Compression

The engine creates Brotli and Gzip copies of selected nonempty files in brotli/ and gzip/. It selects .html, .css, .js, .svg, .xml, .txt, and extensionless names consisting of _c, _j, _s, or _i followed by exactly 16 lowercase hexadecimal characters. robots.txt is excluded. See the compression decision.

Two codecs

1.0.0-beta1 supports no other codecs. Adding one requires measurements and a change to the artifact and hosting contract.

site/html/how_bemjson.html        ← original
site/brotli/how_bemjson.html.br   ← Brotli Q11
site/gzip/how_bemjson.html.gz     ← Gzip 9

Each pre-compressed copy keeps the original's flat filename and adds a codec suffix. The brotli/ and gzip/ directories are not served directly — the nginx configuration closes them off.

Build determinism

The engine normalizes byte 9 (the tenth byte) of the gzip header (the OS field) to 0xFF ("unknown"), so builds on macOS and on Linux use the same OS marker. Identical bytes also require the same compressor/runtime versions. A different OS marker changes the compressed file’s hash and causes unnecessary uploads during manifest-based deployment; the manifest itself remains correct. For the full rationale, see the compression decision.

Brotli has no such header, so it needs no extra normalization.

Server-side negotiation

The hosting configuration looks for explicit br and gzip tokens with nonzero q and selects an available copy in a fixed order: Brotli, then Gzip, then the original. Content-Encoding matches the selected file. A simplified word match does not reproduce these rules.

The rules do not rank codecs by their q values or enable a codec from * alone. The fallback is the original file; identity;q=0 does not make the generated configuration return 406. If a site needs different negotiation behavior, customize its serving rules.

Why compress in advance, not on the fly

On-the-fly compression uses server CPU. Its cost depends on the compression level, caching, and server configuration. Nuckty performs this work during the build and prepares Brotli Q11 and Gzip 9 copies.

At request time, the generated configuration selects a suitable prepared copy or the original. This requires the server modules and settings described in the hosting guide.

What is not compressed

Selection uses filenames, not content inspection. Files with other names, including .json and ordinary binary extensions, receive no compressed copies; neither do robots.txt or empty files. .htaccess and manifest.json are written after compression. Copies use flat directories, so matching basenames from different source directories cause an error.