User repository on GitHub
Prompt Claude Code: tre sistemi costruiti su Jev di TypeSafe — routing del modello, compattazione del contesto e code review
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Jev Claude Code
Three prompts, written in Italian, that you paste into a Claude Code session to build a Jev model router, a Jev-driven context compactor and a diff reviewer that emits a verdict, with nothing to install.
- Stars
- 7
- License
- MIT
- Published by
- Community
- Author
- DarioFontanel
- Last verified
- Oct 2026
About this project
jev-claude-code è una raccolta di tre prompt per Claude Code che ti permettono di integrare Jev, il modello decisionale di TypeSafe AI, nel tuo flusso di lavoro: scelta del modello, compattazione del contesto e code review. Ogni prompt si incolla in una sessione di Claude Code e costruisce il sistema completo nel tuo progetto, verifiche comprese, senza pacchetti da installare.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
Highlights
- The router picks model and effort on the first message of a session and keeps them, because changing either invalidates the prompt cache.
- Compaction asks Jev which tool calls are still needed and drops or truncates the rest, leaving your messages and the kept content verbatim.
- The reviewer sends a diff as 14 typed questions in one call and derives BLOCK, SECURITY REVIEW, NITS or MERGE from thresholds in a JSON file.
- Each prompt regenerates the type definitions from your own build and treats them as authoritative over its own text.
Watch out
Prompts 01 and 02 depend on function hooks, an early-access feature gated behind CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 on Claude Code 2.1.276 or newer, and the README says the interface can change between releases. TypeSafe API access is also early access, and the README says approval can take up to a day. Neither the router nor the compactor reports failure: with no key or no answer from Jev they fall back to native Claude Code behaviour, which can look like the prompt never ran. You are pasting code that generates code into your project, and the verification it offers is a transcript line per system. Prompt 03 needs Python 3.10+ and git, with no external dependencies. 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 paste. Each file in prompts/ holds a description, prerequisites and a full prompt that you copy into a Claude Code session and follow until it reports a summary; it then creates the files and runs the checks itself. There is no package to install. The prompts are independent, so you can take one, two or all three, and deleting a folder under .claude/skills/ switches that system off.
Both the router and the compaction hook hand control back to Claude Code's native behaviour, and the README states that as the design rather than an error path. The same applies when the key is missing. The consequence is that a broken setup looks like a no-op, so the README tells you to verify by restarting claude in the folder and looking for a [jev-router] line in the transcript, jev-compact on the first /compact, or a response to /jev-review.
No. It reports the verdict and asks for confirmation before any change, which the README lists under security along with reading the key only from the environment and never from a versioned file. The `/jev-review` skill has Claude restate the verdict and then read the code only at the points Jev left uncertain, so the expensive reading is targeted at the disagreement rather than the whole diff.
Instead of summarizing the conversation, it asks Jev which tool calls are still needed and removes or truncates the others, so your messages and Claude's stay identical and retained content is verbatim rather than paraphrased. Setup asks you for the context percentage at which to trigger it. Each subagent gets classified separately in the router because it runs with its own context.
Claude Code 2.1.276 or newer, because the router and compaction prompts use function hooks, an early-access feature switched on with CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 in project settings. A TypeSafe key from console.typesafe.ai, where access is early access and can take up to a day after the request. Python 3.10+ and git, but only for the code review prompt, which has no external dependencies.
A Python reviewer sends the diff to Jev with 14 typed questions in a single call and computes the verdict from that: BLOCK, SECURITY REVIEW, NITS or MERGE. The thresholds live in a JSON file the setup writes rather than being hard-coded, so the policy is something you can read and change without regenerating the reviewer.
Changing model or effort invalidates the prompt cache, so the router asks Jev on the first message of a session which of Sonnet, Opus or Fable to use and at what effort level, then holds that choice for the rest of the session. Each subagent is classified when it starts, because it runs with its own context rather than the session's.
More Jev resources like this
20GitHub stars
TypeScript SDK and CLI coding agent where Jev only picks options while your code runs effects, building programs as syntax trees.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
rhighs/jev-codeTypeScript
rhighs/jev-codeTypeScript25GitHub stars
Agent-oriented playbook for using Jev as a cheap, fast decision layer that triages, classifies, scores, and gates work before expensive browser, research, retry, or subagent steps.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
Bodila51/muse-jev-playbookPython
Bodila51/muse-jev-playbookPython13GitHub stars
Pi extension that classifies each task with Jev over OpenRouter and routes it to a Pareto-optimal model balancing measured quality, cost, and latency under a role policy.
Jev Agents & MCP#community#jevbest#catalog-import
16GitHub stars
Shell wrapper that intercepts unknown subcommands, reads the target CLI help output, and asks Jev which documented subcommand matches the intent behind the typo.
Jev Agents & MCP#community#jevbest#catalog-import
Community
syumai/jevyoumeanGo
syumai/jevyoumeanGo23GitHub stars
Coding-agent plugin that routes prompt classification, subagent tiering, edit rule checks, and fast Jev-backed compaction through small typed judgments.
Jev Agents & MCP#community#jevbest#catalog-import
Community
0x7067/claude-jevTypeScript
0x7067/claude-jevTypeScript12GitHub stars
Local OpenAI-compatible proxy that skips repeat model calls on same-intent paraphrases and retried agent turns, judged by Jev or a local System One model.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
kushals256/jevcacheTypeScript
kushals256/jevcacheTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.