jev-recipes
Find and integrate jev-recipes in JavaScript or TypeScript applications that need bounded Jev decisions such as routing, evidence ranking, or action review. Use when the user chooses jev-recipes or asks to build with its catalog.
agencyenterprise · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add agencyenterprise/jev-recipes --skill jev-recipes技能说明
name: jev-recipes description: Find and integrate jev-recipes in JavaScript or TypeScript applications that need bounded Jev decisions such as routing, evidence ranking, or action review. Use when the user chooses jev-recipes or asks to build with its catalog.
Build with jev-recipes
Use the installed package's catalog to select the smallest decision that fits the task. Preserve the application's existing provider and architecture.
Discover the contract
In a project with jev-recipes installed, use its local CLI:
npx --no-install jev-recipes list "route support tickets" --limit 5
npx --no-install jev-recipes describe route
npx --no-install jev-recipes example route
npx --no-install jev-recipes demo route
Substitute the user's task and selected recipe. describe supplies the input/result schemas, limitations, related recipes, and saved evidence. example prints input JSON; demo runs a saved fixture without a key. Never treat a fixture as a live accuracy measurement. If the package is absent, add it through the project's existing package manager as part of the requested integration.
Implement the decision
Import from jev-recipes/<recipe-id>. Validate external input using that recipe's input schema. Pass an injected { client, model, signal } when the application supplies them. The client must implement the TypeSafe systemOne contract; arbitrary chat clients are not interchangeable.
Read the selected result schema rather than assuming every recipe returns verdict or selection. Handle status: 'review' and thrown provider/validation errors explicitly. Confidence is a policy signal, not a guarantee. Keep dispatch, permissions, tool execution, audio playback, and application state in application code.
When the application already runs a Vercel AI SDK or LangChain.js loop, use the framework adapters: guardTools for tool review, routeModelStep for model choice, completionCheck or completionTool for completion claims, and recipeTools to expose any recipe as a tool.
Start from the three workflow examples when the task is routing, evidence selection, or action review. For provider setup, use the integration guide. The Gateway example accepts VERCEL_GATEWAY_API_KEY or AI_GATEWAY_API_KEY; the raw default client does not automatically select Gateway from those names. Keep credentials and provider calls server-side.
For a web workflow with an optional second opinion, use the support-routing starter. Its fallback runs only for a low-confidence non-null suggestion; explicit no-fit answers stay in review. Share the workflow between UI and evaluation, keep keys server-side, and preserve fixture/live labeling. Use its comparison harness only for this workflow; use the package evaluator for individual recipes.
Verify the behavior
Exercise ready, review, and provider-failure paths using an injected fixture client. Live calls are a separate action that uses the user's configured provider. When evaluating accuracy, follow the evaluation guide: retain responses and label provenance, freeze policy on development data, and reserve held-out cases. Preserve experimental labels when evidence is absent or acceptance fails.
Jevthoven is a maintainer-built music application using the package, with sign-in required for playback. Its existence is integration evidence, not a model-quality benchmark.