How to Intercept WordPress 7 AI Connector Requests

A practical guide to hooking into the WP 7 AI client and correctly reading token usage from the SDK.

---

The Three Hooks

WordPress 7 exposes three hooks for AI connector traffic:

Hook	Type	Callback Signature	When It Fires
wp_ai_client_prevent_prompt	filter	(bool $prevent, WP_AI_Client_Prompt_Builder $builder): bool	Before the AI request is sent. Return true to block.
wp_ai_client_before_generate_result	action	(BeforeGenerateResultEvent $event): void	After gating, before the HTTP call to the provider.
wp_ai_client_after_generate_result	action	(AfterGenerateResultEvent $event): void	After a successful response, with token usage available.

Minimal Example

add_filter( 'wp_ai_client_prevent_prompt', function ( bool $prevent, WP_AI_Client_Prompt_Builder $builder ): bool {
    // Block all AI calls from cron.
    if ( wp_doing_cron() ) {
        return true;
    }
    return $prevent;
}, 10, 2 );

add_action( 'wp_ai_client_after_generate_result', function ( AfterGenerateResultEvent $event ): void {
    $usage = $event->getResult()->getTokenUsage();
    error_log( sprintf( 'AI used %d tokens', $usage->getTotalTokens() ) );
}, 10, 1 );

---

WP 7.0 Core Bug: Event Dispatcher Not Injected (Fixed in RC1)

Resolved. WP 7.0 beta shipped without passing the event dispatcher to the SDK PromptBuilder. This caused wp_ai_client_before_generate_result and wp_ai_client_after_generate_result to never fire. AI Valve previously worked around the bug by injecting the dispatcher via reflection at priority 5 of the wp_ai_client_prevent_prompt filter.

As of WP 7 RC1, the WP_AI_Client_Prompt_Builder constructor now passes AiClient::getEventDispatcher() to the SDK PromptBuilder directly. The reflection workaround has been removed from AI Valve.

---

SDK API Reference (Correct Accessor Patterns)

The SDK uses private properties with getter methods. Direct property access ($obj->property) will fail silently or throw errors.

TokenUsage

Returned by $result->getTokenUsage(). Always non-null.

Method	Return
getPromptTokens()	int
getCompletionTokens()	int
getTotalTokens()	int
getThoughtTokens()	?int

Wrong: $usage->promptTokens Right: $usage->getPromptTokens()

Model Metadata

Available on BeforeGenerateResultEvent and AfterGenerateResultEvent via $event->getModel().

$model = $event->getModel();

// Provider info (e.g. "azure_openai")
$provider_id = $model->providerMetadata()->getId();

// Model info (e.g. "gpt-4.1")
$model_id   = $model->metadata()->getId();
$model_name = $model->metadata()->getName();

GenerativeAiResult

Available on AfterGenerateResultEvent via $event->getResult().

$result = $event->getResult();
$usage  = $result->getTokenUsage();       // TokenUsage (non-nullable)

Capability

$capability = $event->getCapability(); // ?string — e.g. "text_generation"

---

Caller Attribution

The AI hooks don’t tell you which plugin made the call. Use debug_backtrace() to walk the call stack and match file paths against WP_PLUGIN_DIR:

$trace = debug_backtrace( DEBUG_BACKTRACE_IGNORE_ARGS, 30 );

foreach ( $trace as $frame ) {
    if ( ! isset( $frame['file'] ) ) {
        continue;
    }
    $path = wp_normalize_path( $frame['file'] );
    if ( str_starts_with( $path, wp_normalize_path( WP_PLUGIN_DIR ) ) ) {
        // Extract plugin slug from the path segment after wp-content/plugins/
        $relative = substr( $path, strlen( wp_normalize_path( WP_PLUGIN_DIR ) ) + 1 );
        $slug     = explode( '/', $relative )[0];
        break;
    }
}

Tip: Stash the detected slug in a static property during prevent_prompt, then reuse it in before_generate_result and after_generate_result for correlation.

---

Complete Logging Example

Putting it all together — a self-contained snippet that gates requests and logs token usage:

add_filter( 'wp_ai_client_prevent_prompt', 'my_gate_prompt', 10, 2 );
add_action( 'wp_ai_client_after_generate_result', 'my_log_usage', 10, 1 );

function my_gate_prompt( bool $prevent, WP_AI_Client_Prompt_Builder $builder ): bool {
    if ( $prevent ) {
        return true;
    }
    // Example: block a specific plugin.
    $slug = detect_caller_plugin(); // Your backtrace logic.
    if ( 'expensive-plugin' === $slug ) {
        return true;
    }
    return false;
}

function my_log_usage( AfterGenerateResultEvent $event ): void {
    $model = $event->getModel();
    $usage = $event->getResult()->getTokenUsage();

    error_log( sprintf(
        '[AI] provider=%s model=%s capability=%s tokens=%d (prompt=%d completion=%d)',
        $model->providerMetadata()->getId(),
        $model->metadata()->getId(),
        $event->getCapability() ?? 'n/a',
        $usage->getTotalTokens(),
        $usage->getPromptTokens(),
        $usage->getCompletionTokens(),
    ) );
}

---

Debugging Checklist

If hooks aren’t firing as expected:

1. Is the AI connector configured? — wp_supports_ai() must return true.
2. Is wp_ai_client_prompt() being used? — Only this function triggers the hooks. Direct SDK usage bypasses them.
3. Are before/after actions silent? — Ensure you’re on WP 7 RC1+. Earlier betas had a core bug where the event dispatcher wasn’t injected (now fixed).
4. Are you getting errors on token access? — Use getter methods, not property access.
5. Is your prevent_prompt filter returning the wrong type? — Must return bool. Returning a non-bool can break the filter chain.
6. Check hook registration: wp eval 'var_dump( has_filter("wp_ai_client_prevent_prompt") );'
7. Check event dispatcher exists: wp eval 'var_dump( WordPress\AiClient\AiClient::getEventDispatcher() );'

📦 Source: soderlind/ai-valve · Edit on GitHub