Release notes

What changed in each release, and what it means for a machine you already run. For the full command reference see the docs.

Currently published v1.0.5-b43 Download

v1.0.5

2026-10-05

A deploy of a service that keeps crashing now fails at once and says so, instead of waiting out the timeout.

A crash-looping service fails the deploy straight away

While wor deploy waits for a service to answer, it now also watches the supervisor's restart count. If PM2 or systemd has restarted the process since the deploy and it is still not ready, the deploy stops at once with is crash-looping and how many restarts it saw, rather than waiting the full 30 seconds and reporting only "nothing is accepting connections".

The automatic revert and the log excerpt work as before. A service that is merely slow to start, with no restarts, still gets the whole ready.timeout. It fails nothing that would have passed: the stay-up check already treated any restart as a failure.

v1.0.4

2026-10-05

A fix for the one thing 1.0.3's wor ssl status told you to do: it now has a command that really registers the missing renewal hook.

wor ssl hook: register the renewal hook without reissuing

If wor ssl status says renewal is NOT WORKING because certbot's renewal config has no hook, run wor ssl hook <host>. It adds the hook with certbot reconfigure, checks that it landed, and leaves the certificate alone.

Needs certbot 2.3 or newer, and port 80 reachable, because certbot checks the change with a dry-run renewal.

wor ssl status no longer points at a fix that does nothing

1.0.3 told you to reissue with wor ssl issue. While the certificate is still valid, certbot keeps it and rewrites nothing, so the hook was never registered and the status did not change. It now points at wor ssl hook, and at ssl issue only when certbot has no renewal config for the host at all.

Removing a host or service tells you how to stop certbot renewing it

For a Let's Encrypt host, wor host remove, wor service remove --cascade and wor ssl remove now print sudo certbot delete --cert-name <host> to copy and run. wor removed only its own copy of the certificate, so certbot kept renewing a name that had gone, and failing.

service remove --cascade also left the host's SSL record and certificate copy behind; it now removes them like host remove. wor does not revoke: that cannot be undone and is for a leaked key.

v1.0.3

2026-10-05

Git from any server, deploys that check the service really came up and put the old version back when it did not, deploy hooks, and node deploys that do not drop requests.

Git from any server, including your own

wor source clone takes an address from any git server: https, ssh, user@host:path, git://, file:// or a local path. A private repository on a self-hosted server (Gitea, Forgejo, GitLab, a bare repository on another machine) works the same as a hosted one: --ssh-key uses an SSH deploy key, which is recorded in the clone so pulls and deploys use it too, and --ca-cert trusts a server whose certificate comes from a private CA. --branch and --depth are new as well.

Addresses are validated (nothing git could read as an option or run as a helper is passed through), credentials in an address are hidden from everything wor prints, and under --non-interactive git never stops to ask a question. There is deliberately no switch to turn certificate verification off.

wor source status

Where a service's source came from and where it stands: the remote, branch, last commit, how far ahead or behind, uncommitted changes and stashes. Read-only, and it makes no network call unless you add --fetch.

A deploy waits until the service really answers

After the restart, a service that has a port is not reported deployed until it accepts connections, optionally answers a URL you name, and then stays up for a few seconds, which catches an app that opens its port and crashes a moment later. If it does not, the deploy fails with the tail of its log. It applies to wor deploy, wor service start and wor service restart. Tune it per service in .wor/deploy.ini (ready.timeout, ready.path, and so on).

This used to be checked only by asking PM2 or systemd whether the process existed, so a service that crashed two seconds after starting was reported as deployed and stayed down until someone ran wor health.

A failed deploy puts the previous version back

When the readiness check fails after a deploy that moved the code, wor resets the tree to the commit that was running, reinstalls and rebuilds it, restarts, and checks again. The deploy is still reported as failed, because the deploy you asked for did not happen.

It only does this when it is safe: no uncommitted changes to tracked files, no --stash, and no pre-restart hook that may have changed a database. Otherwise it says why and gives the way back by hand, wor rollback <target> --to=<commit>. --no-revert or revert = no turns it off.

Deploy hooks

Two optional executable files carry your application's own steps: .wor/hooks/pre-restart runs after the build and before the restart (a database migration; if it fails the deploy stops and the running version is untouched), and .wor/hooks/post-deploy runs after the service answered (a failure only warns). Commit them with the code, so a hook always belongs to the commit that needs it.

A hook runs as the account that runs wor, never elevated, with a curated environment (not yours), no stdin and a time limit. One that is present but cannot run is an error, not a skipped step. --no-hooks skips both.

Node deploys that do not drop requests

A node service can run in PM2 cluster mode and be deployed with pm2 reload, which replaces its workers one at a time while the others keep answering: pm2.instances and pm2.reload in .wor/deploy.ini. In a test, a plain restart refused requests for as long as the app took to start, and a reload refused none.

Opt-in, and your application has to tolerate several workers (no state kept in the process's own memory, no once-only scheduled job inside the app, no sticky websocket sessions). The first deploy after turning it on takes one normal restart. wor service status, health, info and diagnose treat the workers as one service.

wor ssl status checks that renewal works

For a Let's Encrypt host it no longer prints a remembered "enabled". It checks whether certbot's renewal config registers the hook that refreshes wor's copy, and whether wor's copy expires when certbot's certificate does. Where it cannot check (the files are root-only), it says unchecked, never "fine".

A host issued by an older version could have certbot renewing on schedule while wor kept serving the old copy, and the site worked only until that certificate expired. If ssl status reports a stale copy, run wor ssl sync <host>, then reissue with wor ssl issue <host> --provider=letsencrypt to register the hook.

wor rollback --to, and rollback no longer deletes your .wor and .env

--to=<commit|tag> goes back to a known-good version; without it, rollback resets to the branch's own upstream, whatever the remote is called.

Fix: rollback ran git clean -fd, which deletes every untracked file that is not gitignored, and .wor (your web-server snippets and PHP and deploy settings) and .env are untracked in almost every service. The backup it takes first honours .gitignore, so it might hold no copy of exactly those files. They are now kept.

Smaller changes and fixes

  • wor source pull is fast-forward only. wor no longer makes a merge commit on a server; a tree that has commits the remote lacks fails and points at wor rollback.
  • wor source clone can be driven from a program: --env and --wor answer the questions about an existing .env and .wor. Without a terminal they are required.
  • wor host add without a terminal says to pass --domain-type instead of a generic error.
  • wor diagnose now finds PM2's logs, which it never did: it looked in a file name PM2 does not use.
  • wor doctor no longer shows a row of dashes as PM2's version.
  • A deploy.ini, hook or php.ini that is broken is checked in the tree after the pull, so the commit that fixes it can deploy. Before, a committed typo locked the service out of deploying, including the commit that fixed it.

Documentation at /docs/

The documentation was rewritten as one set of pages: install and a quick start, how each feature works, and example output. It replaces the old command reference on the site.

v1.0.2

2026-08-21

Per-service PHP configuration, a self-upgrade command, and a deploy that stops reloading php-fpm when nothing changed.

Per-service PHP settings

A php service with its own pool can now configure itself, in two files inside its own tree: .wor/php.ini for PHP ini settings (memory_limit, upload sizes, timezone) and .wor/php-fpm.ini for the pool's process manager (pm, pm.max_children and friends). No more editing the host's php.ini and hoping no other site minded.

WOR reads these files; PHP does not. A PHP-FPM pool cannot include a php.ini of its own, so WOR parses them and renders directives into the service's pool config — which php-fpm -t validates before anything reloads. A value php-fpm rejects rolls the pool back to exactly what it was, so one bad setting cannot take down the other services sharing that master. Only an allowlist of keys is accepted, and a key outside it is an error rather than a skipped line.

wor service reload

Applies those files on their own: re-renders the pool, validates it, reloads php-fpm and prints what is now in force — without reinstalling dependencies, rebuilding or restarting the service.

Configuration only. Restarting a process is still wor service restart and re-rendering a vhost is still wor host reload. It is also the only way to apply these files to a service whose source is not a git repository, which wor deploy requires.

Deploy no longer reloads php-fpm for nothing

wor deploy re-renders a pooled php service's config on every deploy, so an edit to either settings file ships with the code that needs it. But it now skips the write and the reload entirely when the result would be identical to what is already on disk.

Reloading the shared php-fpm master cycles the workers of every other service under it. A deploy that only changed code has no business doing that.

Settings that were never applied are now visible

wor info lists what each service asks for. wor diagnose and wor health warn when the running pool no longer matches those files — edited and never applied — or when a file no longer parses, which would fail the next deploy.

All three are read-only and never ask for sudo: a pool file they cannot read is reported as unchecked, never as wrong.

wor upgrade

Compares the running binary against the release this site publishes, shows you both, and installs the newer one once you confirm. --yes skips the confirmation.

Installation is handed to the install.sh inside the downloaded archive rather than reimplemented, so there is only one tested install path.

wor service chown

Hands one service's files back to the operator account, or to a user you name — for a tree left owned by root, by a CI rsync, or by an admin who is not the operator.

It changes the owner only, never the group, and re-grants the php-fpm pool afterwards. Rewriting the group would revoke the pool's access to the very files it has to read.

wor source clone asks about .wor

A clone replaces the service's whole tree, and the backup it takes first honours .gitignore — so a gitignored .wor could be lost with no copy anywhere. It now asks first: keep the current one (the default) or take the repository's.

Committing .wor to your repository is the tidy way to version this configuration alongside the code.