Hooks
The plugin exposes three Kirby hooks. They fire for every individual text the pipeline processes – including text inside blocks, structures, layouts, and table cells.
| Hook | When | Return |
|---|---|---|
content-translator.translate:before | Before a unit is sent to the strategy | Modified text (string) |
content-translator.translate:after | After the strategy returns a translation | Modified text (string) |
content-translator.translate:warning | When a unit is dropped (per-unit failure) | n/a (event-style) |
Translator entry points run the full pipeline. Terminology enforcement, logging, or drop alerting wired up here therefore skips Panel AI translations silently.On the paths where they do fire, :before and :after run for every unit holding content, including values no provider ever sees – a field that is just a price or a URL still runs both hooks, with the text unchanged. Inspect $text if a hook should only act on prose.
content-translator.translate:before
Rewrite text on its way to the strategy. The returned string replaces the unit's text and is always sent on – the hook cannot skip a unit or cancel the translation. To keep content out of translation entirely, narrow fieldTypes or list the field in excludeFields.
de, fr, en-gb, …).null if not specified.text for now. Reserved for future expansion.TranslationUnit (text, fieldKey). Lets you branch on the originating field or table cell – but only where there is one, see Field-Aware Preprocessing.ExecutionOptions carrying both targetLanguage and sourceLanguage as TranslationLanguage value objects.apply() matches by parameter name, so a closure only has to declare the payload keys it uses.return [
'hooks' => [
'content-translator.translate:before' => function ($text) {
// Strip internal editorial markers so they never reach the provider
return str_replace(['[draft]', '[review]'], '', $text);
}
]
];
content-translator.translate:after
Postprocess translated text – language-specific formatting, terminology enforcement, logging.
:before rewrites. To see what actually went to the strategy, read $unit->text.text for now.return [
'hooks' => [
'content-translator.translate:after' => function ($text, $originalText, $targetLanguage) {
// Restore untranslatable terms
$protected = ['API', 'CSS', 'HTML', 'JavaScript'];
foreach ($protected as $term) {
$text = preg_replace('/\b' . $term . '\b/i', $term, $text);
}
return $text;
}
]
];
content-translator.translate:warning
Fires once per unit that was dropped. The unit keeps its source text. Silent by default – wire it to logging, Sentry, or Slack to surface drops in production.
How much survives depends on where the drop happened. missing translation, non-string translation, empty translation and placeholder mismatch are caught after the strategy returns, so they apply to every strategy – including a custom one – and only the affected unit is reverted. The remaining reasons come from the strategy itself: CopilotAIStrategy drops per unit and per chunk, so the rest of the batch still gets written, while DeepLStrategy sends the batch in one request and, when it fails, fires a warning for every unit and then throws TranslationException, so nothing from that batch is stored. Treat a DeepL warning as a batch-level failure that happens to be reported per unit.
Drop Reasons
| Reason | Strategy | Cause |
|---|---|---|
<upstream error message> | DeepLStrategy | Upstream batch request threw |
<upstream error message> | CopilotAIStrategy | Provider call threw (rate limit, network, auth) |
response length mismatch | CopilotAIStrategy | AI returned the wrong number of translations |
placeholder mismatch | any | Translation lost, invented or repeated a <cN/> |
non-string translation | any | Strategy returned a non-string entry for the unit |
empty translation | any | Strategy returned a string holding nothing but whitespace |
missing translation | any | Strategy returned fewer entries than it was given units |
Example: Send Drops to Sentry
use Sentry\State\Scope;
use function Sentry\captureMessage;
use function Sentry\withScope;
return [
'hooks' => [
'content-translator.translate:warning' => function ($unit, $reason, $previous) {
withScope(function (Scope $scope) use ($unit, $reason, $previous) {
$scope->setExtra('field', $unit->fieldKey);
$scope->setExtra('text_excerpt', mb_substr($unit->text, 0, 200));
if ($previous !== null) {
$scope->setExtra('upstream_error', $previous->getMessage());
}
captureMessage('Translation dropped: ' . $reason);
});
},
],
];
Field-Aware Preprocessing
The unit payload key lets you branch on the originating field. fieldKey is only filled in when the translation walks a model's content – that is translateContent() and the CLI. Panel translations post a flat list of texts to the batch endpoint, and standalone translateTexts() calls carry no field context either, so fieldKey is null in both cases. The example below stays safe either way, because a null key never matches:
return [
'hooks' => [
'content-translator.translate:before' => function ($text, $targetLanguage, $sourceLanguage, $type, $unit) {
// Tighter prompts for headlines
if (in_array($unit->fieldKey, ['headline', 'title'], true)) {
return trim($text);
}
return $text;
}
]
];
Notes
Hooks fire for every individual text – they may run hundreds of times during a batch translation. Keep them fast.
:after hook applies to the final stored content. Bugs in your hook propagate to disk – review carefully before deploying.