AI

Registro de alterações

1.4.0 – 2026-10-05

Adicionado

• When comment moderation is run, calculate a value score, providing a relevance signal (0–1) for each comment based on how substantive and on-topic it is relative to the post it was left on. Show this score in a new column on the comment screen (#681).
• New experiment, Markdown Feeds, that creates markdown feeds at /feed/markdown/ (available in every feed context) and of individual posts and pages via ?output_format=markdown (#855).
• User-triggered retry for Content Translation if it fails (#941).
• Limit comment moderation bulk analysis queue size (#972).
• Storage and CRUD layer for embedding vectors (new wpai_embeddings table) that can be used to recording the provider and model alongside every vector (#976).
• 9 new default target languages to the Content Translation experiment (#986).
• Guideline categories support to the Content Translation Ability (#987).
• New Vector_Math and Vector_Ranker classes, allowing users to run similarity and ranking queries against vectors (#993).
• Loading indicator for type-ahead suggestion requests (#1029).
• Integrate Alt Text Generation with the Image block’s decorative setting (#1031).

Alterado

• Bump minimum supported WordPress version to 7.0.3 (#1001).
• Display specific taxonomy labels in Content Classification notice error messages instead of generic terminology (#916).
• Show a notice and disable title translation when the post title is too short (#967).
• Update to the latest version of the embedding code from the
  PHP AI Client (#975).
• Our generate_embeddings helper function now requires a
  specific embedding model. Pass model (a model ID or a ModelInterface
  instance) and, for a model ID, the provider that offers it. This is a
  breaking change for anyone that happened to start using this function (#975).
• Update Settings vertical ellipsis icon to tool icon (#1001).
• The core/read-content and core/read-users abilities are renamed to core/content-query and core/users-query. The old names keep working as deprecated aliases for now (#1002).
• Use gpt-image-2.5-flare as our default OpenAI image generation model (#1023).
• The core/read-settings ability has been renamed to core/settings-get. The old name keeps working as a deprecated alias for now (#1087).

Deprecated

• The ai/get-post-details ability. Use the single-post mode of core/content-query instead (#1002).

Removido

• Unnecessary URL validation that is now handled by WordPress core (#1011).
• The Enable AI header toggle from the AI settings page (#985).

Corrigido

• Ensure focus moves to the Accept button when content generation completes in the Content Resizing modal (#917).
• Increase the default request timeout from 5 seconds to 30 seconds when Core revalidates API keys. If this validation request times out, Core will delete the API key so this is an attempt to fix that (#947).
• Cached the active SEO plugin detection – with a TTL and immediate invalidation on any plugin activation/deactivation – so meta description meta-key lookups no longer re-scan active plugins on every request (#973).
• Improves the Content Translation loading animation to preserve existing text and background colors (#977).
• Settings page failing to render on Gutenberg 23.9+ after @wordpress/dataviews was removed from the private-apis allowlist (#989).
• Ensure our guidelines integration reads published guideline rows from the wp_knowledge post type instead of the removed wp_guideline so those work with the latest version of Gutenberg. We don’t migrate data from the old post type and taxonomy to the new so anyone that had previously set guidelines in an older version of Gutenberg will need to re-add those (#988).
• Import core’s prefixed PSR EventDispatcher in the vendored EmbeddingBuilder (#1005).
• Reserve space for type-ahead text generation when caret is at the end (#1008).
• Updated the settings page container min-height so that it spans the full viewport height (#1013).
• Fix focus style on collapsible card by updating the @wordpress/dataviews package to latest version (#1035).
• AI Request Logs now records HTTP error responses (4xx/5xx) from providers as error with the status code and message, instead of success; Gemini text generation requests are no longer mislabeled as Metadata (#1040).
• Hide the “Suggest Reply” button when the inline comment form is in Quick Edit mode (#1049).
• Editorial Notes only reviewing template blocks when “Show template” mode is enabled in the block editor (#1053).
• Insert the generated summary into the post content when Show template is enabled (#1055).
• Aligned the focus styles of Content Classification suggestion pills with the core Button component (#1060).
• Connector Approvals no longer blocks core’s connector key check by attributing it to the provider plugin or to Gutenberg (#1070).
• Abilities Explorer no longer crashes when an ability’s input property lists several types (#1073).
• Preserve HTML entities and text between shortcodes when normalizing content, while still encoding entity-encoded tags (#1077).
• Prevent options.php validation error when saving general settings with abilities active (#1080).
• Ensure post titles, the site tagline and excerpts no longer contain HTML entities in the markdown feeds (#1086).
• Ensure markdown feeds don’t 404 when the rewrite rules were rebuilt without the feed in place (#1110).

Segurança

• Ensure we check for the moderate_comments capability before we allow bulk comment moderation and then check for the edit_comment capability for each individual comment that is being moderated (GHSA-pjw7-q94g-4q34).
• Harden the Playground preview publishing workflow to derive pull request and commit metadata from the triggering workflow run (#978).
• Harden JSON data parsing on admin screen (#991).

1.3.0 – 2026-08-18

Adicionado

• New Experiment: Content Translation; translates Paragraph and Heading blocks—and optionally the post title—into a selected language directly from the post editor (#747).
• New Experiment: Slug Generation; suggest SEO-friendly permalinks that can be set as the slug (#897, #932).
• New Experiment: Custom Abilities. Gates the plugin’s custom WordPress Abilities (ai/get-post-details, ai/get-post-terms, core/read-settings, core/read-users, core/read-content) behind a single opt-in toggle, so enabling it exposes all of them at once via the Abilities API. Note for anyone that is using these Abilities, you’ll need to enable this new experiment first for those to be available (#881).
• New Developer Tool: Import/Export functionality for non-sensitive AI settings (#734).
• Cleanup plugin data when the plugin is deleted (#692).
• AI-specific Site Health integration and status tests (#734).
• New filters, wpai_content_classification_available_terms, wpai_content_classification_min_confidence and wpai_content_classification_candidate_pool_size, to allow more control over Content Classification (#633).
• Prompt template extension points, making it easy for others to filter prompts and prompt builders (#770).
• Brought the embedding code over from the PHP AI Client and load that conditionally so those using the AI plugin can start to take advantage of embedding generation (#892, #946).
• Public WordPress\AI\log_ai_request() API so MCP servers and ability consumers can record requests in the AI Request Log (#914).

Alterado

• Updated all meta keys to use the wpai_ prefix instead of just ai_. Note this changes the prefix on the ai_generated, ai_generated_summary and ai_note meta so if you are directly using those, update to using the wpai_ names (#867).
• Updated preferred models to more recent ones for the three default providers (#913).
• Bump WordPress tested-up-to version 7.1 (#934).
• Improve the relevance of category and tag suggestions produced by the
  Content Classification experiment (#633).
• Editorial Updates now links to the visual revisions screen when reviewing refined content, falling back to the classic revisions screen when visual revisions are unavailable (#861).
• Reordered setting experiments list; grouped linked experiments and sorted editor experiments alphabetically (#871).
• Improved keyboard focus handling when generating, accepting, or dismissing classification suggestions (#873).
• The Abilities Explorer provider filter dropdown now includes custom providers, and the overview statistics count abilities by origin so custom-provider abilities remain in their Core/Plugins/Theme bucket (#884).
• Set focus to the generated title textarea when generating a title (#901).
• The core/read-users ability now returns collections ordered by display name, A to Z (#948).

Deprecated

• The AI_Service class and the get_ai_service() helper introduced in 0.2.1 will be removed in the next major release. Neither is used anywhere in the plugin; experiments and abilities call wp_ai_client_prompt() directly (#905).
• Filter wpai_meta_description_result_temperature is no longer being used and will be removed in the next release (#913).

Removido

• No longer set custom temperature values on any of our requests (#913).

Corrigido

• The AI Request Log REST endpoint now registers its operation filter parameter, so it appears in the REST schema and a non-string value returns a 400 instead of causing a fatal error (#758).
• Inline reply textarea not receiving focus after generating a suggested reply (#877).
• Meta Description suggestions applied on pages and custom post types were lost on save when Yoast SEO was active (#886).
• Improved accessibility and keyboard usability for the request logs provider/model details (#889).
• Improved keyboard and focus handling for the Suggest Reply tone dropdown (#907).
• Synchronized generating state across the inline and modal excerpt generation buttons (#908).
• Ensure caller detection in the encryption experiment properly matches the calling plugin, not the host plugin (#909).
• Synchronized loading state between the sidebar and block toolbar regenerate summary buttons (#912).
• Preserve inline HTML when resizing content (#915).
• Bulk actions no longer re-run when sorting or paginating the list after the action completes (#928).
• Apply editorial updates to blocks that store editable text in the value attribute (#930).

Segurança

• Ensure any content we render from the LLM or content we send to the LLM is properly sanitized (#950).
• Add proper nonce check prior to bulk alt text and summarization generation (GHSA-hfp9-55vw-ccjc).
• When passing a custom image URL to the Alt Text Generation Ability, ensure that URL is public, that it points to an allowed image type and that the final URL we download matches the initial one we verify (GHSA-v2wx-9j88-4rqq).

1.2.0 – 14/07/2026

Adicionado

• Novo experimento: Sugerir resposta. Oferece a moderadores de comentários uma forma rápida de gerar uma resposta a um comentário pela área administrativa (#724).
• Adiciona a opção “Configurações avançadas” às ferramentas do desenvolvedor para mostrar ou ocultar opções de configuração adicionais de recursos e experimentos (#842).
• Adiciona a ação em massa “Gerar síntese com IA” à tabela de lista de posts e páginas, permitindo gerar sínteses de vários posts de uma só vez (#650).
• Adiciona a habilidade core/read-content com modos seguros de post único e consulta (incluindo campos de inclusão e campos opcionais) para permitir acesso somente leitura ao conteúdo (#739).
• Adiciona a habilidade core/read-users, que recupera um único usuário por ID, e-mail, nome de usuário ou nome amigável, ou uma coleção filtrada e paginada de usuários, com campos sensíveis opcionais e protegidos por permissões (#774).
• Adiciona um aviso administrativo embutido quando as aprovações de conectores estão ativadas e nenhum conector de IA foi aprovado, solicitando que administradores aprovem o uso do plugin AI (#830).
• Introduz o novo filtro wp_ai_client_default_request_timeout para permitir a configuração do tempo limite padrão das solicitações. Usa esse filtro para o tempo limite das solicitações de geração de imagens (#862).

Alterado

• A detecção do bloco Síntese do conteúdo agora verifica blocos aninhados (#810).
• Move todas as opções de configuração existentes para uma nova seção “Configurações avançadas”, oculta por padrão (#842).

Corrigido

• Restauração do foco após salvar as configurações de IA (#812).
• Adiciona texto alternativo descritivo às imagens dos cartões de recursos da página inicial da IA para melhorar a acessibilidade dos leitores de tela (#819).
• Impede que os recursos de texto preditivo sejam carregados na interface pública (#820).
• Descartar uma sugestão preditiva com a tecla Esc não deve acionar uma nova solicitação de sugestão (#840).
• Posicionamento do texto fantasma preditivo e sugestões desatualizadas sobrepondo os espaços reservados de blocos vazios (#847).
• Falha do MutationObserver quando o corpo do iframe do editor ainda não está pronto ao usar a geração de títulos (#849).
• Respeita um valor explícito de show_in_abilities nas configurações selecionadas e deixa o controle do sinalizador para o núcleo do WordPress assim que ele for declarado pelo núcleo (#852).
• Registra as configurações iniciais antes que core/read-settings crie snapshots delas (#856).

Entradas antigas do changelog podem ser encontradas no arquivo CHANGELOG.md.