Skip to content

review-notes.md

Latest commit

 

History

History
432 lines (339 loc) · 16.5 KB

review-notes.md

File metadata and controls

432 lines (339 loc) · 16.5 KB

AI Review Notes

Summary

The AI Review Notes experiment adds a block-by-block AI editorial review to the WordPress post editor. Clicking "Generate Review Notes" in the post sidebar triggers the AI to examine each reviewable block and create WordPress Notes directly on the relevant blocks with concise, actionable suggestions across four categories: Accessibility, Readability, Grammar, and SEO.

Overview

For End Users

When enabled, a "Generate Review Notes" button appears in the post status info panel (the sidebar area below the post status). Clicking it triggers a review pass:

  1. The button label updates to show review progress (Reviewing blocks… (2 of 8))
  2. Each content block is sent individually to the AI for analysis
  3. Notes with suggestions appear directly on the blocks inside the Notes panel
  4. After completion, a count of new suggestions is shown beneath the button

Key Features:

  • Block-level Notes with suggestions scoped to each block's content and type
  • Four review categories: Accessibility, Readability, Grammar, SEO
  • Accumulating history: subsequent review runs append replies to existing Note threads rather than creating duplicate threads
  • Prior suggestions are sent back to the AI as context so it avoids repeating itself
  • Blocks whose Note thread has been resolved (marked as approved) are skipped on re-run
  • Works with common block types: paragraphs, headings, images, lists, tables, quotes, and more

For Developers

The experiment consists of:

  1. Experiment Class (WordPress\AI\Experiments\Review_Notes\Review_Notes): Registers the ability, enqueues the block editor asset, and wires server-side hooks for Note author override and block metadata cleanup
  2. Ability Class (WordPress\AI\Abilities\Review_Notes\Review_Notes): Receives a single block's content and returns structured JSON suggestions
  3. React Plugin (src/experiments/review-notes/): Drives the UI and orchestrates block traversal, Note creation, and thread management via WordPress data stores

Architecture & Implementation

Key Hooks & Entry Points

WordPress\AI\Experiments\Review_Notes\Review_Notes::register() wires everything once the experiment is enabled:

  • wp_abilities_api_init → registers the ai/review-notes ability
  • enqueue_block_editor_assets → enqueues the React bundle whenever the block editor loads
  • rest_pre_insert_comment (filter) → maybe_set_ai_author() — overrides the comment author to "WordPress AI" when meta.ai_note is true, so AI-generated Notes are not attributed to the authenticated user's account

Assets & Data Flow

  1. PHP Side:

    • enqueue_assets() loads experiments/review-notes and localizes window.aiReviewNotesData:
      • enabled: Whether the experiment is currently enabled

  2. React Side:

    • index.tsx registers the ai-review-notes plugin
    • ReviewNotesPlugin.tsx renders the button inside PluginPostStatusInfo
    • useReviewNotes.ts hook manages all state and orchestration:
      • Flattens the block tree to get all descendants
      • Filters to reviewable block types with sufficient content (≥ 20 chars), capped at 25 blocks
      • Fetches Notes in two parallel requests:
        • GET /wp/v2/comments?type=note&status=hold&post=<id>&per_page=100 — pending Notes used as context to avoid repeating suggestions
        • GET /wp/v2/comments?type=note&status=approve&post=<id>&per_page=100 — resolved Note IDs; blocks with a resolved Note are skipped entirely
      • Processes blocks in parallel batches of 4, calling the ability for each
      • Creates new Note threads via POST /wp/v2/comments (with meta: { ai_note: true } to trigger the AI author override) and updates block metadata.noteId
      • Subsequent runs append replies to existing Note threads

  3. Ability Execution:

    • Receives one block's content at a time (block type, plain text, post context, prior Notes, review types)
    • Builds a structured prompt and sends it to the AI with the system instruction and a JSON schema for structured output
    • Parses the JSON response, sanitizes each suggestion, and returns { suggestions: [...] }
    • Returns { suggestions: [] } when the AI finds no issues
    • Deduplicates against existing_notes: if a Note already contains a [TYPE] marker for a given review type, that type is skipped in the current run

Block Types Reviewed

const REVIEWABLE_BLOCK_TYPES = [
  'core/paragraph',
  'core/heading',
  'core/list',
  'core/list-item',
  'core/quote',
  'core/verse',
  'core/image',
  'core/table',
  'core/preformatted',
  'core/pullquote',
];

Blocks with fewer than 20 characters of text content are skipped. The review is capped at 25 blocks per run to control cost.

Input Schema

array(
    'block_type'     => array(
        'type'        => 'string',
        'description' => 'The block type, e.g. core/paragraph, core/heading.',
    ),
    'block_content'  => array(
        'type'        => 'string',
        'description' => 'The plain-text content of the block to review.',
    ),
    'context'        => array(
        'type'        => 'string',
        'description' => 'Surrounding content to improve review relevance.',
    ),
    'post_id'        => array(
        'type'        => 'integer',
        'description' => 'ID of the post being reviewed.',
    ),
    'existing_notes' => array(
        'type'        => 'array',
        'items'       => array( 'type' => 'string' ),
        'description' => 'Existing Note texts for this block from prior review runs, used to avoid repeating suggestions.',
    ),
    'review_types'   => array(
        'type'        => 'array',
        'items'       => array( 'type' => 'string', 'enum' => array( 'accessibility', 'readability', 'grammar', 'seo' ) ),
        'description' => 'Review types to perform.',
    ),
)

Output Schema

array(
    'type'       => 'object',
    'properties' => array(
        'suggestions' => array(
            'type'  => 'array',
            'items' => array(
                'type'       => 'object',
                'properties' => array(
                    'review_type' => array( 'type' => 'string' ),
                    'text'        => array( 'type' => 'string' ),
                ),
            ),
        ),
    ),
)

Permissions

The ability's permission_callback has two paths:

  • With a numeric post_id (post ID): Validates that the post exists, the current user has edit_post capability for that specific post, and the post type is registered with show_in_rest => true. Returns false if the post type is not REST-accessible.
  • Without a post ID: Requires current_user_can( 'edit_posts' ).

In both cases, users without the required capability receive an insufficient_capabilities WP_Error.

Note Association

Notes are WP_Comment objects with comment_type = 'note' and status = 'hold'. Block association is maintained via block metadata:

  • New thread: POST /wp/v2/comments with parent: 0 → response id stored in block.attributes.metadata.noteId via updateBlockAttributes
  • Reply: POST /wp/v2/comments with parent: existingNoteId → block metadata unchanged (association already set)
  • AI author: All Notes created by this experiment include meta: { ai_note: true }. The rest_pre_insert_comment filter intercepts this and sets the author to "WordPress AI" with no email, URL, or user ID, so Notes are not attributed to the authenticated user's account.
  • Resolved Notes: Notes with status = 'approve' (resolved) cause their associated block to be skipped entirely on the next review run.

Using the Ability via REST API

Endpoint

POST /wp-json/wp-abilities/v1/abilities/ai/review-notes/run

Authentication

See TESTING_REST_API.md for authentication details (application passwords or cookie + nonce).

Request Examples

Review a Paragraph Block

curl -X POST "https://yoursite.com/wp-json/wp-abilities/v1/abilities/ai/review-notes/run" \
  -u "username:application-password" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "block_type": "core/paragraph",
      "block_content": "The committee was formed by the director in order to study the problem and make recommendations.",
      "review_types": ["readability", "grammar"],
      "existing_notes": [],
      "post_id": 42
    }
  }'

Response:

{
  "suggestions": [
    {
      "review_type": "readability",
      "text": "Rewrite in active voice: \"The director formed a committee to study the problem and make recommendations.\""
    }
  ]
}

Review an Image Block (Accessibility Only)

curl -X POST "https://yoursite.com/wp-json/wp-abilities/v1/abilities/ai/review-notes/run" \
  -u "username:application-password" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "block_type": "core/image",
      "block_content": "",
      "review_types": ["accessibility"],
      "existing_notes": []
    }
  }'

Response (missing alt text):

{
  "suggestions": [
    {
      "review_type": "accessibility",
      "text": "Add descriptive alt text to this image so screen reader users understand its content."
    }
  ]
}

Using WordPress API Fetch (in Gutenberg)

import apiFetch from '@wordpress/api-fetch';

async function reviewBlock( blockType, blockContent, existingNotes = [] ) {
  const result = await apiFetch( {
    path: '/wp-abilities/v1/abilities/ai/review-notes/run',
    method: 'POST',
    data: {
      input: {
        block_type: blockType,
        block_content: blockContent,
        review_types: [ 'accessibility', 'readability', 'grammar', 'seo' ],
        existing_notes: existingNotes,
        context: String( postId ), // numeric post ID as string
      },
    },
  } );

  return result.suggestions; // Array of { review_type, text }
}

Error Responses

Code Meaning
block_content_required block_content was empty
post_not_found The post ID passed does not exist
insufficient_capabilities User lacks edit_posts (or edit_post for the specific post)

Extending the Experiment

Customizing the System Instruction

Edit includes/Abilities/Review_Notes/system-instruction.php to adjust:

  • Which review types apply to which block types
  • How strictly prior suggestions are de-duplicated

Filtering Preferred Models

add_filter( 'wpai_preferred_text_models', function( $models ) {
    return array(
        array( 'openai', 'gpt-4o' ),
        array( 'openai', 'gpt-4o-mini' ),
    );
} );

Disabling the Experiment Programmatically

add_filter( 'wpai_feature_review-notes_enabled', '__return_false' );

Adding Custom Review Types

The review_types input field accepts any string values. Pass additional type names from the JS side and update the system instruction to provide guidance for those types:

// In your custom JS
await runAbility( 'ai/review-notes', {
  block_type: 'core/paragraph',
  block_content: '...',
  review_types: [ 'accessibility', 'readability', 'grammar', 'seo', 'tone' ],
  existing_notes: [],
} );

Then add guidance for the tone type to system-instruction.php.

Testing

Manual Testing Steps

  1. Enable the experiment:

    • Go to Settings → AI
    • Enable the global toggle
    • Enable AI Review Notes
    • Ensure valid AI credentials are configured

  2. Run a review:

    • Create or open a post with a mix of block types (headings, paragraphs, an image without alt text, a list)
    • Open the post sidebar (click the Settings button in the toolbar)
    • Click Generate Review Notes in the post info panel
    • Watch the progress counter advance (Reviewing blocks… 2 of 8)
    • After completion, open the Notes panel (via the block toolbar or the comments icon)
    • Verify Notes appear on relevant blocks, formatted as [REVIEW_TYPE] Suggestion text.
    • Verify Notes show "WordPress AI" as the author rather than your account name

  3. Re-run accumulation:

    • Click Generate Review Notes a second time
    • Verify existing Note threads gain replies rather than new top-level Notes
    • Verify prior suggestions are not repeated

  4. Resolved Notes:

    • Mark a Note as resolved in the Notes panel
    • Run the review again
    • Verify the resolved block is skipped entirely

  5. Note deletion cleanup:

    • Delete a Note from the Notes panel
    • Save the post
    • Verify the deleted block no longer has a noteId in its block metadata (inspect via the Code Editor)

  6. Edge cases:

    • Post with only very short blocks → button completes instantly with "No new suggestions found."
    • All blocks already have Notes → second run skips repeats
    • Disable experiment → button disappears from sidebar

Automated Testing

PHPUnit integration tests:

npm run test:php

Test files:

  • tests/Integration/Includes/Abilities/Review_NotesTest.php — Ability class tests
  • tests/Integration/Includes/Experiments/Review_Notes/Review_NotesTest.php — Experiment class tests

Covers:

  • Input/output schema structure
  • suggestions_schema() OpenAI wrapper structure (name, strict, schema keys; inner type must be object)
  • Empty content validation
  • Mock-based suggestion return and structure
  • Content sanitization
  • Permission callbacks: no post ID path (editor, subscriber, logged-out), and post-specific path (valid post, missing post, insufficient edit_post, non-REST post type)
  • execute_callback with missing post ID → WP_Error
  • get_existing_review_types_from_notes(): type extraction, case normalisation, multiple types per Note, Notes without brackets
  • Experiment hook registration (rest_pre_insert_comment)
  • ai_note comment meta registered with show_in_rest
  • maybe_set_ai_author(): overrides author when ai_note is true, passes through otherwise, handles WP_Error

Playwright E2E tests:

npm run test:e2e:env:start   # Start wp-env + mock server
npm run test:e2e -- --grep "AI Review Notes"
npm run test:e2e:env:stop

Test file: tests/e2e/review-notes.spec.ts

Covers:

  • Button visibility in editor sidebar
  • Button busy/disabled state during review
  • Suggestion count feedback after completion
  • Empty result handling
  • Button hidden when experiment is disabled
  • No-op when post has no reviewable blocks

Notes & Considerations

Requirements

  • WordPress 6.9+ (Notes feature required for block-level comment association)
  • Valid AI credentials configured in Settings → Connectors
  • User must have edit_posts capability (or edit_post for the specific post when a post ID is provided)
  • The block editor must be active (classic editor is not supported)

Performance

  • Each block generates one API call; blocks are processed in parallel batches of 4
  • The review is capped at 25 blocks per run to control cost
  • Blocks with fewer than 20 characters of text are skipped
  • AI temperature is set to 0.7

Note Storage

  • Notes are stored as WordPress comments with comment_type = 'note' and comment_author = 'WordPress AI'
  • Block association is stored in block.attributes.metadata.noteId
  • Block metadata is saved as part of the post content when the editor saves
  • Note threads accumulate across review runs by design
  • Deleting or trashing a root Note automatically clears metadata.noteId from its associated block

Limitations

  • Image block review is limited to alt text presence; it does not analyze the image itself
  • Block metadata (noteId) is only persisted after the post is saved
  • The 25-block cap means very long posts will have only the first 25 reviewable blocks analyzed per run
  • Resolved blocks (approved Notes) are skipped in full; they will not receive new suggestions until the Note is un-resolved or deleted

Related Files

  • Experiment: includes/Experiments/Review_Notes/Review_Notes.php
  • Ability: includes/Abilities/Review_Notes/Review_Notes.php
  • System Instruction: includes/Abilities/Review_Notes/system-instruction.php
  • React Entry: src/experiments/review-notes/index.tsx
  • React Plugin Component: src/experiments/review-notes/components/ReviewNotesPlugin.tsx
  • React Hook: src/experiments/review-notes/hooks/useReviewNotes.ts
  • PHPUnit Tests (Ability): tests/Integration/Includes/Abilities/Review_NotesTest.php
  • PHPUnit Tests (Experiment): tests/Integration/Includes/Experiments/Review_Notes/Review_NotesTest.php
  • E2E Tests: tests/e2e/review-notes.spec.ts
  • Mock Fixtures: tests/e2e-request-mocking/responses/OpenAI/review-notes-suggestions.json