JevLint-LE

Ask the model itself, TypeSafe's Jev or OpenAI's Luna, to check the questions in the given text or paths after linting them: overlapping options, a question that needs counting, a label that does not match its description, and the other checks list_rules marks runsHere: false. This sends the questions to the vendor with the user's own API key and costs money, so call it only when the user asks for a check with Jev or Luna, never on every edit, and use plan_jev first when the cost matters. The key is read from TYPESAFE_API_KEY or OPENAI_API_KEY in the environment the server was started with, or from .env.local or .env in its working directory, and is never taken as an argument. Returns the lint report with the model's findings added, each with the probability the model gave, what was sent and answered, and why the check fell short when it did. Bounded by maxCalls.

A rule's own page, as Markdown: what it catches, an example that is flagged with the message the linter gives, one that is not, how to fix it and how to silence it. Use it before making a fix that fix_text leaves to you, or before silencing a finding.

Apply the safe fixes to a request body or source code passed as text, and return the mended text with how many findings were mended and the report of what is left. A safe fix only mends what the API would refuse: a criteria key written yes where it takes true, a mistyped question type, criteria in the wrong shape. A fix that changes what a working request does, such as adding a fallback option or pinning a model, is never applied here. It stays in the report with its edit marked safe: false, for you to make or to leave. Nothing is written to disk and nothing is sent over the network.

Lint the Jev and OpenAI Decisions questions in files or directories on disk. A directory is searched for the file types the linter reads. Returns the same report as lint_text, one entry per file. It sends nothing over the network.

Lint questions written for a decision model before they are sent: TypeSafe's Jev (System One, /v1/systemone) and OpenAI's Decisions API (gpt-6-luna, /v1/decisions), in either request shape, or the Vercel AI SDK's decide() for both. Pass a request body as JSON, or source code that builds one. Returns each finding with its rule, message, position and docs link, the edit that would mend it when there is one, plus how many questions were found and how many could not be read in full. Use it after writing or changing a question. It sends nothing over the network.

List every rule with its code, name, what it catches, default level, its own page with an example, the vendor page behind it, and whether it runs offline here or only when check_with_jev asks the model.

Say what check_with_jev would send for the given text or paths, and send nothing: how many requests, to which model, about how many input tokens, how many questions are held back for having a part built at runtime, and how many fall past maxCalls. Needs no key. Call it first when the user wants to know the cost of a check before paying for it.

Send one question to Jev several times, as written and with its layout changed, options reversed, labels hidden, levels reversed, criteria removed, to see whether the answer holds. Returns a report as Markdown with a verdict and which variants moved the decision. The text must carry the state the question is asked against, and when it has more than one question, questionId names the one. Asks Jev only, about a dozen requests, with the user's TypeSafe key found the way check_with_jev finds it. It costs money, so call it only when the user asks to probe a question.