[XTR] AI Assistant 1.0.19

In this release, we focused heavily on User Privacy and Data Security. We know that running a private community means keeping your members' data safe and giving them control over how their content is used. With v1.0.19, we are introducing a powerful multi-layered privacy infrastructure.

What's New in this Release?

• Global User Opt-Out: We believe in user autonomy. Your members can now go to their standard XenForo Account -> Privacy settings and check a single box to prevent the AI Assistant from processing their messages. If a user opts out, their content is completely masked and ignored by the AI across the entire forum.
• Automated PII Data Masking: Administrators now have the option to enable a Personally Identifiable Information (PII) filter. When active, the system will automatically scan and hide sensitive data like email addresses and phone numbers before securely sending the context to AI providers.
• Thread-Level Privacy Consistency: When a thread starter disables AI bots for their topic, the AI Assistant is now completely disabled for that thread. We have also updated the user interface so that AI action buttons (like "Summarize Thread" or "Suggest Reply") are strictly hidden in these topics.

We highly recommend all administrators to update to this version to provide a safer and more privacy-compliant experience for their communities.

---

Changelog v1.0.19

New Features:

• Feature: Added a "Do not allow AI to process my messages" option to the native XenForo Account Privacy settings.
• Feature: Added an automated PII (Personally Identifiable Information) masking system to censor emails and phone numbers before AI processing.
• Feature: Added a new admin option to easily toggle the PII data filter on or off globally.

Enhancements & Fixes:

• Change: "Summarize Thread" and "Suggest Reply" buttons are now completely hidden if AI bots are disabled by the thread starter.
• Fix: Resolved an issue where AI bots could still process context from threads that were marked as "AI Disabled".
• Change: Opted-out users will no longer trigger Auto-Replies or Bot Mentions.

• Vision Capability Detection Fixed: Updated the strict regex filtering mechanism that caused the add-on to fail to recognize next-generation AI models (such as gpt-4.1, gpt-5.x, o1, o3, etc.) as vision-capable, which previously resulted in image attachments being incorrectly stripped from the API payload.
• Silent Bot Failure Resolved: Fixed the underlying logic issue where using custom or proxy-based Vision model names caused API requests to fail with a 400 Bad Request error, leading to the bot remaining completely silent and failing to reply to users.

• Expanded Model Support: The AI Assistant now provides full native support for the upcoming gpt-5 series, o1/o3 reasoning models, and flexible custom model names (e.g., gpt-4.1-mini) provided by third-party API proxies. Regardless of the model name, image attachments will now be successfully processed and delivered to the API.

• Fix: Added a specialized filtering method to automatically remove Perplexity/Sonar style footnote markers from AI output.
• Fix: Implemented a sanitization step to purge unclosed or broken BBCode tags generated by AI models.
• Fix: Improved the quoting logic to completely remove media tags from the shortened quote content, preventing broken embeds in the front end.
• Enhancement: Added an "IDENTITY REMINDER" reinforcement block to the system prompt to prevent context loss for models.
• Feature: Added a new global option to manage the default checked/unchecked state of the AI trigger on the thread creation page.
• Update: Updated the thread creation template modifications to reflect the new dynamic checkbox option.

With this update, we have focused on improving the robustness of AI outputs, fixing quote rendering issues, and expanding customization options for administrators based on your feedback.
Key Highlights in this Release:

• Smart Citation Filtering: We have added a smart filter for research-augmented models like Perplexity and Sonar. Ugly, unclickable reference markers (like [1] or [1, 2]) that these models sometimes inject are now automatically stripped before rendering on the forum.
• Media Rendering in Quotes: Fixed a critical issue where the AI would sometimes break MEDIA, IMG, or ATTACH BBCode tags when trying to quote a user. Orphaned or half-written tags generated by the AI are now properly cleaned up and sanitized.
• Enhanced AI Persona Retention: Some AI models with shorter context windows tend to "forget" their assigned persona (e.g., "You are a friendly support assistant") as the conversation gets longer. We've added a strict reinforcement block to the end of the system prompt to guarantee your bot stays in character at all times.
• New Admin Option for Thread Creation: You can now control the default behavior of the "Enable AI Bots" checkbox when users create new threads. By using the new "Enable AI Bots by default" option in the admin panel, you can choose to have it unchecked natively.

• Updated AI Models: The predefined model lists in the system for Claude (claude-sonnet-4-6, claude-opus-4-6, etc.), Google Gemini (gemini-3.1-pro-preview, gemini-3.1-pro-preview-customtools, etc.), and xAI Grok (grok-3, grok-3-mini, grok-code-fast-1, etc.) have been completely refreshed with their latest flagship versions.
• Dynamic Model Fetching (OpenRouter & Hugging Face): The admin CP model lists for OpenRouter and Hugging Face have transitioned from static configurations to a dynamic API fetching system. To heavily optimize server performance and API limits, this vast list is cached securely via the XenForo Data Registry with a 12-hour lifespan.
• Fix - Editor Quote Block Merging Issue: Resolved a crucial JavaScript bug in the reply editor. Previously, when users quoted a post and utilized the "AI Suggestion (Rewrite/Refine)" feature for their draft, the system would destroy the existing BBCode quote blocks and merge all contents into a single paragraph. Structured quote blocks are now flawlessly preserved, isolated before sending, and safely prepended back to the AI's rewritten response.

This release is a stability and architectural update that brings the AI core up to date with the latest flagship generations (such as Claude 4.6, Gemini 3.1, Grok 3) and completely automates expansive networks like OpenRouter and Hugging Face.

Additionally, prioritizing community feedback, this patch rectifies a critical user-experience flaw in the frontend editor where the AI Rewrite feature would corrupt and consume existing BBCode Quote structures.

• Fix: Fixed an issue where the "Reply to unanswered threads (24h)" option in scheduled tasks ignored the time limit.
• Improvement: The scheduler logic has been updated to strictly enforce the "Minimum Thread Age" criteria (default 24h).

This update addresses a critical timing issue in the AI Assistant's "Reply to Unanswered Threads" feature.

What's Changed? In previous versions, even if the "Reply to unanswered threads (24h)" option was enabled in scheduled tasks, the bot could occasionally reply to threads that hadn't yet reached the 24-hour mark. With this release, the scheduler (cron) logic has been rewritten to strictly enforce this rule.

The bot will now only target threads that have absolutely zero replies AND are older than 24 hours (or your configured time). Threads newer than this threshold will be safely ignored.

Post-Update Recommendation: If you rely on this feature, please check your Schedule settings after updating to ensure the "Reply to unanswered threads (24h)" checkbox is enabled.

• New Feature: Added "Reply to unanswered threads (24h)" option for Schedules. Bots can now target threads with zero replies older than 24 hours.
• New Feature: Added "Maximum Thread Age" setting for the unanswered thread logic. This prevents bots from necro-bumping very old threads.
• New Feature: Implemented a distinct reply limit for bot mentions and quotes. Direct interactions now have a separate limit from automated replies.
• Improvement: Moved the max_thread_replies setting from the "Access" tab to the "Activation" tab in Bot settings for better UX.
• Improvement: Refactored the internal trigger logic to strictly distinguish between auto-replies, mentions, and quotes.

This update focuses on making your bots smarter and their interactions more controlled within your community.

Highlights:

1. Reviving Unanswered Threads:
Threads can sometimes slip through the cracks without a reply. With the new "Reply to unanswered threads" scheduler option, your bots can identify threads that have been sitting with 0 replies for over 24 hours and provide a one-time engagement triggers. To prevent "necro-bumping" ancient topics, we've included a "Maximum Thread Age" setting, ensuring the bot only targets relevant, recent discussions (e.g., last 30 days).

2. Granular Interaction Limits:
Previously, if a bot reached its auto-reply limit in a thread, it might stop responding even when users directly asked it a question. We have separated these limits! You can now cap the bot's automated responses (e.g., Max 3) while still allowing it to respond freely when users explicitly mention or quote it.
We also reorganized some bot settings for better usability and refactored the backend trigger logic for improved stability.

• Bot Permissions: Improved the control mechanism ensuring bots only operate in forums authorized by the administrator.
• Quick Reply Buttons: Fixed an issue where "Suggest Reply" and "Tone Selector" buttons would sometimes fail to appear even in authorized forums.
• General Improvements: Stability improvements were made regarding the system's detection of bot settings and forum permissions.

This update addresses an important visibility issue related to AI Assistant forum permissions.

Key Fix:
Previously, when a bot was restricted to specific forums, the AI buttons (Suggest Reply, Tone Selector) in the Quick Reply area would sometimes fail to appear or function correctly.
With this update, the system now detects the current forum and bot permissions with much greater precision. As a result, AI features are displayed flawlessly only in the areas you have designated and where the bot is authorized.

Fix: Added support for OpenAI Reasoning (o1/o3 series) models. Implemented automatic switching from max_tokens to max_completion_tokens and removal of unsupported creativity parameters (temperature, etc.) for these models.
Fix: Resolved an issue where notifications were not sent to users for replies to Profile Posts.
Improvement: Implemented a safeguard against mass token usage in systems with large Knowledge Bases. The system now prevents loading the entire database for short or empty queries/images, significantly reducing API costs.

This update focuses on compatibility with OpenAI's latest models and significant performance optimizations.

Key Changes:

• OpenAI o1 and o3 Model Support: You can now seamlessly use OpenAI's next-generation "Reasoning" models. The add-on automatically detects these models and optimizes API requests according to the new standards (max_completion_tokens, etc.).
• Smart Token Conservation: A critical improvement for users with large Knowledge Bases (RAG). The bot will no longer attempt to load the entire database when the user message is very short or contains only an image. This effectively prevents unnecessary API costs.
• Notification Fix: Resolved the issue where bot replies to profile posts were not triggering user notifications.

• Fix: Refactored bot triggering and forum permission logic. The "Access > Reply forums" setting is now strictly enforced as a prerequisite for bot triggers (mentions/quotes).
• New Feature: Added "Raw Mode" support to the "Summarize Thread" function. Bots with Raw Mode enabled will now provide direct summaries without conversational fillers.
• Improvement: Reduced the delay timer for the "Reply to unanswered threads" action from 24 hours to 15 minutes, making bots significantly more responsive to new discussions.
• Update: Updated admin panel option phrases to clearly explain the distinction between general access permissions and trigger filters.

This update introduces significant improvements to permission management and bot response times.

• Consistent Permissions: Based on feedback, we've refactored forum access logic. Permissions in the Access tab now act as global rules, while settings in the Activation tab act as additional filters. This resolves configuration conflicts.
• Faster Response: Bots configured to reply to unanswered threads now engage within 15 minutes instead of waiting 24 hours.
• Smarter Summaries: The "Summarize Thread" feature now respects your bot's "Raw Mode" setting, delivering concise, professional summaries without unnecessary chatter.

• Critical: Fixed context limit "0" bug - When "Thread context limit" was set to 0, the system incorrectly sent ALL previous messages to the AI instead of none. This caused translation bots and other specialized bots to ignore their instructions and switch to general chat mode. (<MentionReplyCreator.php, PromptBuilder.php>)
• Critical: Fixed conversation context limit bug - Same issue as above, but for Direct Message conversations. (MentionReplyCreator.php)
• Fixed schedule queue handling - Previously, schedules were marked as "completed" even when no suitable thread was found. Now schedules properly retry on the next run if no eligible content is found. (ScheduleRunner.php)
• Fixed array_rand() empty array error - Server error "array_rand(): Argument #1 ($array) cannot be empty" no longer occurs when node_ids array is empty. (ScheduleRunner.php)
• Fixed 'skipped' status not recognized - The ActionLog entity now properly recognizes "skipped" as a valid status value for skip pattern detection. (ActionLog.php)
• Fixed error_message column reference - Corrected wrong column name that was causing errors when logging skipped responses. (MentionReplyCreator.php)

Important Notes

• Translation bots using "Thread context limit: 0" should now work correctly
• "Reply Unanswered Threads" schedule feature is now more reliable

We're releasing version 1.0.9 with critical bug fixes that address issues reported by our customers.

What's Fixed

1. Translation Bot / Specialized Bot Issues
We discovered and fixed a critical bug affecting bots with strict instructions (like translation-only bots).

The Problem:
When you set "Thread context limit" to 0 (meaning "don't show previous messages to AI"), the system was actually sending all previous messages. This caused the AI to see previous conversations and switch to general chat mode, ignoring your specific instructions.

The Solution:
Setting context limit to 0 now properly means "include zero previous messages". Your specialized bots will now correctly follow their Thread Prompt instructions.

2. Scheduled Tasks Reliability

• Reply Unanswered Threads: Now properly retries if no eligible thread is found, instead of marking the schedule as "completed" and waiting for the next interval.
• Server Errors Fixed: No more "array_rand() empty array" errors in the server log.

3. Skip Pattern Feature
Fixed issues with the [SKIP] response pattern feature that allows AI to skip certain posts conditionally.

Thank You

Thank you to all customers who reported these issues. Your feedback helps us improve the add-on for everyone.

If you encounter any issues after updating, please open a support ticket.