super-jev
Decision-to-action harness that puts Jev between an agent and your data to find the file, check the claim and permit actions.
Overview
Super Jev puts one judge, Jev, between a coding agent and your data. The agent asks in its own words; Super Jev ranks and finds the relevant file, checks the claim against it, permits or refuses the action, and remembers what you approved, while the agent stays responsible for the answer. The motivation is that agents burn whole LLM turns on lookups, re-hunt the same answers daily, and state things files never said, whereas Jev is a fast, cheap judge for yes/no and which-one questions.
The daily loop is ask, read the top file, answer, then approve, mark a miss or add a missing fact. Connectors bring data in, and the cache fills only from your approvals: when the same file wins the same question twice on complete searches and passes the content check, the ranked list is saved, with secret scanning and stale-file withholding. Version 1.0 promises skill search, file navigation, connect and bulk prepare, a check gate, a permit gate and the harness loop, and explicitly does not promise unattended answering or semantic cache matching. It integrates with Claude Code via a retrieval rule card and a Stop-hook claim gate, and documents exactly what text leaves the machine.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which file best answers this question? | — |
| Yes / no | Does the evidence file support this claim? | — |
| Yes / no | Is this action permitted? | — |
How it works
- Agent asks a question in its own words
- Jev ranks files by their descriptions and confirms the top file
- Check or permit gates judge claims and actions
- Approvals populate the cache for repeat questions
Strengths
- Cache only from human approvals
- Stale and secret-bearing files withheld
- Offline demo without a key
Caveats
- Edges still want an agent in the loop per the README
- File text and descriptions go to TypeSafe
Getting started
npm run jev to open the app, or npm run demo for an offline run; see docs/GETTING-STARTED.md.At a glance
- Type
- Library
- Status
- beta
- Language
- TypeScript
- Runtime
- CLI/terminal app + Python skill scripts
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Decision harness: turns a Jev answer into a bounded action instead of leaving the caller to interpret it.
yibie/awesome-jevA TypeScript decision-to-action harness connecting evidence, Jev judgments, permitted actions, and outcome records.
logicrw/awesome-jev-projectsA small, extensible decision-to-action harness for TypeSafe Jev.
hellogumbo/awesome-jevPuts one judge between an agent and your data: it finds the file, checks the claim, and permits or refuses the action.
MrJev/awesome-jev
Frequently asked questions
What does Jev decide in Super Jev?
Which file answers a question, whether a claim is supported, and whether an action is permitted.
What gets cached?
Only ranked file lists you approved or that won repeatedly and passed checks.
Does it answer for the agent?
No; the agent stays responsible for the answer.
Related projects
Small stdio MCP server with ask, rank and verify tools backed by Jev, a local event log and installers for several agent harnesses.
MCP serverbetaTypeScriptFree Claude Code skill where Jev sorts and Claude writes: classify, pre-route items so Claude reads only unsure ones, and gate tool calls.
ProjectbetaPythonVerified computer-use agent: a Jev guard checks claimed completions and gates irreversible actions; Jev also takes simple steps.
ProjectbetaPythonSkill pack handing an agent's small decisions to Jev: model routing, skill choice, search, memory, compaction and computer use.
Agent skillbetaPythonClaude Code PreToolUse hook that auto-approves shell commands Jev is confident are read-only, and otherwise shows the normal prompt.
HookbetaPythonClaude Code plugin adding typed Jev decisions to hooks: task kind, command risk, failure class, completion; never grants permissions.
PluginbetaJavaScript