10 个 Jev + Opus 5.5 编码 agent 开源仓库合集

Mika @mika_systems原推发布于 2026.10.09

在 X 上打开原推

this Jev + Opus 5.5 toolbox is f*cking worth stealing a 19-year-old developer from China launched 200 coding AI agents with Opus 5.5 doing the builds and Jev screening their reports. these 10 repos cover the routing, evidence checks and review behind that kind of setup 01 model-router → Jev scores each subagent task; auto mode picks Haiku, Sonnet or Opus before the worker starts https://github.com/Akurganow/model-router 02 effortless → with Jev enabled, each prompt gets an effort decision; uncertain calls go to Haiku, and the bar shows context and cache status https://github.com/HeyCubit/effortless 03 jev-lint → checks newly edited code against team rules, including tests that cannot fail and client-supplied user IDs https://github.com/ckorhonen/jev-lint 04 Jevris → tracks tasks, routing advice and verification receipts beside Claude Code; approved checks produce the evidence for completion https://github.com/CryptVenture/Jevris 05 visual-proof → captures changed Vite screens and uses Jev to assess whether their visible text supports the agent's completion claim https://github.com/communal-hub/visual-proof 06 CloakBrowser Agent → a browser MCP where Jev chooses the next control and action, then returns relevant page sections to the coding agent https://github.com/CloakHQ/CloakBrowser-Agent 07 jev-doc-search → Jev searches a long PDF through its section tree, then selects pages likely to answer the question https://github.com/VectifyAI/jev-doc-search 08 ex-regex → finds and extracts text by meaning; every match points to an exact span in the original input https://github.com/nothans/ex-regex 09 anofox-decide → puts Jev decisions inside DuckDB queries: classify rows, choose a category or score text without generating an answer per row https://github.com/DataZooDE/anofox-decide 10 jevals → checks tool choice, grounding, completeness and scope together; its quickstart packs eight evaluations into one Jev request https://github.com/openlayer-ai/jevals one repo in this list comes with measured results: jev-lint’s author tested the hook across 96 Claude Code runs > without the hook → 2.38 rule violations per task > with the hook → 1.21 > about 49% fewer violations in that experiment that's a concrete job for the cheaper model: check the edit while the coding agent still has the file open keep Opus 5.5 on implementation and difficult reasoning. give Jev bounded questions, then let code enforce the checks and handle uncertain answers pick the tool that fixes a problem in your workflow take a look at what it looks like ↓

这条是引用着另一条写的：x.com/mika_systems/status/2101686148338798610