Repository navigation
AI Workspace: propose changes to existing posts and the site title, applied on approval - #1314
Conversation
…ng content Part of #1287 (PR A, U1). The proposal flow only ever created things, and relied on a per-item token to survive a second approval. Updates to existing content need stronger guarantees. - Approval takes a per-proposal lock with one INSERT IGNORE into the options table, as WP_Upgrader::create_lock() does, and saves each item's result as it is written. add_option() is not used: its ON DUPLICATE KEY UPDATE lets a second claim of a different value succeed. A second approval returns the stored results. A proposal still executing after five minutes reads as failed with its partial results and cannot be approved again. - Approval is refused once the proposal's chat no longer exists. - Proposals are stored past their expiry with a grace period, so an expired one is reported as expired rather than not found. - An optional `edits` argument re-runs edited items through the kind's normalize_item(); kinds that create things reject edits. - Result counts gain updated, unchanged, stale and locked. - Each kind names its propose ability, capability and instruction, and the tool list, turn-ending list and follow-up stripping are built from the kind registry (`wpai_workspace_proposal_kinds`). Update kinds can be turned off with `wpai_workspace_write_proposals_enabled`. - A reopened chat now reports discarded proposals as discarded. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part of #1287 (PR A, U2). ai/propose-drafts, ai/propose-template and ai/ask-user only work inside an AI Workspace turn, and their comments said they were withheld from MCP. They only set show_in_rest to false, though, and the MCP Access experiment exposes every ai/* ability by default regardless of show_in_rest, so MCP clients were offered tools that always refused with workspace_context_required. - Exposure_Overrides::default_exposure() now respects a plugin ability's explicit meta.mcp.public of false. A saved override from the MCP Access screen still wins. - The three workspace-only abilities declare meta.mcp.public as false. - The docs say how they are withheld. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tent Part of #1287 (PR A, U3). The assistant can now propose new values for existing posts through ai/propose-post-changes. Nothing is written until the person approves, as with drafts and templates. This ports #1214's update rules as a proposal kind rather than a registered write ability, which MCP clients could call with no approval. - One item is one field of one post. Published posts can be changed; the card is told the change is live. - Titles must contain no markup and are capped at 500 characters. Excerpt and content go through wp_kses_post for every user. - Content changes are refused for a post whose current content kses would alter (iframes, embeds saved with unfiltered_html), so a rewrite can't silently strip them from a live post. - Writing one field keeps every other text column exactly as stored, so wp_update_post() re-running kses can't strip an iframe when only the title changes. - At approval: a field changed since the proposal is stale, one already at the new value is unchanged, and a post locked by someone else is locked. Capability is re-checked. - An edit may change only the proposed value, never which post or field an item targets. - The model hears the written title or excerpt in full, and content as updated with its length. - The ability is withheld from MCP and REST and refuses outside a turn. Co-authored-by: Vrishabhsk <vrishabhsk@gmail.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part of #1287 (PR A, U4). The assistant can now propose a new site title or tagline through ai/propose-site-identity, for administrators, and nothing changes until the person approves. This ports #1214's settings allowlist as a proposal kind rather than a registered write ability. - Only blogname and blogdescription, at most one item each, and manage_options at propose time and again at approval. - Values are plain text: HTML is stripped, and a value over 200 characters (title) or 300 (tagline) is refused rather than cut. An empty title is refused; an empty tagline is allowed. - At approval: a value changed in Settings since the proposal is stale, and one already at the new value is unchanged (compared as core stores it, so an ampersand doesn't read as a change). - An edit may change only the proposed value. - The ability is withheld from MCP and REST and refuses outside a turn. Co-authored-by: Vrishabhsk <vrishabhsk@gmail.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part of #1287 (PR A, U5). Approval cards for post-field and site identity proposals: - Each change shows the current value next to the proposed one, with content rendered rather than shown as HTML, a notice when the change goes live immediately, and a checkbox. Update cards start with every change selected; the button counts them ("Apply 3 changes"). - Edit turns the proposed value into a field: a text input for titles and the tagline, a textarea for the excerpt, and a markup textarea with a live rendered preview for content. A value the server refuses keeps the card editable, with the server's message by the field. - Outcomes read updated, unchanged, changed since proposed (with a link to the post or Settings) or locked. - A second window shows the stored outcome instead of Apply. An approval refused because another window is applying shows "Applying in another window" and waits for that outcome. A failed approval shows its partial results. A reopened chat shows a discarded change as discarded. - Post content is parsed in an inert DOMParser document and mapped onto the existing markdown node tree through an allowlist, so no markup is handed to the browser and no image is fetched; links are live only on this site's host. - The proposal GET route now returns stored results once a proposal is no longer pending, and an edit error names the item it belongs to. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part of #1287 (PR A). Covers ai/propose-post-changes and ai/propose-site-identity, how changes to existing content are shown, edited and checked at approval, the single-approval lock, and the wpai_workspace_write_proposals_enabled and wpai_workspace_proposal_kinds filters. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part of #1287 (PR A). No behavior change. - The write loop, log envelope and field reader the three writers copied now live in one Proposal_Writes trait; each kind keeps its own write_item(), its deselected outcome and its log details. - The stale check stores a SHA-256 of the value it was proposed over, rather than the whole value, so a proposal of content changes no longer carries up to 20 post bodies through every save. - The tool selector builds its curated list once per pass, and the system instruction computes the proposal abilities once. - The update card memoizes its rendered HTML, so typing in one edit field no longer re-parses every item's content. - Small cleanups: a capability constant instead of a repeated literal, a simpler results tally, one field-label table. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part of #1287 (PR A). From the ce-code-review run 20261106-154448-0faae422. - #1: content changes were refused for any post with an image or separator block, because the "kses would damage this" check compared bytes and kses re-serializes void tags. It now compares tags, attributes, comments and text, and refuses only when kses would remove something (iframes, scripts, event handlers, disallowed protocols). - #2: the approval card dropped markup that survives kses and renders on the live site. Self-closing blocks (rss, latest-posts, ...), media, embeds and form controls are now described in place as inert text, so the person sees what they are approving. Nothing is fetched or run. - #7: the assistant could read post content only as stripped text, so a content change would have dropped the post's blocks. ai/read-content-bodies takes `raw: true` and returns the stored markup for posts the user can edit, and the propose instruction tells the model to build changes from it. - #6: a proposal can no longer name the same field of one post twice. - A model-supplied stale-check snapshot is ignored; the server takes it. - #4: tests for the empty title and content refusals. #10: docs list the two new abilities among those not available over REST or MCP. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
✅ WordPress Plugin Check Report
📊 ReportAll checks passed! No errors or warnings found. 🤖 Generated by WordPress Plugin Check Action • Learn more about Plugin Check |
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## feat/ai-workspace #1314 +/- ##
=======================================================
+ Coverage 86.06% 86.54% +0.48%
- Complexity 5848 6190 +342
=======================================================
Files 215 220 +5
Lines 21250 22321 +1071
=======================================================
+ Hits 18289 19318 +1029
- Misses 2961 3003 +42
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Found in a live run of #1314. Once an update was applied, its card collapsed to "Tagline - Updated" and no longer said what the field now holds; only the assistant's follow-up text did. Each updated or unchanged item now shows "Now: <value>" for titles, excerpts, the site title and the tagline. Content is reported by length only, so it shows nothing extra. Also pluralize the tool-steps summary: "Looked up site content (1 step)" rather than "(1 steps)", which predates this branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The following accounts have interacted with this PR and/or linked issues. I will continue to update these lists as activity occurs. Unlinked AccountsThe following contributors have not linked their GitHub and WordPress.org accounts: @Vrishabhsk. Contributors, please read how to link your accounts to ensure your work is properly credited in WordPress releases. If you're merging code through a pull request on GitHub, copy and paste the following into the bottom of the merge commit message. To understand the WordPress project's expectations around crediting contributors, please review the Contributor Attribution page in the Core Handbook. Props Bot commandsYou can trigger Props Bot actions by commenting on this PR:
|
Two approvals that both read the same stale lock could each delete the other's fresh one with delete_option() and both claim the proposal. The stale row is now deleted only while it still holds the value read, and the claim is taken only when that delete removed a row. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… outcomes - A writer that throws part way now leaves the proposal failed with the results it saved, releases the lock, reports those results to the model, and answers 500 with the stored outcome in the error data. - A discard takes the approval lock and re-reads the proposal before it deletes, so it cannot interleave with an approval that has claimed the proposal but not yet saved it as executing. - Only a pending proposal expires on read. One an approval started on is returned past its expiry, so reopening an applied change shows what was written instead of "expired". Its record is still removed: save() stores the transient for TTL + GRACE after its last save, and the transient lapses on its own after that, so no extra delete is needed. - "Select at least one item." no longer says "to create" for updates. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tent reads - The other columns of a post-field write are put back on wp_insert_post_data for that post's ID only, so a save_post hook that saves another post during the write keeps that post's own columns. - The turn context remembers a hash of each post's content_raw that ai/read-content-bodies returned this turn. A content change proposed for a post saved since that read is refused, telling the model to read it again. Outside a turn, or with no raw read, nothing changes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nged - An approval that fails with a 5xx, no answer, or an answer that cannot be read may have written, so the card reads the proposal back and shows its stored outcome when it is no longer pending. "Nothing was changed" is kept for a 4xx, which the server sends before it claims, and for a proposal still pending. If the read back fails too, the card says the outcome could not be confirmed. - A failed discard reads the proposal back as well, so one another window already applied shows what was applied. - Returning to the card after another window let go keeps the person's selection instead of re-ticking what they unticked. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Block settings sit in their delimiters and are not rendered, so a content change that only alters a wrapping block, such as a Query Loop's query or a binding, looked the same in Current and New. The card now compares the block delimiters of both versions and names the blocks whose settings differ. Also covers the withheld svg, math, canvas, select, input, form, embed and track elements in the renderer spec. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The card now reads a proposal back before saying nothing was created. The test aborted every proposal request, read-back included, so the card rightly said the outcome could not be confirmed. Aborting only the execute request keeps the test about an approval that never got there, and covers the read-back finding the proposal still pending. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The OpenAI connector returns each item of a Responses API reply as its own candidate. When the model says what it is about to do and then calls a tool, the sentence and the call arrive as two candidates, and the client read only the first with toMessage(). The model announced a proposal, the call was lost, and the turn ended with no approval card. Parallel calls after the first were lost the same way. The client now returns every candidate's message, in order, when there is more than one. The turn loop adds each to the history and answers each message's calls in a message of its own, because that connector refuses a function call or result that shares a message when the history is replayed. A reply in one candidate, as Anthropic and Google send, is handled as before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When the OpenAI connector returns a sentence and a tool call as separate
messages, the turn's text was read from the last model message only,
which is the call, so the sentence ("Your site is running PHP 8.3") was
saved and shown as nothing. The text now comes from the whole run of
model messages that makes up the last reply.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Asking for 50 posts cut the request off at core's 30 second default before any reply arrived; in the browser it showed as a network error. - A workspace round now waits at least 120 seconds. The raise is scoped to building that round's request and never shortens a longer default a site set. `wpai_workspace_request_timeout` changes it. - The drafts instruction states that one proposal holds at most 20 drafts, to propose the first 20 when asked for more, and to keep bodies short when proposing many. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
jeffpaul
left a comment
There was a problem hiding this comment.
I'm still getting those NetworkError messages, so not sure if there's something amiss in the Playground environment (compared to say a local env) that may be causing this?
WordPress Playground gives up on any request after 25 seconds, and a
proxy in front of a site can give up before PHP does. The turn keeps
running and is saved, but the browser only saw a network error
("NetworkError when attempting to fetch resource").
Each turn now carries an ID the client chose, kept on its saved entry,
and a first turn names the chat it creates. When the request fails in
transit, the stream breaks, or a gateway answers 502/503/504/524, the
client reads the saved chat every 3 seconds, for up to 10 minutes,
until the turn's entry appears, and shows that instead. Stop still
works while it waits.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Waiting for a saved reply made a request that never left the browser look like a turn still running, for up to 10 minutes. The turn route now marks a named turn running while it runs and done when it ends, and GET /workspace/messages/<turn_id> reports running, done or unknown, for the current user only. The client checks it straight away: a turn the server never saw fails after a second check, about 3 seconds; one that is running is waited on; one that is done is read from the saved chat. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks, Jeff, you were right that it's Playground. Playground's service worker gives up on any request after 25 seconds ( The workspace now recovers that reply instead of showing the error:
Earlier fixes on this PR, in case you didn't see them: the tagline proposal that never arrived was the OpenAI connector splitting the reply into separate pieces (391fb52, 35f756f), and a round now waits up to 120 seconds and proposes at most 20 drafts at a time (0da9f4d). AI assistance: Yes. Tool(s): Claude Code (Claude Opus 5.5). Used for: diagnosing the Playground timeout, writing the fix and tests, and drafting this reply. I've reviewed the changes, understand them and can explain them under review. |
The Anthropic connector caps a reply at 4,096 tokens when the caller sets no limit, and a proposal of twenty drafts stopped partway with "Generation stopped due to token limit (4096) with stop reason max_tokens". Workspace rounds now ask for up to 16,384 output tokens on every provider, filterable with wpai_workspace_max_output_tokens. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
One more fix, found while testing the Playground preview with an Anthropic key. After the reply came back, the 50-post prompt failed with "Generation stopped due to token limit (4096) with stop reason "max_tokens"". The Anthropic connector caps a reply at 4,096 tokens unless the caller asks for more, and a 20-draft proposal is longer than that. OpenAI sets no cap, which is why it only showed up with Anthropic. Fixed in de66f8c: workspace turns now ask every provider for up to 16,384 output tokens, filterable with In the preview for de66f8c, "generate 50 posts about traveling to each US state and then publish them" now ends with a proposal for the first 20 states, Alabama through Maryland, with the remaining 30 offered in two more batches. No NetworkError and no token-limit error. Ready for another look when you have a minute. AI assistance: Yes. Tool(s): Claude Code (Claude Opus 5.5). Used for: diagnosing the token limit, writing the fix and tests, and drafting this reply. I've reviewed the changes, understand them and can explain them under review. |
jeffpaul
left a comment
There was a problem hiding this comment.
Approving and we can come back later to look at how better to handle a response with MANY checkboxes to sort through
#1315) ## What? Related: #1287 The AI Workspace assistant can now answer questions about the site itself, and point people to the right admin screen: - **"What PHP version am I running?"** It reads the site's name, tagline, language, timezone, theme and, for admins, the WordPress and PHP versions. - **"Where do I hide my site from Google?"** It answers with a link to Settings → Reading. This is the second stacked PR for #1287, on top of #1314. In a live run of #1314, the assistant proposed a new tagline and site title without being able to read the current ones ("I can't see your current site title"). This PR fixes that. ## Why? #1287's first version needs the agent to answer site questions and navigation questions directly. Before this, the workspace could read posts but nothing about the site. It also had no way to know which admin screen handles which job. ## How? **`ai/get-site-info`** returns these fields, and no others: - the site name, tagline and home URL; - the language and timezone; - the date and time formats; - the active theme, and whether it's a block theme. The WordPress and PHP versions are included only for users who can see Site Health (`view_site_health_checks`), which on multisite means super admins, as in core. It never returns the admin email, database details or the plugin list. Core's own `core/get-site-info` returns the admin email, so it joins the workspace's `NEVER_ADMITTED` list next to the environment and settings readers. **`ai/find-admin-screens`** matches a request against a curated map of 33 core admin screens on the server, using titles and keywords. There's no model call inside it. It returns up to five screens, best match first, and only screens the current user can open, as `admin_url()` links on the site's own host. - Sites and plugins can add or remove entries with the new `wpai_admin_screens` filter. A filtered entry whose path has a scheme, a leading slash or `..` is dropped. - The Site Editor is offered only on block themes, and the Customizer only on classic ones. **Both are plain read abilities.** Any logged-in user can call them, they're annotated read-only and not open-world, and they're listed over MCP like the content reads. Both join the workspace's curated tool list, because policy admission is off by default. #1250's palette ability, which makes a nested model call over the palette's own commands, stays separate. It comes with the Command Palette PR. Site-info field choices are ported from #1214's `Read_Site`, credited to @Vrishabhsk with Co-authored-by. ### Use of AI Tools AI assistance: Yes. Tool(s): Claude Code (Claude Opus 5.5). Used for: planning, implementing the changes and tests, and code review. I've reviewed the changes, understand them and can explain them under review. They have all been tested locally. ## Testing Instructions 1. Enable AI Workspace and connect a provider. 2. Ask "What PHP version am I running?". The assistant answers with your PHP version. 3. Ask "Where do I hide my site from Google?". The reply links Settings → Reading, and following the link opens it. 4. As an editor, call `ai/get-site-info` through the Abilities Explorer or REST. The response has no version fields. **Tests run** (all passing): - **PHPUnit, full suite:** 3325 tests on WordPress latest. - **`Get_Site_InfoTest`:** `test_administrator_gets_every_field_including_versions`, `test_editor_gets_the_fields_without_versions`, `test_manage_options_without_site_health_access_gets_no_versions`, `test_no_response_carries_admin_email_siteurl_database_or_plugins`, `test_executes_without_any_input`, `test_name_and_tagline_are_decoded`, `test_timezone_falls_back_to_the_utc_offset`, `test_logged_out_visitor_is_refused`, `test_output_schema_marks_the_versions_optional`, `test_registration_has_a_category_and_read_only_annotations`, `test_declared_to_the_model_for_administrator_and_editor_with_admission_off` and `test_listed_for_mcp`. - **`Find_Admin_ScreensTest`:** - Queries: `test_hiding_the_site_from_google_finds_reading_settings_first`, `test_stopping_spam_comments_finds_discussion_settings_first`, `test_changing_the_timezone_finds_general_settings_first`, `test_singular_and_plural_settings_match_alike`, `test_returns_at_most_five_screens` and `test_stopword_only_query_returns_an_empty_list`. - Access: `test_editor_asking_for_plugins_gets_no_plugins_or_settings_screen`, `test_filter_added_screen_appears_only_for_users_with_its_capability`, `test_filter_can_remove_a_core_screen` and `test_every_url_is_on_the_admin_host`. - Themes and matching: `test_block_theme_offers_the_site_editor_not_the_customizer`, `test_classic_theme_offers_the_customizer_not_the_site_editor` and `test_stem_is_idempotent_for_every_word_in_the_map`. - Registration: `test_logged_out_visitor_is_refused`, `test_registration_has_a_category_and_read_only_annotations`, `test_declared_to_the_model_for_administrator_and_editor_with_admission_off` and `test_listed_for_mcp`. - **`Tool_PolicyTest`:** `test_core_get_site_info_is_withheld_with_perfect_annotations`. It failed before `core/get-site-info` was added to `NEVER_ADMITTED`. - **e2e (Chromium):** all `tests/e2e/specs/experiments/ai-workspace*` specs, 126 passed. That includes the new `ai-workspace-navigate.spec.ts`: asking where to change the timezone links Settings → General, and the link opens it. - **Test-first:** tests written before their change failed before it. - **Other checks:** PHPStan finds no errors, phpcs is clean on the changed plugin PHP, and the build passes. - **Not run yet:** a live check against a real provider. **Code review:** six reviewers ran: correctness, security, adversarial, testing, maintainability and agent-native. The orchestrator then reproduced each finding on a test site, instead of running the skill's separate validator. Four reviewers found nothing. The confirmed findings are fixed in 21cf902: - a call with no input failed; - the name and tagline came back HTML-escaped; - the versions were visible to sub-site admins on multisite; - the screen matcher missed singular words; - the theme branch had no test. ## Known limits - An entry a plugin adds through `wpai_admin_screens` is trusted to name the capability its screen needs. A weaker capability gives users a link they can't open. It never gives a link off the site's host. - Keywords are English only. A query in another language matches only the translated screen titles. - The navigation e2e test's link comes from the mocked model reply, so it proves the tool round trip and the link working, while the ranking itself is covered by `Find_Admin_ScreensTest`. ## Screenshots **Asking where to change the timezone links Settings → General**  ## Changelog Entry > Added - AI Workspace can answer questions about the site (name, tagline, theme, and WordPress and PHP versions for admins) and link to the admin screen for a request. **Post-deploy monitoring:** no additional operational monitoring is required. This is an experiment behind the AI Workspace toggle, off by default, and it adds read-only abilities. --- [](https://github.com/EveryInc/compound-engineering-plugin) 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- wp-playground-preview:start --> <a href="https://playground.wordpress.net?blueprint-url=data:application/json,%7B%22preferredVersions%22%3A%7B%22wp%22%3A%22latest%22%7D%2C%22login%22%3Atrue%2C%22landingPage%22%3A%22%2Fwp-admin%2Fadmin.php%3Fpage%3Dai-wp-admin%22%2C%22steps%22%3A%5B%7B%22step%22%3A%22installPlugin%22%2C%22pluginData%22%3A%7B%22resource%22%3A%22url%22%2C%22url%22%3A%22https%3A%2F%2Fgithub.com%2FWordPress%2Fai%2Freleases%2Fdownload%2Fci-artifacts%2Fpr-1315-2113f328ff63bca3163299767db00d0699313222.zip%22%7D%7D%5D%7D" target="_blank" rel="noopener noreferrer"> <img src="https://raw.githubusercontent.com/adamziel/playground-preview/refs/heads/trunk/assets/playground-preview-button.svg" alt="Open WordPress Playground Preview" width="220" height="57" /> </a> <!-- wp-playground-preview:end --> --------- Co-authored-by: Vrishabhsk <vrishabhsk@gmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
## What? Related: #1287, #1250 The Command Palette joins the AI Workspace: - **AI suggestions when a search finds nothing.** Type "hide my site from google" into the palette (Cmd/Ctrl+K). If the palette has no match of its own, it asks the model which of its commands fit and shows up to three, such as Settings → Reading. - **"Ask AI" for bigger questions.** For people who can use the AI Workspace, the palette also offers "Ask AI: <your search>". Choosing it opens the AI Workspace with that text in the composer, ready to send. This is the third stacked PR for #1287, on top of #1315 and #1314. ## Why? #1287 asks for one agent surface, with the palette's "Ask AI" (#1250) folded in rather than shipped as a separate experiment. Before this, a plain-language palette search with no match just showed "No results found", and there was no route from the palette into a conversation. ## How? **AI suggestions** port #1250's ability and loader, credited to @JDeepD with Co-authored-by. - When the palette has settled with no results, its search is at least 3 characters, and the person has stopped typing for 450 ms, the loader sends the search and the palette's own command list (names, labels and keywords) to `ai/command-palette-ai-search`. The model picks up to three command IDs. IDs that weren't sent are dropped, on the server and again in the browser. Failures stay silent, and the palette shows its own empty state. - "Settled" means core's own post and page search has finished too. The loader waits for the palette's empty state, which core shows only once its search loaders are done, so an AI request never races a slow post search. - The payload is cut to the ability's limits (500 commands, and caps on label and keyword length), so a large site doesn't fail validation. A search longer than the ability accepts gets no AI request, but still gets the Ask AI row. - Editors and above get suggestions (`edit_posts`); subscribers don't. - #1250's separate experiment and toggle are gone. The loader is enqueued by the AI Workspace experiment, on screens that load the Command Palette. - The ability is hidden from MCP (`meta.mcp.public = false`), and the workspace's tool policy refuses it as an agent tool. It's annotated non-destructive and idempotent, but not read-only: core's REST run route accepts only GET for read-only abilities, and the palette's command list is too long for a URL. **Ask AI** is offered only to workspace users (`manage_options`, the AI Workspace screen's capability). - It appears under the same conditions as the suggestions, after them, and it stays when the AI request fails. - Choosing it opens the AI Workspace with a `wpai-prompt` query arg. The text is seeded into the composer as plain text, cut to 500 characters, and never sent automatically. The arg is removed from the URL after load, so a reload doesn't seed it again. - A notice above the composer says the message came from a link and should be checked before sending. Anyone can send an admin a link like that. - `get_seed()` gains a prompt field. A prompt seed takes precedence over a post seed. Adds `@wordpress/commands` 1.57.0 as a dependency. ### Use of AI Tools AI assistance: Yes. Tool(s): Claude Code (Claude Opus 5.5). Used for: porting #1250, implementing the changes and tests, and code review. I've reviewed the changes, understand them and can explain them under review. They have all been tested locally. ## Testing Instructions 1. Enable AI Workspace and connect a provider. 2. Press Cmd/Ctrl+K and type "hide my site from google". After a moment, Settings → Reading appears selected, and Enter opens it. 3. Type "which of my posts are missing featured images" and choose "Ask AI: …". The AI Workspace opens with that text in the composer, unsent, with a notice above it. 4. As an editor, repeat step 2: you get suggestions but no Ask AI row. 5. Search for one of your post titles. The palette shows the post, and no AI request is made. **Tests run** (all passing): - **PHPUnit, full suite:** 3376 tests, single site and multisite, on WordPress latest. - **`Command_Palette_AI_SearchTest`** (ported from #1250, extended): including `test_ability_is_hidden_from_mcp_and_refused_as_an_agent_tool`, `test_execute_callback_drops_an_id_that_was_not_sent`, `test_execute_callback_drops_invalid_ids_and_caps_results`, `test_permission_callback_for_subscriber` and `test_choose_commands_falls_back_without_structured_output`. - **`Command_PaletteTest`:** `test_palette_search_has_no_experiment_of_its_own`, `test_loader_is_enqueued_with_the_command_palette_for_an_editor`, `test_loader_is_not_enqueued_for_a_subscriber`, `test_loader_offers_ask_ai_to_a_workspace_user`, `test_loader_does_not_offer_ask_ai_to_an_editor` and `test_loader_is_not_enqueued_when_the_experiment_is_disabled`. - **`Admin_PageTest`:** `test_prompt_seed_carries_the_query`, `test_prompt_seed_is_cut_to_the_cap`, `test_prompt_seed_is_plain_text`, `test_empty_or_whitespace_prompt_gives_no_seed`, `test_non_string_prompt_gives_no_seed`, `test_prompt_seed_takes_precedence_over_a_post_seed` and `test_post_seed_has_an_empty_prompt`. - **e2e (Chromium):** all `ai-workspace*` specs plus the palette spec, 157 passed. - `command-palette-ai-search.spec.js`: finds the right screen when the palette has no results; no suggestions when no screen fits; no AI search when the palette has results; a one or two character search makes no request; only the last search is sent and answered when the search changes; an answer naming a command the palette didn't send is dropped; waits for slow post results before searching with AI; a search longer than the ability accepts shows only Ask AI; the palette's own empty state when the request fails; an editor gets suggestions; a subscriber gets none; no AI row and no request when the experiment is disabled. - `ai-workspace-palette-handoff.spec.ts`: Ask AI opens the workspace with the query in the composer, unsent; after a reload the composer is empty and the URL has no prompt; Ask AI comes after the suggestions; an editor sees suggestions but no Ask AI row; a 2,000-character query is cut to the cap; a prompt from a link shows a notice until it's edited; opening the workspace directly shows no notice; a query with markup arrives as plain text; an empty or whitespace query shows no Ask AI row; when the AI request fails, a workspace user sees only Ask AI. - The timing tests were also run three times each, 21 of 21 passing. - **Test-first:** every new PHPUnit test, and 6 of the 8 first Ask AI e2e tests, failed before their change. - **Other checks:** PHPStan finds no errors, phpcs is clean on the changed plugin PHP, and tsc, lint-js and the build are clean. Style lint reports two errors in `ai-workspace/index.scss` that are already on the base branch. - **Not run yet:** a live check against a real provider, and the PHPUnit suite on WordPress trunk. **Code review:** six reviewers ran on this PR's own diff: correctness, security, adversarial, frontend races, testing and maintainability. Each finding was checked in the code before being fixed. Fixed in a4a84fb and a0d7830: - the AI request could fire before core's post and page search finished, showing AI rows next to real results (two reviewers found this); - "Searching with AI…" could stay on screen after an early exit or a changed search; - a cached answer could show for a search that now has real results; - the command list wasn't cut to the ability's limits, so a large site would fail validation, silently; - a composer filled in from a link had no notice; - test gaps: short searches, a changed search, an unsent command name, a slow post search. ## Known limits - Choosing an AI suggestion records it among the palette's recent commands, where it doesn't resolve, so it can take one of the five recent slots. - A cached answer can name a command that was unregistered after a context change. Selecting it runs the original callback. - Suggestions come from the palette's own command list, as in #1250, not from #1315's admin-screen map. ## Screenshots **A palette search with no match gets an AI suggestion**  **Searching with AI**  **Ask AI opens the workspace with the query in the composer, unsent**  **An editor gets suggestions but no Ask AI row**  ## Changelog Entry > Added - The Command Palette suggests admin screens with AI when a search finds nothing, and offers "Ask AI" to continue the question in the AI Workspace. **Post-deploy monitoring:** no additional operational monitoring is required. This is an experiment behind the AI Workspace toggle, off by default. --- [](https://github.com/EveryInc/compound-engineering-plugin) 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- wp-playground-preview:start --> <a href="https://playground.wordpress.net?blueprint-url=data:application/json,%7B%22preferredVersions%22%3A%7B%22wp%22%3A%22latest%22%7D%2C%22login%22%3Atrue%2C%22landingPage%22%3A%22%2Fwp-admin%2Fadmin.php%3Fpage%3Dai-wp-admin%22%2C%22steps%22%3A%5B%7B%22step%22%3A%22installPlugin%22%2C%22pluginData%22%3A%7B%22resource%22%3A%22url%22%2C%22url%22%3A%22https%3A%2F%2Fgithub.com%2FWordPress%2Fai%2Freleases%2Fdownload%2Fci-artifacts%2Fpr-1316-3b897cb8a6b06ce6e099c27b86176305ebc64e54.zip%22%7D%7D%5D%7D" target="_blank" rel="noopener noreferrer"> <img src="https://raw.githubusercontent.com/adamziel/playground-preview/refs/heads/trunk/assets/playground-preview-button.svg" alt="Open WordPress Playground Preview" width="220" height="57" /> </a> <!-- wp-playground-preview:end --> --------- Co-authored-by: Jaydeep Das <jaydeep.das@rtcamp.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…ase, Wapuu busy states (#1317) ## What? Related: #1287 The AI Workspace gains an audit trail, personal-data export and erase, and Wapuu busy states: - **One audit trail.** Every agent event is one `agent` row in the AI Request Log and one call of a new `wpai_agent_event` action. Events are tool calls (including withheld ones), each item an approval writes, and each step in a proposal's life: created, edited, declined, expired, failed, or refused because its chat was deleted. The action fires even when AI Request Logging is off. - **Chats are personal data.** Tools → Export Personal Data includes the person's saved chats, and Tools → Erase Personal Data deletes them. This works with the experiment off. - **Wapuu.** While a turn runs, the transcript shows a working Wapuu instead of a spinner, then a celebrating one for a moment when the turn finishes, or an oops one next to an error. This is the fourth stacked PR for #1287, on top of #1316, #1315 and #1314. ## Why? #1287 asks that a site owner can see everything the agent did, was asked and was refused, in one place, and that audit plugins can hook it. Before this, workspace rows were mixed in with every other `ability` row, proposal events weren't logged at all, and nothing fired with logging off. Saved chats are personal data that core's export and erase tools didn't know about. And Jeff posted a Wapuu set on #1287 so a long turn looks busy rather than stuck. ## How? **`Agent_Audit`** is the one place agent events are recorded. - It fires `wpai_agent_event( $operation, $status, $context, $error )`, then writes the row through `log_ai_request()` with the new `agent` type. - The two existing log calls, for tool calls in `Turn_Runner` and for written items in `Proposal_Writes`, now go through it. Proposal events are recorded in `Proposal_Store` and `Proposal_Controller`: - created, when a proposal is stored; - edited, when an approval carries edits, naming the edited items; - declined; - expired, once, when an expired proposal is next read, since it's deleted then; - failed, when a write throws or is interrupted for longer than the approval lock; - refused, when an approval comes after its chat was deleted. - Context names the conversation, the proposal and its item keys, the post and field, or the option. It never carries post content or option values. - The hook covers agent turns and approvals only. A direct MCP call to an ability doesn't fire it. Docs say so. - The log screen gains the `denied` status, ported from #1214, and agent rows are labelled "Agent". Rows logged before this keep the `ability` type. - #1214's `wp_after_execute_ability` listener isn't ported: `Turn_Runner` already logs each call, so it would log them twice. **`Conversation_Privacy`** registers an "AI chats" exporter and eraser. - They're hooked from `Main`, next to the chat ownership filter, so they work while the experiment is off. - The export has one item per chat: name, created and updated dates, whether it's archived, and each turn's message and reply. Tool results are left out. It pages at 20 chats. - The eraser deletes all the person's chats. - Chat IDs are read straight from the posts table, as the user-deletion cleanup already does, so archived chats are found even while their status isn't registered. That reader is now shared, as `Conversation_Store::chat_ids()`. - A proposal from an erased chat can't be approved, because an approval is refused once its chat is gone. **Wapuu** sprites replace the transcript's spinner. - **Sprites:** `working` while a turn runs, `celebrating` for 1.8 seconds after one finishes, and `oops` next to an error. They come from the set @jeffpaul posted on #1287. - **Accessibility:** each sprite is an `<img alt="">`. The existing "Responding…" text and error text stay as the status. - **Reduced motion:** an SVG inside an `<img>` doesn't reliably follow the page's setting. So with reduced motion on, the component switches each sprite's animations off in its markup, which gives the same still pose. - **Scope:** only busy, done and error states ship. Per-tool sprites need streaming to know which tool is running. - **Credit:** the sprites are derived from Kazuko Kaneuchi's original Wapuu, GPL-2.0-or-later, and credited in `CREDITS.md`. ### Use of AI Tools AI assistance: Yes. Tool(s): Claude Code (Claude Opus 5.5). Used for: implementing the changes and tests, and code review. I've reviewed the changes, understand them and can explain them under review. They have all been tested locally. ## Testing Instructions 1. Enable AI Workspace and AI Request Logging, and connect a provider. 2. Send a message. A Wapuu shows beside "Responding…", then a celebrating Wapuu for a moment when the reply arrives. 3. Ask for a new tagline. Edit the proposed value and approve it. Ask for a draft and discard it. 4. Open Tools → AI Request Logs and filter by type `agent`. The tool calls, `proposal_created`, `proposal_edited` and `proposal_declined` rows are there, none of them with the tagline's text. 5. Tools → Export Personal Data for your own email: the download has an "AI chats" section with your chats. Erase personal data for a test user who has chats: their chats are gone. 6. With reduced motion turned on in your OS, the Wapuu is still. **Tests run** (all passing): - **PHPUnit, full suite:** 3398 tests, single site and multisite, on WordPress latest. - **`Agent_AuditTest`:** `test_agent_is_a_log_type`, `test_an_event_writes_one_row_and_fires_the_hook`, `test_the_hook_fires_with_logging_off`, `test_a_proposal_event_names_the_proposal_and_its_items_only` and `test_a_proposal_without_a_kind_is_recorded_as_drafts`. - **`Proposal_ControllerTest`:** `test_each_proposal_event_is_recorded_once_in_order`, `test_events_fire_the_hook_with_logging_off`, `test_an_expired_proposal_is_recorded_once`, `test_an_expired_proposal_read_through_the_store_is_recorded_once`, `test_an_interrupted_write_is_recorded_failed_once`, `test_a_writer_that_throws_is_recorded_failed` and `test_an_approval_for_a_deleted_chat_is_recorded_denied`. - **`Turn_ControllerTest`:** `test_a_turn_records_a_withheld_call_and_a_proposal_in_order`, and `test_every_tool_call_writes_exactly_one_log_row` updated for the `agent` type. - **`Post_Field_ProposalTest`:** `test_a_content_update_records_the_field_not_the_content`. - **`AI_Request_Log_ControllerTest`:** `test_get_logs_filters_agent_rows_by_type_and_denied_status`. - **`Conversation_PrivacyTest`:** `test_the_exporter_and_eraser_are_registered_without_the_experiment`, `test_export_returns_every_chat_with_its_prompts_and_replies`, `test_export_with_the_experiment_off_includes_archived_chats`, `test_export_pages_at_twenty_chats`, `test_erase_removes_the_persons_chats_only`, `test_a_proposal_in_an_erased_chat_cannot_be_approved` and `test_an_unknown_email_exports_and_erases_nothing`. - **e2e (Chromium):** all `ai-workspace*` specs, the palette spec and the request-log spec, 168 passed in one run on a fresh site. That includes the new "AI Workspace Wapuu" tests in `ai-workspace.spec.js`: a running turn shows the busy Wapuu beside "Responding…"; with reduced motion the busy Wapuu is a still frame; a failed turn shows the error Wapuu with the error; a finished turn shows the done Wapuu briefly. Each Wapuu test also ran three more times, 12 of 12 passing. - **Test-first:** the audit tests failed before `Agent_Audit` existed. The privacy tests were written before `Conversation_Privacy` but first run after it, so they have no recorded red run. - **Other checks:** PHPStan finds no errors, phpcs is clean on the changed plugin PHP, and tsc, lint-js and the build are clean. Style lint reports errors in `ai-request-logs/index.scss` and `ai-workspace/index.scss` that are already on the base branch. - **Live check:** core's exporter and eraser lists include "AI chats" on a running site, and the screenshots below come from one. - **Not run yet:** a live check against a real provider, and the PHPUnit suite on WordPress trunk. **Code review:** seven reviewers ran on this PR's own diff (correctness, security, adversarial, testing, maintainability, frontend races and API contract), plus an independent validator. The verdict was "Ready to merge" with no actionable findings. The one candidate, that the status filter rejects `skipped`, was rejected because it's already on the base branch. Applied afterwards in eef886a: agent rows were labelled "Text" on the log screen, a docblock still said these rows use the `ability` type, and the store's expiry path had no test. **One test from #1316 hardened** (1060833): its short palette search for "q" matched "AI Request Logs" whenever AI Request Logging was on, which the request-log spec leaves on when both run in one session. It now searches for "~". ## Screenshots **A running turn**  **A finished turn**  **A failed turn**  **Agent events in the AI Request Log**  **AI chats in a personal-data export**  ## Known limits - After an erase, a pending proposal from an erased chat can still be read by its owner, though not approved, until it expires (about 90 minutes). - "Recorded once" for an expired or interrupted proposal holds for requests that run one after another. Two requests reading the same proposal at the same instant can each record it. - A `wpai_agent_event` listener that throws during an approval write marks the proposal failed, as any WordPress hook would. The hook isn't wrapped. - The log screen's status filter doesn't accept `skipped`, which item rows have used since before this PR. - Workspace rows written after this are `agent`, not `ability`, so a view filtered to `ability` no longer shows them. ## Changelog Entry > Added - AI Workspace records every agent tool call and proposal event in the AI Request Log under an `agent` type, with a `wpai_agent_event` hook that fires even when logging is off; saved chats are included in personal-data export and erase; and a Wapuu shows while a turn runs, finishes or fails. **Post-deploy monitoring:** no additional operational monitoring is required. This is an experiment behind the AI Workspace toggle, off by default. The exporter and eraser run only when an administrator processes a personal-data request. --- [](https://github.com/EveryInc/compound-engineering-plugin) 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- wp-playground-preview:start --> <a href="https://playground.wordpress.net?blueprint-url=data:application/json,%7B%22preferredVersions%22%3A%7B%22wp%22%3A%22latest%22%7D%2C%22login%22%3Atrue%2C%22landingPage%22%3A%22%2Fwp-admin%2Fadmin.php%3Fpage%3Dai-wp-admin%22%2C%22steps%22%3A%5B%7B%22step%22%3A%22installPlugin%22%2C%22pluginData%22%3A%7B%22resource%22%3A%22url%22%2C%22url%22%3A%22https%3A%2F%2Fgithub.com%2FWordPress%2Fai%2Freleases%2Fdownload%2Fci-artifacts%2Fpr-1317-77686452fbd3e34a0cb582c18648d7667c78a603.zip%22%7D%7D%5D%7D" target="_blank" rel="noopener noreferrer"> <img src="https://raw.githubusercontent.com/adamziel/playground-preview/refs/heads/trunk/assets/playground-preview-button.svg" alt="Open WordPress Playground Preview" width="220" height="57" /> </a> <!-- wp-playground-preview:end --> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
## What? Related: #1287 The docs pass for the #1287 stack. AI Workspace keeps its name and is documented as the plugin's one agent. The experiment doc, `readme.txt`, `README.md`, the changelog and the settings card now describe what #1314 to #1317 added. No code paths change. This is the fifth and last stacked PR for #1287, on top of #1317, #1316, #1315 and #1314. ## Why? #1287 folds two planned experiments into AI Workspace: Site Agent (#1214) and Command Palette AI Search (#1250). A reader of the docs, readme or changelog should learn that from them. Before this, the settings card and readme still described a workspace that only creates drafts and templates. The changelog had nothing for the write, palette, audit, privacy or Wapuu changes, `README.md` had no AI Workspace entry at all, and "Site Agent" was still listed under Coming Soon. ## How? - **`docs/experiments/ai-workspace.md`** - The summary and key features cover changes to existing content, site answers and screen links, the chats panel, the Wapuu, the audit log and personal-data export. - A new "One agent" section says what Site Agent and Command Palette AI Search proposed and where each now lives. It also says that the name, menu, URL and every identifier, including the experiment ID `ai-workspace`, are unchanged. - The entry-points list names every ability the experiment registers, and `Conversation_Privacy`. - A stale limitation is removed. It said chats have no history screen, but the chats panel ships. - **`readme.txt` and `CHANGELOG.md`** - Added: changes to existing content, the palette's AI search and Ask AI, the `agent` log rows, chats in export and erase, and the Wapuu. - Developer: `wpai_agent_event`, the two new proposal filters, the MCP visibility change, `raw: true`, and the `ability` to `agent` log type change. - `CHANGELOG.md` also gains the two ability lines #1315 added only to `readme.txt`. - **`README.md`:** adds the missing AI Workspace entry. - **Both readme files:** "Site Agent" leaves Coming Soon. - **Settings card:** the AI Workspace description says it answers site questions, proposes changes to posts and the site title and tagline, and adds AI to the command palette. The name stays "AI Workspace". Nothing on #1287 has settled on another name. ### Use of AI Tools AI assistance: Yes. Tool(s): Claude Code (Claude Opus 5.5). Used for: writing the documentation and changelog entries, checking them against the code, and code review. I've reviewed the changes, understand them and can explain them under review. They have all been tested locally. ## Testing Instructions 1. Open Settings → AI. The AI Workspace card reads as in the screenshot below. 2. Read `docs/experiments/ai-workspace.md` → "One agent", and the Unreleased changelog in `readme.txt` and `CHANGELOG.md`. **Tests run** (all passing): - **PHPUnit, full suite:** 3398 tests, single site and multisite, on WordPress latest. Only the settings-card string is code, and no test asserts it. - **Other checks:** phpcs and `php -l` are clean on `AI_Workspace.php`. - **Name check:** a grep for "Site Agent" and "Wordspace" across `includes/`, `src/` and the built assets finds nothing user-visible. - **Claims checked against the code:** - the abilities `AI_Workspace::register()` registers; - `meta.mcp.public` false on the propose and ask-user abilities; - the `wpai_workspace_write_proposals_enabled` and `wpai_workspace_proposal_kinds` filters; - `raw` on `ai/read-content-bodies`; - the chats panel's rename, archive and delete actions; - the three in-page anchors the doc links to. - **Not run:** e2e. No script, style or behavior changes. **Code review:** `ce-code-review` sized this diff to a single-context review: 51 changed lines, two of them code. It found nothing and gave "Ready to merge". Its one residual is applied in 0b06540: the doc's feature bullet now says rows are written while AI Request Logging is on. ## Screenshots **The settings card**  ## Known limits - The changelog lines for #1314, #1316 and #1317 are added here rather than in those PRs, so they land only when this one merges after them. ## Changelog Entry > Changed - AI Workspace is documented as the plugin's one agent: it now includes what the Site Agent and Command Palette AI Search proposals planned, under its existing name. **Post-deploy monitoring:** no additional operational monitoring is required. Documentation and one settings-card string. --- [](https://github.com/EveryInc/compound-engineering-plugin) 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- wp-playground-preview:start --> <a href="https://playground.wordpress.net?blueprint-url=data:application/json,%7B%22preferredVersions%22%3A%7B%22wp%22%3A%22latest%22%7D%2C%22login%22%3Atrue%2C%22landingPage%22%3A%22%2Fwp-admin%2Fadmin.php%3Fpage%3Dai-wp-admin%22%2C%22steps%22%3A%5B%7B%22step%22%3A%22installPlugin%22%2C%22pluginData%22%3A%7B%22resource%22%3A%22url%22%2C%22url%22%3A%22https%3A%2F%2Fgithub.com%2FWordPress%2Fai%2Freleases%2Fdownload%2Fci-artifacts%2Fpr-1318-aa0faf515572fcec68b35f0b8a0634d387875fd1.zip%22%7D%7D%5D%7D" target="_blank" rel="noopener noreferrer"> <img src="https://raw.githubusercontent.com/adamziel/playground-preview/refs/heads/trunk/assets/playground-preview-button.svg" alt="Open WordPress Playground Preview" width="220" height="57" /> </a> <!-- wp-playground-preview:end --> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>


What?
Related: #1287
The AI Workspace can now change things that already exist, not only create drafts and templates. The assistant can propose a new title, excerpt or content for a post, or a new site title or tagline. Nothing changes until you approve, and the card shows each current value next to the proposed one. You can edit a value before approving it.
This is the first of the stacked PRs that build #1287's single agent on top of
feat/ai-workspace(#1004). It ports #1214's write actions. The next PRs add site info and screen finding, then the Command Palette entry, then audit and privacy hooks, then the docs pass for the name.Why?
#1287's first version needs the agent to finish tasks like "change my tagline to fit what I write about" instead of telling the person how. Before this, the workspace's approval flow could only create new posts. Its single-approval guarantee rested on a per-post token, which doesn't exist when you're editing something that's already there.
How?
Writes are proposal types, not abilities.
ai/propose-post-changesandai/propose-site-identityonly store a proposal. The write happens when the person approves it in the card. #1214 registeredai/site-agent-update-postandai/site-agent-update-settingas abilities instead. In this plugin everyai/*ability is offered over MCP by default, so those would have been callable by any MCP client with no approval step. #1214's validation rules are ported into the two new writers.The approval flow is now safe for editing existing content:
INSERT IGNOREinto the options table, the pattern core'sWP_Upgrader::create_lock()uses.add_option()won't do here: it writes withON DUPLICATE KEY UPDATE, which lets a second claim with a different value succeed.Content stays intact:
unfiltered_html, so this check is the only one.ai/read-content-bodiesgainsraw: true. It returns the stored markup, only for posts the user can edit, so a content change can keep the post's blocks.The card hides nothing that would publish. Content is parsed in an inert
DOMParserdocument and drawn through the transcript's existing renderer. Self-closing blocks (core/rss,core/latest-posts), media, embeds and form controls show in place as plain text, so a person sees everything they're approving. Nothing is fetched or run, and links are live only on the site's own host.Workspace-only abilities are off MCP.
ai/propose-*andai/ask-usernow declaremeta.mcp.publicas false. Before, MCP clients were offered tools that always refused outside a workspace turn. An override on the MCP Access screen still wins.Filters:
wpai_workspace_write_proposals_enabled(default true) turns these two proposal types off. That's New Experiment: One AI agent across wp-admin #1287's "allow writes" question: on by default, with no settings toggle.wpai_workspace_proposal_kindsis the list of proposal types. The tool list, the turn-ending rules and the system instruction are now all built from it.Code from #1214 is credited to @Vrishabhsk with Co-authored-by on the two writer commits.
Use of AI Tools
AI assistance: Yes. Tool(s): Claude Code (Claude Opus 5.5). Used for: planning, implementing the changes and tests, and code review.
I've reviewed the changes, understand them and can explain them under review. They have all been tested locally.
Testing Instructions
Tests run (all passing):
Proposal_StoreTest: lock, failed window, expiry grace, edits, kind registry, write gate, and client/server ability parity. Includestest_a_lock_row_written_outside_the_cache_blocks_the_claim, which failed against the oldadd_option()lock.Post_Field_ProposalTest, includingtest_a_title_changed_since_the_proposal_is_stale,test_two_fields_of_one_post_both_write,test_a_post_locked_by_another_user_is_locked,test_content_with_an_iframe_is_refused_but_other_fields_work,test_content_with_an_image_block_accepts_a_change,test_content_with_a_separator_block_accepts_a_change,test_content_kses_would_strip_from_is_refused,test_an_edit_cannot_retarget_the_item,test_a_duplicate_field_of_one_post_is_refused,test_a_model_supplied_snapshot_is_ignoredandtest_a_title_with_a_backslash_survives_the_write.Site_Identity_ProposalTest: 16 tests.Proposal_ControllerTest:test_approval_while_another_request_holds_the_lock_writes_nothing,test_an_interrupted_write_keeps_partial_results_and_reads_as_failed,test_approving_after_the_chat_was_deleted_is_refusedandtest_get_returns_the_stored_results_once_approved.MCP_AdapterTest:test_plugin_ability_declared_hidden_stays_hidden,test_workspace_only_abilities_are_hidden_from_mcp,test_override_exposes_a_plugin_ability_declared_hiddenandtest_workspace_only_ability_refuses_outside_a_turn.Read_Content_BodiesTest:test_raw_returns_stored_markup_to_an_editor_of_the_post,test_raw_is_withheld_from_a_user_who_can_only_read_the_postandtest_raw_is_off_by_default.tests/e2e/specs/experiments/ai-workspace*specs, 135 passed at 4f6931b. They were run with this branch merged into AI Workspace: answer site questions and link to the right admin screen #1315, because this PR's own test site holds the live-check data.ai-workspace-update-proposals.spec.ts(8 tests): current next to new, plus editing; the server's message on an invalid edit; a second page after approving; approving in a second page while the first is still writing; a stale field; deselecting one of three changes; a reopened discarded proposal; the tagline.ai-workspace-html.spec.tsfor the content renderer. It covers every element shown as plain text instead of rendered, and the notice naming blocks whose settings a content change alters.claude-sonnet-5, at 4f6931b, through the workspace REST routes as an administrator): one turn and one approval per write kind, with the posts and options read straight from the database before and after.post_modified), and the edited excerpt was written as edited.wp:paragraphblock.test_a_stale_lock_replaced_before_the_delete_is_not_deleted,test_a_writer_that_throws_part_way_leaves_the_proposal_failed,test_discarding_while_an_approval_holds_the_lock_is_refused,test_an_executed_proposal_read_after_expiry_shows_its_outcome,test_approving_with_nothing_selected_is_refused_for_any_kind,test_a_post_another_hook_updates_during_the_write_keeps_its_own_columnsandtest_content_proposed_after_the_post_changed_since_it_was_read_is_refused.Code review, two rounds:
20261106-154448-0faae422): 8 reviewers and an independent validator. A bug in how the reviewer prompts were built cut off the PR's intent and the location of its diff, so each reviewer worked out the scope on its own. Its findings were checked against the code and are fixed in 25748bd:20261106-212611-pra2rerun): the same 8 reviewers on the full diff at 36820b2, with the intent and diff in every prompt. Each finding was reproduced or traced in the code before being fixed, rather than going through the separate validator. Fixed in 476784a to 4f6931b:The cross-model pass didn't run, because no model CLI from another provider is installed.
Unapplied review findings
includes/Experiments/AI_Workspace/Propose_Post_Changes.php: the four propose abilities repeat the same setup code (turn check, meta block, proposal envelope). A shared base would remove it. Deferred: it follows the existing pattern, and it's a refactor outside this PR's scope.Propose_Post_ChangesandPropose_Site_Identityeach check for a duplicate target themselves. The kind could own a target key, and the store could refuse duplicates once. Deferred: same refactor scope as above.src/experiments/ai-workspace/utils/proposal-card.ts: the client works out which proposal types are updates (UPDATE_KINDS,SITE_FIELDS) on its own. The server could sendis_updateand the target instead. Deferred: it changes the contract between client and server.decline_proposal()doesn't take the approval lock, so a discard can interleave with an approval that has claimed the lock but not yet saved.keep_other_columns()pins the*_save_prefilters for the wholewp_update_post()call. Asave_posthook that writes a different post would receive this post's columns.Screenshots or screencast
Post changes: each current value beside the proposed one
Editing a proposed value before approving
A tagline proposal
After approval: the post is updated
After approval: the new tagline is live
A field changed since the proposal is left alone and reported
Changelog Entry
Post-deploy monitoring: no additional operational monitoring is required. This is an experiment behind the AI Workspace toggle, off by default, and every write needs explicit approval.
🤖 Generated with Claude Code