007AJ AI Ready Full – AI Chatbot, llms.txt & AI Crawler Manager

Changelog:

= 1.0.3 = (current version)

* Fixed: AI responses that come back without usable answer text (empty, refused or filtered, cut off, or unreadable) are now charged to the monthly token budget. The tokens the AI service reports are kept separately from whether the answer was usable and are returned with the error, so they are counted even when the visitor gets an error message.

* Fixed: each reservation is now settled exactly once in one atomic database update keyed by its id: the hold is removed and the billable tokens are added in the same step, so usage can never be dropped or counted twice, also with many questions at the same time.

* Fixed: when a successful response reports no usage (or zero, or can't be read), the whole reservation is charged: the estimate of the exact request that was sent plus its full reply length limit. Timeouts after sending, gateway timeouts and PHP stopping after sending (fatal error, exit) are charged the same way; requests that were never sent are not charged.

* Fixed: AI errors that still report usage are charged. Usage now includes Anthropic cache tokens and Gemini thinking tokens.

* Fixed: a request the AI service could bill now counts toward the site-wide daily cap even when the visitor got an error.

* Fixed: reservations left by a process that died are settled after 10 minutes at their full amount instead of being dropped.

* Fixed: the admin Test connection button and the admin search preview's query embedding are now added to the month's usage.

* Changed: visitors get a clear message when the AI service returns no usable answer ("please rephrase" for refused questions, "please ask again in other words" for empty answers).

* Docs: what is charged in each case, the fallback for responses without usage, and a troubleshooting entry for the new messages.

= 1.0.2 =

* Fixed: the monthly token budget can no longer be overshot by questions running at the same time. Before any paid call, each question reserves a conservative upper bound of its tokens: the full prompt with the conversation history and the largest source passages in the index, the reply length limit sent to the AI service, and the query embedding when embeddings search is on. One database update accepts the reservation only if usage plus every running reservation plus the new one fits the budget. A question that doesn't fit makes no paid call, not even the embedding.

* Fixed: right before the AI call the reservation is set to the estimate of the exact request that will be sent; if it no longer fits, the call is not made, and a request without a reply length limit is never sent. Reported usage, including paid query embeddings, is added with atomic increments.

* Fixed: required lead capture now follows the server in both "before the first question" and "after N questions" modes. When the server asks for details again, the chat forgets the browser's old capture, shows the form, stays paused, and sends the held question once the details are saved. The browser remembers a capture exactly as long as the server (30 days), and answers carry the current lead settings, so cached pages stay in step.

* Changed: the ajai_token_estimate filter can only raise the reservation. Added the ajai_token_count and ajai_query_embedding_tokens filters.

* Docs: how the token budget is enforced and how exact it is, a "Switching themes, deactivation and removing data" section, and more troubleshooting (connection test, chat not appearing, WP-Cron, email delivery).

= 1.0.1 =

* Fixed: parallel chat requests could get past the site-wide daily cap and the monthly token budget. Each AI request now reserves one message and an estimate of its tokens in a single database update before the AI is called (requests still running count against the limits), the real usage is recorded afterwards, and failed requests give the reservation back.

* Fixed: monthly usage and per-visitor counters use atomic database increments instead of read-then-write options, so parallel chats keep every count. Existing usage moves to the new counters table on update.

* Fixed: Export Personal Data now includes every message of the conversations linked to the email's leads, page by page, each conversation once, and nothing unrelated. Erase Personal Data deletes the same messages and the leads, in batches; email matching ignores letter case.

* Fixed: with lead capture set to Required after N questions, the question box is hidden and disabled until the lead is saved, including after closing, reopening, reloading or restarting the chat. The chat API enforces it too. Optional capture keeps its Skip button.

* Fixed: a lead that fails to save now shows an error instead of unlocking the chat.

* Improved: embeddings retry with backoff after key, quota or network errors, and a passage the provider refuses is skipped so it can't block the rest.

* Added: a Get support button and a Support card on the dashboard.

* Docs: wide tables scroll on phones; privacy section now says exactly what is encrypted and what export and erase cover.

= 1.0.0 =

* First release.