Jev 工程化生态：6 个已落地项目盘点

Archive @ArchiveExplorer原推发布于 2026.09.26

在 X 上打开原推

JEV engineering is becoming the standard way serious teams make agents take decisions, and here is what people have already built with it. if you want to actually embed typed decisions into a system instead of forcing an LLM to improvise, copy this: fast-jev-compaction - a Claude Code plugin that replaces vague compaction summaries with JEV decisions. 6.9k stars, 415 forks, every tool call and result is evaluated in a single request, stale context is dropped or truncated, and everything that remains is kept verbatim > https://github.com/tamaratran/fast-jev-compaction typesafe-ai/skills - the official TypeSafe skill for Claude Code and agents through http://skills.sh. 2.2k stars, 124 forks, install it once and the agent starts building Choice, Score, and Noul questions instead of hiding decisions inside prompts > https://github.com/typesafe-ai/skills claude-jev - a Claude Code plugin that sends code review results, debugging hypotheses, architecture options, and search results to a second JEV judge. 6 stars, 5 MCP tools, around 150 ms per response, $0.042 per million input tokens, with thresholds stored in code instead of intuition > https://github.com/buchmark/claude-jev jev-engineering - a gate for tool calls in Claude Code, Codex, Cursor, and any other MCP agent. 2 stars, 2 forks, allow/ask/deny in a median 371 ms at $0.0000189 per check, plus a 300-injection test that shows exactly where the gate fails > https://github.com/eugeniughelbur/jev-engineering jev-tree - recursive Choice for taxonomies that do not fit into one flat question. 9 stars, default fanout of 32, up to 255 options per call, while a 320-leaf incident catalog lost its entire tail when truncated by the Choice limit until tree traversal recovered it > https://github.com/reachjalil/jev-tree open-jev - browser-native TypeScript, 36 stars, 8 forks, one state plus any number of typed questions, one forward pass, and no answer generation. It runs on-device through WebGPU, with WebAssembly as a fallback > https://github.com/nico-martin/open-jev jev-trader - 2.4k stars, 462 forks, one JEV decision per Monad block, roughly every 300 ms. It reads the Kuru MON-USDC order book, chooses buy or sell, and places a post-only limit order one tick inside the touch. A dry run with a mock model shows a 100 ms p50 > https://github.com/jarrodwatts/jev-trader none of these repositories will make a bad question good. They simply show where the decision layer is missing: when the agent summarizes instead of preserving, approves instead of verifying, or chooses a label without a threshold full build in the article, and before giving JEV access to production, run calibration and the injection test first

这条是引用着另一条写的：x.com/archiveexplorer/status/2103590048230695408