Jev Chat Windows — caizili999

Drafting defaults to DeepSeek, OpenRouter, or any OpenAI-compatible address. Only the OpenRouter Jev-kernel tier returns calibrated win probabilities; draft-model judging gives summaries without probabilities, and off skips judging.

It screenshots only your own visible WeChat window with local OCR, never hooks the process or reads its database. Export and knowledge history stay on disk, and idle minutes with no new messages make no model calls.

Open the three separate switches for direct messages, mentions, and unmentioned group chatter, plus per-window authorization, send-key match, input-empty checks, and idle detection. Every block names its reason in status and logs.

Needs Windows 10 1903 or 11 with WeChat Windows 4.x, and unsigned exe builds trip SmartScreen. The packaged UI widgets impose the GPLv3 commercial-license condition on the zip, not on source runs.

Nothing unless the other party sends a new message. Full mode is one drafting call plus one Jev judgment; the self-judge mode is one drafting call plus one judgment that reuses the drafting key; choosing no judgment is the drafting call alone. Ten minutes with nobody talking means ten minutes of zero calls. Thinking mode ships off, because drafting three sentences does not need it and it is several times slower and dearer.

Per-conversation CSVs written under the program's own 聊天记录/ directory, layered as 会话名/日期.csv, with four columns: 时间, 方向, 发送者 and 内容. The 方向 column splits into 对方, 我 and 程序, so a message the program auto-sent is marked separately when you read the log later. Writes are batched at 20 messages or 5 seconds, plus one on exit, so a forced kill loses at most the last few seconds. Nothing leaves the machine and the directory is in .gitignore.

A note hits when any of its tags or its title appears in the conversation title or in the last 6 messages, and at most 5 notes are carried, newest first. The body does not participate in matching, so the words likely to appear in a conversation belong in the tags. Contacts hit when the conversation title equals the name or any alias, ignoring case, group member-count suffixes and zero-width characters, so 测试群(12) and 测试群（12） resolve to the same person. A note marked 常驻 is carried every time.