User repository on GitHub
Claude Code Stop hook that blocks an unverified done: reads the transcript for evidence, asks Jev once, fails open on everything else
- Forks
- 6
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Jev Belay
A Claude Code Stop hook that reads the turn transcript locally and spends one four-question Jev call only when files changed with no passing check since, blocking unverified done claims.
- Stars
- 39
- Language
- JavaScript
- License
- MIT
- Reading time
- 4 min
- Published by
- Community
- Author
- valentynkit
- Last verified
- Sep 2026
About this project
Claude Code says "Done, tests pass." jev-belay checks whether anything ran before it lets that stand. A Stop hook. It reads the turn's transcript locally, and only when files changed and no check has passed since does it spend one Jev call, four questions, $0.00005, on whether the closing message is an unverified "done". If it is, the turn does not end: Claude gets the reason and goes back to run the suite. Every other stop costs nothing. Every error path lets the turn end.
Quoted from the project's README. Read the full README.
Highlights
- Measured AUROC 0.976 for telling a false done from an honest one over 100 labeled stops, versus 0.777 for wording alone.
- One Jev call covers four questions for $0.00005, with 1,222 median input tokens and output free.
- At the shipped 0.70 threshold it blocks 8 turns in 100, 7 of them rightly, and Jev answers in 346 ms at the median.
- Only 17.7% of stops reach the question; every other stop costs a local transcript read and no network.
- A turn with no file edits never reaches the question, so a read-only tests-pass claim passes through.
Quickstart
json
{
"env": { "TYPESAFE_API_KEY": "YOUR_API_KEY" },
"hooks": { "Stop": [{ "hooks": [{ "type": "command", "command": "node ~/src/jev-belay/belay.mjs", "timeout": 25 }] }] }
}Watch out
MIT-licensed. Needs Node 20+, Claude Code 2.1.196 or newer, and a TypeSafe API key; every error path lets the turn end.
Questions about this entry
Short answers drawn from the linked project or page.
That rule has been measured: limpet published a calibration run over 1,500 real stops where it came out at AUROC 0.50, a coin flip, and was left in shadow mode. jev-belay establishes the run facts first with two regex belts over the transcript slice since your last prompt, then asks four questions in one call. On 100 labeled stops that reaches AUROC 0.976, against 0.777 for judging the wording alone.
Seven of those eight were right, and the blast radius is capped: one block per prompt, none within 60 seconds of the last, three per session, so a wrong verdict costs you one extra test run rather than a loop. The honest limit is stated up front: a turn with no file edits never reaches the question, so a read-only "confirmed, tests pass" sails through. Every error path lets the turn end.
No. Only 17.7% of stops reach it, 477 of 2,694 in one real corpus; the rest is a local transcript read with no network. A stop that does reach the question spends one call with four questions for $0.00005 at $0.042 per million input tokens with output free, with 1,222 median input tokens and 346 ms median, 433 ms p90. A thousand checks cost five cents.
It needs Node 20+ and Claude Code 2.1.196 or newer for prompt_id. The closing message is read off the Stop payload where the host sends it, 2.1.263 does, and from a transcript re-read otherwise; older versions fall back to a session-keyed guard that allows one extra possible block per session. JEV_MODEL is pinned to jev-1.13.0 specifically because aliases move.
One request per stop that clears the evidence gate: the model name, the task, and the closing message. Claude Code's transcript slice since your last prompt stays local, with subagent transcripts folded into the turn, and JEV_BELAY_LOG=1 only writes the decision log to ~/.claude/belay/decisions.jsonl on your machine. Without a key the hook exits 0 and does nothing, so a missing key never blocks work.
As a Claude Code plugin: `/plugin marketplace add valentynkit/jev-belay` then `/plugin install jev-belay@jev-belay`, which asks for the TypeSafe key and keeps it in the Keychain. By hand it is one Stop hook line in ~/.claude/settings.json with TYPESAFE_API_KEY in an env block beside it. To see the block path without a key, run `node tools/fake-jev.mjs --port 4321` and point JEV_BASE_URL at it.
Both. The live view shows every verdict as it lands, including the stops the gate let through on a passing check: `node belay.mjs watch`, `node belay.mjs stats` for verdicts, calls, cost and latency, and `node belay.mjs last` for one decision in full. Shadow mode runs the whole pipeline and still lets the turn end, printing only that jev-belay would have blocked this turn.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Vercel's fx safety reviewer, measured at 18× faster
Reddit: Early experiments wiring Jev into Pi as a safety layer and model router.
YouTube: Ray Amjad puts Jev inside an agentic coding loop and looks at what it costs to run.
YouTube: RoboNuggets puts Jev to work inside Claude Code.
More Jev resources like this
A security hook for coding agents that asks Jev three typed questions before every tool call and scans tool results for prompt injection, with adapters for Claude Code, Codex, Copilot CLI, Gemini CLI, Cursor, pi, OpenCode and ACP.
Jev Tools & CLIs#community#javascript#agents
Community
leepokai/jev-guardJavaScript
leepokai/jev-guardJavaScriptA hybrid coding harness for Pi where any LLM writes the code while Jev returns typed probabilities that deterministic policy code turns into allow, confirm, block, warn, or steer.
Jev Tools & CLIs#community#typescript#agents
Community
AbdelStark/bicameralTypeScript
AbdelStark/bicameralTypeScriptA Jev-guided context compaction library for coding agents: instead of summarizing old context, Jev decides which tool calls and results still matter while user and assistant text is kept verbatim.
Jev Tools & CLIs#community#typescript#compaction
A warden for Claude Code and Codex CLI that keeps an append-only ledger of what an agent did, refuses a done claim when no check passed after the last edit, and uses optional Jev Noul judgments.
Jev Tools & CLIs#community#typescript#claude-code
Community
qkal/CannyTypeScript
qkal/CannyTypeScriptA Pi extension that turns constraints stated in conversation into a scoped, replayable policy and checks every side-effecting tool call against it before it runs, with Jev only classifying how each message changes the policy.
Jev Tools & CLIs#community#typescript#guardrails
Community
Nyarlathoteppppp/pi-heedTypeScript
Nyarlathoteppppp/pi-heedTypeScriptCLI and GitHub Action that turn project-defined semantic preferences from AGENTS.md into Jev Noul or Choice checks over code changes, then map each result to an advisory or blocking outcome. Checks can be tuned on labeled diffs.
Jev Tools & CLIs#community#javascript#cli
Community
doeixd/jev-prefJavaScript
doeixd/jev-prefJavaScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.