Jevest
Automated PR review: an LLM agent finds issues, a verifier refutes them, and Jev decides how to triage, route and label.
Overview
Jevest is an automated pull-request review pipeline that places Jev as a cheap decision layer in front of and behind an LLM reviewer. In the default agentic mode, one read-only Claude agent per PR (limited to read, grep and glob on a checkout of the head) finds issues with quoted code evidence, a fresh verifier agent tries to refute each finding, and Jev routes each surviving finding to publish or question. Jev also triages PRs from metadata and profiles each hunk, so format-only hunks never reach the LLM, and spend caps apply. A merge gate only emits a check conclusion; there is no merge call wired.
Each run produces one verdict (fix, questions, clear or unavailable) and one label, with reviews in English or Spanish. Configuration is read from the PR base so a PR cannot change the rules judging it. A rule learned from measurement is that Jev never judges code: asking whether a hunk has a bug failed, and a Jev-only reviewer found 0 of 44 issues on real PRs, while asking what a hunk touches works. On 10 held-out PRs, version 1.0 reports 58.2% weighted recall and 82.4% precision at about $1.60 and 4.5 minutes per PR.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Should this verified finding be published or raised as a question? | publish · question |
| Choice | What kind of change is this hunk / what does it touch? | — |
| Other | PR triage from metadata and merge-gate conclusion | — |
How it works
- Jev triages the PR and profiles hunks; skip format-only hunks
- Read-only LLM agent finds issues with code evidence
- Verifier agent tries to refute each finding
- Jev routes survivors; one verdict and label per run
Strengths
- Reduces noise with evidence checks and refutation
- Spend caps and triage cut cost
- Config read from PR base for safety
- Honest measurement of what Jev should not do
Caveats
- About $1.60 and 4.5 min per PR
- Jev cannot judge code correctness itself
- Recall around 58% on held-out PRs
At a glance
- Type
- Project
- Status
- production
- Language
- TypeScript
- Runtime
- GitHub Action (Node 22) + Claude Code CLI
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Automated PR review pipeline using Jev (TypeSafe AI) as a millisecond decision layer over an LLM reviewer.
hellogumbo/awesome-jev
Frequently asked questions
Does Jev review the code?
No; the author found it cannot judge bugs, so it only triages, routes and labels.
How well does it perform?
58.2% weighted recall and 82.4% precision on 10 held-out PRs, per the README.
Can it merge PRs?
No, the merge gate only emits a check conclusion.
Related projects
Coding-agent hook that asks Jev one yes/no per project rule on every edit or turn and makes the agent fix any violation.
HookbetaTypeScriptClaude Code plugin sending small judgments to Jev: rule checks that block edits, verbatim compaction, subagent sizing and routing hints.
PluginbetaPythonAdvisor for coding agents: git and Stop hooks flag test tampering, secrets and skipped hooks, with optional Jev review findings.
HookbetaClaude Code plugin with rules and checks that make agents prove their work; an optional Jev judge triages review findings.
PluginbetaMarkdown + shellCLI and agent commands auditing a git diff against YAML standards packs, scoring each hunk with Jev so agents fix only flags.
Agent skillbetaTypeScriptCoding-agent tools on Jev with public benchmarks: a strict command gate, a code-review router Action, and use cases.
MCP serverbetaTypeScript