v5.2.0pre1

• Added support for the separate ThunderAI Privacy add-on, which masks personal data locally before it is sent to the AI provider and restores it in the answer. It is turned on with the new Use ThunderAI Privacy option in the advanced options; the switch in the ThunderAI Privacy settings is the same option. If the add-on is not installed, a notice links to it. The local model and its processing live in ThunderAI Privacy; ThunderAI does not bundle its runtime.
• The option can be turned on only once the local model is downloaded: the Download the model button starts the download from ThunderAI, with its progress, and Open ThunderAI Privacy settings opens the add-on's own settings. Administrators can lock the option with a policy and have the model downloaded automatically with the new ai_processing_model_autodownload key.
• While the option is on, a processing failure blocks the operation, and so does a missing ThunderAI Privacy add-on: turn the option off to resume unprocessed requests. Locally restored replies never enter provider conversation history. ThunderAI Privacy works with the API connections only: while the option is on, a prompt sent through ChatGPT Web is blocked with a notice instead of being sent unprocessed. ThunderAI Privacy currently requires Thunderbird 153 or later, 64-bit.
• Custom prompts: added the Insert at cursor position action for the compose window. The answer is inserted where the cursor was when the prompt started, without replacing any text and without asking for confirmation; with no cursor in the email it goes at the beginning, above the quoted text and the signature. Useful for prompts that write new text, for example a whole email from the instructions asked with {%additional_text%} [#937].
• [ChatGPT Web] Fix: after a Retry following "message box not found", if the model cannot be verified the warning and the "Use the current model" button are shown again, so the prompt can still be sent.
• API chat: links in the answers now open in the default browser, and only when they are plain http or https addresses without a user name or password; any other link (mailto:, relative links, other schemes) does nothing, and links can no longer be dragged out of the chat. The chat window also has a stricter Content Security Policy, as an extra safeguard against remote content in the answers [#1001].
• API connections: an answer that stops arriving in the middle is now given up after a while, with a clear error ("The server stopped responding"), instead of staying on "generating" for ever. The Stop button now works at once also while the answer is stalled, keeping the text received so far. The wait is set by the new API stream idle timeout advanced option (120 seconds by default): increase it for reasoning models or slow local models that stay silent for a long time while working [#999].
• [Ollama API, OpenAI Comp API] Safer server addresses: an address with a user name or password is refused, a server on another computer reached over unencrypted http:// is used only after you allow it (asked once per server), and when a server redirects the requests a warning under its address tells you where, so you can enter the final address directly [#1000].
• Spam Filter, Add Tags, Calendar Event, Task and Translate: the settings page now warns when the prompt text no longer contains a valid JSON response format, which the feature needs to read the answer. The text area turns red and the reason, with the position of the error, is shown below it; the text can still be saved. The examples in the default English Spam Filter and Calendar Event prompts are now valid JSON.
• Fix: a prompt using {%mail_typed_text%} in the compose window no longer replaces the text you selected with the whole typed text: the answer now replaces your selection, and the typed text is selected automatically only when nothing is selected [#978].
• Fix: a custom prompt can no longer have the same ID as a built-in prompt, which ran the built-in feature in its place, hid one of the two from the context menu and could delete it when saving the Menu Order page. A custom prompt that already has the ID of a built-in prompt is renamed at startup (and on import) without its prompt_ start, for example prompt_summarize becomes summarize. The ID of a new custom prompt, or a changed one, can no longer start with prompt_, and copying a built-in prompt gives an ID without it [#979].
• Fix: the position, visibility and icon chosen for the organization prompts on the Menu Order page are now kept after a restart, instead of going back to the ones set by the policy. The text and the other options of an organization prompt still come from the policy only [#990].
• Fix: opening the AI chat window no longer leaves a background listener behind when the window is closed before it is ready or when the connection is not fully configured: a small leak that grew each time the window was opened.

To use this release download and install the thunderai-v5.2.0pre1.xpi file.