Context compaction summarizes older messages and keeps the most recent ones verbatim, splitting the chat at a user message. It skipped the chat whenever the first follow-up message sat too close to the end to keep the share set in the Retained Messages admin setting. So a fresh chat with one question and one reply carrying a lot of tool output never compacted: the follow-up failed with a context length error, and regenerating or editing that follow-up failed the same way. Now compaction splits at the first follow-up message in that case, keeping fewer messages than Retained Messages asks for, so the chat keeps working. Chats where a split already met that share compact as before, and compaction between tool calls is unchanged.
Fixes#32781
Since 0.12.3, any chat opened under /s/, whether through a share link or by an admin viewing another user's chat, was fetched again and again from the moment it opened, and each fetch scrolled the page to the bottom. Scrolling up was impossible and older messages never loaded. The chat is now loaded once when the page opens.
Fixes#32657
Since 0.12.3, chat messages show a stray space at the start of every paragraph and inside every italic, bold, struck-through or linked span, so an em dash followed by italic text renders with a visible gap. The edit view has no such space; it only appears in the displayed message. The stray space is gone and formatted text sits flush against its neighbours again.
Fixes#32739
When a tool on a server built with the standalone FastMCP package returned a list, the model got the list twice, once as plain text and once as structured data, so every list result cost two to three times the tokens. That package sends the whole list as a single piece of text, a format Open WebUI did not recognise as a duplicate. The list now reaches the model once, and structured data that the text does not already contain is still passed along.
Fixes#32719
When the voice provider ended a realtime call, the overlay sometimes said "Voice connection failed" instead of "Voice provider connection closed". It happened when microphone audio reached the server at the same moment the provider hung up, so forwarding that audio failed and the generic error was shown. Anything that fails to reach the provider because it just closed the connection now reports that the provider closed the call.
Fixes#32729
When a function asked the user for input or confirmation and an execute event (a script sent to the browser) arrived in the same tab before they answered, clicking Confirm or Cancel closed the dialog but the answer never reached the function, which then waited until it timed out. Now the answer always goes back to the function that opened the dialog.
Fixes#32742
When a model sent the ask_user questions as anything other than a list, for example a JSON-encoded string, the tool answered "ask_user requires 1-3 questions." The model had sent the right number of questions, so it did not understand the error and often repeated the same call in a loop. It now gets a separate error saying the questions must be an array of question objects, and the count error only appears when the count is actually wrong.
Fixes#32662
Closing a page while the server was still sending it data (a file's processing status, a chat export) could interrupt a database read halfway. On SQLite the connection used for it was then never given back, and a later unrelated request could hang forever when that connection was finally cleaned up; once enough were lost, every request needing the database waited 30 seconds and failed with a 503. The connection is now kept and rolled back instead of being thrown away inside the interrupted request, and the session is closed once instead of twice, as either alone still lost connections. Closing 600 chat exports and 300 file status streams mid-load on SQLite used to lose 60 (the whole pool in that test) and 56 connections, now none; Postgres lost none before or after.
Fixes#32609
With ENABLE_FORWARD_USER_INFO_HEADERS turned on, voice calls in Realtime mode connected to the configured realtime provider without any user info headers, so a gateway in front of the provider could not tell which user was on the call. The call connection now sends the same user headers (or the signed user JWT, when configured) as chat requests and realtime text-to-speech already do.
Since 0.12.3 every frontend build first verifies the prebuilt encryption module, and on Windows that check failed before the build started, because it mishandled Windows paths and Git's default Windows line endings changed the source files it hashes. The check now gives the same result on Windows, Linux and macOS, including in existing Windows clones, so npm run build works on Windows again. The stored source fingerprint changes because the check script is one of the files it covers. Past the check, the build ran out of memory under Node's default limit, so it now runs with a 12 GB limit, the same value the Docker CI build uses.
Fixes#32656
When the model asked the user a question, or a tool call was approved or rejected, the reply was saved in a way that broke adding anything to it later in the same turn. If the model then generated an image, the image was created but never shown, the tool returned "unsupported operand type(s) for +: 'NoneType' and 'list'" and the model kept retrying, generating more hidden images. Progress status updates for that reply were also silently dropped. Both now work after a question or tool approval.
Fixes#32678
