The built-in tool for creating automations showed its example schedules with a literal backslash-n between the start time and the rule. Models copied that, and the schedule was rejected with an "Invalid RRULE" error, so they retried until they happened to write it differently. The examples now show a real line break, which is what the schedule parser expects.
Fixes#33730
Six toggles under Admin Settings > Interface > Visible controls (Selected Tools, Skills, Filters, Web Search, Image Generation and Code Interpreter) showed their raw translation key instead of a name. The English labels were missing, so they now display normally.
With request auditing turned on, only fields whose name ends in "password" were masked, so API keys and tokens saved in the admin settings (OpenAI and Ollama connections, Jupyter, web search and others) were written to the audit log as they were. A secret cut off by the body size limit also had its beginning logged. Fields whose name ends in password, secret, token, key or authorization are now masked, including lists of keys and values cut off by the size limit.
Co-authored-by: xfscott <229475069+xfscott@users.noreply.github.com>
Opening a channel waited for the channel details to arrive before it even asked for the messages, so the wait was the two requests added together. Both requests now start at the same time. A channel that fails to load still sends you back to the home page and shows no messages.
Before a chat is stored or returned, the server checks all of its text for null bytes and invalid characters that the database cannot store. That check converted the whole chat to JSON text and scanned it, which on chats with long histories or pasted images took as long as reading the chat from the database, and blocked every other request on that server process meanwhile. The check now inspects each text value in place and skips the slow scan for plain ASCII values such as base64 images. It flags exactly the same chats as before, except that chats containing the six characters \u0000 as ordinary text (common in code) no longer go through a needless cleanup copy on every load.
Measured on SQLite, best of 5 runs, sizes are the stored chat JSON:
| Chat | Check before | Check after | Opening the chat before | Opening the chat after |
| --- | --- | --- | --- | --- |
| 7 MB, text only | 13 ms | 2 ms | 31 ms | 20 ms |
| 49 MB, with images | 107 ms | 4 ms | 211 ms | 108 ms |
| 209 MB, with images | 456 ms | 15 ms | 890 ms | 444 ms |
Tool results sent through an OpenAI-compatible gateway lose their prompt cache marker right before the request leaves, so Anthropic models behind OpenRouter or LiteLLM never cache the prefix in tool loops. Now the marker is kept on a single text block when a tool result carries one. Tool results without a marker are sent exactly as before.
Fixes#33755
Opening a channel looked up every replied-to message one by one, so a page of 50 mostly-reply messages ran well over a hundred database queries. The replied-to messages and their authors are now fetched together, and the response is unchanged. Measured on a local SQLite database with a 50-message page where 46 messages are replies:
| | before | after |
|---|---|---|
| queries | 147 | 3 |
| time (median) | 46 ms | 1.9 ms |
Opening the channel list ran several database queries for every channel, so the load time grew with the number of channels. The last message time, unread count, mute state and direct message participants are now fetched for all channels at once, so the number of queries stays the same no matter how many channels a user has. The response is unchanged.
Measured on SQLite with an admin user in 23 visible channels (50 seeded), 2000 messages per channel:
| | before | after |
|---|---|---|
| database queries | 93 | 9 |
| response time | 255 ms | 47 ms |
Without Task Model Params set, the compaction summary was capped at 1000 tokens. Reasoning models could spend all of that on thinking and return an empty or cut-off summary, which then replaced the chat's context. The summary request now carries no token limit of its own unless Task Model Params sets one, so the provider's normal limits apply.
With the OpenAI image engine and no image model set, requests went out for DALL·E 2, which OpenAI has shut down, so image generation failed out of the box. The default is now GPT-IMAGE 2 and the model list no longer offers the shut-down DALL·E models. Fixes#33578
Converting 16,000 pasted lines to Markdown now takes 74 ms instead of 8.6 seconds, with byte-identical output. The converter got about 4x slower each time the text doubled, so every keystroke in the chat input after a large paste took seconds; its latest release fixes that and this is a plain version bump.
| Pasted lines | Before | After |
|---|---|---|
| 16,000 | 8,634 ms | 74 ms |
| 8,000 | 2,232 ms | 39 ms |
| 4,000 | 575 ms | 17 ms |
| 1,000 | 31 ms | 16 ms |
Related to #12087.
MinerU 4.1 replaced the upload and parse job routes of 4.0 with a new Tasks API, so every document sent to a 4.1 server failed with a 404. The MinerU engine in local mode now detects a 4.1 server and uses the new API, sending the API key only to the MinerU server itself. MinerU 4.0 and 3.x servers keep working as before. Tested against real MinerU 4.1.6 and 4.0.12 servers with an API key set.
Fixes#33718
After a thumbs down, pressing Save in the rating form while the rating's tags were still being generated did nothing: the form showed its thank-you message, but the score, reason and comment were never stored. Save now stores them right away, and clicking a thumb again while a rating is saving is ignored as before.
Fixes#33715
On the Notes, Models, Prompts, Skills, Tools and Automations pages, holding Shift is the quick-delete shortcut that turns each item's menu button into a Delete button. It stopped working once any button on the page had been clicked, for example after switching Notes to Grid view, and now works again. A menu button reached with the Tab key still stays a menu button while Shift is held, so keyboard users do not delete by accident.
Fixes#33714
With Retrieval Query Generation on, an external knowledge base (Qdrant, Milvus or pgvector) ran one search per generated query, one after another, and then kept only the results of the first query. The other searches cost time and added nothing to the answer. The searches now run at the same time and their results are merged by position (the top result of each query, then the second of each, and so on) with duplicates dropped, so every generated query contributes. Results are merged by position because scores cannot be compared the same way across vector databases: for some a higher score is better, for others a lower one.
| 3 queries, simulated search taking 0.3 s | before | after |
| --- | --- | --- |
| time | 0.9 s | 0.3 s |
| results used | first query only | all three queries |
Fixes#33184
Refreshing an OAuth token, or checking a token's client during token exchange, with the client secret sent over HTTP Basic logged aiohttp deprecation warnings on every call. aiohttp removes the deprecated helper in 4.0, after which these requests would fail. The Authorization header is now built with aiohttp's recommended replacement, and what the provider receives is unchanged. Fixes#33297
The external reranker, Exa and Ollama web search wrote the full user query to the server log at info level, and reusing generated search queries logged them as well. Those lines no longer include the query. A failing external reranker also logged the provider URL in its error message along with a stack trace; it now logs only the error type.
Co-authored-by: manooll <16244686+manooll@users.noreply.github.com>
Asking for a line range (for example the whole file by its last line number) returned every requested line, however large, so one tool call could exceed the model's context window and fail the chat. Line ranges now stop at the last whole line that fits the same limit as character reads, and report where to continue from.
Fixes#33607
The sidebar loaded channels first and only then started loading chats, so a slow channel request delayed the chat list on every app open. Both requests are independent and now run at the same time. The chat list shows up as soon as its own data is ready.
When a message is sent in an existing chat, the server loaded the full chat, then loaded it again right after to check that the user has write access. The check now reuses the chat it already loaded. Who can write to a chat does not change. The saving grows with chat size and is small for ordinary chats.
Time saved per message, measured on SQLite, best of 3 runs, sizes are the chat as stored in the database, images embedded:
| Chat | Time saved |
| --- | --- |
| 7 MB, text only | 34 ms |
| 49 MB, with images | 204 ms |
Email addresses with non-ASCII characters before the @ (internationalized addresses per RFC 6531) were blocked by the browser's own email field check, so the admin account and new users could not be created with them. The email fields on the sign-up page and in the add and edit user dialogs now accept them and still show the email keyboard on mobile. The server already accepted these addresses, so no backend change is needed.
Fixes#33681
A terminal connection with Auth set to Session had no API key field, so Verify Connection, the policy settings and the tool list request all went out without the orchestrator's key and were rejected as unauthorized. That left the policy dropdown empty and chats failing with "Terminal server is unavailable". Session connections now have the API key field, and the key is sent for those admin and discovery calls the same way Bearer connections already do. Terminals that users open still sign in with the user's own session.
Fixes#30001Fixes#33579
Typing indicators, sending a message, shared links, comments and file access all check whether the user may access the chat, and each check loaded the full chat, images included, only to look at who owns it and which folder it is in. On large chats that delayed every other request on the same server process by up to a few hundred milliseconds per check. The check now reads just those few fields, and who gets access is unchanged.
Measured on SQLite, best of 5 runs, sizes are the chat as stored in the database:
| Chat | Checked for | Before | After |
| --- | --- | --- | --- |
| 7 MB, text only | the chat owner | 34.6 ms | 0.36 ms |
| 7 MB, text only | a user with access through a shared folder | 35.1 ms | 2.58 ms |
| 57 MB, with images | the chat owner | 249.8 ms | 0.72 ms |
| 57 MB, with images | a user with access through a shared folder | 249.0 ms | 2.80 ms |
With the tabbed layout on, the Temporary Chat button disappeared from a new chat, so a temporary chat could not be started from there. The button now shows in the new chat's tab like it did in the classic navbar.
Our workflows refer to actions by tag, and a tag can be moved to different code after we reviewed it. This adds GitHub's workflow lockfile, which records the exact commit behind each action version, and the workflows now name exact releases (v5.1.0 instead of v5, for example) as the lockfile tool requires. Dependabot already tracks these actions. The regression workflow that follows the tests repo's main branch is not covered.
With Redis enabled, each chat request and each title or tag generation looks up which pipeline filters apply, and every lookup parsed the full model list stored in Redis from scratch. Each worker now keeps the parsed list and parses it again only when the model list changes. The lookup returns the same filters as before, at the cost of about 3 MB of memory per worker for 1000 models.
Measured per lookup against a local Redis; the larger list size in each pair comes from models with profile images:
| Models | Model list size | Before | After |
| --- | --- | --- | --- |
| 200 | 0.17 MB | 0.42 ms | 0.05 ms |
| 200 | 2.2 MB | 1.16 ms | 0.05 ms |
| 1000 | 0.83 MB | 2.43 ms | 0.06 ms |
| 1000 | 11 MB | 6.21 ms | 0.10 ms |
Every time a user loaded Open WebUI, each OAuth MCP server they had not connected yet wrote a warning to the logs ("No OAuth session found for user ..., client_id mcp:<server id>"), so instances with a few such servers filled their logs with it. Not having connected a server yet is a normal state, so the message now logs at debug level and only shows up when debug logging is on. Using an MCP server without connecting it still fails with "Auth required" as before.
When a streamed reply from a Responses API connection stopped early, for example at the token limit or through a content filter, its token counts were lost. The chat showed no usage for that reply and the admin analytics recorded zero tokens for it. Such replies now keep the usage the provider reported, the same as replies that finished normally.
Fixes#33615