User repository on GitHub
Claude Code Stop hook that checks an AI assistant's claims against what it actually read this session, using TypeSafe's Jev as the judge
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Clear Head
Claude Code Stop hook that checks each factual claim in a finished answer against the tool output read during that session, blocking the turn on claims the evidence contradicts or does not support.
- Stars
- 2
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- VladyslavHontar
- Last verified
- Oct 2026
About this project
A Claude Code Stop hook that checks the factual claims in an AI assistant's answer against the evidence it actually read that session, using TypeSafe's Jev as a fast, cheap, calibrated judge. If a claim is contradicted by, or unsupported by, what was read, the turn is blocked with the specific claims listed, and the assistant has to verify them or mark them as assumptions before finishing.
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- Sends only the tool-output lines sharing keywords with each claim, plus paths, commands and doc comments; every payload is logged to sent.jsonl.
- On 1,309 replayed claims a local Kev-4b judge matched the hosted one on 61% of supported and 75% of not-addressed verdicts, 12% on contradictions.
- Backends are explicit: the hosted API by default, or a local Kev server on 127.0.0.1 needing roughly 10 GB of RAM and an 8 GB first-run download.
- Fails open if the checker itself breaks; a dead local server blocks once with its start command rather than passing checks silently.
Watch out
Speed is the real cost of the local backend: an 18-sentence answer took 44 s end to end against about 1 s for the hosted path, and each call is prefill-bound at roughly 1.5 s per claim. Excerpts of your code and session still leave the machine for the hosted API, which the author says is disqualifying for a repo you would not hand to a third party. Kev also has known false positives, such as a negated word in the excerpt reading as a contradiction at 0.98. Thresholds ship as defaults tuned to one project's evidence shape. 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.
Most turns pass silently. A block lists the specific contradicted or unsupported claims and asks you to verify them or mark them as assumptions. If the checker itself throws — network down, bad key, malformed input — it fails open by default so real work is never blocked over a broken checker, and you can invert that with JEV_FAIL_CLOSED. One shell can be exempted with JEV_HOOK=off.
With the hosted API, an 18-sentence answer took about a second. With the local Kev backend it took 44 s, because each claim is a separate call at roughly 1.5 s plus about 10 s for the sentence-classification pass, and batching all claims into one state gave near-identical verdicts while overrunning the row limit. The installer registers a 180 s timeout for that backend instead of 60.
Per claim, only the tool-output lines that share keywords with that claim, plus the list of file paths and commands used and any doc comments found in what was read — not whole files. Every payload is appended to sent.jsonl next to the script so you can inspect exactly what was transmitted. The author is explicit that you should not install it on a repository you would not hand to a third-party API.
Not as a copy. Replaying 1,309 claims with the hosted judge as reference, Kev-4b agreed on 61% of supported, 75% of not-addressed and 12% of contradicted verdicts, and a manual review found the two wrong about equally often. Treat them as different judges. The 0.8B variant is 5x faster and unusable, having called an off-topic excerpt supported with 0.90 confidence.
Clone the repository and run ./install.sh to install for this user across every project, or ./install.sh --project for the current project only. The installer prompts for a TypeSafe API key unless TYPESAFE_API_KEY is already set, and verifies the key before finishing.
./install.sh --kev clones Kev, runs uv sync --extra serve, starts the 4B server on 127.0.0.1:8009 and persists VERIFIER_BACKEND=kev. It needs uv, git, roughly 10 GB of RAM while serving and a first-run download of about 8 GB, and the server does not survive a reboot.
Claims about mutable outside state such as a PR's status, CI, a running process or a remote branch block as [STALE] when the freshest matching tool output is JEV_STALE_TURNS old or missing, regardless of the judge's verdict. It came from a real miss where PRs were described as open when they were not.
More Jev resources like this
9GitHub stars
Hermes Agent plugin exposing Jev as four coding-workflow tools for parallel judgments, binary checks, option routing and rubric scoring, with a bundled question-design skill.
Jev Agents & MCP#community#jevbest#catalog-import
Community
ajensenwaud/hermes-jev-pluginPython
ajensenwaud/hermes-jev-pluginPython3GitHub stars
Hermes plugin that puts an agent's plans, drafts and claimed completions up for review — plan_review, evidence_review, action_verify — plus local approval rules that make no model request.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
bojansandhaus/jev-decisionsPython
bojansandhaus/jev-decisionsPython1GitHub stars
An OpenAI-compatible proxy that spends one Jev call per request on injection, harm and PII checks plus a tier label, then routes to a small or large backend from pure code.
Jev Agents & MCP#community#jevbest#catalog-import
Community
aniruddh-krovvidi/switchboardPython
aniruddh-krovvidi/switchboardPython0GitHub stars
Decision Graph Protocol from Numerous ApS: the app publishes an immutable frame of evidence and offered decisions, an eligible assessor returns a typed assessment, and the app's own guards authorize the commit.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
numerous-com/dgpPython
numerous-com/dgpPython51GitHub stars
Local LinkedIn recruiting workspace on Browser Use's Jev Ultrafast: Jev screens titles, drives navigation, and assesses profile excerpts against your brief while a live browser stays visible; matches and evidence export as JSON.
Jev Agents & MCP#community#jevbest#catalog-import
Community
skeptrunedev/jev-recruiterPython
skeptrunedev/jev-recruiterPython33GitHub stars
Codex Skill plus stdlib-only Python CLI that recommends which installed skill to read next for a task, allowing no-skill or uncertain answers and never executing candidates.
Jev Agents & MCP#community#jevbest#catalog-import
Community
win4r/jev-skill-suggesterPython
win4r/jev-skill-suggesterPythonFrom the Jev community
Posts from builders shipping with Jev right now.