AI-Hub - Releases

Added

• Saving a conversation (the owner, 2026-10-06): /apps/ai_hub/conversation reads a remembered conversation back, …/save writes it to the person's Files as Markdown (AI-Hub/<app>/<date> <time> <first question>.md, in their language and time zone, images as "[n images]"), …/summary has the model sum it up once and saves that, …/forget drops it; from PHP transcript(), saveTranscript(), summarize(). A scenario's optional transcript callable shapes how its turns read when saved.
• Images with a question: an app may send up to 4 PNG, JPEG, GIF or WebP images (5 MB each) with a question, checked again on the server, and every engine passes them on — the Claude, Gemini and OpenAI-compatible APIs, and the Claude command line through a stream-json message. The Gemini command line takes none, and status() says so (images). A remembered conversation keeps a mark in place of the images, never the images themselves. / 質問に画像（PNG・JPEG・GIF・WebP、1 枚 5MB まで、4 枚まで）を付けられるようにした。サーバーでも確かめ、各エンジンに渡す（Claude・Gemini・OpenAI 互換の API、Claude のコマンドラインは stream-json で）。Gemini のコマンドラインは画像を受け取れず、status() の images で分かる。記録する会話には画像そのものでなく印だけを残す。
• Per-app additional prompt: in the administration settings each connected app (and Nextcloud's Task Processing) can be given its own instructions, added to every question that app sends, after the app's own instructions. Text from documents and records still never counts as an instruction. / 管理者設定で、アプリごと（Nextcloud の Task Processing を含む）に追加プロンプトを書けるようにした。そのアプリが送るすべての質問に、アプリ自身の指示の後に加わる。文書やレコードの中の文は従来どおり指示として扱わない。

Changed

• A remembered conversation lasts as long as the login it was held in (the owner, 2026-10-06): it is kept under the login (Nextcloud's auth token of the browser session, or the app password), so a new login starts a new conversation; logging out drops the conversations of that login; one unused for twelve hours (was a day) is dropped. The records kept the old way were removed. A caller with no login of its own is seeded from the history it sends when its record has lapsed. The answer says fresh when it began afresh, so a page can clear the turns it still shows.
• A scenario can carry the app's own rule on who may ask (allow), and AI-Hub checks it on every way in; the OCS endpoint answers only the scenarios an app opened to it (ocs), so an app's gate cannot be gone round by calling AI-Hub directly.
• The Claude API is given room for long answers (16,384 output tokens, up from 4,096), and an answer cut off at the limit is noted in the log.
• The Gemini command line tool is given its prompt on stdin, so a long context no longer runs into the kernel's argument limit.

Fixed

• The Claude command line no longer leaves a file in its home for every question: it is run with --no-session-persistence and with its auto memory off (CLAUDE_CODE_DISABLE_AUTO_MEMORY), and no longer resumes sessions of its own (it is handed the stored transcript, as every engine is), so .claude/projects under its home stops growing.
• A remembered conversation is keyed by a hash of app, user and token: a user id holding _ could collide with another user's conversation.
• When the Claude command line tool cannot resume its session, the dead session id is dropped instead of being tried again (and failing again) on every later turn.
• The per-session working folders of the command line tool are swept with their conversation.

• Each message handed to the model is capped, as the context already was.

• CalcBase, the spreadsheet of the Base series, is listed from the start among the apps that connect, as the others of the series are.

Removed

• The "tools for administrators" setting, which no request ever used; the tools setting that remains applies to Nextcloud's own Task Processing and the connection test, and says so.