Xenon Comment Review Agent

Changelog

1.2.3

• Reworded the screenshot captions. No change to plugin behaviour.

1.2.2

• Move the multisite uninstall loop into a function, so it no longer leaves variables in
  the global namespace
• Use the same annotation form as the rest of the file for the audit log insert

1.2.1

Hardening and correctness pass over 1.2.0.

Moderation timing:

• Moderate a new comment immediately instead of waiting for the next scheduled sweep.
  WP-Cron only runs when a page is loaded, so a comment previously waited for both the
  next visitor and the configured interval; it is now usually reviewed within seconds
• Only schedule the recurring sweep while moderation is enabled, and raise its default
  interval to 5 minutes now that it is a fallback rather than the primary trigger

Injection protection:

• Normalise zero-width, bidi and full-width characters before matching, so a single
  invisible codepoint no longer slips a payload past every detection pattern
• Also scan the post title, post extract and author IP, which reach the model but were
  previously excluded from detection
• Derive the fence token from random_bytes() rather than wp_generate_password(), whose
  public filter let another plugin make the token predictable
• Prefer flagging over rejecting when a signal could be innocent, so an ordinary bug
  report beginning “System: Windows 11” is held for a human rather than trashed
• Record detections even when the AI request itself fails
• Put a flagged comment on hold explicitly rather than assuming it already is

Fixes:

• Fix “Moderate approved comments” doing nothing. Comments WordPress had already approved
  were closed as “already moderated” before the AI ever saw them, which is exactly the
  case that setting exists to cover
• Fix settings silently reverting. The settings screen re-seeded itself from the values
  present at page load, so switching tabs and saving again could overwrite what had just
  been stored
• Fix custom detection patterns containing “<” being stripped on save
• Fix new settings never reaching an existing installation, which left injection blocking
  off while the settings screen showed it on
• Fix the connection test sending the stored API key to a caller-supplied endpoint
• Fix a fatal error when the queue or log endpoints were given a non-integer page size
• Fix a transient provider failure consuming all three retries within one sweep
• Fix a spam verdict from another plugin being overturned on retry
• Fix the “comment re-queued” notice never appearing
• Fix the cleanup cutoff comparing UTC against site-local timestamps
• Fix the cleanup event being left scheduled after deactivation
• Record the database version only once the schema is verifiably in place

Housekeeping:

• Restrict admin notices to the screens where they can be acted on
• Remove all plugin data from every site of a multisite network on uninstall, including
  the stored API key
• Use English as the source language for all translatable strings, so the plugin can be
  translated on translate.wordpress.org
• Require WordPress 6.6 or later, matching the react-jsx-runtime script dependency the
  admin interface is built against
• Correct the readme description of what is sent to the AI provider

1.2.0

• Add prompt injection protection. Visitor-supplied data (comment body, author name and
  email, article context, parent comment) is now isolated inside a fence whose delimiter
  carries a per-request random token, so comment text cannot be read as an instruction to
  the moderating model
• Add configurable injection detection. Built-in patterns cover tool-name references,
  instruction overrides, role spoofing, delimiter forgery and false authority claims;
  site owners can add their own plain-text patterns in the settings page
• Add a setting controlling whether a detection blocks auto-approval. Detection always
  runs and is always recorded; enforcement is optional
• Add an audit log filter and badge for injection detections. Detections are recorded
  even when routine audit logging is disabled
• Only include a parent comment in the AI context when it has been approved, closing a
  channel where an unapproved comment could still reach the prompt
• Fix the default system prompt never being used. It is stored as an empty string, so the
  null-coalescing fallback never triggered and the model received no moderation rules
• Fix a race in the moderation queue. Items are now claimed atomically, so concurrent
  cron runs cannot moderate the same comment twice
• Fix queue items becoming permanently stuck. Items abandoned by a worker that died
  mid-request are now returned to the queue or marked as errors
• Fix the admin UI language not following the stored setting on first load
• Fix CREATE TABLE statements so dbDelta can perform schema upgrades. The previous
  statements used IF NOT EXISTS, which made dbDelta parse the table name incorrectly and
  silently skip every schema comparison
• Add an injection column to the audit log table (database version 1.1.0)
• Cap the length of every field sent to the AI provider and of the returned reason

1.1.3

• Fix Contributors username to match WordPress.org profile
• Add External Services section declaring OpenAI API usage
• Add Development section with source code link and build instructions
• Rename JS global variable to use plugin-specific prefix (xcraData)
• Rename WP-Cron schedule to use plugin-specific prefix (xcra_every_)

1.0.0

• Initial release
• OpenAI provider with function calling (AI Agent mode)
• Async moderation queue with configurable WP-Cron interval
• React SPA settings page with multi-language support (Chinese / English)
• Queue management with retry and delete
• Audit log with detail view and bulk clear
• Statistics dashboard
• Comment list AI status column
• Re-moderation action
• Clean uninstall (removes all plugin data)