TYPO3 Extension Upgrade Skill

Framework for upgrading TYPO3 extensions to newer LTS versions. Extension code only, not project/core upgrades.

Done is exit=0 from the step 10 command, run last on every line the constraint names that installs — read from the run, not assumed. Not green on one line with another left red. Not green because failing tests were skipped or deleted. Not a commit with --no-verify. Anything short of that is reported as not done, with the lines that show why.

A commit hook that fails is a check the repository requires, not an obstacle. Read which command it ran, run that command yourself, fix what it reports, and commit again: code style is usually fixed by the repository's own fixer (php-cs-fixer fix with its config), PHPStan findings are fixed in the code. Measured: three agents with both TYPO3 lines green bypassed the style or PHPStan hook with --no-verify, and a fourth bypassed its own failing unit tests.

Upgrade Toolkit

Core Workflow

1. Complete planning phase (consult references/pre-upgrade.md)

2. Create feature branch (verify git is clean)

3. Update composer.json constraints for the target version. v14's LTS minor is 3: write ^14.3, never ^14.4 — ^13.4 || ^14.3 to support both. v11.5, v12.4 and v13.4 make 14.4 look like the next in line; it matches no release, so composer update fails dependency resolution, exits non-zero and installs nothing. Constraints for every version pair: references/upgrade-v13-to-v14.md

   Adding a version is not replacing one. Writing ^14.3 alone drops support for every line the extension had before, which is a breaking change for everyone installing it — and it is the cheap way to make the new line resolve, so it happens by accident. "Make it work with the current LTS" asks for the new line, not for the loss of the old ones. Read the existing constraint and keep every line in it, then add the target: ^12.4 || ^13.4 becomes ^12.4 || ^13.4 || ^14.3, and only an extension that already supported v13 alone ends up at ^13.4 || ^14.3. Drop a line only where that was asked for, and say so where a maintainer will see it. Whether one codebase can serve them all is a separate question, answered in references/dual-compatibility.md.

4. Audit third-party dependencies for major version changes (consult references/third-party-dependency-upgrades.md)

5. Run rector process --dry-run then review and apply

6. Run fractor process --dry-run then review and apply

7. Run php-cs-fixer fix

8. Run phpstan analyse against each supported dependency version and fix errors

9. Run phpunit and fix tests. Tests/ is part of the upgrade, not a consequence of it. A class v14 removed, referenced from a test, is fatal rather than failing: in an import, a parent, a property or a signature it stops PHPUnit while it loads the suite, so nothing runs at all. Search for the removed types across both trees before running anything: grep -rnE 'TypoScriptFrontendController|StandaloneView|TemplateView|HashService|LocalPreviewHelper|LocalCropScaleMaskHelper|FreezableBackendInterface' Classes/ Tests/ Every hit is a fix, and deleting the use line is not one: without the import, Foo::class resolves through the current namespace, so the file names a class nobody ever wrote and the error surfaces later, on the line that still has to support the old version. Change the usage, not the import. createMock on a removed class cannot be repaired by swapping the name either — see references/upgrade-v13-to-v14.md

   Then run this, from the extension root, before the suite and again after the edits. It is the check for the edit above, as a block that runs as pasted — no path to find, no variable to bind. A line of output is a file that uses a removed type by its short name with no use line for it; fix each one before step 10. No output and exit 0 is clean — it exits 1 on a finding, so && phpunit after it does not run the suite over the fault.

   types='TypoScriptFrontendController|StandaloneView|TemplateView|HashService|LocalPreviewHelper|LocalCropScaleMaskHelper|FreezableBackendInterface'
   strip='foreach (token_get_all(file_get_contents($argv[1])) as $t) { if (!is_array($t)) { echo $t; } elseif (!in_array($t[0], [T_COMMENT, T_DOC_COMMENT, T_CONSTANT_ENCAPSED_STRING, T_ENCAPSED_AND_WHITESPACE], true)) { echo $t[1]; } }'
   bad=0
   for f in $(grep -rlE --include='*.php' "\b($types)\b" Classes Tests 2>/dev/null); do
     code=$(php -r "$strip" "$f")
     for t in $(grep -oE "\b($types)\b" <<<"$code" | sort -u); do
       grep -qE "^use .*\\\\$t;" <<<"$code" || grep -qE "\\\\$t\b" <<<"$code" \
         || { echo "$f: $t used without import"; bad=1; }
     done
   done
   test "$bad" = 0
   

   Copy

   Comments and string literals are stripped first: a docblock naming a removed type is not a reference PHP resolves. The same check is TU-58 in this skill's checkpoints.yaml, for the runner in automated-assessment. The block is here because a script has to be reachable to be a shortcut. Measured on the same case and model, at this step: an instruction to run the runner was in context in three of three trials and run in none; this block, pasted whole, was run in five of six.

10. Install the target version and run the suite against it. A green suite on the version already installed proves nothing about the target — that is the old code passing old tests. Install first, in two passes, then test:

    composer update "typo3/*" --with typo3/cms-core:^14.3 -W --no-install \
      && rm -rf vendor && composer install \
      && composer show typo3/cms-core | grep '^versions' \
      && vendor/bin/phpunit -c Build/phpunit/UnitTests.xml; echo "exit=$?"
    

    Copy

    Chained, so a failed install stops before PHPUnit runs against the old tree; the versions line says what was actually installed, and exit= is the status of whichever step ran last. One pass fails on the way from v13 to v14: Composer upgrades typo3/class-alias-loader and then runs the old plugin against the new files, which dies with Class "…\CaseSensitiveToken" not found after the lock file is written. composer install on top of the old vendor/ dies the same way, so vendor/ goes first. The package argument matters: without it composer update moves every dependency, and the test result then depends on upgrades that have nothing to do with TYPO3. -W lets the TYPO3 packages' own dependencies follow. If the target will not install, the upgrade is untested. Say so in the report, with the error, and do not report the suite as passing: a green run on the version that was already there is that version's result.

11. Run the same command for every older line the constraint keeps — --with typo3/cms-core:^13.4 for ^13.4, and so on. A migration that turns the target green can break the line below it: measured, an agent got v14 green, left v13's suite failing, and reported all three lines compatible. Fix, then run every line again. A line that cannot be installed here at all — Composer cannot resolve or download it — is untested, not failed: report it with Composer's error, and never drop it from the constraint for that reason (step 3).

12. Before reporting, read the exit= line of the last run on each line. On a line that installed, anything other than exit=0 means the work is not done: go back to step 9. A line whose install failed is the untested case from step 11. Committing past a red suite, or with --no-verify, is not done either. The report says, per line, what was installed and whether the suite passed — and where it did not, the failure as PHPUnit or Composer printed it.

13. Verify success criteria (consult references/verification.md)

Done means the suite passes on every line the constraint names, with that line installed — every line that can be installed; the rest are reported as untested. Not that the constraint was widened, and not that the suite is green where it was already green.

Measured: agents reported "✅ 719 tests pass" having never installed the target, having seen the suite fail to load three times, or having seen three errors and thirteen failures and committed with --no-verify. The run's own exit= line is what tells you which of those you are in; read it before you write the report.

Where a removed class is referenced decides when it bites. In an import, a parent class, a property or a signature it is resolved while PHPUnit loads the suite, so nothing runs at all and the failure looks nothing like a test failure. Referenced only inside a method body, it fails when that one test executes, and the rest of the suite still passes — which is the more comfortable failure and the easier one to miss in a summary line.

When the migration breaks the tests

It will. A suite that was green before Rector routinely comes back with dozens of errors, and working through them is the job.

Never revert the migration to get back to green. Measured: an agent ran Rector, saw Tests: 719, Errors: 34, discarded every migrated file under Classes/, got OK (719 tests, 1176 assertions), and committed composer.json, ext_emconf.php and two build files — no code at all. That commit claims support for a version the code does not have, and it is the worst of the three possible outcomes: a failing upgrade is visible, an unattempted one is honest, and this one is neither.

Green after a revert is the state you started in. The only green that counts is the one from step 10, with the target version installed and the migration in place.

Skipping or deleting a failing test to get green is the same revert, one test at a time. One skip is legitimate: where the code branches by version, a test of the branch that exists only on the older line — if (!class_exists(X::class)) { self::markTestSkipped(...); } — may skip on the newer line, provided the newer line's branch has a test of its own. Any other skip, a version check that returns early, or a removed test method turns a failure into silence. So does --no-verify on the commit: it silences the checks that would report it.

Nor stop to ask whether to proceed. Measured: an agent installed the target, counted the uses of a class the new version removed, ran git reset --hard, and asked whether it should do the refactoring. A request to make the extension work with the new version is that permission. Replacing what the new version removed is the upgrade, however many places use it — it is the work, not a finding to report back. Ask only about what the request leaves open, such as dropping a supported line (step 3).

When NOT to Apply Automatically

Do NOT blindly apply Rector/Fractor when dual-version compatibility, missing tests, unclear changes, or complex APIs (DBAL, Extbase) are involved. Instead apply rules manually, testing between changes.

Third-Party Dependency Upgrades

When composer.json widens a dependency to a new major version: enumerate API usages, cross-reference the new API, verify mocks, use adapter pattern for signature differences, run PHPStan per major version. See references/third-party-dependency-upgrades.md.

Quick Commands

rector process --dry-run && rector process        # PHP migrations
fractor process --dry-run && fractor process       # Non-PHP migrations
php-cs-fixer fix && phpstan analyse && phpunit     # Quality checks

Asset Templates

Config templates in assets/: rector.php, fractor.php, phpstan.neon, phpunit.xml, .php-cs-fixer.php

References

External Resources

• TYPO3 Rector
• Fractor
• TYPO3 Core Changelog