Deployment and Operations

Deployment and Operations

For the complete moderation-layer model behind account restrictions, posting bans, and nginx hard bans, see Account Restrictions and IP Bans.

Deployment Flow

deployment is GitHub Actions driven.

current chain:

1. push to main
2. code lint workflow runs
3. if lint passes, deploy to fridge.dev runs from the successful workflow event
4. repo is rsynced to /var/www/fridge.dev
5. the hard-ban source index is built or reused as the PHP-FPM http user
6. the Toast Discord bot is gracefully restarted from the deployed copy
7. the deploy workflow asks Toast to post a patch notice approval preview in Discord channel 1526075637096255548; an admin ✅ reaction posts it to update channel 1455194403642802309 and pings role 1408064850688475197

Deploy Workflow

/.github/workflows/deploy.yml

main details:

• triggered by successful code lint workflow completion
• only deploys pushes to main
• installs rsync and openssh-client
• uses DEPLOY_KEY
• deploy target is deploy@45.76.134.105:/var/www/fridge.dev
• the workflow verifies /var/www/fridge.dev exists and is writable before rsync, and refuses any unexpected target path
• after rsync, the workflow creates data/etc/banlists/index as http, verifies that http can read the source directory and write the index directory, then asks PHP running as http to build or reuse the hard-ban source index so a public request does not pay the one-time rebuild cost
• if index preparation or construction fails, the workflow prints the relevant directory permissions, largest partial index files, and filesystem usage to distinguish ownership from capacity failures
• after rsync, ssh stops any toast GNU screen session owned by deploy, stops any toast session owned by http, prepares others/toast-discord-bot/bot/toast-bot.log for http, then runs /var/www/fridge.dev/others/toast-discord-bot/bot/start.sh as http
• the restart step needs passwordless sudo for deploy to run the Toast bot as http; Toast writes DM history, feed notification state, and patch approval state under /data, which is owned by http:http
• because the http user home is not a normal login home, the workflow sets SCREENDIR=/tmp/toast-screen-http for http-owned screen commands and creates that socket directory as http with mode 700 before start

Toast Patch Notice Commit Format

after a successful main deploy, the workflow sends Toast a list of non-merge commits using each commit's subject (git log %s) and full body (git log %b). Toast turns each commit into Discord patch-note bullets and posts the fully formatted patch notice preview in approval channel 1526075637096255548. Toast reacts to that preview with ✅; when an admin approves with ✅, Toast posts the update to channel 1455194403642802309 and pings role 1408064850688475197.

Pending approval message IDs and payloads are persisted in /data/etc/toast-patch-approvals.json, so older messages remain actionable after a bot restart. The raw reaction handler also fetches uncached approval messages from Discord. For a legacy Toast-authored approval embed created before persistence was added, Toast reconstructs the notice from the embed's commit URL and patch-note fields and sends it to the standard update channel.

• commit subject

• first body note

• second body note

format commits like this when the Discord patch notice should read cleanly:

Short user-facing summary

Concrete patch note detail

Second useful detail, if needed

rules to keep in mind:

• the Discord embed shows the deployed build ID followed by the patch-note fields, without introductory or pull-request source text
• the first bullet is the commit subject and does not include the commit ID
• body notes are separated by blank lines in the commit body
• every body note starts with its own bullet
• long patch notes are split across multiple Discord embed fields instead of being cut off
• Markdown and Discord mentions are escaped by Toast, so write plain text instead of relying on formatting or pings
• merge commits are excluded from the shipped commit list

example:

git commit -m "Add a settings system info dashboard" \
  -m "Give admins a live view of server, PHP, storage, and site health." \
  -m "Load shared runtime scripts through rendered pages so settings controls keep working."

same message as plain text for VS Code's source control commit message box:

Add a settings system info dashboard

Give admins a live view of server, PHP, storage, and site health.

Load shared runtime scripts through rendered pages so settings controls keep working.

that appears in Toast's Discord embed as three patch-note bullets: one for the subject and one for each body note.

admins can manually post the same style of update directly from Discord with Toast's /shareupdate command:

• /shareupdate latest posts the currently deployed HEAD commit from the bot's local repo
• /shareupdate <commit ID> posts a specific 7-40 character commit SHA

the command uses the same embed formatter and update channel as approved deploy notices, so it also pings role 1408064850688475197.

What Does Not Deploy

deployment uses .rsyncignore, so these are excluded:

• /data/**
• sitemap.xml
• repo docs and local config files
• .github/**
• /scripts/**
• local editor/codex folders
• /others/toast-discord-bot/bot/venv/**

that means production runtime data is expected to already exist on the server.

Server Permissions

from README.md:

• project files should belong to deploy:http
• directories should be 755
• files should be 644
• /data and sitemap.xml need http:http ownership for webserver writes
• Toast runs as http in production so it can update /data/etc/toast-dm-history.json and /data/etc/toast-feed-notify-state.json; only toast-bot.log in the bot code directory is made writable for that runtime user

the deploy user needs passwordless sudo for the Toast restart step:

deploy ALL=(http) NOPASSWD: ALL

install that with visudo, preferably as a small file under /etc/sudoers.d/, because typoing sudoers directly is how servers become decorative bricks. the workflow prepares toast-bot.log as deploy with group http and mode 664, so no root sudo is needed for log setup.

Nginx Config Source

the repo-tracked files in .nginx/ are the source for the production nginx config.

• .nginx/nginx.conf corresponds to /etc/nginx/nginx.conf
• .nginx/fridge.dev corresponds to /etc/nginx/sites-enabled/fridge.dev
• production uses these through symlinks, so edits here are real server config edits, not examples

when adding routes, APIs, uploads, redirects, or private data folders, check .nginx/fridge.dev as part of the feature. a correct PHP route can still fail if nginx redirects POSTs, misses a clean-url rewrite, or accidentally exposes/blocklists the wrong /data path.

legacy fridg3.org, www.fridg3.org, and m.fridg3.org redirects are handled in Cloudflare, not nginx. the redirect must append legacy_domain=fridg3.org; the frontend consumes that marker for the one-time rebrand popup and then removes it from the URL.

Nginx Clean URLs

production nginx needs explicit rewrites for PHP routes that accept path-style ids. without these, nginx falls through to the root /index.php fallback before the route can parse the URL.

the generic location / fallback should route missing paths to /error/404, not /index.php, so nonexistent URLs do not quietly render the homepage.

POST-only API directory routes also need POST-safe rewrites when called without index.php; otherwise nginx can normalize the directory URL with a redirect and the browser may retry as GET. /api/dev-bootstrap and /api/toast-feed-generate are included in that rewrite list.

the contact route is configured POST-safe at /contact, old /email paths redirect to /contact, and /data/contact/ is blocked from direct web access. /data/guestbook and /data/guestbook/ are also blocked because entry files contain moderation-only IP metadata; the public guestbook remains available through its PHP routes. account form routes such as /account/login, /account/change-password, and /account/admin/edit are also rewritten directly to their PHP handlers so POST bodies are not lost to trailing-slash redirects.

site-wide hard bans are stored in data/etc/hard-banned-ips.txt, augmented by read-only .txt source files recursively discovered beneath data/etc/banlists/ and containing exact IPs or IPv4/IPv6 CIDR subnets, and enforced by nginx auth_request through the internal /_hard-ban-check location. PHP-FPM compiles source lists into fixed-width binary range buckets beneath data/etc/banlists/index/; its source-stat signature automatically invalidates the index when a list changes, and a lock prevents concurrent rebuilds. build-time external sorting merges overlapping ranges with bounded memory, then steady-state checks take the lock-free ready path and binary-search one bucket. index construction and its streaming fallback use fixed-size token chunks, keeping memory bounded even for large files or lines. the index directory must be writable by http; deleting it or updating a source makes the next deployment prewarm or request rebuild it. a denied subrequest returns 401, which nginx converts into a 302 redirect to /error/blacklisted; that route, files beneath its directory, and font files beneath /resources explicitly disable the authorization check so the redirect cannot loop and the stripped Blackprint page can render. browser/IP associations live in data/etc/hard-ban-identities.json; the global strictIdentityEnforcement policy lives in data/etc/hard-ban-settings.json and defaults to enabled. disabling it stops identity-based denial and future-IP propagation while continuing to record recognized observations. the physical checker route, hard-ban data files, settings, source-list directory, and binary index must remain inaccessible to clients. this requires nginx's standard ngx_http_auth_request_module.

the upload API posts to /tools/upload/?api=*; keep the exact /tools/upload nginx rewrite so stale no-slash requests hit PHP directly instead of losing their POST body to a trailing-slash redirect. cursed but real.

mdpaste share links use /tools/mdpaste/s/{id} and need this block before the generic location / fallback. keep the regexes quoted, because nginx treats unquoted {16} like cursed config syntax.

# mdpaste clean URLs
location ~ "^/tools/mdpaste/s/[a-fA-F0-9]{16}/?$" {
    rewrite "^/tools/mdpaste/s/([a-fA-F0-9]{16})/?$" /tools/mdpaste/s/index.php?id=$1 last;
}
location /tools/mdpaste/s/ { try_files $uri $uri/ /tools/mdpaste/s/index.php?$args; }
location /tools/mdpaste/   { try_files $uri $uri/ /tools/mdpaste/index.php?$args; }

Backup Workflow

/.github/workflows/backup-data.yml

what it does:

1. ssh to the server
2. remove stale temporary backup zips from /home/deploy
3. verify deploy can read/traverse /var/www/fridge.dev/data
4. zip /var/www/fridge.dev/data into a temporary archive under /home/deploy
5. download the archive to the runner
6. upload it to Google Drive using rclone
7. keep only the 10 newest backups
8. delete temp archives from runner and server

triggers:

• manual workflow_dispatch
• scheduled daily cron at 0 0 * * *

required secrets:

• DEPLOY_KEY
• GDRIVE_BACKUP_FOLDER_ID
• RCLONE_CONFIG

setup notes live in /.github/workflows/backup-data-setup.md.

if archive creation fails with zip exit code 18, at least one path under /data was unreadable to deploy. run the unreadable-path check from /.github/workflows/backup-data-setup.md, then fix ownership/permissions before rerunning the workflow.

the backup and developer-data workflows also refuse to run if TARGET is anything other than /var/www/fridge.dev, so a stale workflow variable cannot accidentally back up or publish the wrong site tree.

Developer Data Copy

/.github/workflows/publish-dev-data.yml

on the same daily schedule as the private backup workflow, this workflow:

1. copies production /var/www/fridge.dev/data into a temporary server workspace
2. runs /.github/scripts/sanitize-dev-data.php against the copy
3. zips the sanitized directory as DD-MM-YY_hh-mm-ss.zip
4. uploads it to the public Google Drive developer data folder
5. keeps only the 10 newest zip files in that folder
6. removes temporary server and runner files

the sanitizer currently clears accounts, login/page-view/IP/rate-limit logs, guestbook IP ownership and entry IP metadata, feed guest reply IPs/browser tokens, shared posting ban lists, the site-wide hard-ban list and browser/IP associations, blanks Toast bot and Groq credentials, blanks Toast private lore, clears Toast DM/notification state and browser notification state, clears webhooks, removes upload room tokens, clears encrypted mdpaste records, clears encrypted chat data and local chat keys, replaces the off-topic Discord archive with an empty placeholder, and replaces private journal drafts with a harmless placeholder draft. the development archive additionally excludes data/etc/hard-banned-ips.txt and data/etc/hard-ban-identities.json entirely.

setup notes live in /.github/workflows/publish-dev-data-setup.md.

Sitemap Generation

sitemap.xml is not deployed from git. it is generated by /api/sitemap, which means:

• the file must be writable by the server
• the server copy is the one that matters

Operational Truths

• this repo is source code, not a full backup
• /data is operational state
• if prod data disappears, git will not magically save you
• if file permissions are wrong, deploys and runtime writes will get weird fast