This is a prior step to integrate with WordPress AI providers shipped on 7.0

@claude how you feel about leaving here a review? :)

Claude finished @AllTerrainDeveloper's task in 2m 40s —— View job

Code Review — PR #298

• Gather context and understand the PR
• Read changed PHP files (analysis.php, hooks.php, jobs.php, search.php, migrations.php)
• Read new test files
• Read migration doc
• Analyze and write review

Overall verdict: looks solid. The architectural simplification is well-motivated and executed cleanly. The core loop (remove bulk reindex → remove post/term analysis → switch to native WP search → migrate cron cruft) is a tight diff with no surprises. A few things worth fixing before merge, and a few observations.

Bugs / correctness

1. Stale file header docblock in search.php (line 10)

The top-of-file docblock still reads:

* Each tool fetches 10 entities with their _desktop_mode_ai_analysis meta so the model
* can compare AI-generated summaries to the user's description.

That's the old behaviour. The tools now return real titles + excerpts via WP_Query s= / get_comments search= with no dependency on the meta. Fix this →

2. analysis.php — dead 'post'/'term' branches in public helpers

desktop_mode_ai_save_meta() (line 159) and desktop_mode_ai_get_meta() (line 189) still have 'post' and 'term' switch cases. Since the hooks that called them are gone these branches are dead within the plugin itself. They also carry a misleading docstring: 'post' | 'term' | 'comment'. If the intent is to keep them as a public escape-hatch for third-party plugins that might store arbitrary meta under the key, the docblock should say so explicitly; if they're truly dead, drop them. Right now they imply the plugin still does post/term analysis.

Performance

3. N+1 queries in search_fetch_comments() (line 454)

For every comment in the batch a get_post() call fetches the parent post title:

foreach ( $comments as $comment ) {
    $parent_post  = get_post( $comment->comment_post_ID );

With batch size 10 that's up to 10 extra queries per tool call. Comments on the same parent post (common) pay the penalty multiple times since WP's object cache is not seeded. A single get_posts(['post__in' => array_unique($post_ids)]) before the loop eliminates this. (Same pattern in search_fetch_comments_by_post, though there the parent is fetched once already, so that one is fine.) Fix this →

Code quality / minor

4. @since 0.14.0 tags throughout files that ship at 0.11.0

analysis.php, hooks.php, jobs.php, and search.php all carry @since 0.14.0 on their functions. The PR bumps the plugin to 0.11.0, not 0.14.0. These were presumably leftovers from a previous version-numbering plan. Not a runtime issue, but it undermines the accuracy of the inline docs.

5. desktop_mode_ai_run_search() — docblock stub left in place (line 650)

There is a dangling orphan doc comment immediately before desktop_mode_ai_progress_message():

/**
 * Runs the agentic content-search loop.
 * …
 * @return array|WP_Error
 */
/**
 * Returns a friendly progress message for a tool name…

The first docblock belongs to desktop_mode_ai_run_search() but desktop_mode_ai_progress_message() sits between it and the function it describes. The diff (removed a duplicate dispatch-tool docblock) shows awareness of the duplication pattern — this one slipped through. Fix this →

Security

No new issues. Good handling throughout:

• sanitize_text_field() on the query arg before it reaches WP_Query / get_comments.
• get_php_error_log capability check (manage_options) sits in the dispatcher before desktop_mode_ai_fetch_error_log() is ever invoked.
• The streaming endpoint validates nonce + is_user_logged_in() + current_user_can('read') before touching any input.
• follow_up.result is JSON-encoded and bounded to 4 KB (desktop_mode_ai_followup_outcome_max_chars) before being forwarded to OpenAI, preventing unbounded token inflation from a misbehaving plugin.

Tests

The three new test files are well-structured and cover the right things:

• aiNativeSearch.php — proves content with no analysis meta is findable and that the keyword actually filters. The assertion assertArrayNotHasKey('ai_summary', $match) is a nice explicit guard against the old shape creeping back.
• aiCommentOnlyMigration.php — covers the happy path, the non-interference case (comment job left alone), and the full runner dispatch + stamp.
• commentsAiSpamScore.php — the assertSame(50, $score) in test_harmful_verdict_adds_20 is tight; consider whether that test should assert >= 50 in case the filter stacks with other contributors, or document that the fixture comment is designed to start with exactly score 30.

One gap worth adding: test_search_comments_by_post with a post_id = 0 input to confirm the error-path return is well-formed (the early-return at line 506 is there but untested).

Migration

Migration v2 (desktop_mode_migrate_unschedule_post_term_ai) is correct and minimal — wp_unschedule_hook clears all queued events for the hook regardless of argument signature, which is exactly right here since we want a full sweep. Leaving existing post/term _desktop_mode_ai_analysis meta in place is the right call (harmless, avoids a potentially large table scan, matches the documented behaviour in migration-ai-comment-only.md).

The migration doc itself is clear and the hook-removal table is the right level of detail for plugin authors.

Summary: Three items worth addressing before merge — the stale header docblock (#1), clarifying or removing the dead post/term branches in the meta helpers (#2), and the misplaced run_search docblock (#5). The N+1 query (#3) is a real efficiency win for comment-heavy sites and easy to fix. The @since tag noise (#4) is low priority but would be nice to clean up in the same pass.