nginx.conf

load_module modules/ngx_http_brotli_filter_module.so;
load_module modules/ngx_http_brotli_static_module.so;

daemon off;
user nginx nginx;
worker_processes 1;
error_log stderr warn;
worker_rlimit_nofile 8192;
pid /var/run/nginx.pid;

events {
  worker_connections  1024;
}

http {
  include /etc/nginx/mime.types;
  charset utf-8;
  sendfile on;
  tcp_nopush on;
  gzip_static on;
  brotli_static on;
  port_in_redirect off;
  server_name_in_redirect off;

  # Redirects
  map_hash_bucket_size 256;

  map $request_uri $uri_without_slash {
    "~(?<u>(.+))/$" $u;
    default $request_uri;
  }

  map $uri_without_slash $redirect_uri {
    include redirects.map;
    default "";
  }

  # Content negotiation: clients asking for markdown (Claude sends
  # `Accept: text/markdown`) get the llms.txt rendition of the same URL.
  map $http_accept $negotiated_index {
    ~*text/markdown llms.txt;
    default index.html;
  }

  server {
    listen 8080;
    access_log /dev/stdout;
    root /app/public;

    add_header Vary Accept;

    location ~ ^/openapi(.*)\.json$ {
      add_header Access-Control-Allow-Origin *;
      add_header Vary Accept;
    }

    # bin/compress.js writes a .br/.gz sibling for every file. gzip_static and
    # brotli_static read those from disk while serving the original URL, so the
    # siblings need no URL of their own. A genuine .br/.gz download would need
    # an exception here.
    location ~ \.(br|gz)$ {
      internal;
    }

    # The markdown rendition Hugo writes alongside each index.html. `internal`
    # keeps it reachable only via the negotiation above, so every page has
    # exactly one public URL. mime.types would serve it as text/plain.
    location ~ (^|/)llms\.txt$ {
      internal;
      types { }
      default_type text/markdown;
      charset_types text/markdown;
    }

    if ($redirect_uri != "") {
      rewrite ^(.*)$ $redirect_uri redirect;
    }

    index $negotiated_index index.html;
    error_page 404 /404.html;
  }
}