Repository navigation
0.12.0 - #1454
Merged
Merged
0.12.0#1454
Conversation
Live updates are now published on one Redis channel per room when WEBSOCKET_MANAGER=redis, so each instance only decodes updates for rooms it has members in. Adds the env var reference entry and a scaling guide note that all instances must run the same mode, so changing the setting or upgrading from before 0.11.5 needs every instance stopped or updated together.
Moves the WEBSOCKET_REDIS_ROOM_CHANNELS note in the scaling guide from a Step 2 bullet into its own Step 3 section with the measured CPU cost it removes, and trims the same-mode warning in both places to the setting change.
…and ComfyUI USER_AGENT now also covers the Attach Webpage pre-check. The Pyodide library table lists openpyxl, python-pptx and python-docx, which install offline on import. Admin model reordering works in a filtered view, with hidden models keeping their place. The ComfyUI page names the output nodes images are collected from, now including Save Image (Advanced).
The Interface Settings Access note now names the model selector's Set as default as where the default model is saved, since the Interface page no longer carries it. The SSO page notes that an expiring OAuth token is refreshed once across concurrent requests, with a Redis lock spanning workers and replicas, which keeps sessions working with providers that rotate refresh tokens.
…, notes and channels Terminal skills are read from /skills/read?name=<name>, served by Open Terminal from v0.13.0. Admin terminals are called by the server, personal terminals from the user's browser, so a terminal on the user's own machine works with a remote Open WebUI. The Last 24 hours chart now honors the group filter. Stream Chat Response resolves model, then chat, then account. Note PDF exports include pasted images, and Ctrl+Enter to Send covers channel messages and thread replies.
The env reference and hardening guide described OAUTH_BLOCKED_GROUPS as denying sign-in to members of the listed groups. It has only ever excluded matching groups from OAuth group sync: users are never added to or removed from them, and from 0.11.5 they are never auto-created either. The description now says that and lists the supported match forms, and the group creation and SSO pages note the creation part.
…/ skills and folder share links max_tokens sent through the API reaches Ollama as num_predict. POST /api/v1/retrieval/process/web returns 400 Could not read content from <url> for a link that cannot be fetched, with process=false too. Tool call reconciliation keeps each call only with its own result from the same reply. Skills can be picked from the / menu as well as with $, and deleting a chat together with its folder also removes its share links.
The parent Sharing permissions for models, knowledge, prompts, tools, skills, notes and folders read as if they allowed or blocked sharing. They set whether the Access Control panel shows its access list. The pages now say that and name the Public Sharing and Allow Sharing With Users and Groups permissions as the ones that decide which grants are accepted.
The outlet filter examples only changed a reply's content, which updates the stored text but not what the chat shows, so they now edit the reply's output text too. The migration guide now shows pending migrations with the Alembic history and previews an upgrade by running it on a copy of the database.
Tavily searches can now run at a chosen depth, from ultra-fast to advanced, set in the admin web search settings or through an environment variable. The Tavily page and the settings reference now describe it.
With the setting off, admins are now refused on every chat action for another user's chat, not only when opening it, and the description lists those actions. It also states plainly that the setting is a product-level privacy control and not a security boundary, since an admin can always reach chats through the database, Functions and Tools or the server.
The entry only said it sets the default locale, which left open whether it beats the browser language. It now says the setting wins over the browser language on a first visit, that a ?lang= link or a language picked in Settings wins over it, and that browsers which already visited keep their saved language.
Every banner can be closed by the user. The dismissible setting, labelled Remember Dismissal in the admin panel, only decides whether a closed banner stays closed or shows again on the next page load.
A background sub-agent keeps the tool servers an admin added, OpenAPI and MCP alike. What it loses are the tool servers a user added in their own Settings, which run from the browser.
The Database download in Admin Settings > Database streams webui.db as it sits on disk and leaves out the webui.db-wal file, so with SQLite in WAL mode the copy can miss recent data. The backup and database pages and the ENABLE_ADMIN_EXPORT entry now point admins to copying the files themselves or using pg_dump on PostgreSQL.
…kers Both are read fresh on every turn on purpose, so the cached prefix changes whenever the terminal's AGENTS.md or its saved skills change mid-chat.
On a host that uses Tailscale's DNS, the c-ares resolver rejects that server's cached answers, so lookups hang and connections fail with the async DNS resolver setting on. The setting's warning now names this and says to keep it off on such hosts.
ENABLE_ORJSON now defaults to True on dev, so its reference entry, the multi-replica, performance and scaling pages give v0.11.5 as the point it turns on by default and say that setting it to False goes back to the standard library. The advice to turn it on stays for v0.11.0 to v0.11.4.
Upgrading Milvus installs to 0.11.5 copies every collection once at startup, so the update guide and the Milvus settings now carry a prominent warning: back up first, plan hours of startup time and free disk space for a second copy, and use Milvus 2.5 or newer. MCP tool calls now honour the tool server timeout, temporary chats offer no sub-agents or timers, channels queue messages while one is still sending, and the analytics token total splits into input and output on hover.
The env reference listed the web loader timeout as a float with a default of None. It is stored as text with an empty default, so a config import has to send it as a string such as "30".
Deploying with
|
| Status | Name | Latest Commit | Updated (UTC) |
|---|---|---|---|
| ✅ Deployment successful! View logs |
open-webui-docs-search | 2d1f5b6 | Oct 10 2026, 10:06 PM |
The next release was renamed from 0.11.5 to 0.12.0, so every mention of 0.11.5 in the docs now says 0.12.0.
Admins can now give a model simple controls, such as Thinking or Verbosity, that let people pick from approved options in chat without touching raw parameters. The model page explains how to set them up and how they appear in chat, the chat parameters page adds them to the precedence order, and the API reference shows how to pick an option in a request.
Voice Mode can now run calls through a realtime voice model that handles small talk itself and hands real questions to the chat's model, with transcripts saved into the chat. The voice mode page covers setup, how a call works, what is saved and the limits, the settings reference lists the new options, and the text-to-speech docs add OpenAI Realtime as an engine that reads text aloud.
The realtime call screen now centres on a voice orb with a status line, mute, end call and a retry button when the call drops, so the voice mode page describes it. The group editor's Users tab now switches between direct and inherited members with a dropdown, and the groups page matches.
# Conflicts: # docs/features/workspace/models.md
…at and filters after approval Models gain a Skills switch under Builtin Tools that stops the model from finding and loading skills on its own. On PostgreSQL, migrations now run inside DATABASE_SCHEMA, and startup explains when the schema is missing or inconsistent. Read-only chats in shared folders can be cloned into your own chats, and request filters now run before an approved tool executes.
A model can now carry a rigged 3D character that replaces the orb during realtime calls, with its mouth following the voice and optional animations and gestures. The voice mode page covers setup in the model editor, the file requirements, animations and gestures and what people see in a call, and the models page points to it.
Anyone with read access to a shared note can now open the files attached to it, and the note's chat uses them. The notes page says so, and that this covers files uploaded by the note's owner with read access only.
…creation The admin model list now has Available and Unavailable views, with Reset for base models and Delete for leftovers. Links that load a page, a YouTube video or start a call now ask first, ?q= only fills the input, and creating a note from a link asks before it is created. Dates follow the interface language and the streaming cursor now pulses.
… settings and sub-agent context MCP servers without OAuth discovery documents now fall back to the standard endpoints, and structured tool results reach the model. Outlet filters receive the finish reason, Open Terminal connections gain switches for the working directory and the user shell tools, realtime calls work with any OpenAI Realtime compatible server, and sub-agents and timers rebuild their instructions from the chat's own context.
Shared chats and folders can now let the people they are shared with reply in the same chat, with each message showing its sender, a typing indicator and owner settings kept private. Only a folder's owner or an admin can change its sharing, admins read other users' chats read-only in the app, the profile card shows local time and last activity, and turning Notes off also closes the notes API. The URL parameter sections are renumbered after the prefilled prompt section was removed.
Saving a function whose new code fails to load is now refused in the editor, with the error shown there and the saved version left running. The functions page said such a save switched the function off. It now describes the editor behaviour, and keeps the switch-off for code that fails when it is loaded again later, such as after a restart.
Knowledge bases now open as a file browser with a side preview that switches between the original file and its indexed text, which can be edited and downloaded. The directory sections now match the tree view and the actual delete behaviour. Exa can also read pages as the web loader, and the loader timeout applies to it.
Saves of workspace models, tools, functions, prompts and skills are now kept as versions that can be compared, set live or deleted, and the workspace pages, plugin docs and schema reference describe how. Skills now carry supporting files next to SKILL.md, import and export as ZIP or JSON, and can be created from chat without a terminal. List menus gain an Access entry, and the Docker pages give the measured 0.12.0 download sizes.
Tool Search keeps large tool definitions out of each request until the model looks them up, so the tools page, settings reference and prompt caching page now explain it and its four settings. Compaction can now also run between the tool rounds of one long reply, and temporary chats ask for tool approval through a live Allow or Deny dialog.
The model editor now groups its settings into folding rows, and model controls can be edited by anyone with write access and translated per language. Skills can be imported from a URL and shared to or opened from the community site. Also covered: the HTML shell no longer being cached by default, the Redis permissions room channels need, failed-chat notifications on provider errors, Docling page numbers and the renamed Import from URL menu.
…chats A pass over every change since 0.11.4 found pages that no longer matched: admin settings paths still naming the Experience group, the tool info popup, plugin module names, the code interpreter library list, the SSO error text, orjson described as opt-in and the web search error. It also adds the MCP tool specs endpoint, provider error codes passed through the API, model ID suggestions and the token exchange claim handling. Shared and read-only chats can now be forked as well as cloned, and enforced temporary chats refuse saved chats on every route.
A second pass over every change since 0.11.4 found the realtime call docs out of date (the built-in prompt, the missing Stop button, a new transcription error), old model editor wording on the tools and filter pages, and stale audio and model export notes. It also adds terminal AGENTS.md, MCP elicitation and server instructions, the image size option, duplicate knowledge content rejection and when changed Valves block a save.
… key generator With MFA on, accounts awaiting approval get no session, so the custom pending overlay only shows while MFA is off, and the legacy OAuth ID token cookie is only set when the MFA step is skipped. The Channels permission covers the whole Channels feature, and open-webui serve now generates its secret key with a secure generator.
# Conflicts: # docs/getting-started/quick-start/tab-docker/ManualDocker.md
Adds entries for the OAuth sign-in timeout and token endpoint auth method, the chat share permission, source metadata keys and the security header variables, and corrects the prompt suggestion and image API version defaults. The OAuth client timeout now says it applies to MCP tool server OAuth, the device type entry says it is detected automatically, and the unused fonts directory entry is removed.
A review of every 0.12.0 docs change found skills described as injected in full despite the length cap, model controls losing to an Ollama model's own think, format and keep_alive values, outdated prompt caching rows and an old claim that outlet filters skip the API. It also tightens repeated wording on the skills, models and tools pages, adds links between permissions, features and settings, and points several links at the right section.
A review of every 0.12.0 docs change found a ComfyUI setup list broken by an edit, the Web API speech engine wrongly said to block Realtime calls, and repeated or code-heavy wording on the voice mode page. It adds a tip to adjust the built-in voice prompt when calls feel cumbersome, says arena models are off by default, and adds missing links between chat features, audio settings, permissions and troubleshooting.
Three audits of every 0.12.0 docs change found wrong or incomplete statements, repeated wording, missing links and outdated paths. This applies their fixes: sharing and permission details, auth and env reference corrections, consistent admin settings paths, a reordered voice mode page, new troubleshooting entries and accurate image descriptions.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.