10 个把 Jev 接进 Opus 编码流程的开源仓库

Zero @zer0point_eth原推发布于 2026.10.13

在 X 上打开原推

HOLY SH*T! 10 Jev repos that turn Opus 5.5 into a cheaper, safer coding setup Opus 5.5 builds. but you don’t want it burning tokens on every small decision around the build. these 10 repos hand those decisions to Jev, a fast, cheap model that only answers bounded questions: which model should take this task, did the work really get finished, does this edit break a team rule 01 model-router → hooks into subagent spawning in Claude Code and asks Jev how hard the task is: lookups go to Haiku, bounded judgment to Sonnet, open-ended work to Opus. the default mode only logs its pick, auto mode actually sets the model https://github.com/Akurganow/model-router 02 effortless → a bar above the Claude Code prompt that sets reasoning effort for every message and shows context fill plus a prompt-cache countdown, with one-click handoff or compact. Haiku is the default judge, Jev is an optional faster one that Haiku backs up when it’s unsure https://github.com/HeyCubit/effortless 03 jev-lint → after every edit by Claude Code or Codex, a hook checks the new lines against your team’s rules in about 0.3 s, things like tests that can never fail or user IDs trusted from the client, so the agent fixes them while the file is still open https://github.com/ckorhonen/jev-lint 04 Jevris → a local control layer that sits next to Claude Code, Codex, Kilo, OpenCode or Antigravity. facts, permissions and verification stay in plain code, Jev only answers bounded questions, and a task counts as done only when independent evidence says so. release candidate, install from source for now https://github.com/CryptVenture/Jevris 05 visual-proof → for Vite apps: a watcher re-shoots the screens you touched on every save, and finish packs the stills (plus a replay video if ffmpeg is installed) into a proof block, failing if a changed screen has no clean frame. optional model checks compare the stills with the claim you wrote, and Jev prunes routes and writes captions https://github.com/communal-hub/visual-proof 06 CloakBrowser Agent → give it a goal in plain language: Jev picks each click or keystroke in about 0.3 s, a stealth Chromium performs it with human-like input, and the coding agent gets back only the relevant page sections as markdown. MCP server, CLI and Python API, CAPTCHAs not handled https://github.com/CloakHQ/CloakBrowser-Agent 07 jev-doc-search → no vector DB: PageIndex turns a long PDF into a tree of sections, and Jev picks one branch per level until it lands on the pages that answer the question. tested on a 93-page and a 318-page annual report https://github.com/VectifyAI/jev-doc-search 08 ex-regex → a Python library shaped like re, except the match test is meaning: sub, findall and extract take a plain-English description. a finder proposes spans, Jev scores them, and every result is a verbatim slice of your input with offsets https://github.com/nothans/ex-regex 09 anofox-decide → a DuckDB extension that adds decide_probability, decide_choice and decide_table to SQL, so you can score or classify rows with Jev. remote calls stay off until you opt in, and local models can run fully offline https://github.com/DataZooDE/anofox-decide 10 jevals → agent evals and guardrails from a single request: the quickstart packs six checks (tool choice, use of tool results, grounding, scope, indirect injection, PHI) into one call, and the same checks can act as gates that allow, escalate or block a tool call. alpha https://github.com/openlayer-ai/jevals the one repo here with a clean before/after is jev-lint. its author ran the same tasks with the hook off and on: Claude Code, 96 runs → 2.38 rule violations per task without the hook, 1.21 with it Codex, 48 runs → 1.67 without, 0.92 with it that’s roughly half the violations, for about $0.04 extra per task. a violation caught at the edit costs a few seconds. the same one caught in review costs a whole fix and re-review round that’s the job for the cheap fast model: check the edit while the agent still has the file open keep Opus 5.5 for building and hard reasoning. let Jev answer small, bounded questions, and let plain code enforce the result and handle anything the model isn’t sure about pick the repo that removes your biggest bottleneck

这条是引用着另一条写的：「Jev 给 Claude Code 加决策与安全闸门」