diff --git a/CHANGELOG.md b/CHANGELOG.md index cc7730251..9433fb186 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -6,6 +6,13 @@ All notable changes to this project will be documented in this file, per [the Ke ### Added - New Experiment: AI Workspace; a dedicated `Tools → AI Workspace` screen where you hold a multi-step, site-aware conversation with an assistant. The assistant can search your content under your own capabilities, renders post results as a table, and writes nothing until you approve a proposal, which shows the exact values -- including the post status -- that will be written ([#282](https://github.com/WordPress/ai/issues/282)). - New `ai/search-content` Ability that runs a bounded, capability-filtered full-text search across the post types exposed to abilities, returning titles and excerpts rather than post bodies. Registered by the AI Workspace experiment, so it is also available to the REST `abilities` endpoints, the Abilities Explorer and MCP clients ([#282](https://github.com/WordPress/ai/issues/282)). +- New `ai/get-site-info` Ability that returns the site's name, tagline, home URL, language, timezone, date and time formats, active theme and whether it is a block theme, plus the WordPress and PHP versions for users who can view Site Health (`view_site_health_checks`). It never returns the admin email, the WordPress address, database details or the plugin list, which is why the AI Workspace offers it while withholding the broad `core/get-environment-info` and `core/settings-get` ([#1287](https://github.com/WordPress/ai/issues/1287)). +- New `ai/find-admin-screens` Ability that answers "where do I…" with links: given a short description such as "hide my site from google", it returns up to five matching admin screens from a curated map of core screens, only those the current user can open, each linked with `admin_url()`. Matching runs on the server with no model call, and the map is filterable through `wpai_admin_screens` ([#1287](https://github.com/WordPress/ai/issues/1287)). +- The AI Workspace can propose changes to things that already exist: a post's title, excerpt or content (`ai/propose-post-changes`), and the site title or tagline (`ai/propose-site-identity`). Nothing changes until the person approves. The card shows each current value beside the new one, says when a change goes live as soon as it is approved, and lets the person edit a value first. Approving is strictly one-time, and a value that changed since it was proposed is left alone and reported. This brings in the Site Agent's write actions from [#1214](https://github.com/WordPress/ai/pull/1214) as part of the AI Workspace ([#1287](https://github.com/WordPress/ai/issues/1287)). +- The command palette (Cmd/Ctrl+K) suggests admin screens with AI when a search finds nothing of its own, and offers "Ask AI: " to carry the question into the AI Workspace, unsent. It is part of the AI Workspace experiment rather than an experiment of its own, bringing in [#1250](https://github.com/WordPress/ai/pull/1250) ([#1287](https://github.com/WordPress/ai/issues/1287)). +- Every AI Workspace tool call, including a withheld one, and every proposal event (created, edited, declined, expired, failed, or refused because its chat was deleted) is recorded in the AI Request Log under a new `agent` type, and the log screen gains a `denied` status. No row records post content or option values ([#1287](https://github.com/WordPress/ai/issues/1287)). +- Saved AI Workspace chats are included in `Tools → Export Personal Data` and removed by `Tools → Erase Personal Data`, archived chats included, whether or not the experiment is on ([#1287](https://github.com/WordPress/ai/issues/1287)). +- A Wapuu shows while an AI Workspace turn runs, finishes or fails, from the set Jeffrey Paul (@jeffpaul) posted, derived from Kazuko Kaneuchi's original Wapuu. With reduced motion it is a still frame ([#1287](https://github.com/WordPress/ai/issues/1287)). - The AI Workspace's tool surface is no longer a fixed list. An ability exposed to the surface -- through `meta.public`, or by naming the `ai-workspace` channel -- and annotated `readonly: true`, `destructive: false`, `open_world: false` -- can be offered to the assistant alongside the three that ship with it. Admission defaults to refusing: an ability that is not public, and has not opted in, is never offered. Declaration-based admission is itself switched **off** until [#354](https://github.com/WordPress/ai/issues/354) settles the channel name, so the shipped surface is still the three curated abilities; `add_filter( 'wpai_workspace_tool_admission_enabled', '__return_true' )`, or the `WPAI_WORKSPACE_TOOL_ADMISSION` constant, turns it on to try it. A short list of core abilities that read personal data, settings or environment detail is never admitted whatever they declare, filterable through `wpai_workspace_withheld_abilities`. Exposure and the annotations are both the ability author's self-attestation, and the effect class is a declared intent rather than an enforced property: the site owner's Explorer controls and the execute-time `permission_callback` remain the boundaries ([#1003](https://github.com/WordPress/ai/issues/1003)). - The Abilities Explorer shows which abilities the AI Workspace assistant can call, the exact description text the model receives, and why any other ability is excluded. A site owner can remove an ability from the surface or switch the policy off entirely and return to the three that ship ([#1003](https://github.com/WordPress/ai/issues/1003)). @@ -14,6 +21,11 @@ All notable changes to this project will be documented in this file, per [the Ke - New temporary filter `wpai_workspace_tool_admission_enabled` (and the matching `WPAI_WORKSPACE_TOOL_ADMISSION` constant) switches declaration-based admission on ahead of [#354](https://github.com/WordPress/ai/issues/354); both are removed when it lands. Separate from the site owner's admission-policy off switch on the Abilities Explorer, which stays ([#1003](https://github.com/WordPress/ai/issues/1003)). - The site owner's tool removals are applied wherever the workspace builds its candidate list, rather than through a filter registered by the workspace experiment's bootstrap, so a removal holds on a site that runs the Abilities Explorer without the workspace ([#1003](https://github.com/WordPress/ai/issues/1003)). - New filters for the AI Workspace: `wpai_workspace_tool_candidates` (the ability allowlist offered to the model), `wpai_workspace_max_rounds`, `wpai_workspace_system_instruction`, `wpai_workspace_stream_emitter` and `wpai_has_function_calling_support` ([#282](https://github.com/WordPress/ai/issues/282)). +- New `wpai_agent_event` action, fired for every AI Workspace agent event with its operation, status, context and error message, whether or not AI Request Logging is on. It covers agent turns and approvals, not direct MCP calls to an ability ([#1287](https://github.com/WordPress/ai/issues/1287)). +- New filters `wpai_workspace_write_proposals_enabled` (default `true`), which turns off the proposals that change existing content, and `wpai_workspace_proposal_kinds`, the list of proposal types the tools, turn rules and system instruction are built from ([#1287](https://github.com/WordPress/ai/issues/1287)). +- The AI Workspace's own abilities (`ai/propose-*` and `ai/ask-user`) declare `meta.mcp.public` as false, so MCP clients are no longer offered tools that only work inside a workspace turn. An override on the MCP Access screen still wins ([#1287](https://github.com/WordPress/ai/issues/1287)). +- `ai/read-content-bodies` accepts `raw: true` to return a post's stored block markup, only for posts the user can edit ([#1287](https://github.com/WordPress/ai/issues/1287)). +- AI Workspace tool-call and write rows in the AI Request Log now use the `agent` type; rows logged before keep the `ability` type ([#1287](https://github.com/WordPress/ai/issues/1287)). - Documented how to run the PHP and e2e suites in one session: `npm run test:php` reinstalls WordPress in the shared test environment and deactivates the plugins, so they must be re-activated before `npm run test:e2e` ([#282](https://github.com/WordPress/ai/issues/282)). ## [1.5.0] - 2026-10-29 diff --git a/README.md b/README.md index e69b816d3..86e20566e 100644 --- a/README.md +++ b/README.md @@ -57,6 +57,7 @@ Everything is opt-in and off by default. Features are stable and supported; Expe * **[Abilities Explorer](docs/experiments/abilities-explorer.md)** – Discover, inspect, test, and document all abilities registered via the WordPress Abilities API. * **[AI Request Logging](docs/experiments/ai-request-logging.md)** – Logs AI requests for observability and debugging. +* **[AI Workspace](docs/experiments/ai-workspace.md)** - Hold a multi-step conversation about your own site under Tools. The assistant searches and reads your content, answers questions about the site, links the admin screen for a task, and proposes new drafts and templates or changes to posts and the site title and tagline, which happen only once you approve them. It also adds an AI fallback and an "Ask AI" row to the command palette. * **[Alt Text Generation](docs/experiments/alt-text-generation.md)** - Generates descriptive alt text for images using AI vision models. * **[Block Container Label Generation](docs/experiments/block-container-label-generation.md)** - Generates short List View labels for Group, Columns, and other container blocks. * **[Comment Moderation](docs/experiments/comment-moderation.md)** (runs automatically) - Automatically moderate comments based on toxicity detection and sentiment analysis, and give each comment a value score. Includes guest comments by default. @@ -108,7 +109,6 @@ Overview of planned features: * **AI Playground** – Experiment with different AI models and providers. * **Content Assistant** – AI-powered writing and editing in Gutenberg. -* **Site Agent** – Natural language WordPress administration. * **Workflow Automation** – AI-driven task automation. ## Developer Experience diff --git a/docs/experiments/ai-workspace.md b/docs/experiments/ai-workspace.md index f64a45ef2..825349619 100644 --- a/docs/experiments/ai-workspace.md +++ b/docs/experiments/ai-workspace.md @@ -2,7 +2,7 @@ ## Summary -The AI Workspace experiment adds a conversation surface on a dedicated admin screen at `Tools → AI Workspace`, where a site owner can ask questions about their own site. The assistant is given a small allowlist of WordPress Abilities as callable tools; every call runs through `WP_Ability::execute()`, so it can never reach content the requesting user could not read. It cannot write anything on its own: creating drafts or a block template is a two-step proposal that a person approves before the server writes. Replies arrive in the transcript whole, once the model has finished; token-by-token streaming from the provider is planned for a follow-up once the PHP AI Client ships streaming ([WordPress/php-ai-client#255](https://github.com/WordPress/php-ai-client/pull/255)). +The AI Workspace experiment adds a conversation surface on a dedicated admin screen at `Tools → AI Workspace`, where a site owner can ask questions about their own site. The assistant is given a small allowlist of WordPress Abilities as callable tools; every call runs through `WP_Ability::execute()`, so it can never reach content the requesting user could not read. It cannot change anything on its own: creating drafts or a block template, and changing a post's title, excerpt or content or the site title and tagline, are two-step proposals that a person approves before the server writes. Replies arrive in the transcript whole, once the model has finished; token-by-token streaming from the provider is planned for a follow-up once the PHP AI Client ships streaming ([WordPress/php-ai-client#255](https://github.com/WordPress/php-ai-client/pull/255)). ## Overview @@ -18,12 +18,25 @@ Enable the experiment at `Settings → AI` and open `Tools → AI Workspace`. Th - Post lists returned by a tool render as a read-only DataViews table inside the message, with editor links for posts the user can edit - Draft creation as a proposal: the assistant proposes, the person picks which items to create, and only then is anything written - Block template creation as a proposal, on a block theme: the assistant proposes one template, the person sees its markup, and approving creates it for the active theme -- Saved conversations: each chat is kept until its owner deletes it, and a "New space" control starts a new one +- Changes to existing content as a proposal: a post's title, excerpt or content, or the site title and tagline, shown as the current value beside the new one, editable before approval (see [Changes to existing content](#changes-to-existing-content)) +- Answers about the site itself, such as its PHP version or theme, and links to the admin screen for a task, such as Settings → Reading for "hide my site from Google" +- Saved conversations: each chat is kept until its owner deletes it, a "New space" control starts a new one, and a chats panel lists, renames, archives and deletes them +- A Wapuu beside the status while a turn runs, finishes or fails, still when reduced motion is on +- Every tool call and proposal event recorded in the AI Request Log while AI Request Logging is on, with a hook for audit plugins either way, and chats included in personal-data export and erase (see [Auditing with the AI Request Log](#auditing-with-the-ai-request-log)) - An action in the block editor that opens the workspace seeded with the current post's identity -- An AI fallback in the command palette (Cmd/Ctrl+K): a plain-language search the palette finds nothing for suggests the screens that answer it, and an "Ask AI" row carries the search into the workspace (see [The command palette](#the-command-palette)) +- An AI fallback in the command palette (Cmd/Ctrl+K): a plain-language search the palette finds nothing for suggests the screens that answer it, and an "Ask in AI Workspace" row carries the search into the workspace (see [The command palette](#the-command-palette)) The screen degrades with an explanatory state rather than failing when the plugin has no valid AI credentials, or when no configured connector exposes a model that supports function calling. +### One agent + +AI Workspace is the plugin's one agent surface. Two proposals that were heading for separate experiments are part of it instead ([#1287](https://github.com/WordPress/ai/issues/1287)): + +- **Site Agent** ([#1214](https://github.com/WordPress/ai/pull/1214)) proposed an agent that reads site information and changes posts and settings after approval. Its site-information read and its post, site title and tagline writes are here as `ai/get-site-info`, `ai/propose-post-changes` and `ai/propose-site-identity`; its plugin-list and environment reads are not, since the workspace withholds them. The writes are proposal types rather than registered write abilities, so no MCP client can call them without a person approving in the workspace. +- **Command Palette AI Search** ([#1250](https://github.com/WordPress/ai/pull/1250)) proposed an AI fallback for the command palette. It is here as the palette's AI suggestions and its "Ask in AI Workspace" row, enabled with this experiment rather than with a toggle of its own (see [The command palette](#the-command-palette)). + +The name, menu entry, URL and every identifier are unchanged: the experiment ID stays `ai-workspace`, so a site that had it on keeps it on. + ### Starter prompts The empty state offers four prompts. Choosing one fills the composer, to edit or send. Each asks only for what the tools above can do, so each ends in something to act on: @@ -365,7 +378,7 @@ With the experiment on, the command palette gets an AI fallback on every admin s - **Not an agent tool.** `ai/command-palette-ai-search` makes its own model call over a list only the browser can send, so it is registered with `meta.mcp.public = false` and annotated open-world, and the tool policy never offers it to the assistant. It is not marked read-only although it writes nothing, because the REST run route accepts only GET for a read-only ability and the command list does not fit in a URL. - **What leaves the site.** The search and the labels and keywords of the commands the user can run are sent to the configured provider. No site content is. -### Ask AI +### Ask in AI Workspace For users who can open the workspace (`manage_options`), the loader adds an "Ask in AI Workspace: " row under the same conditions as the suggestions: the palette found nothing, the search has at least three characters, and the debounce has settled. It comes after any suggestions, and it stays when the AI request fails, so a failed request leaves only that row. The server tells the loader whether to offer it; nobody else is told where the workspace is. @@ -379,11 +392,11 @@ Choosing it opens the workspace with the search in the `wpai-prompt` query argum - Registers `Admin_Page`, which adds the `Tools` submenu, enqueues the React bundle, and re-checks `manage_options` in the render callback so a direct call can never emit the app shell or its localized data. - Registers `Show_In_Abilities`, so the curated core post types are exposed before the abilities build their input schemas. -- Registers `Search_Content` (`ai/search-content`), `Read_Content_Bodies` (`ai/read-content-bodies`), `List_Terms` (`ai/list-terms`), `Propose_Drafts` (`ai/propose-drafts`) and `Propose_Template` (`ai/propose-template`). +- Registers `Search_Content` (`ai/search-content`), `Read_Content_Bodies` (`ai/read-content-bodies`), `List_Terms` (`ai/list-terms`), `Get_Site_Info` (`ai/get-site-info`), `Find_Admin_Screens` (`ai/find-admin-screens`), `Propose_Drafts` (`ai/propose-drafts`), `Propose_Template` (`ai/propose-template`), `Propose_Post_Changes` (`ai/propose-post-changes`), `Propose_Site_Identity` (`ai/propose-site-identity`) and `Ask_User` (`ai/ask-user`). - Registers `REST\Turn_Controller`, `REST\Stream_Responder` and `REST\Proposal_Controller`. -- Registers the `wpai_chat` post type and `wpai_archived` status (`Conversation_Store::register_post_type()`) and `REST\Conversation_Controller`. The chat ownership filter and user-removal cleanup (`Conversation_Store::register_lifecycle_hooks()`) are hooked by the plugin's `Main` class instead, so they hold while the experiment is off. +- Registers the `wpai_chat` post type and `wpai_archived` status (`Conversation_Store::register_post_type()`) and `REST\Conversation_Controller`. The chat ownership filter and user-removal cleanup (`Conversation_Store::register_lifecycle_hooks()`), and the personal-data exporter and eraser (`Conversation_Privacy::register()`), are hooked by the plugin's `Main` class instead, so they hold while the experiment is off. - Hooks `admin_enqueue_scripts` to add the block editor handoff action on `post.php` and `post-new.php`, for users who can open the workspace. The handoff carries the post's identity and a flattened, clamped title, and nothing else — the workspace reads any body through the same permission-checked tool path. -- Registers `ai/command-palette-ai-search` and hooks `admin_enqueue_scripts` (priority 20, after the palette) to add the command palette loader for users who can `edit_posts`, on screens where `wp-commands` is enqueued. The loader is told the workspace URL only for users who can `manage_options`, who get the Ask AI row. +- Registers `ai/command-palette-ai-search` and hooks `admin_enqueue_scripts` (priority 20, after the palette) to add the command palette loader for users who can `edit_posts`, on screens where `wp-commands` is enqueued. The loader is told the workspace URL only for users who can `manage_options`, who get the "Ask in AI Workspace" row. ### The turn loop @@ -677,7 +690,6 @@ One thing bites in practice: ### Limitations - **Bodies are read five at a time, and never in bulk.** `ai/read-content-bodies` takes explicit IDs and caps a call at five posts, so a question that would need dozens of bodies at once cannot be answered in one pass. There is no way to ask for "every post in this category" as bodies. -- **Chats have no history screen yet.** Chats are saved and can be listed, renamed, archived and deleted through the REST API, but the workspace does not show them yet; a follow-up adds the history panel. - **Older tool results are summarized when a chat is continued.** Beyond the replay window the model sees nothing, and within it only the latest turn's tool results are replayed whole. - **No mobile-optimized layout.** The screen is built for a desktop admin. - The transcript shows tool activity as a collapsible step listing each call, rather than a one-line retrieval trace; a result set narrowed by a permission check is not itself reported. diff --git a/includes/Experiments/AI_Workspace/AI_Workspace.php b/includes/Experiments/AI_Workspace/AI_Workspace.php index 168fc3c0a..60230ead2 100644 --- a/includes/Experiments/AI_Workspace/AI_Workspace.php +++ b/includes/Experiments/AI_Workspace/AI_Workspace.php @@ -95,7 +95,7 @@ public static function get_id(): string { protected function load_metadata(): array { return array( 'label' => __( 'AI Workspace', 'ai' ), - 'description' => __( 'Adds a dedicated AI Workspace screen under Tools, where you hold a multi-step conversation about your own site. The assistant searches and reads your content under your own capabilities, and writes nothing until you approve a proposal showing the exact values, including the post status, that will be written.', 'ai' ), + 'description' => __( 'Adds a dedicated AI Workspace screen under Tools, where you hold a multi-step conversation about your own site. The assistant searches and reads your content under your own capabilities, answers questions about the site, and changes nothing until you approve a proposal showing the exact values that will be written: new drafts and templates, or changes to a post or the site title and tagline. It also adds AI suggestions and an "Ask in AI Workspace" row to the Command Palette.', 'ai' ), 'category' => Experiment_Category::ADMIN, 'capability' => 'text_generation', ); diff --git a/readme.txt b/readme.txt index 922193aa5..ad363da0c 100644 --- a/readme.txt +++ b/readme.txt @@ -42,7 +42,7 @@ Experiments marked "runs automatically" do their work without you clicking anyth * **Abilities Explorer** – Browse and test registered AI abilities from a dedicated admin screen. * **AI Request Logging** – Logs AI requests, including short previews of prompts and responses, for observability and debugging. -* **AI Workspace** – Hold a multi-step conversation about your own site from a dedicated admin screen under Tools. It searches and reads your content under your own capabilities and writes nothing until you approve a proposal showing the exact values that will be written. +* **AI Workspace** – Hold a multi-step conversation about your own site from a dedicated admin screen under Tools. It searches and reads your content under your own capabilities, answers questions about the site and links the admin screen for a task, and changes nothing until you approve a proposal showing the exact values that will be written. It also adds an AI fallback and an "Ask in AI Workspace" row to the Command Palette. * **Alt Text Generation** – Generate descriptive alt text for images to improve accessibility. * **Block Container Label Generation** – Generates short List View labels for Group, Columns, and other container blocks. * **Comment Moderation** (runs automatically) – Checks each new comment for toxicity and sentiment, gives it a value score, and can hold comments for moderation. Includes comments from site visitors by default. @@ -74,7 +74,6 @@ We're actively developing new features to enhance your WordPress workflow: * **AI Playground** – Experiment with different AI models and providers. * **Content Assistant** – AI-powered writing and editing in Gutenberg. -* **Site Agent** – Natural language WordPress administration. * **Workflow Automation** – AI-driven task automation. **Roadmap** @@ -235,6 +234,11 @@ Go to `Plugins` and deactivate AI. If you can't reach the admin, deactivate it w - New `ai/search-content` Ability that runs a bounded, capability-filtered full-text search across the post types exposed to abilities, returning titles and excerpts rather than post bodies ([#282](https://github.com/WordPress/ai/issues/282)). - New `ai/get-site-info` Ability that returns the site's name, tagline, home URL, language, timezone, date and time formats, active theme and whether it is a block theme, plus the WordPress and PHP versions for users who can view Site Health (`view_site_health_checks`). It never returns the admin email, the WordPress address, database details or the plugin list, which is why the AI Workspace offers it while withholding the broad `core/get-environment-info` and `core/settings-get` ([#1287](https://github.com/WordPress/ai/issues/1287)). - New `ai/find-admin-screens` Ability that answers "where do I…" with links: given a short description such as "hide my site from google", it returns up to five matching admin screens from a curated map of core screens, only those the current user can open, each linked with `admin_url()`. Matching runs on the server with no model call, and the map is filterable through `wpai_admin_screens` ([#1287](https://github.com/WordPress/ai/issues/1287)). +- The AI Workspace can propose changes to things that already exist: a post's title, excerpt or content (`ai/propose-post-changes`), and the site title or tagline (`ai/propose-site-identity`). Nothing changes until the person approves. The card shows each current value beside the new one, says when a change goes live as soon as it is approved, and lets the person edit a value first. Approving is strictly one-time, and a value that changed since it was proposed is left alone and reported. This brings in the Site Agent's write actions from [#1214](https://github.com/WordPress/ai/pull/1214) as part of the AI Workspace ([#1287](https://github.com/WordPress/ai/issues/1287)). +- The Command Palette (Cmd/Ctrl+K) suggests admin screens with AI when a search finds nothing of its own, and offers "Ask in AI Workspace: " to carry the question into the AI Workspace, unsent. It is part of the AI Workspace experiment rather than an experiment of its own, bringing in [#1250](https://github.com/WordPress/ai/pull/1250) ([#1287](https://github.com/WordPress/ai/issues/1287)). +- Every AI Workspace tool call, including a withheld one, and every proposal event (created, edited, declined, expired, failed, or refused because its chat was deleted) is recorded in the AI Request Log under a new `agent` type, and the log screen gains a `denied` status. No row records post content or option values ([#1287](https://github.com/WordPress/ai/issues/1287)). +- Saved AI Workspace chats are included in `Tools -> Export Personal Data` and removed by `Tools -> Erase Personal Data`, archived chats included, whether or not the experiment is on ([#1287](https://github.com/WordPress/ai/issues/1287)). +- A Wapuu shows while an AI Workspace turn runs, finishes or fails, from the set Jeffrey Paul (@jeffpaul) posted, derived from Kazuko Kaneuchi's original Wapuu. With reduced motion it is a still frame ([#1287](https://github.com/WordPress/ai/issues/1287)). - The AI Workspace's tool surface is no longer a fixed list. An ability exposed to the surface - through `meta.public`, or by naming the `ai-workspace` channel - and annotated `readonly: true`, `destructive: false`, `open_world: false` - can be offered to the assistant alongside the three that ship with it. Admission defaults to refusing: an ability that is not public, and has not opted in, is never offered. Declaration-based admission is itself switched **off** until [#354](https://github.com/WordPress/ai/issues/354) settles the channel name, so the shipped surface is still the three curated abilities; `add_filter( 'wpai_workspace_tool_admission_enabled', '__return_true' )`, or the `WPAI_WORKSPACE_TOOL_ADMISSION` constant, turns it on to try it. A short list of core abilities that read personal data, settings or environment detail is never admitted whatever they declare, filterable through `wpai_workspace_withheld_abilities`. Exposure and the annotations are both the ability author's self-attestation, and the effect class is a declared intent rather than an enforced property: the site owner's Explorer controls and the execute-time `permission_callback` remain the boundaries ([#1003](https://github.com/WordPress/ai/issues/1003)). - The Abilities Explorer shows which abilities the AI Workspace assistant can call, the exact description text the model receives, and why any other ability is excluded. A site owner can remove an ability from the surface or switch the policy off entirely and return to the three that ship ([#1003](https://github.com/WordPress/ai/issues/1003)). @@ -244,6 +248,11 @@ Go to `Plugins` and deactivate AI. If you can't reach the admin, deactivate it w - New temporary filter `wpai_workspace_tool_admission_enabled` (and the matching `WPAI_WORKSPACE_TOOL_ADMISSION` constant) switches declaration-based admission on ahead of [#354](https://github.com/WordPress/ai/issues/354); both are removed when it lands. Separate from the site owner's admission-policy off switch on the Abilities Explorer, which stays ([#1003](https://github.com/WordPress/ai/issues/1003)). - The site owner's tool removals are applied wherever the workspace builds its candidate list, rather than through a filter registered by the workspace experiment's bootstrap, so a removal holds on a site that runs the Abilities Explorer without the workspace ([#1003](https://github.com/WordPress/ai/issues/1003)). - New filters for the AI Workspace: `wpai_workspace_tool_candidates` (the ability allowlist offered to the model), `wpai_workspace_max_rounds`, `wpai_workspace_system_instruction`, `wpai_workspace_stream_emitter` and `wpai_has_function_calling_support` ([#282](https://github.com/WordPress/ai/issues/282)). +- New `wpai_agent_event` action, fired for every AI Workspace agent event with its operation, status, context and error message, whether or not AI Request Logging is on. It covers agent turns and approvals, not direct MCP calls to an ability ([#1287](https://github.com/WordPress/ai/issues/1287)). +- New filters `wpai_workspace_write_proposals_enabled` (default `true`), which turns off the proposals that change existing content, and `wpai_workspace_proposal_kinds`, the list of proposal types the tools, turn rules and system instruction are built from ([#1287](https://github.com/WordPress/ai/issues/1287)). +- The AI Workspace's own abilities (`ai/propose-*` and `ai/ask-user`) declare `meta.mcp.public` as false, so MCP clients are no longer offered tools that only work inside a workspace turn. An override on the MCP Access screen still wins ([#1287](https://github.com/WordPress/ai/issues/1287)). +- `ai/read-content-bodies` accepts `raw: true` to return a post's stored block markup, only for posts the user can edit ([#1287](https://github.com/WordPress/ai/issues/1287)). +- AI Workspace tool-call and write rows in the AI Request Log now use the `agent` type; rows logged before keep the `ability` type ([#1287](https://github.com/WordPress/ai/issues/1287)). = 1.5.0 - 2026-10-29 =