User repository on GitHub
Clean, traceable state briefings for TypeSafe's Jev model
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jevbrief
A pip-installable Python package that sits between a noisy source and Jev: ten adapters condense logs, diffs, tool lists, traces, alerts and screens into small state with a reason for every drop, then record a replayable trace.
- Stars
- 6
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- Parthkomalwad
- Last verified
- Oct 2026
About this project
Other tools show what Jev decided. jevbrief shows what Jev was told, what it wasn't told, and why. Documentation · Quick start · Adapters · Benchmarks · Contributing.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Ten adapters ship in the base package: tools, steps, ci, pr, otel, json and android need no extras; web and nes need [web] or [nes].
- Every drop carries a reason code and the decision is recorded, so `jevbrief ask --view` replays what Jev saw rather than asserting it.
- `select_tools` ranks locally with no API key, `pick_tool` spends one Jev call, and `check_progress(history, goal).stuck` flags a circling agent.
- The README's own table shows pr accuracy unchanged at 55% and vulns accuracy falling from 63% to 59%, below always answering no.
Watch out
The benchmark table needs reading with its footnotes: three runs per task, median input tokens, and several sets the author built alongside the adapters rather than collected independently. Two results are regressions the README states plainly, including the flaky-error question dropping from 93% to 67%. Ten adapters is a lot of surface for one package, so depth per adapter is worth checking against your own source before committing. It is a community project and not affiliated with TypeSafe. 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.
The token reduction is large and easy to check: ci drops from 30,502 to 988 input tokens and otel from 29,678 to 1,070. Accuracy moves independently, and not always up. pr held at 55% while tokens fell from 2,599 to 1,652, so on that adapter the only win was cost. The README reports both arms using the same Jev and the same question, with the raw arm sending what a naive integration would send.
Yes, and the README says so. On vulns, jevbrief scored 59% against 63% for raw alerts, and below the 71% you get from always answering not exploitable. It did find 76% of exploitable cases against 32%, but with twice the false alarms. The README also notes the synthetic sets were built alongside the adapters, which is the caveat that applies to every number in the table.
Adapters keep what matters, drop the rest with a reason for every drop, send one clear question to Jev, and record the whole decision. That record is what `--view` replays, and the contributing guide asks for a trace attached to any wrong-decision report. A worked example in the README turns 1,395 raw log lines into five groups and has Jev pick the config reload failure.
Not for the ranking. `select_tools(tools, goal)` is described as local, free and needing no API key, returning a ranking. `pick_tool` is the call that spends a Jev request and returns the pick plus confidence, and `ask` is the one that calls Jev and can replay with `--view`. `inspect` also needs no key and costs nothing, which is the sensible first run on a source you do not trust yet.
pip install jevbrief brings the tools, ci, pr, otel, json, android, and vulns adapters. Add [web] for Playwright page reading, [nes] for the NES adapter, or [all] for everything. Set TYPESAFE_API_KEY in the environment or a .env file; inspect needs no key and costs nothing.
select_tools ranks tools locally with no API key, and inspect drops noise with a reason code without calling Jev. ask, pick_tool, and check_progress call Jev and consume budget. The tools can be your functions, MCP servers, LangChain, CrewAI, OpenAI, or Anthropic tools.
Both arms use the same Jev model and the same question, with three runs per task and median input tokens. The raw arm sends what a naive integration would send; for example the ci adapter went from 88% to 100% accuracy while input tokens dropped from 30,502 to 988.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython63GitHub stars
Claude Code kit that guards PreToolUse calls with a free pre-filter plus one Jev question for ambiguous cases, and adds sub-agent sizing, rechecks, search, daemon speedups, and monitoring.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
jonathanavis96/jev-kitPython
jonathanavis96/jev-kitPython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPythonFrom 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.