>>> [!note] Migrated issue <!-- Drupal.org comment --> <!-- Migrated from issue #3591602. --> Reported by: [tim bozeman](https://www.drupal.org/user/2241356) >>> <h3 id="summary-problem-motivation">Problem/Motivation</h3> <p>Some AI-initiated work (e.g., image generation via text-to-image models) takes too long to complete within the chat turn's HTTP request. The AI agent needs to commit the entity and respond to the user immediately, while the slow work completes asynchronously and updates the page when finished.</p> <p>Without a deferred processing system, the options are:</p> <ul> <li>Block the chat response until generation completes (unacceptable UX &mdash; 10-30s wait with no feedback)</li> <li>Kick off image generation on the back end and require the user to manually refresh after some unknown delay</li> </ul> <h3 id="summary-proposed-resolution">Proposed resolution</h3> <p>Introduce a first-class deferred operation system with two client-side execution strategies:</p> <h4>Server-side deferred flow</h4> <ol> <li>AI tool persists entity; field handler emits <code>DeferredOperation</code> value objects for work it cannot resolve synchronously</li> <li>Event subscriber stores operations in PrivateTempStore and queues a <code>TriggerDeferredCommand</code></li> <li>Client shows a loading indicator (shimmer animation on placeholder), POSTs to <code>DeferredProcessingController::process</code></li> <li>Controller dispatches each operation to the matching tagged processor service</li> <li>On success: returns AJAX commands to re-render affected components with highlighting</li> <li>On retryable failure (e.g., content moderation rejection): client POSTs to retry endpoint, processor rewrites the prompt via LLM, tries again</li> </ol> <h4>Chat-driven followup flow</h4> <ol> <li>Server queues a <code>TriggerFollowupChatCommand</code> with an LLM-facing message</li> <li>Client sends a synthetic chat turn to the DeepChat API (user never sees it)</li> <li>Loops on <code>should_continue</code> (mirrors the agent polling loop)</li> <li>On completion, POSTs to <code>RefreshComponentsController</code> to re-render with highlights</li> </ol> <h4>Image generation processor</h4> <p>The canonical deferred processor implementation:</p> <ul> <li>Calls the configured <code>text_to_image</code> AI provider</li> <li>Creates a media entity from the generated image</li> <li>Updates the element on the entity</li> <li>Implements <code>RetryableProcessorInterface</code> &mdash; on content moderation failure, uses an LLM to rewrite the rejected prompt and retries (up to 2x)</li> </ul> <h4>Image generation field handler</h4> <p>An <code>entity_blueprint</code> field handler (<code>ImageGenerationHandler</code>) that intercepts entity_reference fields targeting media:</p> <ul> <li>When the AI sends <code>{_generate: "description"}</code>, sets a configured placeholder media and emits a <code>DeferredOperation</code></li> <li>On update: leaves the field unchanged and emits the deferred operation (element is updated and swaped later)</li> <li>Provides guidance text to the AI describing the <code>_generate</code> syntax</li> </ul> <h3 id="summary-ui-changes">User interface changes</h3> <ul> <li>Placeholder components show a shimmer animation (dashed outline + gradient sweep) while generation is in progress</li> <li>After generation completes, the component re-renders with the real image and receives a highlight animation</li> </ul>