Removes the notes on which release added, changed or fixed what across the docs, along with entries that only described problems in old releases. The permissions page now says each fact once, with admin exemption stated at the top instead of in every row.
Image settings only require each key or URL while its feature is on, and WebP images can be edited. The compaction threshold field and the title, tag and follow-up switches show only while their feature is on, folders get a description, knowledge file previews show the indexed text and refuse empty edits, and more screen reader and keyboard improvements are listed.
Three audits of every 0.12.0 docs change found wrong or incomplete statements, repeated wording, missing links and outdated paths.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.
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.
Relative links on index pages broke when the page was opened through a link without a trailing slash, sending readers to pages that do not exist, for example from the troubleshooting page into Open Terminal. The links on those pages now use full paths, so they work however the page was reached.
The Open Terminal pages linked openwebui.com/enterprise for the Terminals license, which returns 404. They now point at the enterprise page on the docs site, the same URL the rest of the docs already use.
Relative Markdown links written without a file extension or a trailing
slash are dead when clicked from a leaf page. Docusaurus renders the
correct href for the server-side pass. The hydrated router then
re-resolves the original relative path against the page URL rather than
the file path, so a click navigates somewhere the markup never pointed:
../security-policy -> /security/accepted-risks/security-policy
./rag -> /troubleshooting/web-search/rag
Both render Page Not Found. Hovering, middle-clicking and pasting the
URL all work, because each of those uses the href, which is correct.
Only a plain left click reproduces it, which is also why the build does
not catch it: onBrokenLinks is set to throw and passes, since it
validates the rendered href.
Adding the extension makes Docusaurus resolve the link at build time, so
the router never re-resolves it. That form was already used by 109 links
in this repo.
This covers 119 links across 64 files, every one confirmed to have an
existing target. Links already carrying an extension or a trailing
slash are untouched, as are relative links on index pages, where the
file directory and the page URL are the same and resolution agrees.
Verified by building the site and clicking links on pages in each
affected section: security, features, ecosystem and troubleshooting.
Brings the model-name refresh from main (PR #1347) onto dev, which had moved 90 commits ahead.
Ten files conflicted where both sides had touched the same lines. Resolved by keeping dev's corrected admin panel paths (Settings > Admin > Interface / Models / Audio / Connections, which main's audit predates) together with main's current model names. Where dev had rewritten the prose outright, dev's version stands: the multi-model merge no longer claims to use the Task Model, and the task model recommendations in performance.md now point at the Task Models page instead of repeating a list.
Group permissions live on the admin panel's own Users page, not inside the settings modal, so paths that routed a reader through Settings were sending them somewhere the option does not exist. Every path to that page now names it the same way, with one separator style rather than two.
Documents the sidebar on a touchscreen: a swipe follows your finger and settles on release or on a flick, the chat menu button is always visible because there is no hover to reveal it, and reopening the sidebar keeps the chat list you had already scrolled through rather than starting again from the top. The share instructions no longer tell a phone user to hover the mouse pointer.
Paths written as Admin Settings followed by a tab are the same destination as Settings then Admin then that tab, so they are written the one way throughout. Paths to the admin panel's own top-level pages are a different destination and are left as they are.
The sharing permission table now uses the labels the admin panel actually shows, so a reader can match a row to the switch in front of them. The claim that the group sharing default does not persist is removed, verified against the code as no longer true.
Records that a knowledge base search no longer holds up the worker that issued it, so a slow vector store costs the chat that searched rather than everyone connected to that worker, and that the timeout setting stops a request being held open rather than stopping the whole application from stalling.
A message sent while the shared model list is being rebuilt no longer fails or runs against a mix of old and new entries. That applies to chats using a person's own direct connections on a Redis-backed deployment, which is where the model list is shared between replicas.
Follow-up sweep over the mentions that were stale but not broken.
Prose and comparison pages move to the current generation: GPT-5.6,
Claude Opus 5 / Sonnet 5, Gemini 3.5 Pro. Context-window guidance that
described GPT-5 as a 400k-token model now reflects the 1M-token window,
and the RAG page no longer points at GPT-4 Turbo for large contexts.
API reference and plugin examples used gpt-4o, gpt-4-turbo, llama3.1 and
llama3.2:latest as placeholder model ids. These are illustrative rather
than functional, but they read as recommendations, so they now use
gpt-5.6-sol and qwen3.6:27b. The context-budget filter example in the
context-window guide is updated the same way, including the prefix-match
commentary that referenced specific ids.
Llama model references are replaced throughout: the Continue.dev config
example, the autocomplete and tokenizer lists, the model-identity FAQ,
and the Ollama pull example (now muse-glimmer:30b-q4_K_M). References to
llama.cpp and llama-server as a runtime are untouched, as is the note
that Chatterbox TTS is built on a Llama backbone, since that describes a
third-party model's architecture. The customer case studies are left
alone: they record what those customers actually deployed.
DALL-E is no longer named as the OpenAI image engine, since the images
API now spans gpt-image-2, gpt-image-1 and DALL-E.
Verified with a full Docusaurus build.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016jT6A7usMVzSMwciLU9PxS
A sweep over the plugin documentation, checking every factual claim against the backend and the frontend that consumes it, plus the pages elsewhere that repeat those facts.
The corrections that change what a reader would write: the Action return contract, where every example returned a bare content dictionary that the frontend ignores, so the examples silently did nothing; the confirmation and input examples that used the event emitter instead of the event caller and therefore could never receive an answer; the code execution example that was Python in a slot that runs JavaScript in the browser; a metadata field used on two pages to tell WebUI requests from API ones that has never existed; the Responses API stream event shape, which was undocumented, along with the fact that editing only the deltas is undone before the message is saved; and reserved arguments whose type, shape or hook scope was wrong.
Also refreshed against renamed UI (Integrations, Model Defaults, Import From Link, the iframe sandbox toggles), removed a citation flag that has done nothing since the 2025 tools refactor, corrected the event caller timeout, which has no default rather than 300 seconds and returns an error dictionary rather than raising, and documented that a function whose module fails to load is switched off automatically.
Classic298
Timothy Jaeryang Baek
silentoplayz
Claude