier checks each one against the code, and Jev decides in milliseconds what to triage, route and label. v1.0.0 (2026-10-02) makes agentic review the default and is the first release measured end to end on real pull requests.

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

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Defaults to one read-only agent per PR with quoted evidence, a refuting verifier, and Jev routing survivors to publish or question.
Keeps a legacy per-hunk LLM mode that re-judges each finding as real defect, style-only, or actionable.
Emits one verdict per run with fixed labels and a check that is red only when something must be fixed.
Reads strict config from the PR base and fails closed when Jev is unavailable.

Watch out

Expects GitHub Actions, a TypeSafe key, and a Claude provider token for agentic mode; measure cost on your own repos. 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

chess-with-jev-

Hybrid chess experiment pairing Jev strategic candidate selection with a deterministic alpha-beta search engine inside a Chrome extension that plays on Chess.com with live telemetry.