nc://cjk-sanitizer

CJK Sanitizer: strips the Chinese and Cyrillic characters a Chinese-trained model leaks into its answer. A Nodecode rewrite of evanyudis/cjk_sanitizer by evanyudis.

Get it

Tell the agent "add the cjk-sanitizer cell", or run nodecode add nc://cjk-sanitizer.

Set it up

Nothing is required: enabled, it strips from the first answer. The cjk-sanitizer section of the config refines it — strip_chinese, strip_cyrillic, strip_punctuation, collapse_space, blank_lines (the most consecutive newlines kept, 2 by default), strip_reasoning (the thinking stream too, off by default). config.example.jsonc is the section written out.

Use it

Ask a model served by a Chinese provider something and watch its answer: where a Chinese-trained model drifts into hanzi mid-sentence, or into Cyrillic in a long reasoning chain, the characters never arrive. Nothing else about the answer changes — Korean Hangul, Japanese kana and every Latin word are as the model said them, and the strip happens on the stream, so a shell, a channel, a transcript and the history the next round reads back all see the cleaned text.

• /cjk reports the settings and what this process has stripped — deltas read, characters taken out, deltas dropped whole.
• /cjk off pauses it until /cjk on or a restart; /cjk reset clears the counters.
• The tool cjk_strip cleans text handed to it: cjk_strip with the text and, optionally, which scripts to take out.

What it does not do: it does not touch the text of a tool result, a recalled memory or a file the model read — only what the model itself says. A leak the provider splits across two fragments is cleaned per fragment, and the whitespace tidying is per fragment for the same reason. A Japanese answer that uses kanji loses them too: the CJK unified ideographs are the block Chinese and Japanese share, which is why the original's own "Japanese is preserved" claim only ever held for kana.

What is in it

License

MIT. The original is Copyright (c) 2026 Evan Yudistira: see LICENSE and NOTICE.