CJK Sanitizer — Hermes Agent Plugin

Strip unwanted Chinese (CJK unified ideographs) and Cyrillic characters from LLM output. Korean Hangul and Japanese Kana are intentionally preserved for legitimate use.

Why

If you run LLMs from Chinese providers (MiniMax, DeepSeek, GLM, Qwen, etc.), you've seen it: Chinese characters leaking into output mid-sentence, sometimes subtly, sometimes page-long blocks of gibberish. It's a known quirk with models that have significant CJK training data bleed.

This plugin hooks into Hermes Agent's transform_llm_output lifecycle and strips those characters silently before they reach your terminal or chat.

What It Strips

Script	Range	Stripped?
CJK Unified Ideographs (Chinese)	\u4E00-\u9FFF	✅
CJK Extension A	\u3400-\u4DBF	✅
CJK Radicals	\u2E80-\u2EFF	✅
CJK Compatibility Ideographs	\uF900-\uFAFF	✅
Cyrillic (Russian, etc.)	\u0400-\u04FF	✅
Korean Hangul	\uAC00-\uD7AF	❌ Preserved
Japanese Hiragana	\u3040-\u309F	❌ Preserved
Japanese Katakana	\u30A0-\u30FF	❌ Preserved

Why the name "CJK" then? The regex targets the CJK unified ideograph block which is shared across Chinese/Japanese/Korean, but the output filtering only strips what looks like a glitch — Chinese characters and Cyrillic. Genuine Korean and Japanese text passes through untouched.

How It Works

Hermes Agent has a plugin system with lifecycle hooks. This plugin registers a transform_llm_output hook that runs every time the LLM produces a response:

1. LLM responds → hook fires
2. Regex matches Chinese/Cyrillic character runs
3. Strips them and collapses whitespace
4. If nothing changed, returns None (no-op)
5. If something was stripped, returns the cleaned text

The plugin is context-aware: if your conversation is in a language that genuinely uses these scripts (e.g., you're learning Chinese), nothing gets stripped.

Current behavior: The _sanitize callback always runs. A future update will add prompt-based detection to skip stripping when the user's input contains those scripts.

Installation

Prerequisites

• Hermes Agent installed and running

1. Install with one command

hermes plugins install evanyudis/cjk_sanitizer --enable

This clones the repo into ~/.hermes/plugins/cjk_sanitizer/ and adds it to your plugins.enabled config automatically.

2. Restart Hermes Agent

hermes gateway restart

Or start a new CLI session — plugins auto-discover on startup.

Verify it's loaded

hermes plugins list

You should see cjk_sanitizer with a checkmark or "loaded" status.

Alternative: Manual install

If hermes plugins install isn't available (older version), you can do it manually:

git clone https://github.com/evanyudis/cjk_sanitizer.git ~/.hermes/plugins/cjk_sanitizer

Then edit ~/.hermes/config.yaml:

plugins:
  enabled:
    - cjk_sanitizer

Then restart.

Updating

hermes plugins update cjk_sanitizer

Usage

There's nothing to configure. Once the plugin is enabled, it works automatically on every LLM response. Chinese characters and Cyrillic text in model output will be stripped before you see them.

If you need to temporarily disable it without removing it:

plugins:
  disabled:
    - cjk_sanitizer

(Leave it out of enabled and it won't load.)

What This Fixes

• MiniMax M2.7 — known to occasionally emit Chinese mid-sentence
• DeepSeek models — can drift into Chinese in long reasoning chains
• GLM / Qwen — bilingual by nature, sometimes don't switch cleanly
• Any provider serving China-trained models — same quirks apply

What This Does NOT Fix

• Bad translation quality
• Hallucinations
• Language mixing in your prompt (that's a prompting issue)
• Korean or Japanese text leak (those are preserved by design)

Contributing

Open source is a team sport. Here's how to contribute.

Getting Started

1. Fork the repo
2. Clone your fork: git clone https://github.com/<your-username>/cjk_sanitizer.git
3. Create a branch: git checkout -b feature/your-feature-name
4. Make your changes
5. Test with your Hermes Agent setup
6. Push and open a PR

Development Setup

The plugin lives in a single file (__init__.py) — no build step, no dependencies beyond Python stdlib (re). To test changes locally:

# Symlink your dev version
ln -sf /path/to/your/fork ~/.hermes/plugins/cjk_sanitizer

# Restart gateway to pick up changes
hermes gateway restart

Pull Request Guidelines

• Keep it focused — one change per PR
• Update the description in plugin.yaml if your change affects scope
• Preserve backward compatibility — don't break existing installs
• Test edge cases — empty responses, already-clean text, mixed scripts
• Use English in all code, comments, and PR descriptions

What Needs Help

• Context-aware detection — skip stripping when user prompt contains CJK/Cyrillic scripts (language learning use case)
• Configurable character ranges — let users specify which scripts to strip via plugin.yaml
• Korean/Japanese opt-in — add option to strip those too for users who want it
• Test suite — even basic pytest cases would help catch regressions
• More LLM testing — we've tested with MiniMax and DeepSeek. Reports from other providers welcome.

Architecture

~/.hermes/plugins/cjk_sanitizer/
├── __init__.py      # Plugin logic: regex patterns + hook registration
├── plugin.yaml      # Manifest: name, version, hooks declared
├── README.md        # This file
├── LICENSE          # MIT
└── .gitignore

The plugin system is documented in the Hermes repo. Plugins go in ~/.hermes/plugins/<name>/, the manifest is plugin.yaml, and the entry point is __init__.py with a register(ctx) function.

License

MIT — do what you want, just don't sue me. See LICENSE.

---

Made by Evan Yudistira. Not affiliated with Nous Research.