Files
summercms/docs/setup/upgrading.md
Jakub Zych 33460607db feat(summer): help, examples and hints for every tool command
- build runs go build under a spinner on a terminal and ends with
  'Built bin/<name> in <time>'; make:* print 'Created <path>' per file
- make:*, plugin:add, new, parity:*, docs:build, docs:sync and every
  delegate get long help and examples; delegates note that global options
  pass through to the application binary
- make:* failures (no summer.yaml, invalid plugin ID or name, existing
  artifact, unknown plugin) explain themselves and suggest a command
- docs:build and docs:sync return bonfire.Reported after the problem list
- TestCommandsDocumented covers the tool commands too
- docs: Upgrading to 0.4, delegates and global options, make: results and
  hints, the checklist for a finished command
2026-10-21 13:25:53 +02:00

13 KiB

title, description, section, order
title description section order
Upgrading Move an application to a newer framework release, what the version numbers promise, and what to check before you take 0.2 and 0.4. setup 60

Upgrading

An application created by summer new builds against a framework checkout next to it; the update script moves that checkout to the newest release tag, reinstalls the summer tool and rebuilds the application (see Updating an application). An application or plugin that requires a tagged framework version in its go.mod moves with the go command instead:

go get git.golem15.com/golem15/summercms@v0.2.0
go mod tidy

Then rebuild, run migrate with the new binary and restart it.

Version numbers

SummerCMS is in the 0.x series, and the version tells you whether to read before you upgrade:

  • The minor version (0.1 to 0.2) goes up when an upgrade can refuse a configuration that the previous release accepted, or changes a documented behaviour. Read the section for that version below before you deploy it.
  • A patch version (0.2.0 to 0.2.1) is safe to take without reading: it fixes bugs and adds features, and an application that started on the previous patch starts on it too.

Applications pin exact versions, so a new release never reaches an application until it moves its go.mod or its checkout.

Upgrading to 0.4

0.4 rebuilds the console: Artisan-style list and help with examples, global options on every command, a logging pipeline with daily files, startup summaries and shutdown reports. Plugins compile unchanged. After you move the framework version, run summer build to regenerate main.go: the new generated main sets up logging before the configuration is read, parses the global options and returns the exit codes below. Then run bin/<app> list once with the new binary: it boots every plugin and runs the checks below.

Boot refusals

Duplicate command names. Two commands with the same name, from two plugins or from a plugin and the framework, stop the binary at start-up with bonfire.ErrDuplicateCommand naming the command. Remove or rename one of them; a plugin that appends centrifugo.Commands or flare.Commands must be the only one that does.

Reserved flag names. A command flag named like a global option (verbose, quiet, ansi, no-ansi, no-interaction, env, or the shorthands v, q, n) stops the binary with bonfire.ErrReservedFlag. Rename the flag.

Behaviour changes

  • Daily log files. Logs go to the console and, from warnings up, to storage/logs/summer-YYYY-MM-DD.log, kept for 14 days. The directory must be writable by the user that runs the binary; when it is not, the binary logs one warning and keeps logging to the console. To keep the old console-only behaviour, set log.default to console (SUMMER_LOG__DEFAULT=console). See Logging.
  • Component loggers. Every log record carries a component attribute. GORM's lines move to component db and record-not-found is no longer logged; River's lines to component river (warnings and up by default). Raise or lower a component with log.components, for example components: {db: debug}.
  • The request log is on by default. Every request writes one Info record http request from component http with the route pattern (never the path), status, duration and request_id. Silence it with components: {http: warn} in config/log.yaml, or turn it off with http.access_log: false.
  • serve prints a startup summary. On a terminal it is a block; off a terminal serve prints nothing on stdout and logs a startup summary record and a listening on <addr> record on stderr instead, so a script that waits for listening on 127.0.0.1:<port> in the combined output still finds it. Ctrl+C prints a shutdown report and exits 0; a failed drain exits 1. queue:work and schedule:run do the same.
  • Exit codes. A failed command exits 1 and a usage error (an unknown command or flag, a wrong number of arguments) exits 2. Errors print as a block on stderr with a hint where the framework knows the fix, and -v shows the wrapped errors.
  • -q keeps data. -q hides information and success lines, spinners and progress bars, but still prints what a command exists to produce: the token of admin:token:create, the key of key:generate and every --json output.
  • Production confirmation. migrate, migrate:rollback, deferred:purge and queue:clear ask before they run when the environment is production and they run on an interactive terminal; --force skips the question. Off a terminal, as in deploy scripts over SSH, nothing changes.
  • summer dev runs serve. It used to start the binary without a command, which printed the help and stopped; it now runs bin/<app> serve with the global options you give summer dev.
  • Result lines. summer build ends with Built bin/<app> in <time>, the make: commands print Created <path> per file, migrate prints Applied N migrations or Nothing to migrate, and queue:clear prints Cleared N jobs from <queue>. Adjust scripts that matched the old lower-case lines.

Upgrading to 0.2

0.2 tightens two configuration checks so that settings which leaked credentials stop the start-up instead. Most applications are not affected: the first concerns applications that load an auth.cookie block, the second applications that send credentialed CORS answers. Start the new binary once in each environment before you switch traffic to it; a refused configuration fails serve at start-up with an error that names the key.

Boot refusals

A token cookie that is not Secure in production. In the production environment, bouncer.LoadCookieConfig refuses a token cookie whose Secure attribute resolves to false. That covers auth.cookie.secure: false and, since 0.2, an unset auth.cookie.secure with an http or empty app.url, from which the cookie inherited Secure=false before. The error names auth.cookie.secure and app.url.

It affects applications whose plugins load an auth.cookie block for a browser frontend (the cookie transport). To fix it, set app.url to the application's public https URL, or set auth.cookie.secure: true explicitly. See Cookie transport.

Credentialed CORS with a wildcard host or an unanchored pattern. With supports_credentials: true, globally under http.cors or in any profile under http.cors_profiles.<name>, these entries now fail the start-up, and the error names the key with its index, such as http.cors.allowed_origins[1] or http.cors_profiles.spa.allowed_origins_patterns[0]:

  • an allowed_origins entry with a * that is not the whole leftmost host label after a literal scheme, such as https://*, *://*, https://*example.com or http*://www.example.com (https://*.example.com stays valid);
  • an allowed_origins_patterns entry not anchored with ^ and $, in every alternative of the expression;
  • any origin rule that matches an origin outside the configured domains, such as https://*.com or ^https?://.*$.

It affects applications that send credentialed CORS answers, to a browser frontend on another origin that calls the API with cookies. To fix it, list the frontend's exact origins, or write anchored patterns such as ^https://[a-z]+\.example\.com$. Without credentials these entries load as before. See CORS and CORS profiles.

Behaviour changes

These need no configuration change, but a test or an operator may notice them:

  • Site-cache purges are queued. After an admin save commits, the purge request to the site is queued and sent by a background worker, with at most one request in flight per purger; saves committed meanwhile collapse into one follow-up request. The save no longer waits for the site. A test that asserts a purge calls cooler.SiteCachePurger.Wait first, and serve drains the queue on shutdown. See Site cache purge.
  • Protected uploads need a private bucket. While storage.uploads.public_url is set, attach.Store refuses a protected (is_public false) upload with attach.ErrProtectedNeedsPrivateBucket before it writes anything, and serve logs one warning at start-up that the bucket must hold public files only. Keep protected relations on a bucket without public_url, or make the relation public. See Public URL base.
  • The Journal plugin's purge keys are deprecated. golem15.journal.site_cache.purge_url and golem15.journal.site_cache.purge_token still work as aliases of the framework's site_cache.purge_url and site_cache.purge_token, which win when both are set; the plugin logs a deprecation warning. Move the values to the site_cache keys.
  • Cookie defaults. A bouncer.CookieConfig built by hand with every field zero writes SameSite=Lax and Secure; a config with any field set keeps its own Secure value.
  • The API documentation page at /api-docs runs only scripts that carry the page's per-request nonce. See API documentation page.
  • Turnstile without a secret. When turnstile.secret is empty and routes name turnstile:open or turnstile:closed, the start-up logs one warning with the number of such routes. See Turnstile.
  • CORS profile headers on every answer. A route that names cors:<profile> carries the profile's headers on every response, including refusals from guards, throttle, Turnstile and the body limit, and recovered panics.

What 0.2 adds since 0.1

The 0.1 series grew through many patch releases. These are the larger features that arrived after its first public releases, with the page that documents each: