zhhz

Self-contained Simplified/Traditional Chinese converter — a pure-Rust, data-embedded reimplementation of OpenCC.

zhhz converts between Simplified and Traditional Chinese (plus Taiwan, Hong Kong, and Japanese-shinjitai variants) using the OpenCC dictionaries, and detects the script variant of Chinese text. All dictionaries are embedded in the binary at compile time — one ~1.86 MB static binary (or 588 KB xz-compressed) with no runtime download and no separate data directory.

The name is a palindrome: zh hanzi, and zhuan huan han zi (转换汉字, "convert Chinese characters").

Why

OpenCC is the de-facto Chinese-conversion library. zhhz is a from-scratch Rust reimplementation built around the same dictionaries, designed to be:

• One self-contained binary. Data is embedded via include_str!; nothing is fetched or installed alongside it.
• Memory-safe by construction — pure Rust in the conversion core.
• Friendly to custom conversion words at the highest priority, for terminology, branding, or domain vocabulary.
• Tracked against upstream data via a pinned, reproducible sync script (scripts/sync-opencc.sh).

Designed for AI agents

zhhz is built first and foremost for AI agents (Claude, Cursor, custom LLM pipelines, batch jobs). The CLI is deliberately minimal:

• No TUI, no progress bars, no spinners. Output is plain text on stdout; errors go to stderr. An agent can capture both and parse deterministically.
• stdin / stdout friendly. Pipe text in, get text out. Files are positional arguments; - means stdin.
• Stable, predictable, safe. Same input → byte-identical output every time. No network, no filesystem writes unless asked (--in-place), no temp files, no background processes.
• Batch / filelist from stdin (chardet-style): <files>..., --files-from <PATH|->, -0 / --null, recursive directory walking.
• Single self-contained binary. No native deps, no data files to ship alongside. Drop it in a container and it just works.

If you want a fancy interactive experience, this is the wrong tool — use OpenCC or a web demo. If you want a thing you can shell out to from a script or hand to an agent, this is it.

Install

Cargo

cargo install zhhz

Direct binary

curl -L https://github.com/ljh-sh/zhhz/releases/latest/download/zhhz-x86_64-unknown-linux-musl.tar.xz | tar xJ -
sudo mv zhhz-x86_64-unknown-linux-musl/bin/zhhz /usr/local/bin/

Build from source

Requires Rust 1.74+.

git clone https://github.com/ljh-sh/zhhz
cd zhhz
cargo build --release   # binary at target/release/zhhz

npm

npm install zhhz

Same conversion core, compiled to WebAssembly. Zero native deps; the OpenCC dictionaries are baked into the .wasm. See docs/npm.md for the full API and examples/node-usage/ for a runnable demo. The npm API surface is strictly richer than opencc-js (adds detect(), introspection, Converter factory class, semantic region flags).

Usage

echo '汉字' | zhhz                       # default s2t:  漢字
echo '漢字' | zhhz -c t2s                # t2s:          汉字
echo '信息' | zhhz -c s2twp              # s2twp:        資訊
zhhz -c s2t input.txt                   # convert a file
zhhz -c s2t -i input.txt                # rewrite in place
zhhz --list                             # list all configs

Configs (mirrors OpenCC):

| config | direction | |--------|-----------| | s2t / t2s | Simplified ↔ Traditional (OpenCC standard) | | s2tw / tw2s | Simplified ↔ Traditional (Taiwan) | | s2twp / tw2sp | …with Taiwan phrases | | s2hk / hk2s | Simplified ↔ Traditional (Hong Kong) | | s2hkp / hk2sp | …with Hong Kong phrases | | t2tw / tw2t | Traditional (standard) ↔ Taiwan | | t2hk / hk2t | Traditional (standard) ↔ Hong Kong | | t2jp / jp2t | Japanese Kyūjitai ↔ Shinjitai |

Or use semantic region flags (--from / --to):

echo '汉字'   | zhhz --from cn-s --to cn-t      # 漢字
echo '信息'   | zhhz --from cn-s --to cn-tw     # 資訊 (Taiwan phrases)
echo '鼠标'   | zhhz --from cn-s --to cn-tw     # 滑鼠
echo '漢字'   | zhhz --from cn-tw --to cn-s     # simplified
echo '万与两' | zhhz --from jp-n --to cn-t      # 萬與兩

Regions: cn-s / cn-t / cn-tw / cn-hk / jp-t / jp-n.

Detect the script variant of Chinese text

echo '汉字计算机软件' | zhhz detect          # cn-s    57   -
echo '漢字計算機軟體' | zhhz detect          # cn-t    66   -
echo 'こんにちは世界' | zhhz detect          # jp-n    50   -
zhhz detect corpus.txt                      # cn-s    ...  corpus.txt
zhhz detect                                 # detect content piped on stdin

Output is tab-separated: <region>\t<confidence>\t<path>. Confidence is 0–100 (share of signature characters in the input). Region codes are the same six listed above, or unknown when there are no CJK characters / kana.

zhhz detect mirrors chardet's CLI: <files>... to detect each path, - (or no args) to detect stdin content, --files-from <PATH|-> to read a newline-separated list of paths, -0 / --null for NUL-separated lists, and recursive directory walking.

Custom dictionaries

A custom dictionary is a TSV file (key<TAB>value); # lines are ignored. Entries override the built-in tables at the highest priority:

# mywords.txt
# key	value
软件	軟體
独家	獨家

echo '买软件吃独家' | zhhz -c s2t --dict mywords.txt   # 買軟體喫獨家

Library

use zhhz::{Config, Converter};

let c = Converter::new(Config::S2t);
assert_eq!(c.convert("汉字"), "漢字");

// Custom words override the built-in tables.
let c = Converter::with_custom(Config::S2t, &[("软件".into(), "軟體".into())]);
assert_eq!(c.convert("买软件"), "買軟體");

The engine is pure Rust with a tiny dependency tree (serde_json, anyhow) and no filesystem or network access, so it is straightforward to bind from WASM and Python (both are on the roadmap).

Node.js / npm

npm install zhhz

import { convert, detect, Converter, listConfigs } from "zhhz";

console.log(convert("汉字", "s2t"));            // 漢字
console.log(detect("他去了西維珍尼亞州"));      // { region: "cn-hk", confidence: 70 }

const c = new Converter("s2twp");
console.log(c.convert("信息"));                 // 資訊
console.log(c.convertWithCustom("买软件", [["软件", "軟體"]])); // 買軟體

console.log(listConfigs()); // 16 OpenCC config names

The npm package ships the same engine compiled to WebAssembly; dictionaries are embedded, so there is no data directory to ship alongside and no network fetch at runtime. The surface is strictly richer than opencc-js (adds detect(), introspection, factory instance, semantic region flags). See docs/npm.md for the full reference and examples/node-usage/ for a runnable demo.

How it works

zhhz reproduces OpenCC's pipeline exactly:

1. Segment the input with forward maximum matching (FMM) against the segmentation dictionary group.
2. Convert each segment through an ordered chain of dictionary groups; each stage re-walks its segment with longest-prefix matching, emitting the first candidate on a match.

The group match semantics match OpenCC's PrefixMatch: the highest-priority dictionary with any prefix wins (priority dominates length across dictionaries; length dominates only within one dictionary).

The OpenCC build system generates five dictionaries at build time (reversed variant tables, a tofu-risk subset, and a regional-phrase projection). build.rs reproduces all five deterministically from the vendored source data, so data/ stays a pure mirror of upstream.

Data and licensing

Dictionary data is vendored from BYVoid/OpenCC (see data/UPSTREAM for the pinned commit) and is Apache-2.0, same as the source code. Re-vendor the latest upstream data with:

scripts/sync-opencc.sh            # master HEAD
scripts/sync-opencc.sh 1.3.1      # a specific tag/commit

Roadmap

• [x] Pure-Rust engine, all 16 OpenCC configs, embedded data, custom words
• [x] WASM build + npm package (wasm32-unknown-unknown) — npm install zhhz
• [ ] Differential-fuzz harness proving output parity vs the opencc CLI
• [ ] Python native extension (PyO3 / maturin)
• [ ] Compact dictionary representation (FST / double-array) for smaller binaries

See ROADMAP.md.

Contributing

See CONTRIBUTING.md. Issues and PRs are welcome.

Security

See SECURITY.md. For vulnerabilities, email [email protected] rather than opening a public issue.

License

Apache 2.0 — see LICENSE. Dictionary data is Apache-2.0, vendored from OpenCC.