wp-plugin-architecture

Designs and reviews the internal architecture of a WordPress plugin — folder layout, PSR-4 one-class-per-file discipline, Schema/Constants placement, composition-root vs singleton decisions, condition

SKILL.md

GitHub

WordPress plugin: internal architecture

How the plugin organizes itself inside includes/ (or wherever the PSR-4 root lives) once the bootstrap (see wp-plugin-bootstrap) and lifecycle (see wp-plugin-lifecycle) are in place. The bootstrap is the launcher; this skill is the engine layout.

Out of scope: bootstrap-file content, activation / deactivation / uninstall, cron specifics, REST endpoint design — covered by sibling skills.

When to use this skill

Trigger when ANY of the following is true:

  • Scaffolding the includes/ (or src/) folder of a new plugin.
  • Reviewing the class layout in a PR — folder structure, where things live, what's reused.
  • Deciding whether a meta key / option name belongs in a class const, a Schema.php, or a PHP enum.
  • Reviewing asset enqueue code — wrong hook, missing dependencies, unconditional loading.
  • The user asks "should this be a singleton" or "where should this constant live".

Folder layout — by-type or by-feature

There are two reasonable organizational schemes for includes/. Pick one and stay consistent.

By-type (group classes by their WP role). Works for small-to-medium plugins (≤ 15 classes, 1-3 features):

includes/
├── Plugin.php              # composition root / wiring
├── Schema.php              # central constants
├── Actions/                # JFB / WC actions
├── Events/                 # JFB events
├── Settings/               # admin settings
├── Rest/                   # REST controllers
└── Api/                    # external HTTP clients

By-feature (group everything that belongs to a feature together). Scales better past 3-4 distinct features:

includes/
├── Plugin.php
├── Schema.php
├── Verdict/                # AI Verdict feature
│   ├── VerdictAction.php
│   ├── VerdictTrueEvent.php
│   └── VerdictFalseEvent.php
├── Enrichment/
│   ├── EnrichmentAction.php
│   └── EnrichmentDoneEvent.php
├── Settings/
│   └── SettingsTab.php
└── Updater/
    └── UpdateChecker.php

The wrong move is mixing both in the same plugin. A reader gets confused, an AI gets lost, and refactors become tedious. By-type is fine until it isn't — when a third feature ships and Actions/ has 9 classes from 3 unrelated domains, switch the whole plugin to by-feature.

Hard rule across both styles: one class per file, file name matches class name (VerdictAction.php → class VerdictAction), PSR-4 maps <RootNamespace>\Verdict\VerdictAction to includes/Verdict/VerdictAction.php. The kebab-case class-foo-bar.php filename is a holdover from PHPCS-WPCS-old; modern plugins use PascalCase that matches the class.

Centralization — Schema / Constants is non-negotiable

Every string that appears more than once — meta key, option name, custom hook name, cron event name, CPT slug, capability slug, transient prefix — lives in one place. Three patterns that all work:

Class const on the feature class

class UsageTracker {
    public const OPTION_PREFIX = 'myplugin_usage_';

    public static function current_month_key(): string {
        return self::OPTION_PREFIX . gmdate( 'Y_m' );
    }
}

When the constant logically belongs to one feature, define it there. Cleanest scope.

Dedicated Schema class for cross-feature constants

namespace MyPlugin;

final class Schema {
    public const META_FORM_SETTINGS = '_myplugin_form_settings';
    public const OPTION_GLOBAL      = 'myplugin_global_settings';
    public const CPT_LOG            = 'myplugin_log';
    public const CRON_DAILY         = 'myplugin_daily_cleanup';
    public const HOOK_BEFORE_REQUEST = 'myplugin/before_request';

    private function __construct() {} // not instantiable
}

Then everywhere: Schema::META_FORM_SETTINGS instead of the literal string. Renaming becomes one-line; typos become impossible (the class autoloader catches them).

PHP 8.1+ enum for typed value sets

For a fixed value set (failure modes, output types, etc.), use enums only when the plugin's declared minimum PHP version is 8.1 or higher:

enum FailureMode: string {
    case Halt        = 'halt';
    case Permissive  = 'permissive';
    case Restrictive = 'restrictive';
}

// Then instead of error-prone string parsing:
$mode = FailureMode::tryFrom( $raw ) ?? FailureMode::Halt;

Type-safe, enumerable, documents itself.

For PHP 7.4 / 8.0-compatible plugins, use class constants plus explicit validation instead:

final class FailureMode {
    public const HALT        = 'halt';
    public const PERMISSIVE  = 'permissive';
    public const RESTRICTIVE = 'restrictive';

    public static function normalize( string $raw ): string {
        return in_array( $raw, self::all(), true ) ? $raw : self::HALT;
    }

    public static function all(): array {
        return array( self::HALT, self::PERMISSIVE, self::RESTRICTIVE );
    }
}

The hard rule: if you find yourself typing the same magic string in two files, that's the moment to centralize. The wrong direction is "I'll only have it in two places, no need yet" — git grep next year proves the lie.

Singleton discipline — composition root yes, everything else no

The Plugin class is often the composition root: one object wires services, hooks, controllers, settings, and integrations after bootstrap. It MAY expose a small instance() helper when the surrounding plugin style already uses that pattern, but a singleton is not required. A plain new Plugin(...)->register() from the bootstrap is usually easier to test.

Beyond the composition root, don't make everything a singleton by reflex.

  • A SettingsRepository doesn't need getInstance() — instantiate it where you need it (new SettingsRepository()); it's cheap.
  • A logger MAY be a singleton if it holds connection state (a buffer, a remote handler) — but most plugin loggers wrap error_log, which is itself globally available. No state, no singleton.
  • An external API client (Stripe, OpenAI, Slack) is NOT a singleton. Inject the API key + dependencies at construction; pass it where it's needed.

The price of singleton-everywhere: tests can't substitute mocks, dependencies become hidden, and you accumulate a Plugin::instance()->get_storage_manager()->get_provider_registry() getter chain that nobody can refactor.

When in doubt: write the class as a regular class. Promote to singleton only when there's a concrete reason (genuine global state, expensive lazy initialization shared across many call sites, or compatibility with an existing plugin API).

Asset enqueueing — conditional, on the right hook

The wrong way: enqueue every script and stylesheet on every page load via wp_enqueue_scripts. The right way:

  1. Pick the right hook for the right context. Verified in WP source:

    • Frontend pages: wp_enqueue_scripts (wp-includes/script-loader.php:2311)
    • wp-admin pages: admin_enqueue_scripts (wp-admin/admin-header.php:123) — receives $hook_suffix argument identifying the current admin page
    • Block editor (Gutenberg): enqueue_block_editor_assets
    • Front + back of blocks: enqueue_block_assets
    • Login screen: login_enqueue_scripts
    • Customizer preview: customize_preview_init
  2. Gate by context inside the hook callback:

    add_action( 'admin_enqueue_scripts', static function ( string $hook_suffix ): void {
        // Only on the plugin's own settings page
        if ( $hook_suffix !== 'settings_page_myplugin' ) {
            return;
        }
        wp_enqueue_script( /* ... */ );
    } );
    

    For frontend, gate by is_singular() / has_block( 'myplugin/contact' ) / specific shortcodes.

  3. Always declare dependencies — even ones that "feel obvious". wp-element, wp-components, wp-i18n, wp-hooks, your own scripts. Without proper deps the script may run before the dependency is defined and break.

  4. Cache-bust deterministically. Two valid patterns:

    • filemtime( $script_path ) — perfect during development (every save invalidates cache).
    • The plugin VERSION constant — best for production (predictable, changes on release).
    • A hybrid: defined('SCRIPT_DEBUG') && SCRIPT_DEBUG ? filemtime() : VERSION.
  5. $args array since WP 6.3 — the 5th parameter overloads from a boolean $in_footer to an array supporting strategy ('defer' or 'async'), in_footer, and fetchpriority (since 6.9). Use 'strategy' => 'defer' instead of jQuery-era ready-handlers when feasible.

wp_add_inline_script preferred over wp_localize_script for new code

wp_localize_script (since 2.2) creates a JavaScript object from an associative array. It still works, but for arbitrary runtime config it is a legacy-shaped helper: it requires an array payload and creates a top-level global. Use wp_set_script_translations() for JavaScript translations.

wp_add_inline_script (since 4.5) is the modern path: any JS string, before or after the registered script. Use wp_json_encode() to serialize structured data:

wp_add_inline_script(
    'myplugin-editor',
    'window.MyPluginConfig = ' . wp_json_encode( array(
        'restUrl' => esc_url_raw( rest_url( 'myplugin/v1/' ) ),
        'nonce'   => wp_create_nonce( 'wp_rest' ),
    ) ) . ';',
    'before'
);

Both functions still work; for a new plugin, default to wp_add_inline_script for configuration and wp_set_script_translations() for script translations.

Hook naming — prefix everything

Custom action / filter hooks the plugin emits (so other developers can wire into your behavior) follow ONE naming convention, picked once and never deviated from:

// Slash-separated (common in modern plugins and integrations)
do_action( 'myplugin/before_request', $payload );
apply_filters( 'myplugin/api_response', $response, $request );

// Underscore-separated (classic; common in WordPress core)
do_action( 'myplugin_before_request', $payload );

Both are common in the ecosystem, but WordPress core itself mostly uses underscore-separated hook names such as pre_get_posts and rest_pre_serve_request. If the project uses WPCS rules that discourage slashes in hook names, pick underscores. The hard rule: prefix every custom hook with the plugin slug. do_action( 'before_request', ... ) is a foot-gun — collisions with other plugins are guaranteed at scale.

When you DO emit a custom hook, give it a docblock right above:

/**
 * Fires before the AI request is sent.
 *
 * @since 1.0.0
 *
 * @param array $payload The request payload (mutable downstream).
 */
do_action( 'myplugin/before_request', $payload );

@since lets users pin their integration; AI assistants reading the source rely on these to suggest correct hooks.

Critical rules

  • One class per file, file name matches class name, PSR-4 namespace mirrors the folder path.
  • Pick by-type or by-feature, never both in the same plugin. Switch styles by refactoring the whole includes/ at once.
  • Centralize every repeated string (meta keys, option names, hook names, cron events, CPT slugs) in a class const, a Schema class, or a PHP 8.1+ enum when the plugin requires PHP 8.1+.
  • Plugin class is a composition root; singleton is optional. Default all other classes to regular classes; promote to singleton only with a concrete reason.
  • Enqueue conditionally on the right hook (wp_enqueue_scripts / admin_enqueue_scripts($hook_suffix) / enqueue_block_editor_assets); never enqueue unconditionally in init.
  • wp_add_inline_script over wp_localize_script for shipping config / JSON to JS. Both work; the inline-script path is the modern default.
  • Prefix every custom hook with the plugin slug. No collision-prone bare names.
  • $args array in wp_enqueue_script (since 6.3) — use 'strategy' => 'defer' where possible.

Common mistakes

// WRONG — magic strings scattered across files
update_post_meta( $post_id, '_myplugin_settings', $value ); // file A
get_post_meta( $post_id, '_myplugin_setings', true );        // file B (typo!)

// RIGHT — one source of truth
update_post_meta( $post_id, Schema::META_SETTINGS, $value );

// WRONG — enqueue on init unconditionally
add_action( 'init', function () {
    wp_enqueue_script( 'myplugin-frontend', /* ... */ );
} );
// queues assets in broad request contexts and bypasses page-specific gates

// RIGHT — frontend hook + context gate
add_action( 'wp_enqueue_scripts', function () {
    if ( ! has_block( 'myplugin/contact' ) ) return;
    wp_enqueue_script( 'myplugin-frontend', /* ... */ );
} );

// WRONG — singleton by reflex
final class HttpClient {
    private static ?self $instance = null;
    public static function instance(): self { /* ... */ }
}
// 1 plugin = 1 client, can't test, can't swap base URL per call site

// RIGHT — regular class with explicit dependencies
$client = new HttpClient( $api_key, $base_url );
$client->post( '/things', $payload );

// WRONG — bare hook name
do_action( 'before_save', $data );  // collides with 5 other plugins

// RIGHT
do_action( 'myplugin/before_save', $data );

Cross-references

  • Run wp-plugin-bootstrap for the main file (header, autoload, requirements, Plugin bootstrapping).
  • Run wp-plugin-lifecycle for activation / deactivation / uninstall.php.
  • Run wp-security-audit on any classes that handle user input — the architecture is irrelevant if the controllers leak.

What this skill does NOT cover

  • Dependency-injection containers (PHP-DI, Symfony Container, etc.). For most WP plugins they're overkill; constructor injection without a container is plenty.
  • Service layer architecture beyond "regular class, not singleton" — domain-driven design, hexagonal, etc. are valid choices but are a different conversation.
  • Block / Gutenberg-specific architecture (block.json, Edit / Save components) — separate skill territory.
  • Test architecture — what kind of tests, where they live, how they boot WP. Adjacent topic, separate skill.
  • Performance tuning of asset enqueue (concatenation, defer, preload) beyond what the WP 6.3+ $args array enables.

References