User repository on GitHub
Judge agent steps — ok / retry / escalate / stop via TypeSafe Jev
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Harnessjudge
Paste a trace of an agent step, get one of four dispositions back: ok, retry, escalate or stop. Aimed at agent builders putting a supervision step around an agent loop, with the judging powered by TypeSafe Jev and a two-command setup.
- Stars
- 0
- Language
- TypeScript
- Published by
- Community
- Author
- ndolinschi
- Last verified
- Oct 2026
About this project
After an agent step (paste a trace) → ok / retry / escalate / stop — for agent builders, powered by TypeSafe Jev.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- Input is a trace you paste after an agent step finishes, not a live hook, so nothing runs mid-step.
- Output is one of four labels: ok, retry, escalate, stop.
- Setup is npm i && npm run dev, with no config file, environment variables or API key described.
- Framed for agent builders, so the judgement call sits with TypeSafe Jev rather than in the agent's own prompt.
Watch out
The README is 168 characters: a one-line description, the four output labels and the npm i && npm run dev command. Not documented: which Jev question types or schema the judge uses, how a trace is parsed or how large a trace may get, what separates retry from escalate from stop, whether state carries between pastes, any model or provider configuration, how the Jev API key is supplied, tests, or a license. Because the trace is pasted by hand, nothing in your agent loop is automatic: you are the hook, and nothing here shows how the four labels should map to control flow in a real runner. 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.
You are the integration. The README describes pasting a trace after an agent step, with no hook, CLI mode, server or API endpoint documented. There is no automatic interception, so wiring it into a running agent, and deciding when a step is finished, is work the project does not do for you.
The README does not define them. It lists ok, retry, escalate and stop as the four outcomes of judging a step and stops there, with no guidance on what each implies for the caller, how many retries are reasonable, or who handles an escalation. Picking semantics for those labels is left to you.
It does not say. The entire setup section is npm i && npm run dev, with no API key, base URL, environment variable or config file mentioned, so how the TypeSafe Jev call is authenticated and pointed at a backend is undocumented, as is which model or question schema it uses. Nothing about judging a step is specified beyond the four output labels either.
Scope is stated as after an agent step, with a trace as the input. There is no documented trace format, no size limit, no truncation behaviour and no handling for multi-step or multi-agent transcripts, so a long session trace is unverified territory and probably the first thing to test.
npm i && npm run dev. The README documents no config file, environment variables, or API key, and there is no license, test suite, or trace format described.
It is cataloged as TypeScript. The README itself names no framework, runtime version, or dependencies beyond the two setup commands.
It is cataloged with zero stars and a 168-character README, with no license, tests, screenshots, or changelog, so its completeness is unknown.
More Jev resources like this
2GitHub stars
SIEGE puts a live support agent with seven real tools in front of 200 attackers; every proposed call passes a TypeSafe System One action gate, and a defender loop rewrites the gate from each breach it observes.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
vnmoorthy/siegeTypeScript
vnmoorthy/siegeTypeScript22GitHub stars
Independent open-source local decision model inspired by Jev, with a spark-s1 family plus smaller control models for agent gates, served locally with probabilities and policy actions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
abhishek085/open-spark-jevPython
abhishek085/open-spark-jevPython3GitHub stars
Browser agent where Jev makes every navigation decision and agent-browser performs clicks and typing, stopping when the named goal page is reached or blocked.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
kyrylosyzonenko/jev-browseJavaScript
kyrylosyzonenko/jev-browseJavaScript2GitHub stars
Next.js spectator demo in which two independent Jev agents drive separate Snake browsers through same-origin iframes, clicking indexed direction buttons only, with each board state streamed to a server-only endpoint that returns the chosen target.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
raihankhan-rk/jevarenaTypeScript
raihankhan-rk/jevarenaTypeScript2GitHub stars
An agent tool and MCP call gate: paste a planned call plus its context and it returns allow, ask_human or deny with risk scores derived from a TypeSafe Jev decision call.
Jev Demos & Experiments#community#jevbest#catalog-import
Communityndolinschi/toolgateTypeScript
ndolinschi/toolgateTypeScript2GitHub stars
An MCP stdio proxy in front of the official Playwright MCP server that triages page state, masks prompt injection, trims oversized page dumps by goal and risk-gates actions, exposing the same tools and schemas.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
krw82/jev-playwright-mcpTypeScript
krw82/jev-playwright-mcpTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.