Email Ops

It lists every configured account and sorts each mail into ACT, WAIT, FYI or NOISE. It reads every mail in full, including the noise, gives a recommendation, asks one multiple-choice question per mail, saves a reply as a draft when you say so, and files handled mail into a folder.

Every mail text goes through the guard before the agent sees it. The guard asks Jev seven questions in one call: one overall question about whether the text contains an instruction aimed at an AI assistant that does not come from its own user, five hazard questions, and a harm score from 0 to 3. The verdict is printed as one line above the mail text.

No. The repository contains no code that can send mail, and a test checks the scripts for that. The draft command writes a reply into the Drafts folder with In-Reply-To and References set so it threads; you edit and send it in your own mail client.

The verdict is REVIEW when the overall Noul is 0.3 or more, any hazard Noul is 0.9 or more, or hidden characters are found. It is BLOCK when it is REVIEW and the harm score is 2 or more. These are the config defaults and can be changed per install.

The guard masks IBAN and BIC, VAT, EORI and UID numbers, common API keys and tokens, passwords including those inside URLs, private key blocks, and phone numbers. Masked text is text Jev never sees, so a mail with masked secrets counts as not fully checked and is never shown as OK.

Long mails are checked in pieces of 6,000 characters that overlap by half, so every passage of up to 3,000 characters lies whole in one piece. A cost cap per mail, default 0.02 USD, limits how much is checked. A partly checked mail is never shown as OK.

Spam, Junk and Trash are never read. They are refused by folder name in eight languages (English, German, French, Italian, Spanish, Dutch, Polish, Swedish) and by the server's special-use flags, which the script requests explicitly when the server supports that.