0.9.3 (#24482 )

* refac * refac * refac * Merge pull request #24356 from Classic298/patch-1 doc/chore: Update SECURITY.md * refac * refac * refac * refac * refac * chore: Update SECURITY.md (#24363) * Update SECURITY.md * Update SECURITY.md * Implement asynchronous database ping for health checks (#24380) * refac * refac * refac * refac * refac * refac * refac * refac * refac * refac * fix: prevent STT from blocking the uvicorn event loop (#24338) The transcription endpoint was async but called the synchronous transcribe() function directly, blocking the single-threaded uvicorn event loop for the entire duration of inference. This caused all HTTP and WebSocket connections to stall for every user on the instance during STT processing. - Add asyncio import - Use async UploadFile.read() instead of synchronous file.file.read() - Offload the blocking transcribe() call via asyncio.to_thread() Closes #24169 * refac * fix: open file content in new window when clicking file name in FileItemModal (#24125) Previously, clicking the file name link did not open the file content because the condition checked `!isPDF && item.url`, which failed for `type === 'file'` items that use an ID-based URL path. Update the condition to trigger on `item.type === 'file' || item.url`, and resolve the correct URL by extracting `fileId` from `item.id` or `item.tempId` instead of using `item.url` directly as the file identifier. * refac * refac * refac * refac * refac * refac * refac * refac * refac * Refactor file processing to use asyncio for transcribing, improving concurrency. (#24379) * Apply validate_profile_image_url to ChannelWebhookForm.profile_image_url (#24370) * refac * refac * refac * refac * refac * fix: stream GET /chats/all to prevent OOM on large chat histories (#24461) Convert the /chats/all endpoint from loading all user chats into memory at once to a streaming NDJSON response that fetches chats in batches of 100. This prevents Out-of-Memory crashes for users with large chat histories. Backend: Added async generator that paginates through chats with short-lived DB sessions per batch (critical for SQLite lock release). Frontend: Updated getAllChats to consume the NDJSON stream via ReadableStream reader, accumulating results for the export file. Ref: open-webui#22206 * refac * refac * refac * refac * refac * refac * Enhance CommitSessionMiddleware to allow health probes to bypass session management, ensuring faster and more reliable responses. (#24384) * refac * refac * refac * refac * refac * refac * refac: apply DOMPurify to excel and office HTML render assignments (#24468) * I18n/improve chinese translation (#24194) * i18n: improve zh-CN translation * i18n: improve zh-TW translation * perf(prompts): filter prompt list in SQL instead of N+1 has_access loop (#24288) get_prompts_by_user_id used to fetch every active prompt (with users + all access grants), then call AccessGrants.has_access() once per prompt that the user did not own. With 600+ prompts this issued ~600 extra round-trips per request and explained the multi-second delay reported in the GET /api/v1/prompts and /api/v1/prompts/tags endpoints for non-admin users. Push the access check into a single SQL query via the existing AccessGrants.has_permission_filter (EXISTS subquery), so only accessible rows come back from the DB. Users and access grants for the surviving rows are still batch-fetched, no N+1 anywhere on this path. Co-authored-by: Claude <noreply@anthropic.com> * refac * refac * Update catalan translation.json (#24174) * perf(prompts): make /tags fetch only the tags column with SQL access filter (#24287) Non-admin GET /api/v1/prompts/tags went through get_prompts_by_user_id, which loaded every active prompt with its full content/data/meta plus owner records and all access grants, then ran one has_access query per prompt that wasn't owned by the caller - all so the endpoint could collapse the result to a sorted tag list. With 600 prompts this took several seconds while the admin path (a single SELECT) returned in <1s. Add Prompts.get_tags_by_user_id which selects only the tags column and applies the same EXISTS-based access filter used by /list. Also tighten the admin get_tags to project just the tags column instead of full rows. The endpoint is now one DB query (plus one for groups), no row hydration, no N+1. Co-authored-by: Claude <noreply@anthropic.com> * refac * refac * i18n: Add Tagalog (Filipino) translation (#24254) Signed-off-by: Adam Tao <tcx4c70@gmail.com> Co-authored-by: Tim Baek <tim@openwebui.com> Co-authored-by: joaoback <156559121+joaoback@users.noreply.github.com> Co-authored-by: Algorithm5838 <108630393+Algorithm5838@users.noreply.github.com> Co-authored-by: Kylapaallikko <Kylapaallikko@users.noreply.github.com> Co-authored-by: Teay <pythontogoplease@gmail.com> Co-authored-by: tcx4c70 <tcx4c70@gmail.com> Co-authored-by: goodbey857 <76645482+goodbey857@users.noreply.github.com> Co-authored-by: Jacob Leksan <63938553+jmleksan@users.noreply.github.com> Co-authored-by: RomualdYT <romuald@gameurnews.fr> Co-authored-by: Lucas <lucas@vanosenbruggen.com> Co-authored-by: Classic298 <27028174+Classic298@users.noreply.github.com> Co-authored-by: Constantine <Runixer@gmail.com> * refac * refac * refac * style(env): satisfy ruff (datetime alias, line length, identity check) (#24118) * Korean Translation Update (#24087) Signed-off-by: Adam Tao <tcx4c70@gmail.com> Co-authored-by: Tim Baek <tim@openwebui.com> Co-authored-by: joaoback <156559121+joaoback@users.noreply.github.com> Co-authored-by: Algorithm5838 <108630393+Algorithm5838@users.noreply.github.com> Co-authored-by: Kylapaallikko <Kylapaallikko@users.noreply.github.com> Co-authored-by: Teay <pythontogoplease@gmail.com> Co-authored-by: tcx4c70 <tcx4c70@gmail.com> Co-authored-by: goodbey857 <76645482+goodbey857@users.noreply.github.com> Co-authored-by: Jacob Leksan <63938553+jmleksan@users.noreply.github.com> Co-authored-by: RomualdYT <romuald@gameurnews.fr> Co-authored-by: Lucas <lucas@vanosenbruggen.com> Co-authored-by: Classic298 <27028174+Classic298@users.noreply.github.com> Co-authored-by: Constantine <Runixer@gmail.com> * refac * refac * refac * refac * refac * refac * refac * refac * refac * feat: brave search llm context * refac * refac * refac * refac * refac * refac * refac * fix(mcp): remove asyncio.wait_for/shield from MCP cleanup in chat handler (#24105) asyncio.wait_for() and asyncio.shield() create new asyncio Tasks which violate anyio cancel-scope task-ownership rules. The MCPClient's exit_stack contains anyio resources (streamable_http transport) that use anyio cancel scopes. When exited from a different task, anyio raises 'Attempted to exit a cancel scope that isn't the current task's current cancel scope' as a BaseException. This BaseException propagates through the finally block, discards the completed response return value, and surfaces as a 500 Internal Server Error / 'No response returned.' - silently swallowing successful MCP tool calls and blocking the chat endpoint. Fix: call client.disconnect() directly in a simple loop. MCPClient.disconnect() already catches BaseException internally (see prior commit), so no wrapper is needed. Signed-off-by: Adam Tao <tcx4c70@gmail.com> Co-authored-by: Tim Baek <tim@openwebui.com> Co-authored-by: joaoback <156559121+joaoback@users.noreply.github.com> Co-authored-by: Algorithm5838 <108630393+Algorithm5838@users.noreply.github.com> Co-authored-by: Kylapaallikko <Kylapaallikko@users.noreply.github.com> Co-authored-by: Teay <pythontogoplease@gmail.com> Co-authored-by: tcx4c70 <tcx4c70@gmail.com> Co-authored-by: goodbey857 <76645482+goodbey857@users.noreply.github.com> Co-authored-by: Jacob Leksan <63938553+jmleksan@users.noreply.github.com> Co-authored-by: RomualdYT <romuald@gameurnews.fr> Co-authored-by: Lucas <lucas@vanosenbruggen.com> Co-authored-by: Classic298 <27028174+Classic298@users.noreply.github.com> Co-authored-by: Constantine <Runixer@gmail.com> Co-authored-by: Circe (Claude Code Sonnet 4.6) <circe@athena-council.org> Co-authored-by: Claude <noreply@anthropic.com> * fix:image url validation and signout post (#24420) * refac(routers): reject external URLs in profile/model image handlers * refac(ui): centralize image URL validation in safeImageUrl helper * refac(auths): make signout POST-only * refac: gate external profile image redirect behind ENABLE_PROFILE_IMAGE_URL_FORWARDING Restore the 302 redirect for external http(s) profile image URLs in the user and model profile-image endpoints, but gate it behind a new ENABLE_PROFILE_IMAGE_URL_FORWARDING env flag (default: True). Existing deployments that rely on external profile image forwarding continue to work unchanged. Operators who want to suppress the redirect (to prevent client-side IP/UA/Referer leaks) can set the flag to False. * refac * refac * refac * refac * refac * refac * refac * refac * refac * chore: format * chore: changelog (#24358) * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * changelog * doc: changelog * refac * refac * refac * chore: format * refac --------- Signed-off-by: Adam Tao <tcx4c70@gmail.com> Co-authored-by: Classic298 <27028174+Classic298@users.noreply.github.com> Co-authored-by: Jacob Leksan <63938553+jmleksan@users.noreply.github.com> Co-authored-by: Athanasios Oikonomou <athoik@gmail.com> Co-authored-by: Shirasawa <764798966@qq.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Aleix Dorca <aleixdorca@mac.com> Co-authored-by: Vincent Agra <agravj007@gmail.com> Co-authored-by: joaoback <156559121+joaoback@users.noreply.github.com> Co-authored-by: Algorithm5838 <108630393+Algorithm5838@users.noreply.github.com> Co-authored-by: Kylapaallikko <Kylapaallikko@users.noreply.github.com> Co-authored-by: Teay <pythontogoplease@gmail.com> Co-authored-by: tcx4c70 <tcx4c70@gmail.com> Co-authored-by: goodbey857 <76645482+goodbey857@users.noreply.github.com> Co-authored-by: RomualdYT <romuald@gameurnews.fr> Co-authored-by: Lucas <lucas@vanosenbruggen.com> Co-authored-by: Constantine <Runixer@gmail.com> Co-authored-by: Shamil <ashm.tech@proton.me> Co-authored-by: Cyp <cypher9715@naver.com> Co-authored-by: looselyhuman <fieldian@gmail.com> Co-authored-by: Circe (Claude Code Sonnet 4.6) <circe@athena-council.org>
refac
2026-05-09 03:17:07 -04:00 · 2026-05-09 16:11:19 +09:00 · 2026-05-09 16:08:22 +09:00 · 2026-05-09 16:06:09 +09:00 · 2026-05-09 16:05:21 +09:00 · 2026-05-09 15:46:33 +09:00
200 changed files with 12546 additions and 4937 deletions
@@ -5,6 +5,156 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

+## [0.9.3] - 2026-05-09
+
+### Added
+
+- 🔇 **Voice Mode mute control.** Voice Mode now includes a dedicated mute toggle with an "M" shortcut and auto-unmute after assistant playback, so you can prevent accidental interruptions from background noise without leaving the call overlay. [Commit](https://github.com/open-webui/open-webui/commit/072d2000f35a9f7b96342fa9bb28f925a92e7b4c), [#23832](https://github.com/open-webui/open-webui/issues/23832)
+- 🚀 **Faster prompt list loading.** Prompt and prompt-tag pages now load much faster for non-admin users, even with large prompt libraries, because accessible prompts are filtered efficiently in a single database query. [#24288](https://github.com/open-webui/open-webui/pull/24288), [#24258](https://github.com/open-webui/open-webui/discussions/24258)
+- ⚡ **Faster chat history loading.** Chat history maps now load from normalized message records when available, reducing overhead for large conversations while preserving fallback behavior for legacy chats. [Commit](https://github.com/open-webui/open-webui/commit/485d689cfd1ef8b9e7f77cd7b535b8b8747dff1f), [#23159](https://github.com/open-webui/open-webui/pull/23159)
+- 🗑️ **Delete from conversation menu.** You can now delete the current conversation directly from the chat menu with a confirmation step, so cleanup is faster without searching through the full chat list. [Commit](https://github.com/open-webui/open-webui/commit/ef6d4f2d6c4b79c7e12e864a4fcb6a57ee84e5d4), [#24329](https://github.com/open-webui/open-webui/issues/24329)
+- ⬆️ **Scroll to Top shortcut.** Long conversations now include a Scroll to Top action in the chat menu when you are away from the top, making it much faster to jump back to the beginning of a chat. [Commit](https://github.com/open-webui/open-webui/commit/cdfcbc4af6e9aec835b88dc1806a2a46711e6947), [#24133](https://github.com/open-webui/open-webui/issues/24133)
+- 📅 **Calendar creation flow.** Users can now create calendars from a dedicated modal and a quick-add action in the calendar sidebar, making calendar setup faster from the calendar workspace. [Commit](https://github.com/open-webui/open-webui/commit/34146ab60f5dc1a2f8bdda8e61ce02797233a25d), [Commit](https://github.com/open-webui/open-webui/commit/1baf73bdd56f4e5ded12a4bd3c168f4d2a70b840)
+- 🧭 **Unified model unload controls.** Administrators can now unload running models from the model selector across supported providers, with loaded-state indicators shown for Ollama and llama.cpp models. [Commit](https://github.com/open-webui/open-webui/commit/4fe2de78643c2213652190d2820f4e8d9f4f89cc)
+- ⚡ **Health check responsiveness.** Health and readiness probes now avoid blocking database calls and skip sync session commit handling on probe paths, improving responsiveness and reducing false unready transitions during database pressure. [#24380](https://github.com/open-webui/open-webui/pull/24380), [#24384](https://github.com/open-webui/open-webui/pull/24384)
+- 🎛️ **Playground controls panel.** The Playground now includes a dedicated Controls toggle so you can adjust parameters like temperature and related settings per chat run without changing model-level defaults. [Commit](https://github.com/open-webui/open-webui/commit/c6763521c00f042a28829e33fb6f1b7355054046), [#24103](https://github.com/open-webui/open-webui/issues/24103)
+- 🎙️ **STT file extension controls.** Administrators can now configure which audio file extensions are accepted for speech-to-text uploads, helping enforce safer and more predictable upload policies. [Commit](https://github.com/open-webui/open-webui/commit/4754ece4a2de5bba85a1d53af2dc8d24fdfb58be)
+- 📷 **Remembered call camera selection.** Voice call overlay now remembers your last selected camera and restores it automatically when available, so you do not need to reselect it every time you start voice mode. [Commit](https://github.com/open-webui/open-webui/commit/5c3edc2539ac4d92c4cc2d37079549995203238a), [#24416](https://github.com/open-webui/open-webui/issues/24416)
+- 👥 **User group prompt variable.** System and template prompts now support the "{{USER_GROUPS}}" variable, which expands to the user’s group memberships so prompts can adapt to role- or access-based context automatically. [Commit](https://github.com/open-webui/open-webui/commit/c1202a23277abb8e7080271a929dcc9d29b67e66), [#24462](https://github.com/open-webui/open-webui/issues/24462)
+- 🔐 **Public chat sharing permission control.** Administrators can now control whether users are allowed to create publicly shareable chats through a dedicated permission setting. [Commit](https://github.com/open-webui/open-webui/commit/ef6d4f2d6c4b79c7e12e864a4fcb6a57ee84e5d4)
+- 🔐 **Profile image forwarding control.** Administrators can now disable external profile image URL forwarding with the "ENABLE_PROFILE_IMAGE_URL_FORWARDING" setting to prevent browser metadata leaks to third-party servers. [#24420](https://github.com/open-webui/open-webui/pull/24420)
+- 🏷️ **Dynamic header template variables.** Administrators can now use chat, message, and user template variables in custom connection and tool server headers so each request can carry per-conversation context automatically. [Commit](https://github.com/open-webui/open-webui/commit/9907c0a25ae830d134af70022238715f834d20c6), [#24164](https://github.com/open-webui/open-webui/pull/24164)
+- 🛂 **MCP OAuth server URL setting.** Static OAuth tool server setups can now define a separate OAuth server URL, making discovery and client registration work when authentication endpoints are hosted separately from the tool server URL. [Commit](https://github.com/open-webui/open-webui/commit/9907c0a25ae830d134af70022238715f834d20c6), [#24164](https://github.com/open-webui/open-webui/pull/24164), [#24216](https://github.com/open-webui/open-webui/issues/24216)
+- ⚡ **Faster memory query performance.** Per-user memory lookups and deletions now run much faster at scale because the memory user filter is indexed for existing and new installations. [Commit](https://github.com/open-webui/open-webui/commit/38a382ef888685650135d61dcc8ec0e29eb65573), [#23836](https://github.com/open-webui/open-webui/pull/23836)
+- 🚀 **Smarter function dependency installs.** Function dependencies are now skipped when they were already preinstalled and unchanged, reducing first-load delays and repeated package installation churn after startup. [Commit](https://github.com/open-webui/open-webui/commit/ae43562b869b24699408e5ab107261a0a8bdb4bc), [#24166](https://github.com/open-webui/open-webui/pull/24166)
+- 🔎 **Brave LLM Context web search.** Administrators can now choose Brave LLM Context as a web search provider to retrieve richer grounded passages with a configurable context token budget. [Commit](https://github.com/open-webui/open-webui/commit/6700f7bb72d14a3f8dbb72dfa064cae3b3dc29ac), [#24120](https://github.com/open-webui/open-webui/issues/24120)
+- 🗂️ **Open Terminal date sorting.** Open Terminal now includes sort controls for name and date, with directory-first ordering and modified-time visibility to make file browsing faster. [Commit](https://github.com/open-webui/open-webui/commit/6bdc2ffa79d72daf78981209c9c5292c697cbfe5), [#24425](https://github.com/open-webui/open-webui/issues/24425)
+- 🎤 **Voice mode prompt toggle.** Administrators can now explicitly enable or disable the Voice Mode custom prompt behavior from Interface settings, giving finer control over how voice replies are guided. [Commit](https://github.com/open-webui/open-webui/commit/17893038869e3a763a8b34457f723b9666804e27)
+- 🧮 **LaTeX copy shortcut.** You can now click rendered LaTeX expressions to copy the raw formula to your clipboard, making it easier to reuse equations outside chat. [Commit](https://github.com/open-webui/open-webui/commit/064fdecb675c176a04b024c16ce179f4dda45236), [#24244](https://github.com/open-webui/open-webui/pull/24244)
+- ✨ **Smoother rich text editing.** The message composer now defers formatting toolbar refresh work to the next animation frame, reducing typing jank while formatting controls stay accurate. [Commit](https://github.com/open-webui/open-webui/commit/794b97025d4c56f91d49c9d1ec4775d2ea07b53a), [#24013](https://github.com/open-webui/open-webui/pull/24013)
+- 🖼️ **Arena model profile images.** Arena models can now reliably display configured profile images instead of falling back to the default icon. [Commit](https://github.com/open-webui/open-webui/commit/1dee67b64d0b34e70bac949682b216c0aaec8152), [#24412](https://github.com/open-webui/open-webui/issues/24412)
+- 🔄 **Replaceable tool embed updates.** Pipes and Tools can now overwrite previously emitted rich-UI embeds in-place by passing a `replace` flag on the `embeds` event, enabling live dashboards and progress panels that update without stacking duplicate entries.
+- ✏️ **Assistant response editing and continuation.** You can now edit and restructure assistant output items — including reasoning blocks, tool calls, and text content — from a dedicated editor view, and continue generating from the edited state so the model receives full prior context.
+- 🔄 **General improvements.** Various improvements were implemented across the application to enhance performance, stability, and security.
+- 🌐 **Translation updates.** Translations for Chinese, Catalan, Filipino, and Korean were enhanced and expanded.
+
+### Fixed
+
+- 🧵 **Background code execution reliability.** Code execution no longer hangs indefinitely when you switch conversations or browser tabs during a run, and disconnected or inactive sessions now fail with a clear timeout error instead of endless processing. [Commit](https://github.com/open-webui/open-webui/commit/552bbcecfae5ae273ab98e2ce3e540d0771aa964), [#24089](https://github.com/open-webui/open-webui/issues/24089)
+- 🎙️ **Voice recording MIME fallback support.** Voice recording now tries a broader set of browser-supported audio formats and resets halted audio playback cleanly, improving microphone capture reliability in browsers where recording previously failed to start. [Commit](https://github.com/open-webui/open-webui/commit/8ffc3d746f20007e9eb4e3ae4f152f383bc371e1), [#24162](https://github.com/open-webui/open-webui/issues/24162)
+- 🧠 **Direct-connection task generation reliability.** Title, tags, follow-up, emoji, query, and related task-generation endpoints now work correctly when chats use direct-connection models instead of failing with model-not-found errors. [Commit](https://github.com/open-webui/open-webui/commit/1b4cd705d0b9a51a5e3a7851ec012fb3141eb0a9), [Commit](https://github.com/open-webui/open-webui/commit/005df577fec16733a64edbec8a1b46f42f4e9a43), [#24092](https://github.com/open-webui/open-webui/issues/24092)
+- 🔧 **Parameterized URL tool readiness.** New chats now wait for model defaults to finish applying before auto-submit, preventing early requests that can miss configured external tools. [Commit](https://github.com/open-webui/open-webui/commit/212bb68a66435dc1803a6d67cb4ea584d3455fb7), [#24176](https://github.com/open-webui/open-webui/issues/24176)
+- 🚦 **MCP cleanup response reliability.** Successful native MCP tool calls no longer get replaced by a 500 "No response returned" error during cleanup, so valid chat responses are now returned consistently. [#24105](https://github.com/open-webui/open-webui/pull/24105)
+- 🧵 **Active task state recovery.** Chat input no longer stays blocked by unrelated background tasks after a response is already complete, and interrupted assistant replies are now marked done more reliably. [Commit](https://github.com/open-webui/open-webui/commit/04bd0425ead28185bcd124e77892e31209a6e15b), [#23264](https://github.com/open-webui/open-webui/pull/23264)
+- 📌 **Per-user note pinning behavior.** Pinned notes are now tracked per user instead of with a shared note-level flag, so one person’s pin changes no longer affect everyone else. [Commit](https://github.com/open-webui/open-webui/commit/33e588cf09b294f0abe08b9566efa8545a7dbf92)
+- 🧱 **Custom header value coercion.** Custom header values are now converted to text before requests are sent, preventing request failures when non-text values are configured. [Commit](https://github.com/open-webui/open-webui/commit/9907c0a25ae830d134af70022238715f834d20c6), [#24164](https://github.com/open-webui/open-webui/pull/24164)
+- 🔗 **HTTP share link copy fallback.** Copy Link now works reliably on HTTP deployments by using a selection-based fallback when secure clipboard APIs are unavailable. [Commit](https://github.com/open-webui/open-webui/commit/f70b0da1563ffa0a8daecbe71cbc30fd8cf834c4), [#24135](https://github.com/open-webui/open-webui/issues/24135)
+- 🧵 **Regeneration loading lock recovery.** Chats no longer get stuck in a permanent loading state after failed regenerations because invalid message-tree references are repaired before rendering. [Commit](https://github.com/open-webui/open-webui/commit/ee3b82926b37843f2771c6a8d432781a557ea96a), [#24424](https://github.com/open-webui/open-webui/issues/24424)
+- 📸 **Complete chat image capture.** Downloaded chat snapshots now include all messages more reliably through visibility overrides and layout timing improvements during capture. [Commit](https://github.com/open-webui/open-webui/commit/34146ab60f5dc1a2f8bdda8e61ce02797233a25d), [Commit](https://github.com/open-webui/open-webui/commit/1baf73bdd56f4e5ded12a4bd3c168f4d2a70b840), [#24088](https://github.com/open-webui/open-webui/issues/24088)
+- 🗓️ **Calendar deletion lock handling.** Calendar deletion now avoids SQLite write-lock contention by revoking calendar access grants in a separate transaction after calendar and event removal. [Commit](https://github.com/open-webui/open-webui/commit/1d892ce2c513c4d933c902de5e5d76c317a06dd2)
+- 🧩 **Filter and internal tool coexistence.** Internal tools now remain available when filters add provider-native tools, so filter-added tools no longer replace the built-in tool set during request processing. [Commit](https://github.com/open-webui/open-webui/commit/02f9fe78907c2ecf6f1d93646cbfa2173409bbe8), [#24237](https://github.com/open-webui/open-webui/issues/24237)
+- 🛠️ **OpenAPI tool spec compatibility.** OpenAPI tool integrations now handle null or non-operation path entries more safely and parse path-level parameters consistently, preventing crashes and improving tool execution reliability across imperfect OpenAPI specs. [Commit](https://github.com/open-webui/open-webui/commit/2ba6b423aa0c9c800bd96cb638c6ade867cac0f6), [Commit](https://github.com/open-webui/open-webui/commit/5b80932e5951786bb348b91589e8d87753f18905), [#24376](https://github.com/open-webui/open-webui/pull/24376)
+- 🧰 **OpenAPI tool schema parsing.** OpenAPI tool imports now ignore non-method path item fields and correctly resolve nested composition schemas, preventing invalid tool parsing for compatible specs. [Commit](https://github.com/open-webui/open-webui/commit/85c7373f68ac3e39a9cd37e63b6926b13fb8b8cc), [#23254](https://github.com/open-webui/open-webui/pull/23254)
+- 🌍 **Web search proxy compatibility.** DuckDuckGo search now respects configured proxy environments more reliably, and trust-env behavior defaults to enabled so proxied web loading does not fail unexpectedly. [Commit](https://github.com/open-webui/open-webui/commit/bb0e6cb1085aa3c3da66a5f5ea1cecff7e9b5297), [#23810](https://github.com/open-webui/open-webui/pull/23810)
+- 🧾 **Final markdown render flush.** Streaming markdown now forces an immediate final parse when generation completes, preventing stale or partially rendered final output. [Commit](https://github.com/open-webui/open-webui/commit/29f6c72e879d67f23021938e24a21914cc9fb120), [#24088](https://github.com/open-webui/open-webui/issues/24088)
+- 🛡️ **Webhook avatar URL validation.** Channel webhook profile image URLs are now validated before saving, preventing invalid or unsafe avatar URLs from being accepted. [#24370](https://github.com/open-webui/open-webui/pull/24370)
+- 📝 **System prompt editor scroll stability.** Editing large system prompts no longer jumps the page back to the top, so you can continue editing long model prompts without losing your place. [Commit](https://github.com/open-webui/open-webui/commit/c978a788c8315e37357c93c1b605a2831fc77485), [#23999](https://github.com/open-webui/open-webui/issues/23999)
+- 🔎 **Knowledge content search matching.** Knowledge file search now matches both file titles and file content, so relevant files are easier to find even when the keyword is not in the filename. [Commit](https://github.com/open-webui/open-webui/commit/11e076817ae5db34621ce03136353248f7377d97), [#24297](https://github.com/open-webui/open-webui/pull/24297)
+- ⚡ **Faster prompt tag loading.** Prompt tag filters now load much faster for non-admin users by fetching only accessible tags directly, avoiding per-prompt permission checks and unnecessary prompt data loading. [#24287](https://github.com/open-webui/open-webui/pull/24287), [#24258](https://github.com/open-webui/open-webui/discussions/24258)
+- 🧾 **Citation overflow badge readability.** Citation overflow badges now keep multi-digit counts readable in a single compact bubble, preventing wrapped or cramped display when many sources are attached. [Commit](https://github.com/open-webui/open-webui/commit/23ff9943a9fc8c314100fa074157853fbece1a55), [#24391](https://github.com/open-webui/open-webui/pull/24391)
+- 🌐 **Yandex result parsing guard.** Yandex web search no longer fails when some XML fields are missing in individual results, so valid search responses continue to return usable sources instead of dropping to no results. [Commit](https://github.com/open-webui/open-webui/commit/9386fc83a3eff3e55cc157ac8c15c337e3d822c1), [#24243](https://github.com/open-webui/open-webui/issues/24243)
+- 🎧 **Safer voice transcription uploads.** Empty or failed voice conversions are now rejected with a clear error instead of continuing as malformed audio, reducing failed transcription attempts from corrupted or near-empty recordings. [Commit](https://github.com/open-webui/open-webui/commit/072d2000f35a9f7b96342fa9bb28f925a92e7b4c)
+- 🎚️ **Safer chunked STT processing.** Chunked transcription now limits worker concurrency when no external STT engine is configured, reducing failed transcription behavior caused by overly parallel local processing. [Commit](https://github.com/open-webui/open-webui/commit/55e7c7854bba5182803239c903a0ac2d14426a4c)
+- 📈 **Imported chat analytics coverage.** Imported ChatGPT conversations now carry proper model and timestamp metadata and reliably write imported messages into analytics-backed storage, so imported chats are reflected correctly in Admin Analytics totals and model usage views. [Commit](https://github.com/open-webui/open-webui/commit/4d766a3edfa116abcefe7168f1d1284683b860b2), [#24263](https://github.com/open-webui/open-webui/issues/24263)
+- 📎 **Knowledge collection persistence.** Knowledge collections selected with the chat input selector now remain attached after reloads and chat switches, so attached context no longer disappears between sessions. [Commit](https://github.com/open-webui/open-webui/commit/7c398a625a8d51f79d80217f6d329fc30c72b782), [#24142](https://github.com/open-webui/open-webui/issues/24142)
+- 🧹 **Embedding model name trimming.** Embedding model names entered in Documents settings now automatically trim surrounding whitespace, preventing silent embedding failures caused by accidental trailing spaces. [Commit](https://github.com/open-webui/open-webui/commit/6082e1adaebc8aa3e7f55265c8dc2dbe130c0446), [#24090](https://github.com/open-webui/open-webui/issues/24090)
+- 🔊 **PCM TTS playback compatibility.** Text-to-speech audio returned as PCM is now converted to MP3 before delivery, so speech playback works correctly with providers that return raw PCM audio. [Commit](https://github.com/open-webui/open-webui/commit/ff791b4814fc1453df2235ea78016d7015aa6806), [#24143](https://github.com/open-webui/open-webui/issues/24143)
+- 🪟 **Windows PostgreSQL startup compatibility.** Windows pip installs using PostgreSQL now start reliably with psycopg async by using a compatible event loop policy instead of the default Proactor loop. [Commit](https://github.com/open-webui/open-webui/commit/7eaecbad5a0913ed04ca3bc10c930bb051dd2bd9), [#24152](https://github.com/open-webui/open-webui/issues/24152)
+- ⏱️ **MCP OAuth timeout control.** OAuth token exchanges for MCP tool server connections now respect the configurable client timeout setting, reducing callback failures with slower providers. [Commit](https://github.com/open-webui/open-webui/commit/cde72dab71671645e119564ca9747ce25dd590ad), [#24138](https://github.com/open-webui/open-webui/issues/24138)
+- 📄 **PDF text search restoration.** PDF previews now include a proper text layer so browser text selection and find-in-page search work again instead of rendering only image-like pages. [Commit](https://github.com/open-webui/open-webui/commit/bc4d6eef33dcb92719b07483cdb1d63ebf250721), [#24149](https://github.com/open-webui/open-webui/issues/24149)
+- 🔑 **Android password autofill support.** Password inputs now expose the expected field name metadata, improving password manager autofill reliability on Android login pages. [Commit](https://github.com/open-webui/open-webui/commit/60ea4214aa42f1ad22142f1a43535007a2293d16), [#24137](https://github.com/open-webui/open-webui/issues/24137)
+- 🎤 **Non-blocking STT processing.** Speech-to-text transcription no longer blocks the server event loop during both live transcription and uploaded audio file processing, so other users can continue using chats and live connections under concurrent load. [#24338](https://github.com/open-webui/open-webui/pull/24338), [#24379](https://github.com/open-webui/open-webui/pull/24379), [#24169](https://github.com/open-webui/open-webui/issues/24169)
+- 🌐 **SearXNG language parameter handling.** Web searches now send clean multi-language values without trailing separators, so SearXNG requests no longer fail when multiple languages are selected. [Commit](https://github.com/open-webui/open-webui/commit/6dff85b9d205cfc4bc2845dac40909b8d859910c), [#24198](https://github.com/open-webui/open-webui/issues/24198)
+- 📂 **File modal open-link behavior.** Clicking a file name in the file details modal now opens the correct file content in a new tab for uploaded file items instead of failing to open. [#24125](https://github.com/open-webui/open-webui/pull/24125)
+- 📎 **Chat attachment display recovery.** Files attached by chat tools now appear reliably in assistant responses, including non-image file attachments that were previously hidden. [Commit](https://github.com/open-webui/open-webui/commit/7eeff2fdf945024585a01b72071a61971afc844d), [#24332](https://github.com/open-webui/open-webui/pull/24332)
+- 🧱 **Channel embed rendering guard.** Channel message embeds now appear only for model-generated messages and are suppressed in reply previews, preventing unintended embed expansion in regular user posts. [Commit](https://github.com/open-webui/open-webui/commit/e1dce9914745de9b4d2c67b1deddde3472ce4dfa)
+- 🛡️ **Safer image URL handling.** Untrusted external image URLs are now blocked in profile and rich-text image rendering paths, preventing unintended client-side requests to attacker-controlled domains. [#24420](https://github.com/open-webui/open-webui/pull/24420)
+- 🛡️ **Sanitized spreadsheet HTML previews.** Spreadsheet previews now sanitize generated HTML before rendering, reducing the risk of unsafe content being executed when opening office files in chat and file modals. [#24468](https://github.com/open-webui/open-webui/pull/24468)
+- 🧰 **Multi-worker tool update consistency.** Updated tool code now refreshes correctly across workers without requiring a full service restart, so chats no longer run stale tool versions after edits. [Commit](https://github.com/open-webui/open-webui/commit/3309f5d9f11f521c0ee97b64c59a83e3cf390bde), [#24400](https://github.com/open-webui/open-webui/issues/24400), [#24433](https://github.com/open-webui/open-webui/pull/24433)
+- 🧩 **Default model metadata env parsing.** The "DEFAULT_MODEL_METADATA" environment setting is now parsed and applied correctly, including when persistent config is disabled, so configured model capability defaults are no longer ignored at startup. [Commit](https://github.com/open-webui/open-webui/commit/0103d7e82cccbd5c4b1c8daabcb3e5160fa74a97), [#24319](https://github.com/open-webui/open-webui/issues/24319)
+- 🔄 **Config import and Redis consistency.** Imported settings now remain effective after import because configuration values are immediately synchronized to Redis, preventing stale cached values from overriding imported permissions and settings. [Commit](https://github.com/open-webui/open-webui/commit/55a572cd398c9b4e6118728f8f129941437aa225), [Commit](https://github.com/open-webui/open-webui/commit/1c1c8b18e5cc90ca3c6961a4c193a4363febbc83), [#24346](https://github.com/open-webui/open-webui/issues/24346)
+- 🔔 **LDAP signup webhook parity.** New accounts created through LDAP now trigger the same signup webhook notifications as password and OAuth signups, so downstream provisioning and audit automations receive consistent events. [Commit](https://github.com/open-webui/open-webui/commit/fd3368c0bff168417e3c49ffd73491c344702339), [#24377](https://github.com/open-webui/open-webui/issues/24377)
+- 🦆 **DDGS auto-backend compatibility.** Web search now handles DDGS automatic backend selection correctly and safely falls back on empty or rate-limited responses, preventing search failures in newer DDGS versions. [Commit](https://github.com/open-webui/open-webui/commit/9adc0c442a57eaa88a5f30c2b2cb393623154e20), [#24188](https://github.com/open-webui/open-webui/issues/24188)
+- 🤖 **Automation update tool reliability.** Updating existing automations in chat now works correctly instead of failing with a missing method error. [Commit](https://github.com/open-webui/open-webui/commit/f39f4a86aedc2769d8268670a020b1f3c16776dd), [#24405](https://github.com/open-webui/open-webui/issues/24405#issuecomment-4408011166)
+- 📅 **Calendar event permission checks.** Calendar event update and delete actions now handle ownership and access checks more reliably, returning clean access-denied results when appropriate. [Commit](https://github.com/open-webui/open-webui/commit/2977910ffd9d2369dfa504aa6ab12745b3dbd19a)
+- 🛡️ **Safer cached file delivery.** Cached files that are not recognized as image, audio, or video now download as attachments instead of rendering inline, reducing the risk of unsafe browser content handling. [Commit](https://github.com/open-webui/open-webui/commit/4754ece4a2de5bba85a1d53af2dc8d24fdfb58be)
+- 📊 **Streaming token analytics accuracy.** Admin Analytics now records and aggregates token usage correctly for streaming chats across Responses API and OpenAI-compatible providers, including fallback handling for provider usage formats that use prompt and completion token keys. [Commit](https://github.com/open-webui/open-webui/commit/989d5fd4e2ce285edf4475a1e13f0981a78d3821), [Commit](https://github.com/open-webui/open-webui/commit/a32d26e61d24d9f63650faed5cb8909ed90af661), [#24217](https://github.com/open-webui/open-webui/issues/24217), [#24294](https://github.com/open-webui/open-webui/issues/24294), [#24241](https://github.com/open-webui/open-webui/issues/24241)
+- 🔗 **Admin shared chat links.** Admin users can now open and clone shared chat links reliably without 401 errors because shared links are now resolved by share ID first, with safe fallback behavior for direct chat ID access. [Commit](https://github.com/open-webui/open-webui/commit/cde21b9f6dc11575a668484f42440824ec5a4fae), [#24311](https://github.com/open-webui/open-webui/issues/24311), [#24096](https://github.com/open-webui/open-webui/issues/24096)
+- 💾 **Chat settings persistence.** System prompts and other chat-level settings now persist correctly after creating a new chat and reloading, preventing prompt loss in affected conversations. [Commit](https://github.com/open-webui/open-webui/commit/86df8bf27e1b84abbe2eeedcc8650df59c7d23d6), [#24193](https://github.com/open-webui/open-webui/issues/24193), [#24270](https://github.com/open-webui/open-webui/issues/24270)
+- 💾 **Chat control autosave persistence.** Changes to chat controls like system prompt, parameters, and attached files are now autosaved on existing chats, so edits are no longer lost when you refresh or navigate away before sending a message. [Commit](https://github.com/open-webui/open-webui/commit/a938c8ae2e45a00d2f06151fdaeaee94e54a8095), [#23897](https://github.com/open-webui/open-webui/pull/23897)
+- ☁️ **OneDrive option visibility.** OneDrive personal and business upload options now appear only when their respective client IDs are configured, preventing unavailable options from showing in attachment menus. [Commit](https://github.com/open-webui/open-webui/commit/b72019db393a658ca0ceecdcc59b70f6cc5dcd40), [#24411](https://github.com/open-webui/open-webui/issues/24411)
+- 🧠 **Reasoning content leakage prevention.** Tool-call round-trip messages no longer wrap reasoning text in `<think>` tags inside the content field, preventing raw markup from leaking into chat output for models whose templates don't strip think tags (e.g. Gemma 4). [#23844](https://github.com/open-webui/open-webui/issues/23844)
+- 🖥️ **Terminal sidebar auto-open guard.** The terminal sidebar no longer auto-opens on chat load when OpenTerminal is disabled, because stale terminal IDs saved on models or in localStorage are now validated against available terminal servers before use.
+- 🔁 **Single-confirmation connection deletion.** Deleting OpenAI, Ollama, tool server, and terminal server connections now shows exactly one confirmation dialog instead of two, because redundant outer confirmation wrappers were removed from all connection components.
+- 🧵 **Reliable background task cleanup.** The chat task lifecycle now deregisters completed tasks before checking for remaining siblings, eliminating the off-by-one timing issue that could leave the stop button stuck or dismiss the sidebar activity spinner too early.
+
+### Changed
+
+- ⚠️ **Database Migrations**: This release includes database schema changes; we strongly recommend backing up your database and all associated data before upgrading in production environments. If you are running a multi-worker, multi-server, or load-balanced deployment, all instances must be updated simultaneously, rolling updates are not supported and will cause application failures due to schema incompatibility.
+- 🚪 **Signout request method.** The signout endpoint now requires POST instead of GET, so custom clients and integrations must update logout calls accordingly. [#24420](https://github.com/open-webui/open-webui/pull/24420)
+
+## [0.9.2] - 2026-04-24
+
+### Added
+
+- 🧠 **PaddleOCR-vl document extraction.** Administrators can now use PaddleOCR-vl as a content extraction engine for document processing, with configurable API URL and token settings in document retrieval configuration. [#23945](https://github.com/open-webui/open-webui/pull/23945)
+- 🔥 **Firecrawl v2 API.** Firecrawl web loading now uses the v2 API directly with proper retry logic, exponential backoff on rate limits, and configurable timeout handling, improving reliability for both cloud and self-hosted Firecrawl setups. [#23934](https://github.com/open-webui/open-webui/pull/23934)
+- ⏰ **Calendar event reminder customization.** Calendar events now support a configurable `reminder_minutes` parameter, allowing models to set custom reminder durations instead of the default 10-minute notification.
+- 🔑 **Custom API key header.** Administrators can now configure a custom header name for API key authentication via the `CUSTOM_API_KEY_HEADER` environment variable, enabling compatibility with reverse proxies that use the `Authorization` header for their own authentication.
+- 🔌 **OAuth session disconnection.** Users can now disconnect OAuth sessions for specific providers (e.g., MCP connections) through a new API endpoint, enabling cleaner re-authentication workflows.
+- 📚 **Source overflow indicator.** The Sources button now shows a +N badge when more than three sources are available, so hidden sources are clearly indicated in chat responses. [#23918](https://github.com/open-webui/open-webui/pull/23918)
+- ⚡ **Model list performance.** Model list API responses now strip base64 profile image data from paginated results, and model tags are fetched via a dedicated efficient query instead of loading all models. This significantly reduces payload sizes and improves workspace Models page responsiveness.
+- ⚡ **Model avatar cache reuse.** Default model profile images now redirect to a shared static path instead of reading files from disk per-request, reducing repeated I/O and improving loading efficiency when multiple models use the fallback icon. [#24015](https://github.com/open-webui/open-webui/pull/24015)
+- 🚀 **Faster splash image loading.** Splash screen images are now prioritized earlier during page load with preload links, improving first-load LCP behavior and reducing delayed image discovery. [#24011](https://github.com/open-webui/open-webui/pull/24011)
+- 🧵 **Streaming markdown performance stability.** Streaming responses now stay more memory-efficient by preventing repeated cleanup callback registration during markdown updates. [#24048](https://github.com/open-webui/open-webui/pull/24048)
+- 📊 **Telemetry gauge reliability.** OpenTelemetry user gauge callbacks now use synchronous database queries directly, eliminating cross-thread async bridging issues that could cause silent failures in metric collection.
+- 🔄 **General improvements.** Various improvements were implemented across the application to enhance performance, stability, and security.
+- 🌐 **Translation updates.** Translations for Finnish, Korean, Portuguese (Brazil), and Dutch were enhanced and expanded.
+
+### Fixed
+
+- 🔧 **MCP task cancellation stability.** Interrupted MCP tool calls no longer cause CPU spikes or runaway cleanup behavior. MCP client disconnection now runs in the same asyncio task as connection, respecting cancel scope constraints, and chat-active events are properly shielded during cancellation.
+- 🧠 **Persistent chat skill injection.** Skills mentioned in persisted chats now inject into the system prompt reliably. Skill ID extraction from `<$skillId|label>` message tags is now handled server-side, and tags are stripped before messages reach the model.
+- 🗄️ **Async database driver migration.** The async database backend now uses psycopg (v3) instead of asyncpg, eliminating brittle SSL parameter translation and supporting native libpq connection strings including `sslmode`, `options`, and `target_session_attrs` without any stripping or conversion.
+- 🐳 **Docker ARM64 reliability.** Docker images built for arm64 via QEMU cross-compilation no longer produce 0-byte corrupted Python dependencies. `UV_LINK_MODE=copy` is now set in the Dockerfile to force reliable file installation.
+- 🛠️ **Throttle request handling.** Request handling no longer fails when user activity status updates are throttled with a non-zero interval. [#23979](https://github.com/open-webui/open-webui/pull/23979)
+- ✍️ **Rich text extension conflicts.** Rich text editing no longer triggers duplicate extension conflicts for lists and code blocks, improving editor stability. [#24009](https://github.com/open-webui/open-webui/pull/24009)
+- 🔇 **Fetch URL null content guard.** The `fetch_url` built-in tool now safely handles `None` content returned by web loaders instead of crashing with a `TypeError`.
+- 🌐 **OAuth discovery fallback.** OAuth protected resource discovery now falls back to well-known RFC 9728 URIs when the `WWW-Authenticate` header doesn't contain a `resource_metadata` link, improving compatibility with more MCP server implementations.
+- 🔐 **Session token resolution.** Session user endpoints now gracefully handle missing `Authorization` headers by falling back to cookie and request state tokens, preventing errors when used behind forward-auth proxies.
+- 🚫 **Direct API error responses.** Chat completion requests without a WebSocket channel (direct API calls) now return proper HTTP error responses instead of silently returning null on failure.
+- 📡 **Cancelled response stream cleanup.** Cancelled chat generation now explicitly closes the upstream response body iterator, preventing orphaned async generators from spinning in anyio internals.
+- 🔒 **Model profile image path safety.** Model profile image endpoints now validate and sanitize static asset redirect paths, preventing path traversal through encoded dots or malicious URL patterns.
+- 📊 **RAG template validation UI.** The Documents settings page now displays a warning when RAG templates contain multiple `[context]` or `{{CONTEXT}}` placeholders, helping administrators avoid accidental redundant context injection.
+- 🧩 **Automation model detection.** The `create_automation` tool now correctly detects the current model ID even when `model_id` is not yet set in metadata, falling back to the model dict.
+- 🔄 **MCP resource content handling.** MCP tool results with the `resource` content type are now correctly detected and their `resource.text` payload is extracted, instead of being silently ignored.
+- 🔄 **Ollama and OpenAI metadata forwarding.** Ollama and OpenAI proxy routes now forward request metadata to downstream handlers, ensuring consistent context propagation.
+- 🧹 **Browser-native message virtualization.** The custom JavaScript-based message culling system (spacers, height caching, scroll listeners) was replaced with CSS `content-visibility: auto`, letting the browser natively skip rendering of off-screen messages without destroying component trees. This eliminates scroll jump artifacts and mount/destroy thrashing while preserving memory efficiency in long conversations.
+- 📻 **Redis notification compatibility.** Redis pub/sub now handles missing or incompatible `client_name` support more gracefully, preventing connection errors with certain Redis configurations.
+
+### Changed
+
+- ⚙️ **psycopg v3 async driver.** The async database driver has been migrated from `asyncpg` to `psycopg` (v3). This is a transparent change for most deployments, but custom connection strings with `asyncpg`-specific parameters may need adjustment.
+- 🔑 **Brotli dependency update.** Brotli has been updated to address CVE-2025-6176.
+- 🖥️ **Windows startup script.** The Windows startup batch script has been updated for improved compatibility.
+
+## [0.9.1] - 2026-04-21
+
+### Fixed
+
+- 🐛 **Missing `aiosqlite` dependency.** Fixed a startup crash (`ModuleNotFoundError: No module named 'aiosqlite'`) when installing Open WebUI via `pip` or `uv` by adding the missing `aiosqlite` package to `pyproject.toml`. The dependency was listed in `requirements.txt` but not in the published package metadata, so it was not installed automatically. [#23916](https://github.com/open-webui/open-webui/issues/23916)
+- 🐛 **Missing `asyncpg` dependency.** Added the missing `asyncpg` package to `pyproject.toml` to prevent the same startup crash for PostgreSQL users. Like `aiosqlite`, it was present in `requirements.txt` but absent from the published package dependencies.
+
 ## [0.9.0] - 2026-04-20

 ### Added
@@ -135,6 +135,9 @@ RUN apt-get update && \
 # install python dependencies
 COPY --chown=$UID:$GID ./backend/requirements.txt ./requirements.txt

+# Set UV_LINK_MODE to copy to prevent 0-byte file corruption in QEMU arm64 cross-builds
+ENV UV_LINK_MODE=copy
+
 RUN set -e; \
    pip3 install --no-cache-dir uv; \
    if [ "$USE_CUDA" = "true" ]; then \
@@ -47,7 +47,7 @@ For more information, be sure to check out our [Open WebUI Documentation](https:

 - 💾 **Persistent Artifact Storage**: Built-in key-value storage API for artifacts, enabling features like journals, trackers, leaderboards, and collaborative tools with both personal and shared data scopes across sessions.

- 📚 **Local RAG Integration**: Dive into the future of chat interactions with groundbreaking Retrieval Augmented Generation (RAG) support using your choice of 9 vector databases and multiple content extraction engines (Tika, Docling, Document Intelligence, Mistral OCR, External loaders). Load documents directly into chat or add files to your document library, effortlessly accessing them using the `#` command before a query.
+- 📚 **Local RAG Integration**: Dive into the future of chat interactions with groundbreaking Retrieval Augmented Generation (RAG) support using your choice of 9 vector databases and multiple content extraction engines (Tika, Docling, Document Intelligence, Mistral OCR, PaddleOCR-vl, External loaders). Load documents directly into chat or add files to your document library, effortlessly accessing them using the `#` command before a query.

 - 🔍 **Web Search for RAG**: Perform web searches using 15+ providers including `SearXNG`, `Google PSE`, `Brave Search`, `Kagi`, `Mojeek`, `Tavily`, `Perplexity`, `serpstack`, `serper`, `Serply`, `DuckDuckGo`, `SearchApi`, `SerpApi`, `Bing`, `Jina`, `Exa`, `Sougou`, `Azure AI Search`, and `Ollama Cloud`, injecting results directly into your chat experience.

@@ -1,6 +1,7 @@
 import base64
 import os
 import random
+import sys
 from pathlib import Path
 from typing import Annotated

@@ -68,12 +69,18 @@ def serve(
    import open_webui.main  # noqa: F401
    from open_webui.env import UVICORN_WORKERS  # Import the workers setting

+    # On Windows, uvicorn's default loop factory hardcodes ProactorEventLoop,
+    # which is incompatible with psycopg v3 async.  Setting loop='none' lets
+    # asyncio.run() respect the WindowsSelectorEventLoopPolicy set in db.py.
+    loop = 'none' if sys.platform == 'win32' else 'auto'
+
    uvicorn.run(
        'open_webui.main:app',
        host=host,
        port=port,
        forwarded_allow_ips='*',
        workers=UVICORN_WORKERS,
+        loop=loop,
    )


@@ -328,6 +328,17 @@ class AppConfig:
        except Exception as e:
            log.error(f'Failed to async-persist config key {key}: {e}')

+    def _sync_to_redis(self):
+        """Push all in-memory config values to Redis, e.g. after a bulk import."""
+        if not self._redis or not ENABLE_PERSISTENT_CONFIG:
+            return
+        for key, pc in self._state.items():
+            redis_key = f'{self._redis_key_prefix}:config:{key}'
+            try:
+                self._redis.set(redis_key, json.dumps(pc.value))
+            except Exception as e:
+                log.error(f'Failed to sync config key {key} to Redis: {e}')
+
    def __getattr__(self, key):
        if key not in self._state:
            raise AttributeError(f"Config key '{key}' not found")
@@ -1197,6 +1208,12 @@ TOOL_SERVER_CONNECTIONS = PersistentConfig(
    tool_server_connections,
 )

+OAUTH_CLIENT_TIMEOUT = PersistentConfig(
+    'OAUTH_CLIENT_TIMEOUT',
+    'oauth.client.timeout',
+    os.environ.get('OAUTH_CLIENT_TIMEOUT', ''),
+)
+
 ####################################
 # TERMINAL_SERVER
 ####################################
@@ -1299,10 +1316,16 @@ MODEL_ORDER_LIST = PersistentConfig(
    [],
 )

+try:
+    default_model_metadata = json.loads(os.environ.get('DEFAULT_MODEL_METADATA', '{}'))
+except Exception as e:
+    log.exception(f'Error loading DEFAULT_MODEL_METADATA: {e}')
+    default_model_metadata = {}
+
 DEFAULT_MODEL_METADATA = PersistentConfig(
    'DEFAULT_MODEL_METADATA',
    'models.default_metadata',
-    {},
+    default_model_metadata,
 )

 try:
@@ -1477,6 +1500,10 @@ USER_PERMISSIONS_CHAT_EDIT = os.environ.get('USER_PERMISSIONS_CHAT_EDIT', 'True'

 USER_PERMISSIONS_CHAT_SHARE = os.environ.get('USER_PERMISSIONS_CHAT_SHARE', 'True').lower() == 'true'

+USER_PERMISSIONS_CHAT_ALLOW_PUBLIC_SHARING = (
+    os.environ.get('USER_PERMISSIONS_CHAT_ALLOW_PUBLIC_SHARING', 'False').lower() == 'true'
+)
+
 USER_PERMISSIONS_CHAT_EXPORT = os.environ.get('USER_PERMISSIONS_CHAT_EXPORT', 'True').lower() == 'true'

 USER_PERMISSIONS_CHAT_STT = os.environ.get('USER_PERMISSIONS_CHAT_STT', 'True').lower() == 'true'
@@ -1557,6 +1584,7 @@ DEFAULT_USER_PERMISSIONS = {
        'public_skills': USER_PERMISSIONS_WORKSPACE_SKILLS_ALLOW_PUBLIC_SHARING,
        'notes': USER_PERMISSIONS_NOTES_ALLOW_SHARING,
        'public_notes': USER_PERMISSIONS_NOTES_ALLOW_PUBLIC_SHARING,
+        'public_chats': USER_PERMISSIONS_CHAT_ALLOW_PUBLIC_SHARING,
    },
    'access_grants': {
        'allow_users': USER_PERMISSIONS_ACCESS_GRANTS_ALLOW_USERS,
@@ -2051,6 +2079,12 @@ VOICE_MODE_PROMPT_TEMPLATE = PersistentConfig(
    os.environ.get('VOICE_MODE_PROMPT_TEMPLATE', ''),
 )

+ENABLE_VOICE_MODE_PROMPT = PersistentConfig(
+    'ENABLE_VOICE_MODE_PROMPT',
+    'task.voice.prompt.enable',
+    os.environ.get('ENABLE_VOICE_MODE_PROMPT', 'True').lower() == 'true',
+)
+
 DEFAULT_VOICE_MODE_PROMPT_TEMPLATE = """You are a friendly, concise voice assistant.

 Everything you say will be spoken aloud.
@@ -2627,13 +2661,17 @@ ENABLE_ONEDRIVE_INTEGRATION = PersistentConfig(
 )


-ENABLE_ONEDRIVE_PERSONAL = os.environ.get('ENABLE_ONEDRIVE_PERSONAL', 'True').lower() == 'true'
-ENABLE_ONEDRIVE_BUSINESS = os.environ.get('ENABLE_ONEDRIVE_BUSINESS', 'True').lower() == 'true'
-
 ONEDRIVE_CLIENT_ID = os.environ.get('ONEDRIVE_CLIENT_ID', '')
 ONEDRIVE_CLIENT_ID_PERSONAL = os.environ.get('ONEDRIVE_CLIENT_ID_PERSONAL', ONEDRIVE_CLIENT_ID)
 ONEDRIVE_CLIENT_ID_BUSINESS = os.environ.get('ONEDRIVE_CLIENT_ID_BUSINESS', ONEDRIVE_CLIENT_ID)

+ENABLE_ONEDRIVE_PERSONAL = os.environ.get('ENABLE_ONEDRIVE_PERSONAL', 'True').lower() == 'true' and bool(
+    ONEDRIVE_CLIENT_ID_PERSONAL
+)
+ENABLE_ONEDRIVE_BUSINESS = os.environ.get('ENABLE_ONEDRIVE_BUSINESS', 'True').lower() == 'true' and bool(
+    ONEDRIVE_CLIENT_ID_BUSINESS
+)
+
 ONEDRIVE_SHAREPOINT_URL = PersistentConfig(
    'ONEDRIVE_SHAREPOINT_URL',
    'onedrive.sharepoint_url',
@@ -2827,6 +2865,18 @@ MISTRAL_OCR_API_KEY = PersistentConfig(
    os.getenv('MISTRAL_OCR_API_KEY', ''),
 )

+PADDLEOCR_VL_BASE_URL = PersistentConfig(
+    'PADDLEOCR_VL_BASE_URL',
+    'rag.paddleocr_vl_base_url',
+    os.getenv('PADDLEOCR_VL_BASE_URL', 'http://localhost:8080'),
+)
+
+PADDLEOCR_VL_TOKEN = PersistentConfig(
+    'PADDLEOCR_VL_TOKEN',
+    'rag.paddleocr_vl_token',
+    os.getenv('PADDLEOCR_VL_TOKEN', ''),
+)
+
 BYPASS_EMBEDDING_AND_RETRIEVAL = PersistentConfig(
    'BYPASS_EMBEDDING_AND_RETRIEVAL',
    'rag.bypass_embedding_and_retrieval',
@@ -3240,7 +3290,7 @@ ENABLE_WEB_LOADER_SSL_VERIFICATION = PersistentConfig(
 WEB_SEARCH_TRUST_ENV = PersistentConfig(
    'WEB_SEARCH_TRUST_ENV',
    'rag.web.search.trust_env',
-    os.getenv('WEB_SEARCH_TRUST_ENV', 'False').lower() == 'true',
+    os.getenv('WEB_SEARCH_TRUST_ENV', 'True').lower() == 'true',
 )


@@ -3298,6 +3348,12 @@ BRAVE_SEARCH_API_KEY = PersistentConfig(
    os.getenv('BRAVE_SEARCH_API_KEY', ''),
 )

+BRAVE_SEARCH_CONTEXT_TOKENS = PersistentConfig(
+    'BRAVE_SEARCH_CONTEXT_TOKENS',
+    'rag.web.search.brave_search_context_tokens',
+    int(os.getenv('BRAVE_SEARCH_CONTEXT_TOKENS', '8192')),
+)
+
 KAGI_SEARCH_API_KEY = PersistentConfig(
    'KAGI_SEARCH_API_KEY',
    'rag.web.search.kagi_search_api_key',
@@ -3935,6 +3991,19 @@ AUDIO_STT_SUPPORTED_CONTENT_TYPES = PersistentConfig(
    ],
 )

+AUDIO_STT_ALLOWED_EXTENSIONS = PersistentConfig(
+    'AUDIO_STT_ALLOWED_EXTENSIONS',
+    'audio.stt.allowed_extensions',
+    [
+        ext.strip()
+        for ext in os.environ.get(
+            'AUDIO_STT_ALLOWED_EXTENSIONS',
+            'mp3,wav,m4a,webm,ogg,flac,mp4,mpga,mpeg',
+        ).split(',')
+        if ext.strip()
+    ],
+)
+
 AUDIO_STT_AZURE_API_KEY = PersistentConfig(
    'AUDIO_STT_AZURE_API_KEY',
    'audio.stt.azure.api_key',
@@ -1,21 +1,21 @@
+import datetime as dt
 import importlib.metadata
 import json
 import logging
 import os
 import pkgutil
-import sys
+import re
 import shutil
+import sys
 import traceback
-from datetime import datetime, timezone
+from pathlib import Path
 from typing import Any
 from uuid import uuid4
-from pathlib import Path
-from cryptography.hazmat.primitives import serialization
-import re
-

 import markdown
 from bs4 import BeautifulSoup
+from cryptography.hazmat.primitives import serialization
+
 from open_webui.constants import ERROR_MESSAGES

 ####################################
@@ -43,7 +43,8 @@ except ImportError:

 DOCKER = os.environ.get('DOCKER', 'False').lower() == 'true'

-# device type embedding models - "cpu" (default), "cuda" (nvidia gpu required) or "mps" (apple silicon) - choosing this right can lead to better performance
+# device type for embedding models - "cpu" (default), "cuda" (nvidia gpu required), or "mps" (apple silicon)
+# choosing this correctly can lead to better performance
 USE_CUDA = os.environ.get('USE_CUDA_DOCKER', 'false')

 if USE_CUDA.lower() == 'true':
@@ -87,7 +88,7 @@ class JSONFormatter(logging.Formatter):

    def format(self, record: logging.LogRecord) -> str:
        log_entry: dict[str, Any] = {
-            'ts': datetime.fromtimestamp(record.created, tz=timezone.utc).isoformat(timespec='milliseconds'),
+            'ts': dt.datetime.fromtimestamp(record.created, tz=dt.UTC).isoformat(timespec='milliseconds'),
            'level': _LEVEL_MAP.get(record.levelname, record.levelname.lower()),
            'msg': record.getMessage(),
            'caller': record.name,
@@ -180,7 +181,7 @@ def parse_section(section):

 try:
    changelog_path = BASE_DIR / 'CHANGELOG.md'
-    with open(str(changelog_path.absolute()), 'r', encoding='utf8') as file:
+    with open(str(changelog_path.absolute()), encoding='utf8') as file:
        changelog_content = file.read()

 except Exception:
@@ -248,6 +249,17 @@ ENABLE_STAR_SESSIONS_MIDDLEWARE = os.environ.get('ENABLE_STAR_SESSIONS_MIDDLEWAR

 ENABLE_EASTER_EGGS = os.environ.get('ENABLE_EASTER_EGGS', 'True').lower() == 'true'

+####################################
+# ENABLE_PROFILE_IMAGE_URL_FORWARDING
+####################################
+
+# When True (default), the user and model profile-image endpoints
+# honour external http(s) URLs stored in profile_image_url by issuing a
+# 302 redirect to the original origin.  Set to False to suppress the
+# redirect (prevents client-side IP/UA/Referer leaks to attacker-
+# controlled origins) and fall through to the default image instead.
+ENABLE_PROFILE_IMAGE_URL_FORWARDING = os.environ.get('ENABLE_PROFILE_IMAGE_URL_FORWARDING', 'True').lower() == 'true'
+
 ####################################
 # WEBUI_BUILD_HASH
 ####################################
@@ -339,7 +351,7 @@ DATABASE_SCHEMA = os.environ.get('DATABASE_SCHEMA', None)

 DATABASE_POOL_SIZE = os.environ.get('DATABASE_POOL_SIZE', None)

-if DATABASE_POOL_SIZE != None:
+if DATABASE_POOL_SIZE is not None:
    try:
        DATABASE_POOL_SIZE = int(DATABASE_POOL_SIZE)
    except Exception:
@@ -525,6 +537,12 @@ WEBUI_AUTH_TRUSTED_NAME_HEADER = os.environ.get('WEBUI_AUTH_TRUSTED_NAME_HEADER'
 WEBUI_AUTH_TRUSTED_GROUPS_HEADER = os.environ.get('WEBUI_AUTH_TRUSTED_GROUPS_HEADER', None)
 WEBUI_AUTH_TRUSTED_ROLE_HEADER = os.environ.get('WEBUI_AUTH_TRUSTED_ROLE_HEADER', None)

+# Custom header name for API key authentication.  Defaults to 'x-api-key'.
+# Useful when Open WebUI sits behind a reverse proxy / API gateway that
+# already uses the Authorization header for its own authentication — set
+# this to a unique header (e.g. 'X-OpenWebUI-Key') so the middleware
+# checks the custom header instead and avoids the 401 short-circuit.
+CUSTOM_API_KEY_HEADER = os.environ.get('CUSTOM_API_KEY_HEADER', 'x-api-key')

 ENABLE_PASSWORD_VALIDATION = os.environ.get('ENABLE_PASSWORD_VALIDATION', 'False').lower() == 'true'
 PASSWORD_VALIDATION_REGEX_PATTERN = os.environ.get(
@@ -646,7 +664,7 @@ if LICENSE_PUBLIC_KEY:
 -----BEGIN PUBLIC KEY-----
 {LICENSE_PUBLIC_KEY}
 -----END PUBLIC KEY-----
-""".encode('utf-8')
+""".encode()
    )


@@ -284,7 +284,7 @@ async def generate_function_chat_completion(request, form_data, user, models: di
        if params:
            system = params.pop('system', None)
            form_data = apply_model_params_to_body_openai(params, form_data)
-            form_data = apply_system_prompt_to_body(system, form_data, metadata, user)
+            form_data = await apply_system_prompt_to_body(system, form_data, metadata, user)

    pipe_id = get_pipe_id(form_data)
    function_module = await get_function_module_by_id(request, pipe_id)
@@ -1,7 +1,7 @@
 import os
+import sys
 import json
 import logging
-import ssl as _stdlib_ssl
 from contextlib import asynccontextmanager, contextmanager
 from typing import Any, Optional
 from urllib.parse import parse_qs, urlencode, urlparse, urlunparse
@@ -37,90 +37,84 @@ from typing_extensions import Self
 log = logging.getLogger(__name__)


-def extract_ssl_mode_from_url(url: str) -> tuple[str, str | None]:
+# ── SSL URL normalization (used by sync engine & Alembic migrations) ─
+#
+# psycopg2 (sync) needs ``sslmode=`` in the connection string (it does
+# not recognise the bare ``ssl=`` key that some ORMs emit).  The helpers
+# below strip all SSL-related query params, normalise them, and
+# reattach them in the canonical libpq form.
+#
+# The **async** engine now uses psycopg (v3), which speaks libpq
+# natively, so it needs no translation at all — the DATABASE_URL is
+# passed through as-is.
+# ─────────────────────────────────────────────────────────────────────
+
+
+def _pop_first(params: dict[str, list[str]], key: str) -> str | None:
+    """Pop a single-valued query param, returning ``None`` if absent."""
+    values = params.pop(key, None)
+    return values[0] if values else None
+
+
+def _is_postgres_url(url: str) -> bool:
+    """Return True if *url* looks like a PostgreSQL connection string."""
+    return bool(url) and any(url.startswith(p) for p in ('postgresql://', 'postgresql+', 'postgres://'))
+
+
+def extract_ssl_params_from_url(url: str) -> tuple[str, dict[str, str]]:
    """Strip SSL query-string parameters from a PostgreSQL URL.

-    asyncpg and psycopg2 use different query-string keys for SSL
-    (``ssl`` vs ``sslmode``).  This helper removes **both** from the
-    URL so that each driver can receive the correct parameter through
-    its own mechanism (query-string re-injection for psycopg2,
-    ``connect_args`` for asyncpg).
-
-    Returns
-    -------
-    (url_without_ssl, ssl_mode)
-        *url_without_ssl* is the original URL with ``ssl`` / ``sslmode``
-        query parameters removed.  *ssl_mode* is the extracted mode
-        string (e.g. ``'require'``), or ``None`` if neither parameter
-        was present.
-
-    Non-PostgreSQL URLs are returned unchanged with ``ssl_mode=None``.
+    Returns ``(url_without_ssl, ssl_dict)`` where *ssl_dict* maps
+    canonical libpq key names (``sslmode``, ``sslrootcert``, …) to
+    their values.  Non-PostgreSQL URLs are returned unchanged with an
+    empty dict.
    """
-    if not url or not any(url.startswith(prefix) for prefix in ('postgresql://', 'postgresql+', 'postgres://')):
-        return url, None
+    if not _is_postgres_url(url):
+        return url, {}

    parsed = urlparse(url)
-    query_params = parse_qs(parsed.query, keep_blank_values=True)
+    qp = parse_qs(parsed.query, keep_blank_values=True)

-    # Prefer sslmode (libpq canonical) over the asyncpg-only ssl key.
-    ssl_mode: str | None = None
-    for key in ('sslmode', 'ssl'):
-        values = query_params.pop(key, None)
-        if values and ssl_mode is None:
-            ssl_mode = values[0]
+    # Prefer sslmode (libpq canonical) over the bare ``ssl`` key.
+    sslmode_val = _pop_first(qp, 'sslmode')
+    ssl_val = _pop_first(qp, 'ssl')
+    ssl_mode = sslmode_val or ssl_val

-    if ssl_mode is None:
-        # Nothing to strip — return the URL untouched.
-        return url, None
+    ssl_dict: dict[str, str] = {}
+    if ssl_mode:
+        ssl_dict['sslmode'] = ssl_mode
+    for key in ('sslrootcert', 'sslcert', 'sslkey', 'sslcrl'):
+        val = _pop_first(qp, key)
+        if val:
+            ssl_dict[key] = val

-    # Rebuild the query string without the SSL keys.
-    remaining_query = urlencode(query_params, doseq=True)
-    url_without_ssl = urlunparse(parsed._replace(query=remaining_query))
-    return url_without_ssl, ssl_mode
+    if not ssl_dict:
+        return url, ssl_dict
+
+    cleaned_query = urlencode(qp, doseq=True)
+    return urlunparse(parsed._replace(query=cleaned_query)), ssl_dict


-def build_asyncpg_ssl_args(ssl_mode: str | None) -> dict:
-    """Convert a libpq-style SSL mode value to asyncpg ``connect_args``.
+def reattach_ssl_params_to_url(url_without_ssl: str, ssl_dict: dict[str, str]) -> str:
+    """Re-append SSL query-string parameters to a cleaned PostgreSQL URL.

-    Returns a dict suitable for unpacking into
-    ``create_async_engine(..., connect_args=...)``.
+    Used for psycopg2/libpq consumers that expect ``sslmode`` and the
+    certificate-file keys in the connection string.
    """
-    if ssl_mode is None:
-        return {}
-
-    mode = ssl_mode.lower()
-    if mode == 'disable':
-        return {'connect_args': {'ssl': False}}
-    if mode in ('allow', 'prefer'):
-        # asyncpg has no direct equivalent — omit to let it try without.
-        return {}
-    if mode == 'require':
-        # SSL required but no certificate verification (matches libpq).
-        ctx = _stdlib_ssl.create_default_context()
-        ctx.check_hostname = False
-        ctx.verify_mode = _stdlib_ssl.CERT_NONE
-        return {'connect_args': {'ssl': ctx}}
-    if mode in ('verify-ca', 'verify-full'):
-        # Full verification — use the system trust store.
-        ctx = _stdlib_ssl.create_default_context()
-        if mode == 'verify-ca':
-            ctx.check_hostname = False
-        return {'connect_args': {'ssl': ctx}}
-
-    # Unknown value — pass through as-is and let asyncpg decide.
-    return {'connect_args': {'ssl': ssl_mode}}
-
-
-def reattach_ssl_mode_to_url(url_without_ssl: str, ssl_mode: str | None) -> str:
-    """Re-append ``sslmode=<value>`` to a cleaned PostgreSQL URL.
-
-    Used for psycopg2 / libpq consumers that expect the canonical
-    ``sslmode`` query-string key.
-    """
-    if ssl_mode is None:
+    if not ssl_dict:
        return url_without_ssl
-    separator = '&' if '?' in url_without_ssl else '?'
-    return f'{url_without_ssl}{separator}sslmode={ssl_mode}'
+
+    parts = [f'{k}={v}' for k, v in ssl_dict.items() if v]
+    if not parts:
+        return url_without_ssl
+
+    sep = '&' if '?' in url_without_ssl else '?'
+    return f'{url_without_ssl}{sep}{"&".join(parts)}'
+
+
+# Backwards-compatible aliases for external callers.
+extract_ssl_mode_from_url = extract_ssl_params_from_url
+reattach_ssl_mode_to_url = reattach_ssl_params_to_url


 class JSONField(types.TypeDecorator):
@@ -150,9 +144,10 @@ class JSONField(types.TypeDecorator):
 def handle_peewee_migration(DATABASE_URL):
    db = None
    try:
-        # Normalize SSL params so psycopg2 always sees `sslmode=` (never `ssl=`).
-        url_without_ssl, ssl_mode = extract_ssl_mode_from_url(DATABASE_URL)
-        normalized_url = reattach_ssl_mode_to_url(url_without_ssl, ssl_mode)
+        # Normalize SSL params so psycopg2 always sees `sslmode=` (never `ssl=`)
+        # and cert-file params are preserved in the connection string.
+        url_without_ssl, ssl_params = extract_ssl_params_from_url(DATABASE_URL)
+        normalized_url = reattach_ssl_params_to_url(url_without_ssl, ssl_params)

        # Replace the postgresql:// with postgres:// to handle the peewee migration
        db = register_connection(normalized_url.replace('postgresql://', 'postgres://'))
@@ -179,32 +174,36 @@ if ENABLE_DB_MIGRATIONS:
    handle_peewee_migration(DATABASE_URL)


-# Normalize SSL params from the URL once; each engine branch re-injects
-# the driver-appropriate form.
-DATABASE_URL_WITHOUT_SSL, DATABASE_SSL_MODE = extract_ssl_mode_from_url(DATABASE_URL)
+# Normalize SSL params from the URL once; the sync engine needs them
+# reattached in canonical libpq form for psycopg2.
+_url_without_ssl, _ssl_dict = extract_ssl_params_from_url(DATABASE_URL)

-# For psycopg2 (sync engine), re-append sslmode=<value>.
-SQLALCHEMY_DATABASE_URL = (
-    reattach_ssl_mode_to_url(DATABASE_URL_WITHOUT_SSL, DATABASE_SSL_MODE) if DATABASE_SSL_MODE else DATABASE_URL
-)
+# For psycopg2 (sync engine), re-append sslmode + cert-file params.
+SQLALCHEMY_DATABASE_URL = reattach_ssl_params_to_url(_url_without_ssl, _ssl_dict) if _ssl_dict else DATABASE_URL


 def _make_async_url(url: str) -> str:
-    """Convert a sync database URL to its async driver equivalent."""
+    """Convert a sync database URL to its async driver equivalent.
+
+    The async engine uses psycopg (v3) which speaks libpq natively,
+    so all standard connection-string parameters (``sslmode``,
+    ``options``, ``target_session_attrs``, etc.) are passed through
+    without any translation.
+    """
    if url.startswith('sqlite+sqlcipher://'):
-        # SQLCipher has no async driver — not supported for async
        raise ValueError(
            'sqlite+sqlcipher:// URLs are not supported with async engine. '
            'Use standard sqlite:// or postgresql:// instead.'
        )
    if url.startswith('sqlite:///') or url.startswith('sqlite://'):
        return url.replace('sqlite://', 'sqlite+aiosqlite://', 1)
+    # psycopg v3 — auto-selects async mode with create_async_engine
    if url.startswith('postgresql+psycopg2://'):
-        return url.replace('postgresql+psycopg2://', 'postgresql+asyncpg://', 1)
+        return url.replace('postgresql+psycopg2://', 'postgresql+psycopg://', 1)
    if url.startswith('postgresql://'):
-        return url.replace('postgresql://', 'postgresql+asyncpg://', 1)
+        return url.replace('postgresql://', 'postgresql+psycopg://', 1)
    if url.startswith('postgres://'):
-        return url.replace('postgres://', 'postgresql+asyncpg://', 1)
+        return url.replace('postgres://', 'postgresql+psycopg://', 1)
    # For other dialects, return as-is and let SQLAlchemy handle it
    return url

@@ -329,10 +328,19 @@ get_db = contextmanager(get_session)
 # ASYNC ENGINE (used for ALL runtime database operations)
 # ============================================================

-# Use the SSL-stripped URL for asyncpg — SSL is injected via connect_args.
-ASYNC_SQLALCHEMY_DATABASE_URL = _make_async_url(
-    DATABASE_URL_WITHOUT_SSL if DATABASE_SSL_MODE else SQLALCHEMY_DATABASE_URL
-)
+# psycopg (v3) speaks libpq natively — the full DATABASE_URL is passed
+# through as-is.  SSL params, ``options``, ``target_session_attrs``, etc.
+# all work without any stripping or translation.
+ASYNC_SQLALCHEMY_DATABASE_URL = _make_async_url(SQLALCHEMY_DATABASE_URL)
+
+# psycopg v3 cannot run in async mode under Windows' default
+# ProactorEventLoop — switch to SelectorEventLoop before creating
+# the async engine.  This runs at import time, which is early enough
+# to cover every entry point (workers, reload, direct invocations).
+if sys.platform == 'win32' and _is_postgres_url(DATABASE_URL):
+    import asyncio
+
+    asyncio.set_event_loop_policy(asyncio.WindowsSelectorEventLoopPolicy())

 if 'sqlite' in ASYNC_SQLALCHEMY_DATABASE_URL:
    # Generous default — async coroutines + no session sharing = high connection demand.
@@ -350,10 +358,6 @@ if 'sqlite' in ASYNC_SQLALCHEMY_DATABASE_URL:
    def _set_sqlite_pragmas(dbapi_connection, connection_record):
        _apply_sqlite_pragmas(dbapi_connection)
 else:
-    # Inject asyncpg-compatible SSL connect_args when the user specified
-    # sslmode/ssl in DATABASE_URL.
-    asyncpg_ssl_args = build_asyncpg_ssl_args(DATABASE_SSL_MODE)
-
    if isinstance(DATABASE_POOL_SIZE, int):
        if DATABASE_POOL_SIZE > 0:
            async_engine = create_async_engine(
@@ -363,20 +367,17 @@ else:
                pool_timeout=DATABASE_POOL_TIMEOUT,
                pool_recycle=DATABASE_POOL_RECYCLE,
                pool_pre_ping=True,
-                **asyncpg_ssl_args,
            )
        else:
            async_engine = create_async_engine(
                ASYNC_SQLALCHEMY_DATABASE_URL,
                pool_pre_ping=True,
                poolclass=NullPool,
-                **asyncpg_ssl_args,
            )
    else:
        async_engine = create_async_engine(
            ASYNC_SQLALCHEMY_DATABASE_URL,
            pool_pre_ping=True,
-            **asyncpg_ssl_args,
        )


@@ -199,6 +199,7 @@ from open_webui.config import (
    AUDIO_STT_ENGINE,
    AUDIO_STT_MODEL,
    AUDIO_STT_SUPPORTED_CONTENT_TYPES,
+    AUDIO_STT_ALLOWED_EXTENSIONS,
    AUDIO_STT_OPENAI_API_BASE_URL,
    AUDIO_STT_OPENAI_API_KEY,
    AUDIO_STT_AZURE_API_KEY,
@@ -303,6 +304,8 @@ from open_webui.config import (
    DOCUMENT_INTELLIGENCE_MODEL,
    MISTRAL_OCR_API_BASE_URL,
    MISTRAL_OCR_API_KEY,
+    PADDLEOCR_VL_BASE_URL,
+    PADDLEOCR_VL_TOKEN,
    RAG_TEXT_SPLITTER,
    ENABLE_MARKDOWN_HEADER_TEXT_SPLITTER,
    TIKTOKEN_ENCODING_NAME,
@@ -342,6 +345,7 @@ from open_webui.config import (
    BING_SEARCH_V7_ENDPOINT,
    BING_SEARCH_V7_SUBSCRIPTION_KEY,
    BRAVE_SEARCH_API_KEY,
+    BRAVE_SEARCH_CONTEXT_TOKENS,
    EXA_API_KEY,
    PERPLEXITY_API_KEY,
    PERPLEXITY_MODEL,
@@ -476,6 +480,7 @@ from open_webui.config import (
    IMAGE_PROMPT_GENERATION_PROMPT_TEMPLATE,
    TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE,
    VOICE_MODE_PROMPT_TEMPLATE,
+    ENABLE_VOICE_MODE_PROMPT,
    QUERY_GENERATION_PROMPT_TEMPLATE,
    AUTOCOMPLETE_GENERATION_PROMPT_TEMPLATE,
    AUTOCOMPLETE_GENERATION_INPUT_MAX_LENGTH,
@@ -578,6 +583,8 @@ from open_webui.utils.redis import get_redis_connection
 from open_webui.tasks import (
    redis_task_command_listener,
    list_task_ids_by_item_id,
+    has_active_tasks,
+    cleanup_task,
    create_task,
    stop_task,
    stop_item_tasks,
@@ -704,30 +711,34 @@ async def lifespan(app: FastAPI):

    # Pre-fetch tool server specs so the first request doesn't pay the latency cost
    if len(app.state.config.TOOL_SERVER_CONNECTIONS) > 0:
+        mock_request = Request(
+            {
+                'type': 'http',
+                'asgi.version': '3.0',
+                'asgi.spec_version': '2.0',
+                'method': 'GET',
+                'path': '/internal',
+                'query_string': b'',
+                'headers': Headers({}).raw,
+                'client': ('127.0.0.1', 12345),
+                'server': ('127.0.0.1', 80),
+                'scheme': 'http',
+                'app': app,
+            }
+        )
+
        log.info('Initializing tool servers...')
        try:
-            mock_request = Request(
-                {
-                    'type': 'http',
-                    'asgi.version': '3.0',
-                    'asgi.spec_version': '2.0',
-                    'method': 'GET',
-                    'path': '/internal',
-                    'query_string': b'',
-                    'headers': Headers({}).raw,
-                    'client': ('127.0.0.1', 12345),
-                    'server': ('127.0.0.1', 80),
-                    'scheme': 'http',
-                    'app': app,
-                }
-            )
            await set_tool_servers(mock_request)
            log.info(f'Initialized {len(app.state.TOOL_SERVERS)} tool server(s)')
+        except Exception as e:
+            log.warning(f'Failed to initialize tool servers at startup: {e}')

+        try:
            await set_terminal_servers(mock_request)
            log.info(f'Initialized {len(app.state.TERMINAL_SERVERS)} terminal server(s)')
        except Exception as e:
-            log.warning(f'Failed to initialize tool/terminal servers at startup: {e}')
+            log.warning(f'Failed to initialize terminal servers at startup: {e}')

    # Mark application as ready to accept traffic from a startup perspective.
    app.state.startup_complete = True
@@ -1023,6 +1034,8 @@ app.state.config.DOCUMENT_INTELLIGENCE_KEY = DOCUMENT_INTELLIGENCE_KEY
 app.state.config.DOCUMENT_INTELLIGENCE_MODEL = DOCUMENT_INTELLIGENCE_MODEL
 app.state.config.MISTRAL_OCR_API_BASE_URL = MISTRAL_OCR_API_BASE_URL
 app.state.config.MISTRAL_OCR_API_KEY = MISTRAL_OCR_API_KEY
+app.state.config.PADDLEOCR_VL_BASE_URL = PADDLEOCR_VL_BASE_URL
+app.state.config.PADDLEOCR_VL_TOKEN = PADDLEOCR_VL_TOKEN
 app.state.config.MINERU_API_MODE = MINERU_API_MODE
 app.state.config.MINERU_API_URL = MINERU_API_URL
 app.state.config.MINERU_API_KEY = MINERU_API_KEY
@@ -1098,6 +1111,7 @@ app.state.config.YACY_PASSWORD = YACY_PASSWORD
 app.state.config.GOOGLE_PSE_API_KEY = GOOGLE_PSE_API_KEY
 app.state.config.GOOGLE_PSE_ENGINE_ID = GOOGLE_PSE_ENGINE_ID
 app.state.config.BRAVE_SEARCH_API_KEY = BRAVE_SEARCH_API_KEY
+app.state.config.BRAVE_SEARCH_CONTEXT_TOKENS = BRAVE_SEARCH_CONTEXT_TOKENS
 app.state.config.KAGI_SEARCH_API_KEY = KAGI_SEARCH_API_KEY
 app.state.config.MOJEEK_SEARCH_API_KEY = MOJEEK_SEARCH_API_KEY
 app.state.config.BOCHA_SEARCH_API_KEY = BOCHA_SEARCH_API_KEY
@@ -1285,6 +1299,7 @@ app.state.config.IMAGES_EDIT_COMFYUI_WORKFLOW_NODES = IMAGES_EDIT_COMFYUI_WORKFL
 app.state.config.STT_ENGINE = AUDIO_STT_ENGINE
 app.state.config.STT_MODEL = AUDIO_STT_MODEL
 app.state.config.STT_SUPPORTED_CONTENT_TYPES = AUDIO_STT_SUPPORTED_CONTENT_TYPES
+app.state.config.STT_ALLOWED_EXTENSIONS = AUDIO_STT_ALLOWED_EXTENSIONS

 app.state.config.STT_OPENAI_API_BASE_URL = AUDIO_STT_OPENAI_API_BASE_URL
 app.state.config.STT_OPENAI_API_KEY = AUDIO_STT_OPENAI_API_KEY
@@ -1357,6 +1372,7 @@ app.state.config.QUERY_GENERATION_PROMPT_TEMPLATE = QUERY_GENERATION_PROMPT_TEMP
 app.state.config.AUTOCOMPLETE_GENERATION_PROMPT_TEMPLATE = AUTOCOMPLETE_GENERATION_PROMPT_TEMPLATE
 app.state.config.AUTOCOMPLETE_GENERATION_INPUT_MAX_LENGTH = AUTOCOMPLETE_GENERATION_INPUT_MAX_LENGTH
 app.state.config.VOICE_MODE_PROMPT_TEMPLATE = VOICE_MODE_PROMPT_TEMPLATE
+app.state.config.ENABLE_VOICE_MODE_PROMPT = ENABLE_VOICE_MODE_PROMPT


 ########################################
@@ -1520,6 +1536,108 @@ async def get_base_models(request: Request, user=Depends(get_admin_user)):
    return {'data': models}


+class ModelUnloadForm(BaseModel):
+    model: str
+
+
+@app.post('/api/models/unload')
+async def unload_model(request: Request, form_data: ModelUnloadForm, user=Depends(get_admin_user)):
+    """
+    Unified model unload endpoint.
+    Resolves the provider that owns the model and calls its native unload mechanism.
+    Supports: Ollama (keep_alive=0) and llama.cpp (/models/unload).
+    """
+    model_id = form_data.model
+
+    # --- Ollama provider ---
+    ollama_models = getattr(request.app.state, 'OLLAMA_MODELS', None) or {}
+    if model_id in ollama_models:
+        url_indices = ollama_models[model_id].get('urls', [])
+        errors = []
+        for idx in url_indices:
+            url = request.app.state.config.OLLAMA_BASE_URLS[idx]
+            api_config = request.app.state.config.OLLAMA_API_CONFIGS.get(
+                str(idx),
+                request.app.state.config.OLLAMA_API_CONFIGS.get(url, {}),
+            )
+            key = api_config.get('key', None)
+
+            prefix_id = api_config.get('prefix_id', None)
+            actual_model = model_id
+            if prefix_id and actual_model.startswith(f'{prefix_id}.'):
+                actual_model = actual_model[len(f'{prefix_id}.') :]
+
+            payload = json.dumps({'model': actual_model, 'keep_alive': 0, 'prompt': ''})
+
+            try:
+                timeout = aiohttp.ClientTimeout(total=30)
+                async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session:
+                    headers = {
+                        'Content-Type': 'application/json',
+                        **({'Authorization': f'Bearer {key}'} if key else {}),
+                    }
+                    async with session.post(
+                        f'{url}/api/generate',
+                        data=payload,
+                        headers=headers,
+                    ) as r:
+                        if not r.ok:
+                            errors.append({'url_idx': idx, 'error': await r.text()})
+            except Exception as e:
+                log.exception(f'Failed to unload model on Ollama node {idx}: {e}')
+                errors.append({'url_idx': idx, 'error': str(e)})
+
+        if errors:
+            raise HTTPException(
+                status_code=500,
+                detail=f'Failed to unload model on {len(errors)} node(s): {errors}',
+            )
+        return {'status': True}
+
+    # --- OpenAI-compatible providers ---
+    openai_models = getattr(request.app.state, 'OPENAI_MODELS', None) or {}
+    if model_id in openai_models:
+        model_info = openai_models[model_id]
+        idx = model_info.get('urlIdx')
+        api_config = request.app.state.config.OPENAI_API_CONFIGS.get(str(idx), {})
+        provider = api_config.get('provider', '')
+        base_url = request.app.state.config.OPENAI_API_BASE_URLS[idx]
+        key = (
+            request.app.state.config.OPENAI_API_KEYS[idx] if idx < len(request.app.state.config.OPENAI_API_KEYS) else ''
+        )
+
+        if provider == 'llama.cpp':
+            root_url = base_url.rstrip('/').removesuffix('/v1')
+            try:
+                timeout = aiohttp.ClientTimeout(total=30)
+                async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session:
+                    headers = {
+                        'Content-Type': 'application/json',
+                        **({'Authorization': f'Bearer {key}'} if key else {}),
+                    }
+                    async with session.post(
+                        f'{root_url}/models/unload',
+                        json={'model': model_id},
+                        headers=headers,
+                    ) as r:
+                        if not r.ok:
+                            detail = await r.text()
+                            raise HTTPException(status_code=r.status, detail=detail)
+                        return await r.json()
+            except HTTPException:
+                raise
+            except Exception as e:
+                log.exception(f'Failed to unload model via llama.cpp: {e}')
+                raise HTTPException(status_code=500, detail=str(e))
+        else:
+            raise HTTPException(
+                status_code=400,
+                detail=f'Provider "{provider or "default"}" does not support model unloading',
+            )
+
+    raise HTTPException(status_code=404, detail=f'Model "{model_id}" not found')
+
+
 ##################################
 # Embeddings
 ##################################
@@ -1647,6 +1765,7 @@ async def chat_completion(
            'chat_id': form_data.pop('chat_id', None),
            'user_message': user_message,
            'user_message_id': user_message.get('id') if user_message else None,
+            'assistant_message_id': form_data.pop('assistant_message_id', None),
            'session_id': form_data.pop('session_id', None),
            'folder_id': form_data.pop('folder_id', None),
            'filter_ids': form_data.pop('filter_ids', []),
@@ -1719,6 +1838,7 @@ async def chat_completion(
                                ]
                                if user_message_id
                                else [],
+                                'files': metadata.get('files') or [],
                                'tags': [],
                                'timestamp': int(time.time() * 1000),
                            },
@@ -1751,6 +1871,16 @@ async def chat_completion(
                            detail=ERROR_MESSAGES.DEFAULT(),
                        )

+                    # Persist chat-level files (knowledge collections, docs, etc.)
+                    # The old frontend saveChatHandler did this on every message;
+                    # now the backend owns persistence.
+                    chat_files = metadata.get('files')
+                    if chat_files is not None:
+                        existing_chat = await Chats.get_chat_by_id(chat_id)
+                        if existing_chat:
+                            updated = {**existing_chat.chat, 'files': chat_files}
+                            await Chats.update_chat_by_id(chat_id, updated)
+
                    # Save user message to DB
                    user_message = metadata.get('user_message') or {}
                    if user_message and user_message.get('id'):
@@ -1832,7 +1962,7 @@ async def chat_completion(
    except HTTPException:
        raise
    except Exception as e:
-        log.debug(f'Error processing chat metadata: {e}')
+        log.warning(f'Error processing chat metadata: {e}')
        raise HTTPException(
            status_code=status.HTTP_400_BAD_REQUEST,
            detail=str(e),
@@ -1865,17 +1995,16 @@ async def chat_completion(
        except asyncio.CancelledError:
            log.info('Chat processing was cancelled')
            try:
-                event_emitter = await get_event_emitter(metadata)
-                if event_emitter:
-                    await asyncio.shield(
-                        event_emitter(
-                            {'type': 'chat:tasks:cancel'},
-                        )
-                    )
-            except Exception as e:
+
+                async def emit_cancel_event():
+                    event_emitter = await get_event_emitter(metadata)
+                    if event_emitter:
+                        await event_emitter({'type': 'chat:tasks:cancel'})
+
+                await asyncio.shield(emit_cancel_event())
+            except Exception:
                pass
-            finally:
-                raise  # re-raise to ensure proper task cancellation handling
+            raise  # re-raise to ensure proper task cancellation handling
        except Exception as e:
            error_detail = e.detail if isinstance(e, HTTPException) else str(e)
            log.error('Error processing chat payload: %s', error_detail)
@@ -1906,37 +2035,54 @@ async def chat_completion(

                except Exception:
                    pass
+            else:
+                # No chat_id/message_id → legacy/direct API path with no
+                # WebSocket error channel.  We must surface the error as
+                # a proper HTTP response; without this the function would
+                # return None which FastAPI serializes as null.  #23924
+                raise HTTPException(
+                    status_code=status.HTTP_400_BAD_REQUEST,
+                    detail=error_detail,
+                )
        finally:
-            # Clean up MCP clients.  Shield the entire block from
-            # CancelledError so disconnect() can finish even when the
-            # task is being stopped.  Each client is isolated so one
+            # Clean up MCP clients.  Each client is isolated so one
            # failure doesn't skip the rest.
+            #
+            # NOTE: asyncio.wait_for() / asyncio.shield() must NOT be used
+            # here — they create new asyncio Tasks, which violate anyio
+            # cancel-scope task-ownership rules when the MCPClient's
+            # exit_stack contains anyio transport resources (streamable_http).
+            # Exiting those cancel scopes from the wrong task raises
+            # "Attempted to exit a cancel scope that isn't the current
+            # task's current cancel scope", which propagates as a
+            # BaseException through the finally block, discards the response
+            # return value, and surfaces as a 500 "No response returned."
+            # MCPClient.disconnect() already catches BaseException internally.
            try:
                if mcp_clients := metadata.get('mcp_clients'):
-
-                    async def _cleanup_mcp():
-                        for client in reversed(list(mcp_clients.values())):
-                            try:
-                                await client.disconnect()
-                            except Exception as e:
-                                log.debug(f'Error disconnecting MCP client: {e}')
-
-                    await asyncio.wait_for(
-                        asyncio.shield(_cleanup_mcp()),
-                        timeout=10.0,
-                    )
-            except asyncio.TimeoutError:
-                log.warning('MCP client cleanup timed out after 10 s')
-            except Exception as e:
+                    for client in reversed(list(mcp_clients.values())):
+                        try:
+                            await client.disconnect()
+                        except BaseException as e:
+                            log.debug(f'Error disconnecting MCP client: {e}')
+            except BaseException as e:
                log.debug(f'Error cleaning up MCP clients: {e}')
-            # Emit chat:active=false when task completes
+
+            # Deregister this task, then emit chat:active=false if no others remain
            try:
-                if metadata.get('chat_id'):
-                    event_emitter = await get_event_emitter(metadata, update_db=False)
-                    if event_emitter:
-                        await event_emitter({'type': 'chat:active', 'data': {'active': False}})
-            except Exception as e:
-                log.debug(f'Error emitting chat:active: {e}')
+                chat_id = metadata.get('chat_id')
+                task_id = metadata.get('task_id')
+                if chat_id and task_id:
+                    await cleanup_task(request.app.state.redis, task_id, chat_id)
+                    if not await has_active_tasks(request.app.state.redis, chat_id):
+                        event_emitter = await get_event_emitter(metadata, update_db=False)
+                        if event_emitter:
+                            try:
+                                await asyncio.shield(event_emitter({'type': 'chat:active', 'data': {'active': False}}))
+                            except asyncio.CancelledError:
+                                pass
+            except Exception:
+                pass

    # Fan out: one task per model
    if metadata.get('session_id') and metadata.get('chat_id'):
@@ -1984,6 +2130,7 @@ async def chat_completion(
                ),
                id=chat_id,
            )
+            per_model_metadata['task_id'] = task_id
            task_ids.append(task_id)

        # Emit chat:active=true
@@ -2690,6 +2837,14 @@ async def get_opensearch_xml():
    return Response(content=xml_content, media_type='application/xml')


+def _sync_db_ping() -> None:
+    ScopedSession.execute(text('SELECT 1;')).all()
+
+
+async def async_db_ping() -> None:
+    await asyncio.to_thread(_sync_db_ping)
+
+
@app.get('/health')
 async def healthcheck():
    return {'status': True}
@@ -2711,7 +2866,7 @@ async def readiness_check():

    # Check database connectivity
    try:
-        ScopedSession.execute(text('SELECT 1;')).all()
+        await async_db_ping()
    except Exception as e:
        log.warning(f'Readiness check DB ping failed: {e!r}')
        raise HTTPException(
@@ -2738,7 +2893,7 @@ async def readiness_check():

@app.get('/health/db')
 async def healthcheck_with_db():
-    ScopedSession.execute(text('SELECT 1;')).all()
+    await async_db_ping()
    return {'status': True}


@@ -2756,7 +2911,13 @@ async def serve_cache_file(
        raise HTTPException(status_code=404, detail='File not found')
    if not os.path.isfile(file_path):
        raise HTTPException(status_code=404, detail='File not found')
-    return FileResponse(file_path)
+
+    mime, _ = mimetypes.guess_type(file_path)
+    inline_safe = mime and mime.split('/', 1)[0] in {'image', 'audio', 'video'}
+    headers = {'X-Content-Type-Options': 'nosniff'}
+    if not inline_safe:
+        headers['Content-Disposition'] = f'attachment; filename="{os.path.basename(file_path)}"'
+    return FileResponse(file_path, headers=headers)


 def swagger_ui_html(*args, **kwargs):
@@ -5,7 +5,7 @@ from alembic import context
 from open_webui.models.auths import Auth
 from open_webui.models.calendar import Calendar, CalendarEvent, CalendarEventAttendee  # noqa: F401
 from open_webui.env import DATABASE_URL, DATABASE_PASSWORD, LOG_FORMAT
-from open_webui.internal.db import extract_ssl_mode_from_url, reattach_ssl_mode_to_url
+from open_webui.internal.db import extract_ssl_params_from_url, reattach_ssl_params_to_url
 from sqlalchemy import engine_from_config, pool, create_engine

 # this is the Alembic Config object, which provides
@@ -37,9 +37,9 @@ target_metadata = Auth.metadata

 DB_URL = DATABASE_URL

-# Normalize SSL query params for psycopg2 (Alembic uses psycopg2, not asyncpg).
-url_without_ssl, ssl_mode = extract_ssl_mode_from_url(DB_URL)
-DB_URL = reattach_ssl_mode_to_url(url_without_ssl, ssl_mode) if ssl_mode else DB_URL
+# Normalize SSL query params for psycopg2 (Alembic uses psycopg2 for sync migrations).
+url_without_ssl, ssl_params = extract_ssl_params_from_url(DB_URL)
+DB_URL = reattach_ssl_params_to_url(url_without_ssl, ssl_params) if ssl_params else DB_URL

 if DB_URL:
    config.set_main_option('sqlalchemy.url', DB_URL.replace('%', '%%'))
@@ -0,0 +1,80 @@
+"""add pinned_note table
+
+Revision ID: 4de81c2a3af1
+Revises: 56359461a091
+Create Date: 2026-05-09 04:29:27.651341
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+import open_webui.internal.db
+
+
+# revision identifiers, used by Alembic.
+revision: str = '4de81c2a3af1'
+down_revision: Union[str, None] = '56359461a091'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+import uuid
+import time
+from sqlalchemy import select, update, insert
+from sqlalchemy.sql import table, column
+
+
+def upgrade() -> None:
+    op.create_table(
+        'pinned_note',
+        sa.Column('id', sa.Text(), nullable=False),
+        sa.Column('user_id', sa.Text(), nullable=False),
+        sa.Column('note_id', sa.Text(), sa.ForeignKey('note.id', ondelete='CASCADE'), nullable=False),
+        sa.Column('created_at', sa.BigInteger(), nullable=False),
+        sa.PrimaryKeyConstraint('id'),
+        sa.UniqueConstraint('user_id', 'note_id', name='uq_pinned_note'),
+    )
+
+    conn = op.get_bind()
+
+    note_table = table('note', column('id', sa.Text), column('user_id', sa.Text), column('is_pinned', sa.Boolean))
+
+    pinned_note_table = table(
+        'pinned_note',
+        column('id', sa.Text),
+        column('user_id', sa.Text),
+        column('note_id', sa.Text),
+        column('created_at', sa.BigInteger),
+    )
+
+    notes = conn.execute(select(note_table.c.id, note_table.c.user_id).where(note_table.c.is_pinned == True)).fetchall()
+
+    if notes:
+        now = int(time.time_ns())
+        conn.execute(
+            insert(pinned_note_table),
+            [{'id': str(uuid.uuid4()), 'user_id': note[1], 'note_id': note[0], 'created_at': now} for note in notes],
+        )
+
+    with op.batch_alter_table('note', schema=None) as batch_op:
+        batch_op.drop_column('is_pinned')
+
+
+def downgrade() -> None:
+    with op.batch_alter_table('note', schema=None) as batch_op:
+        batch_op.add_column(sa.Column('is_pinned', sa.Boolean(), nullable=True))
+
+    conn = op.get_bind()
+
+    note_table = table('note', column('id', sa.Text), column('is_pinned', sa.Boolean))
+
+    pinned_note_table = table('pinned_note', column('note_id', sa.Text))
+
+    notes = conn.execute(select(pinned_note_table.c.note_id)).fetchall()
+
+    for note in notes:
+        conn.execute(update(note_table).where(note_table.c.id == note[0]).values(is_pinned=True))
+
+    op.drop_table('pinned_note')
@@ -0,0 +1,22 @@
+"""Add memory user_id index
+
+Revision ID: a0b1c2d3e4f5
+Revises: 4de81c2a3af1
+Create Date: 2025-09-15 03:00:00.000000
+
+"""
+
+from alembic import op
+
+revision = 'a0b1c2d3e4f5'
+down_revision = '4de81c2a3af1'
+branch_labels = None
+depends_on = None
+
+
+def upgrade():
+    op.create_index('ix_memory_user_id', 'memory', ['user_id'])
+
+
+def downgrade():
+    op.drop_index('ix_memory_user_id', table_name='memory')
@@ -395,14 +395,16 @@ class CalendarTable:
                # Delete events
                await db.execute(delete(CalendarEvent).filter(CalendarEvent.calendar_id == id))

-                # Delete access grants
-                await AccessGrants.revoke_all_access('calendar', id, db=db)
-
                # Delete calendar
                await db.execute(delete(Calendar).filter(Calendar.id == id))
                await db.commit()
-                return True
-        except Exception:
+
+            # Revoke access grants in a separate transaction to avoid
+            # write-lock contention on SQLite when session sharing is off.
+            await AccessGrants.revoke_all_access('calendar', id)
+            return True
+        except Exception as e:
+            log.exception(f'Failed to delete calendar {id}: {e}')
            return False


@@ -4,6 +4,8 @@ import time
 import uuid
 from typing import Optional

+from open_webui.utils.validate import validate_profile_image_url
+
 from sqlalchemy import select, delete, update, func, case, or_, and_
 from sqlalchemy.ext.asyncio import AsyncSession
 from open_webui.internal.db import Base, JSONField, get_async_db_context
@@ -13,7 +15,7 @@ from open_webui.models.access_grants import (
    AccessGrants,
 )

-from pydantic import BaseModel, ConfigDict, Field
+from pydantic import BaseModel, ConfigDict, Field, field_validator
 from sqlalchemy.dialects.postgresql import JSONB


@@ -244,6 +246,13 @@ class ChannelWebhookForm(BaseModel):
    name: str
    profile_image_url: Optional[str] = None

+    @field_validator('profile_image_url', mode='before')
+    @classmethod
+    def check_profile_image_url(cls, v: Optional[str]) -> Optional[str]:
+        if v is None:
+            return v
+        return validate_profile_image_url(v)
+

 class ChannelTable:
    async def _get_access_grants(self, channel_id: str, db: Optional[AsyncSession] = None) -> list[AccessGrantModel]:
@@ -48,6 +48,25 @@ def get_usage(data: dict) -> Optional[dict]:
    return normalize_usage(usage) if usage else None


+def _token_columns(dialect: str):
+    """Return (input_tokens, output_tokens) SQL column expressions.
+
+    Falls back to OpenAI-style keys (prompt_tokens / completion_tokens)
+    when the normalized keys are absent.
+    """
+    if dialect == 'sqlite':
+        extract = lambda key: cast(func.json_extract(ChatMessage.usage, f'$.{key}'), Integer)
+    elif dialect == 'postgresql':
+        extract = lambda key: cast(func.json_extract_path_text(ChatMessage.usage, key), Integer)
+    else:
+        raise NotImplementedError(f'Unsupported dialect: {dialect}')
+
+    return (
+        func.coalesce(extract('input_tokens'), extract('prompt_tokens')),
+        func.coalesce(extract('output_tokens'), extract('completion_tokens')),
+    )
+
+
 ####################
 # ChatMessage DB Schema
 ####################
@@ -222,6 +241,79 @@ class ChatMessageTable:
            messages = result.scalars().all()
            return [ChatMessageModel.model_validate(message) for message in messages]

+    # DB column names that differ from the JSON message keys.
+    DB_TO_JSON_KEY_MAP = {
+        'parent_id': 'parentId',
+        'model_id': 'model',
+        'status_history': 'statusHistory',
+        'created_at': 'timestamp',
+    }
+    # DB-internal columns excluded from the reconstructed message dict.
+    EXCLUDED_COLUMNS = frozenset({'id', 'chat_id', 'user_id', 'updated_at'})
+
+    async def get_messages_map_by_chat_id(self, chat_id: str, db: Optional[AsyncSession] = None) -> Optional[dict]:
+        """Build a {message_id: message_dict} map from chat_message rows.
+
+        Returns the same shape as chat.history.messages so callers
+        (get_message_list, middleware) work unchanged.  Returns None if
+        no rows exist for the chat (caller should fall back to the
+        embedded JSON blob for legacy chats).
+        """
+        async with get_async_db_context(db) as db:
+            result = await db.execute(select(ChatMessage).filter_by(chat_id=chat_id))
+            rows = result.scalars().all()
+
+        if not rows:
+            return None
+
+        # Strip the composite-id prefix ("{chat_id}-") to recover the
+        # original message_id used as map key.
+        prefix = f'{chat_id}-'
+        prefix_len = len(prefix)
+        col_keys = [c.key for c in ChatMessage.__table__.columns]
+
+        messages_map: dict[str, dict] = {}
+        for row in rows:
+            msg_id = row.id[prefix_len:] if row.id.startswith(prefix) else row.id
+
+            msg: dict = {'id': msg_id}
+            for key in col_keys:
+                if key in self.EXCLUDED_COLUMNS:
+                    continue
+                val = getattr(row, key)
+                if val is None:
+                    continue
+                json_key = self.DB_TO_JSON_KEY_MAP.get(key, key)
+                msg[json_key] = val
+
+            # Ensure content always has a value
+            msg.setdefault('content', '')
+
+            # Mirror usage into info.usage for callers that read it there
+            if 'usage' in msg:
+                msg['info'] = {'usage': msg['usage']}
+
+            messages_map[msg_id] = msg
+
+        # Reconstruct childrenIds from parentId links so that the map
+        # is fully navigable (callers like the frontend rely on this).
+        for msg_id, msg in messages_map.items():
+            parent_id = msg.get('parentId')
+            if parent_id and parent_id in messages_map:
+                parent = messages_map[parent_id]
+                children = parent.get('childrenIds')
+                if children is None:
+                    parent['childrenIds'] = [msg_id]
+                elif msg_id not in children:
+                    children.append(msg_id)
+
+        # Ensure every message has a childrenIds list (leaf nodes get [])
+        for msg in messages_map.values():
+            if 'childrenIds' not in msg:
+                msg['childrenIds'] = []
+
+        return messages_map
+
    async def get_messages_by_user_id(
        self,
        user_id: str,
@@ -343,20 +435,7 @@ class ChatMessageTable:
            bind = await db.connection()
            dialect = bind.dialect.name

-            if dialect == 'sqlite':
-                input_tokens = cast(func.json_extract(ChatMessage.usage, '$.input_tokens'), Integer)
-                output_tokens = cast(func.json_extract(ChatMessage.usage, '$.output_tokens'), Integer)
-            elif dialect == 'postgresql':
-                input_tokens = cast(
-                    func.json_extract_path_text(ChatMessage.usage, 'input_tokens'),
-                    Integer,
-                )
-                output_tokens = cast(
-                    func.json_extract_path_text(ChatMessage.usage, 'output_tokens'),
-                    Integer,
-                )
-            else:
-                raise NotImplementedError(f'Unsupported dialect: {dialect}')
+            input_tokens, output_tokens = _token_columns(dialect)

            stmt = select(
                ChatMessage.model_id,
@@ -404,20 +483,7 @@ class ChatMessageTable:
            bind = await db.connection()
            dialect = bind.dialect.name

-            if dialect == 'sqlite':
-                input_tokens = cast(func.json_extract(ChatMessage.usage, '$.input_tokens'), Integer)
-                output_tokens = cast(func.json_extract(ChatMessage.usage, '$.output_tokens'), Integer)
-            elif dialect == 'postgresql':
-                input_tokens = cast(
-                    func.json_extract_path_text(ChatMessage.usage, 'input_tokens'),
-                    Integer,
-                )
-                output_tokens = cast(
-                    func.json_extract_path_text(ChatMessage.usage, 'output_tokens'),
-                    Integer,
-                )
-            else:
-                raise NotImplementedError(f'Unsupported dialect: {dialect}')
+            input_tokens, output_tokens = _token_columns(dialect)

            stmt = select(
                ChatMessage.user_id,
@@ -366,20 +366,20 @@ class ChatTable:
            await db.commit()

            # Dual-write messages to chat_message table
-            try:
-                for form_data, chat_obj in zip(chat_import_forms, chats):
-                    history = form_data.chat.get('history', {})
-                    messages = history.get('messages', {})
-                    for message_id, message in messages.items():
-                        if isinstance(message, dict) and message.get('role'):
+            for form_data, chat_obj in zip(chat_import_forms, chats):
+                history = form_data.chat.get('history', {})
+                messages = history.get('messages', {})
+                for message_id, message in messages.items():
+                    if isinstance(message, dict) and message.get('role'):
+                        try:
                            await ChatMessages.upsert_message(
                                message_id=message_id,
                                chat_id=chat_obj.id,
                                user_id=user_id,
                                data=message,
                            )
-            except Exception as e:
-                log.warning(f'Failed to write imported messages to chat_message table: {e}')
+                        except Exception as e:
+                            log.warning(f'Failed to write imported message {message_id} for chat {chat_obj.id}: {e}')

            return [ChatModel.model_validate(chat) for chat in chats]

@@ -393,7 +393,6 @@ class ChatTable:
                chat_item.updated_at = int(time.time())

                await db.commit()
-                await db.refresh(chat_item)

                return ChatModel.model_validate(chat_item)
        except Exception:
@@ -461,6 +460,18 @@ class ChatTable:
            return row[0] or 'New Chat'

    async def get_messages_map_by_chat_id(self, id: str) -> Optional[dict]:
+        """Message map for walking history (see ``get_message_list``).
+
+        Prefer ``chat_message`` rows to avoid loading the large ``chat``
+        JSON blob; fall back to embedded history when no rows exist
+        (legacy chats).
+        """
+        # Fast path: build from normalized chat_message rows.
+        messages_map = await ChatMessages.get_messages_map_by_chat_id(id)
+        if messages_map is not None:
+            return messages_map
+
+        # No rows — fall back to the embedded JSON blob for legacy chats.
        chat = await self.get_chat_by_id(id)
        if chat is None:
            return None
@@ -4,7 +4,7 @@ import time
 from typing import Optional
 import uuid

-from sqlalchemy import select, delete, update, or_, func
+from sqlalchemy import select, delete, update, or_, func, cast
 from sqlalchemy.ext.asyncio import AsyncSession
 from open_webui.internal.db import Base, JSONField, get_async_db_context

@@ -313,11 +313,16 @@ class KnowledgeTable:
                    permission='read',
                )

-                # Apply filename search
+                # Apply filename / content search
                if filter:
                    q = filter.get('query')
                    if q:
-                        stmt = stmt.filter(File.filename.ilike(f'%{q}%'))
+                        stmt = stmt.filter(
+                            or_(
+                                File.filename.ilike(f'%{q}%'),
+                                cast(File.data['content'], Text).ilike(f'%{q}%'),
+                            )
+                        )

                # Order by file changes
                stmt = stmt.order_by(File.updated_at.desc(), File.id.asc())
@@ -467,7 +472,12 @@ class KnowledgeTable:
                if filter:
                    query_key = filter.get('query')
                    if query_key:
-                        stmt = stmt.filter(or_(File.filename.ilike(f'%{query_key}%')))
+                        stmt = stmt.filter(
+                            or_(
+                                File.filename.ilike(f'%{query_key}%'),
+                                cast(File.data['content'], Text).ilike(f'%{query_key}%'),
+                            )
+                        )

                    view_option = filter.get('view_option')
                    if view_option == 'created':
@@ -19,7 +19,7 @@ class Memory(Base):
    __tablename__ = 'memory'

    id = Column(String, primary_key=True, unique=True)
-    user_id = Column(String)
+    user_id = Column(String, index=True)
    content = Column(Text)
    updated_at = Column(BigInteger)
    created_at = Column(BigInteger)
@@ -143,6 +143,8 @@ class ModelAccessListResponse(BaseModel):


 class ModelForm(BaseModel):
+    model_config = ConfigDict(extra='ignore')
+
    id: str
    base_model_id: Optional[str] = None
    name: str
@@ -389,6 +391,52 @@ class ModelsTable:

            return ModelListResponse(items=models, total=total)

+    async def get_model_meta_by_id(self, id: str, db: Optional[AsyncSession] = None) -> Optional[tuple[dict, int]]:
+        """Return (meta, updated_at) for a model, skipping access grant resolution."""
+        try:
+            async with get_async_db_context(db) as db:
+                result = await db.execute(select(Model.meta, Model.updated_at).filter_by(id=id))
+                return result.first()
+        except Exception:
+            return None
+
+    async def get_all_tags(
+        self,
+        user_id: str,
+        is_admin: bool = False,
+        db: Optional[AsyncSession] = None,
+    ) -> set[str]:
+        """Extract unique tag names from model meta, querying only the meta column."""
+        async with get_async_db_context(db) as db:
+            stmt = select(Model.meta).filter(Model.base_model_id != None)
+
+            if not is_admin:
+                user_groups = await Groups.get_groups_by_member_id(user_id, db=db)
+                user_group_ids = [group.id for group in user_groups]
+
+                filter_dict = {'user_id': user_id}
+                if user_group_ids:
+                    filter_dict['group_ids'] = user_group_ids
+
+                stmt = self._has_permission(db, stmt, filter_dict, permission='read')
+
+            result = await db.execute(stmt)
+            rows = result.scalars().all()
+
+            tags_set: set[str] = set()
+            for meta in rows:
+                if not meta:
+                    continue
+                for tag in meta.get('tags', []):
+                    try:
+                        name = tag.get('name') if isinstance(tag, dict) else str(tag)
+                        if name:
+                            tags_set.add(name)
+                    except Exception:
+                        continue
+
+            return tags_set
+
    async def get_model_by_id(self, id: str, db: Optional[AsyncSession] = None) -> Optional[ModelModel]:
        try:
            async with get_async_db_context(db) as db:
@@ -13,7 +13,7 @@ from open_webui.models.access_grants import AccessGrantModel, AccessGrants


 from pydantic import BaseModel, ConfigDict, Field
-from sqlalchemy import BigInteger, Column, Text, JSON
+from sqlalchemy import BigInteger, Column, Text, JSON, ForeignKey

 ####################
 # Note DB Schema
@@ -29,7 +29,6 @@ class Note(Base):
    title = Column(Text)
    data = Column(JSON, nullable=True)
    meta = Column(JSON, nullable=True)
-    is_pinned = Column(Boolean, default=False, nullable=True)

    created_at = Column(BigInteger)
    updated_at = Column(BigInteger)
@@ -52,6 +51,15 @@ class NoteModel(BaseModel):
    updated_at: int  # timestamp in epoch


+class PinnedNote(Base):
+    __tablename__ = 'pinned_note'
+
+    id = Column(Text, primary_key=True)
+    user_id = Column(Text, nullable=False)
+    note_id = Column(Text, ForeignKey('note.id', ondelete='CASCADE'), nullable=False)
+    created_at = Column(BigInteger, nullable=False)
+
+
 ####################
 # Forms
 ####################
@@ -100,6 +108,7 @@ class NoteTable:
        access_grants: Optional[list[AccessGrantModel]] = None,
        db: Optional[AsyncSession] = None,
    ) -> NoteModel:
+        # We exclude access_grants to inject them
        note_data = NoteModel.model_validate(note).model_dump(exclude={'access_grants'})
        note_data['access_grants'] = (
            access_grants if access_grants is not None else await self._get_access_grants(note_data['id'], db=db)
@@ -314,15 +323,28 @@ class NoteTable:
            await db.commit()
            return await self._to_note_model(note, db=db) if note else None

-    async def toggle_note_pinned_by_id(self, id: str, db: Optional[AsyncSession] = None) -> Optional[NoteModel]:
+    async def toggle_note_pinned_by_id(
+        self, id: str, user_id: str, db: Optional[AsyncSession] = None
+    ) -> Optional[NoteModel]:
        try:
            async with get_async_db_context(db) as db:
                result = await db.execute(select(Note).filter(Note.id == id))
                note = result.scalars().first()
                if not note:
                    return None
-                note.is_pinned = not note.is_pinned
-                note.updated_at = int(time.time_ns())
+
+                # Check if already pinned
+                pin_result = await db.execute(select(PinnedNote).filter_by(user_id=user_id, note_id=id))
+                pinned_note = pin_result.scalars().first()
+
+                if pinned_note:
+                    await db.execute(delete(PinnedNote).filter_by(user_id=user_id, note_id=id))
+                else:
+                    new_pin = PinnedNote(
+                        id=str(uuid.uuid4()), user_id=user_id, note_id=id, created_at=int(time.time_ns())
+                    )
+                    db.add(new_pin)
+
                await db.commit()
                return await self._to_note_model(note, db=db)
        except Exception:
@@ -338,7 +360,12 @@ class NoteTable:
            user_groups = await Groups.get_groups_by_member_id(user_id, db=db)
            user_group_ids = [group.id for group in user_groups]

-            stmt = select(Note).filter(Note.is_pinned == True).order_by(Note.updated_at.desc())
+            stmt = (
+                select(Note)
+                .join(PinnedNote, PinnedNote.note_id == Note.id)
+                .filter(PinnedNote.user_id == user_id)
+                .order_by(PinnedNote.created_at.desc())
+            )
            stmt = self._has_permission(db, stmt, {'user_id': user_id, 'group_ids': user_group_ids}, permission)

            result = await db.execute(stmt)
@@ -351,11 +378,17 @@ class NoteTable:
        try:
            async with get_async_db_context(db) as db:
                await AccessGrants.revoke_all_access('note', id, db=db)
+                await db.execute(delete(PinnedNote).filter(PinnedNote.note_id == id))
                await db.execute(delete(Note).filter(Note.id == id))
                await db.commit()
                return True
        except Exception:
            return False

+    async def get_pinned_note_ids(self, user_id: str, db: Optional[AsyncSession] = None) -> list[str]:
+        async with get_async_db_context(db) as db:
+            result = await db.execute(select(PinnedNote.note_id).filter_by(user_id=user_id))
+            return result.scalars().all()
+

 Notes = NoteTable()
@@ -320,6 +320,19 @@ class OAuthSessionTable:
            log.error(f'Error deleting OAuth sessions by user ID: {e}')
            return False

+    async def delete_sessions_by_user_id_and_provider(
+        self, user_id: str, provider: str, db: Optional[AsyncSession] = None
+    ) -> bool:
+        """Delete all OAuth sessions for a specific user and provider"""
+        try:
+            async with get_async_db_context(db) as db:
+                result = await db.execute(delete(OAuthSession).filter_by(user_id=user_id, provider=provider))
+                await db.commit()
+                return result.rowcount > 0
+        except Exception as e:
+            log.error(f'Error deleting OAuth sessions for user {user_id} and provider {provider}: {e}')
+            return False
+
    async def delete_sessions_by_provider(self, provider: str, db: Optional[AsyncSession] = None) -> bool:
        """Delete all OAuth sessions for a provider"""
        try:
@@ -231,24 +231,51 @@ class PromptsTable:
    async def get_prompts_by_user_id(
        self, user_id: str, permission: str = 'write', db: Optional[AsyncSession] = None
    ) -> list[PromptUserResponse]:
-        prompts = await self.get_prompts(db=db)
-        user_groups = await Groups.get_groups_by_member_id(user_id, db=db)
-        user_group_ids = {group.id for group in user_groups}
+        async with get_async_db_context(db) as db:
+            user_groups = await Groups.get_groups_by_member_id(user_id, db=db)
+            user_group_ids = [group.id for group in user_groups]

-        result = []
-        for prompt in prompts:
-            if prompt.user_id == user_id:
-                result.append(prompt)
-            elif await AccessGrants.has_access(
-                user_id=user_id,
-                resource_type='prompt',
-                resource_id=prompt.id,
-                permission=permission,
-                user_group_ids=user_group_ids,
+            query = select(Prompt).filter(Prompt.is_active == True).order_by(Prompt.updated_at.desc())
+            query = AccessGrants.has_permission_filter(
                db=db,
-            ):
-                result.append(prompt)
-        return result
+                query=query,
+                DocumentModel=Prompt,
+                filter={'user_id': user_id, 'group_ids': user_group_ids},
+                resource_type='prompt',
+                permission=permission,
+            )
+
+            result = await db.execute(query)
+            accessible_prompts = result.scalars().all()
+
+            if not accessible_prompts:
+                return []
+
+            prompt_ids = [p.id for p in accessible_prompts]
+            owner_ids = list({p.user_id for p in accessible_prompts})
+
+            users = await Users.get_users_by_user_ids(owner_ids, db=db)
+            users_dict = {u.id: u for u in users}
+            grants_map = await AccessGrants.get_grants_by_resources('prompt', prompt_ids, db=db)
+
+            results = []
+            for prompt in accessible_prompts:
+                user = users_dict.get(prompt.user_id)
+                results.append(
+                    PromptUserResponse.model_validate(
+                        {
+                            **(
+                                await self._to_prompt_model(
+                                    prompt,
+                                    access_grants=grants_map.get(prompt.id, []),
+                                    db=db,
+                                )
+                            ).model_dump(),
+                            'user': user.model_dump() if user else None,
+                        }
+                    )
+                )
+            return results

    async def search_prompts(
        self,
@@ -632,12 +659,38 @@ class PromptsTable:
    async def get_tags(self, db: Optional[AsyncSession] = None) -> list[str]:
        try:
            async with get_async_db_context(db) as db:
-                result = await db.execute(select(Prompt).filter_by(is_active=True))
-                prompts = result.scalars().all()
+                result = await db.execute(select(Prompt.tags).filter(Prompt.is_active == True))
                tags = set()
-                for prompt in prompts:
-                    if prompt.tags:
-                        for tag in prompt.tags:
+                for (tag_list,) in result.all():
+                    if tag_list:
+                        for tag in tag_list:
+                            if tag:
+                                tags.add(tag)
+                return sorted(list(tags))
+        except Exception:
+            return []
+
+    async def get_tags_by_user_id(self, user_id: str, db: Optional[AsyncSession] = None) -> list[str]:
+        try:
+            async with get_async_db_context(db) as db:
+                user_groups = await Groups.get_groups_by_member_id(user_id, db=db)
+                user_group_ids = [group.id for group in user_groups]
+
+                query = select(Prompt.tags).filter(Prompt.is_active == True)
+                query = AccessGrants.has_permission_filter(
+                    db=db,
+                    query=query,
+                    DocumentModel=Prompt,
+                    filter={'user_id': user_id, 'group_ids': user_group_ids},
+                    resource_type='prompt',
+                    permission='read',
+                )
+
+                result = await db.execute(query)
+                tags = set()
+                for (tag_list,) in result.all():
+                    if tag_list:
+                        for tag in tag_list:
                            if tag:
                                tags.add(tag)
                return sorted(list(tags))
@@ -23,9 +23,9 @@ from open_webui.retrieval.loaders.external_document import ExternalDocumentLoade
 from open_webui.retrieval.loaders.mistral import MistralLoader
 from open_webui.retrieval.loaders.datalab_marker import DatalabMarkerLoader
 from open_webui.retrieval.loaders.mineru import MinerULoader
+from open_webui.retrieval.loaders.paddleocr_vl import PaddleOCRVLLoader

-
-from open_webui.env import GLOBAL_LOG_LEVEL, REQUESTS_VERIFY
+from open_webui.env import GLOBAL_LOG_LEVEL, REQUESTS_VERIFY, AIOHTTP_CLIENT_SESSION_SSL

 logging.basicConfig(stream=sys.stdout, level=GLOBAL_LOG_LEVEL)
 log = logging.getLogger(__name__)
@@ -205,6 +205,7 @@ class DoclingLoader:
                    **self.params,
                },
                headers=headers,
+                verify=AIOHTTP_CLIENT_SESSION_SSL,
            )
        if r.ok:
            result = r.json()
@@ -399,6 +400,12 @@ class Loader:
                api_key=self.kwargs.get('MISTRAL_OCR_API_KEY'),
                file_path=file_path,
            )
+        elif self.engine == 'paddleocr_vl' and self.kwargs.get('PADDLEOCR_VL_TOKEN') != '':
+            loader = PaddleOCRVLLoader(
+                api_url=self.kwargs.get('PADDLEOCR_VL_BASE_URL'),
+                token=self.kwargs.get('PADDLEOCR_VL_TOKEN'),
+                file_path=file_path,
+            )
        else:
            if file_ext == 'pdf':
                loader = PyPDFLoader(
@@ -0,0 +1,125 @@
+import base64
+import os
+import requests
+import logging
+import sys
+from typing import List
+
+from langchain_core.documents import Document
+from open_webui.env import GLOBAL_LOG_LEVEL
+
+logging.basicConfig(stream=sys.stdout, level=GLOBAL_LOG_LEVEL)
+log = logging.getLogger(__name__)
+
+
+class PaddleOCRVLLoader:
+    """Loader that uses PaddleOCR-vl API to extract text from PDF/images."""
+
+    def __init__(
+        self,
+        api_url: str,
+        token: str,
+        file_path: str,
+    ):
+        if not api_url or not token:
+            raise ValueError('PaddleOCR-vl API URL and Token are required.')
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f'File not found at {file_path}')
+
+        self.api_url = api_url.rstrip('/')
+        self.token = token
+        self.file_path = file_path
+        self.file_name = os.path.basename(file_path)
+
+    def load(self) -> List[Document]:
+        log.info(f'Processing with PaddleOCR-vl: {self.file_path}')
+
+        try:
+            with open(self.file_path, 'rb') as file:
+                file_bytes = file.read()
+                file_data = base64.b64encode(file_bytes).decode('ascii')
+        except Exception as e:
+            log.error(f'Failed to read file {self.file_path}: {e}')
+            raise
+
+        headers = {'Authorization': f'token {self.token}', 'Content-Type': 'application/json'}
+
+        # Detect fileType based on file extension
+        ext = self.file_path.lower().split('.')[-1]
+        image_extensions = ['png', 'jpg', 'jpeg', 'bmp', 'tiff', 'webp']
+        file_type = 1 if ext in image_extensions else 0
+
+        payload = {
+            'file': file_data,
+            'fileType': file_type,
+            'useDocOrientationClassify': False,
+            'useDocUnwarping': False,
+            'useChartRecognition': False,
+        }
+
+        try:
+            response = requests.post(f'{self.api_url}/layout-parsing', json=payload, headers=headers)
+            response.raise_for_status()
+
+            result = response.json().get('result', {})
+            layout_results = result.get('layoutParsingResults', [])
+
+            documents = []
+            total_pages = len(layout_results)
+            skipped_pages = 0
+
+            for i, res in enumerate(layout_results):
+                markdown_text = res.get('markdown', {}).get('text', '')
+
+                if isinstance(markdown_text, str):
+                    cleaned_content = markdown_text.strip()
+                else:
+                    cleaned_content = str(markdown_text).strip()
+
+                if not cleaned_content:
+                    skipped_pages += 1
+                    continue
+
+                documents.append(
+                    Document(
+                        page_content=cleaned_content,
+                        metadata={
+                            'page': i,
+                            'page_label': i + 1,
+                            'total_pages': total_pages,
+                            'file_name': self.file_name,
+                            'processing_engine': 'paddleocr-vl',
+                        },
+                    )
+                )
+
+            if skipped_pages > 0:
+                log.info(f'PaddleOCR-vl: Processed {len(documents)} pages, skipped {skipped_pages} empty pages.')
+
+            if not documents:
+                log.warning('No valid text content found by PaddleOCR-vl.')
+                return [
+                    Document(
+                        page_content='No valid text content found in document',
+                        metadata={
+                            'error': 'no_valid_pages',
+                            'file_name': self.file_name,
+                            'processing_engine': 'paddleocr-vl',
+                        },
+                    )
+                ]
+
+            return documents
+
+        except Exception as e:
+            log.error(f'Error calling PaddleOCR-vl: {e}')
+            return [
+                Document(
+                    page_content=f'Error during OCR processing: {e}',
+                    metadata={
+                        'error': 'processing_failed',
+                        'file_name': self.file_name,
+                        'processing_engine': 'paddleocr-vl',
+                    },
+                )
+            ]
@@ -114,6 +114,8 @@ def build_loader_from_config(request):
        DOCUMENT_INTELLIGENCE_MODEL=config.DOCUMENT_INTELLIGENCE_MODEL,
        MISTRAL_OCR_API_BASE_URL=config.MISTRAL_OCR_API_BASE_URL,
        MISTRAL_OCR_API_KEY=config.MISTRAL_OCR_API_KEY,
+        PADDLEOCR_VL_BASE_URL=config.PADDLEOCR_VL_BASE_URL,
+        PADDLEOCR_VL_TOKEN=config.PADDLEOCR_VL_TOKEN,
        MINERU_API_MODE=config.MINERU_API_MODE,
        MINERU_API_URL=config.MINERU_API_URL,
        MINERU_API_KEY=config.MINERU_API_KEY,
@@ -564,6 +566,13 @@ async def query_collection(
            log.exception(f'Error when querying the collection: {e}')
            return None, e

+    # Sanitize: filter out None/empty queries to prevent embedding crashes
+    # (e.g. when get_last_user_message returns None)
+    queries = [q for q in queries if q]
+    if not queries:
+        log.warning('query_collection: all queries were None or empty, returning empty results')
+        return {'distances': [[]], 'documents': [[]], 'metadatas': [[]]}
+
    # Generate all query embeddings (in one call)
    query_embeddings = await embedding_function(queries, prefix=RAG_EMBEDDING_QUERY_PREFIX)
    log.debug(f'query_collection: processing {len(queries)} queries across {len(collection_names)} collections')
@@ -0,0 +1,66 @@
+import logging
+import time
+from typing import Optional
+
+import requests
+from open_webui.retrieval.web.main import SearchResult, get_filtered_results
+
+log = logging.getLogger(__name__)
+
+
+def search_brave_llm_context(
+    api_key: str,
+    query: str,
+    count: int,
+    filter_list: Optional[list[str]] = None,
+    context_tokens: int = 8192,
+) -> list[SearchResult]:
+    """Search using Brave's LLM Context API and return pre-extracted, relevance-scored
+    page content ready for LLM consumption.
+
+    Uses /res/v1/llm/context instead of /res/v1/web/search. Same API key, same pricing.
+    Returns full extracted passages per URL rather than short snippets, eliminating
+    the need for post-search scraping.
+
+    Args:
+        api_key (str): A Brave Search API key (same key as web search)
+        query (str): The query to search for
+        count (int): Maximum number of results to return
+        filter_list (list[str], optional): Domain filter list
+        context_tokens (int): Maximum total tokens to retrieve (1024–32768, default 8192)
+    """
+    url = 'https://api.search.brave.com/res/v1/llm/context'
+    headers = {
+        'Accept': 'application/json',
+        'Accept-Encoding': 'gzip',
+        'X-Subscription-Token': api_key,
+    }
+    params = {
+        'q': query,
+        'count': count,
+        'maximum_number_of_tokens': context_tokens,
+    }
+
+    response = requests.get(url, headers=headers, params=params)
+
+    # Handle 429 rate limiting - same rate limits as web search
+    if response.status_code == 429:
+        log.info('Brave LLM Context API rate limited (429), retrying after 1 second...')
+        time.sleep(1)
+        response = requests.get(url, headers=headers, params=params)
+
+    response.raise_for_status()
+
+    json_response = response.json()
+    results = json_response.get('grounding', {}).get('generic', [])
+    if filter_list:
+        results = get_filtered_results(results, filter_list)
+
+    return [
+        SearchResult(
+            link=result['url'],
+            title=result.get('title'),
+            snippet='\n\n'.join(result.get('snippets', [])),
+        )
+        for result in results[:count]
+    ]
@@ -1,4 +1,5 @@
 import logging
+import urllib.request
 from typing import Optional

 from open_webui.retrieval.web.main import SearchResult, get_filtered_results
@@ -25,17 +26,25 @@ def search_duckduckgo(
    Returns:
        list[SearchResult]: A list of search results
    """
-    # Use the DDGS context manager to create a DDGS object
+    # The ddgs library (primp-based) does not auto-detect proxy env vars.
+    # Resolve via stdlib getproxies() — same pattern as the other loaders.
+    env_proxies = urllib.request.getproxies()
+    proxy = env_proxies.get('https') or env_proxies.get('http')
    search_results = []
-    with DDGS() as ddgs:
+    with DDGS(proxy=proxy) as ddgs:
        if concurrent_requests:
            ddgs.threads = concurrent_requests

        # Use the ddgs.text() method to perform the search
        try:
-            search_results = ddgs.text(query, safesearch='moderate', max_results=count, backend=backend)
+            kwargs = {'safesearch': 'moderate', 'max_results': count}
+            if backend and backend != 'auto':
+                kwargs['backend'] = backend
+            results = ddgs.text(query, **kwargs)
+            search_results = results if results is not None else []
        except RatelimitException as e:
            log.error(f'RatelimitException: {e}')
+            search_results = []
    if filter_list:
        search_results = get_filtered_results(search_results, filter_list)

@@ -1,52 +1,229 @@
+from __future__ import annotations
+
 import logging
-from typing import Optional, List
+import time
+from typing import TYPE_CHECKING, Any

 import requests
-from open_webui.retrieval.web.main import SearchResult, get_filtered_results
+from langchain_core.documents import Document
+
+if TYPE_CHECKING:
+    from open_webui.retrieval.web.main import SearchResult

 log = logging.getLogger(__name__)

+DEFAULT_FIRECRAWL_API_BASE_URL = 'https://api.firecrawl.dev'
+FIRECRAWL_RETRY_STATUS_CODES = {429, 500, 502, 503, 504}
+FIRECRAWL_MAX_RETRIES = 2
+
+
+def build_firecrawl_url(base_url: str | None, path: str) -> str:
+    base_url = (base_url or DEFAULT_FIRECRAWL_API_BASE_URL).rstrip('/')
+    path = path.lstrip('/')
+
+    if base_url.endswith('/v2'):
+        return f'{base_url}/{path}'
+
+    return f'{base_url}/v2/{path}'
+
+
+def build_firecrawl_headers(api_key: str | None) -> dict[str, str]:
+    return {
+        'Content-Type': 'application/json',
+        'Authorization': f'Bearer {api_key or ""}',
+    }
+
+
+def get_firecrawl_timeout_seconds(timeout: Any) -> float | None:
+    if timeout in (None, ''):
+        return None
+
+    try:
+        timeout = float(timeout)
+    except (TypeError, ValueError):
+        return None
+
+    return timeout if timeout > 0 else None
+
+
+def get_firecrawl_scrape_timeout_ms(timeout: Any) -> int | None:
+    timeout_seconds = get_firecrawl_timeout_seconds(timeout)
+    if timeout_seconds is None:
+        return None
+
+    # Firecrawl v2 expects scrape timeouts in milliseconds.
+    return min(300000, max(1000, int(timeout_seconds * 1000)))
+
+
+def get_firecrawl_client_timeout_seconds(timeout: Any, fallback: float = 60) -> float:
+    # Keep the local HTTP timeout slightly above Firecrawl's scrape timeout.
+    return (get_firecrawl_timeout_seconds(timeout) or fallback) + 10
+
+
+def get_firecrawl_retry_delay(headers: Any, attempt: int) -> float:
+    retry_after = headers.get('Retry-After') if headers else None
+    if retry_after:
+        try:
+            return min(10.0, max(0.0, float(retry_after)))
+        except (TypeError, ValueError):
+            pass
+
+    return min(8.0, float(2**attempt))
+
+
+def request_firecrawl_json(
+    method: str,
+    url: str,
+    *,
+    headers: dict[str, str],
+    json: dict[str, Any] | None = None,
+    timeout: float | None = None,
+    verify: bool = True,
+) -> dict[str, Any]:
+    last_error = None
+
+    for attempt in range(FIRECRAWL_MAX_RETRIES + 1):
+        try:
+            response = requests.request(
+                method,
+                url,
+                headers=headers,
+                json=json,
+                timeout=timeout,
+                verify=verify,
+            )
+
+            if response.status_code in FIRECRAWL_RETRY_STATUS_CODES and attempt < FIRECRAWL_MAX_RETRIES:
+                delay = get_firecrawl_retry_delay(response.headers, attempt)
+                log.warning(
+                    'Firecrawl %s %s returned HTTP %s; retrying in %.1fs',
+                    method,
+                    url,
+                    response.status_code,
+                    delay,
+                )
+                time.sleep(delay)
+                continue
+
+            response.raise_for_status()
+            return response.json()
+        except (requests.ConnectionError, requests.Timeout) as e:
+            last_error = e
+            if attempt >= FIRECRAWL_MAX_RETRIES:
+                break
+
+            delay = get_firecrawl_retry_delay(None, attempt)
+            log.warning('Firecrawl %s %s failed; retrying in %.1fs: %s', method, url, delay, e)
+            time.sleep(delay)
+
+    if last_error:
+        raise last_error
+
+    raise RuntimeError(f'Firecrawl {method} {url} failed without a response')
+
+
+def get_firecrawl_result_url(result: dict[str, Any]) -> str:
+    metadata = result.get('metadata') or {}
+    return (
+        result.get('url')
+        or result.get('link')
+        or metadata.get('url')
+        or metadata.get('sourceURL')
+        or metadata.get('source_url')
+        or ''
+    )
+
+
+def scrape_firecrawl_url(
+    firecrawl_url: str,
+    firecrawl_api_key: str,
+    url: str,
+    *,
+    verify_ssl: bool = True,
+    timeout: Any = None,
+    params: dict[str, Any] | None = None,
+) -> Document | None:
+    payload = {
+        'url': url,
+        'formats': ['markdown'],
+        'skipTlsVerification': not verify_ssl,
+        'removeBase64Images': True,
+        **(params or {}),
+    }
+    scrape_timeout_ms = get_firecrawl_scrape_timeout_ms(timeout)
+    if scrape_timeout_ms is not None:
+        payload['timeout'] = scrape_timeout_ms
+
+    response = request_firecrawl_json(
+        'POST',
+        build_firecrawl_url(firecrawl_url, 'scrape'),
+        headers=build_firecrawl_headers(firecrawl_api_key),
+        json=payload,
+        timeout=get_firecrawl_client_timeout_seconds(timeout),
+        verify=verify_ssl,
+    )
+    data = response.get('data') or {}
+    content = data.get('markdown') or ''
+    if not isinstance(content, str) or not content.strip():
+        return None
+
+    metadata = data.get('metadata') or {}
+    document_metadata = {'source': get_firecrawl_result_url(data) or url}
+    if metadata.get('title'):
+        document_metadata['title'] = metadata['title']
+    if metadata.get('description'):
+        document_metadata['description'] = metadata['description']
+
+    return Document(page_content=content, metadata=document_metadata)
+

 def search_firecrawl(
    firecrawl_url: str,
    firecrawl_api_key: str,
    query: str,
    count: int,
-    filter_list: Optional[List[str]] = None,
-) -> List[SearchResult]:
+    filter_list: list[str] | None = None,
+) -> list[SearchResult]:
    try:
-        url = firecrawl_url.rstrip('/')
-        response = requests.post(
-            f'{url}/v1/search',
-            headers={
-                'Content-Type': 'application/json',
-                'Authorization': f'Bearer {firecrawl_api_key}',
-            },
+        response = request_firecrawl_json(
+            'POST',
+            build_firecrawl_url(firecrawl_url, 'search'),
+            headers=build_firecrawl_headers(firecrawl_api_key),
            json={
                'query': query,
                'limit': count,
                'timeout': count * 3000,
+                'ignoreInvalidURLs': True,
            },
            timeout=count * 3 + 10,
        )
-        response.raise_for_status()
-        data = response.json().get('data', {})
-
-        results = [
-            SearchResult(
-                link=r.get('url', ''),
-                title=r.get('title', ''),
-                snippet=r.get('description', ''),
-            )
-            for r in data.get('web', [])
-        ]
+        data = response.get('data') or {}
+        results = data.get('web') or []

        if filter_list:
+            from open_webui.retrieval.web.main import get_filtered_results
+
            results = get_filtered_results(results, filter_list)

-        results = results[:count]
-        log.info(f'FireCrawl search results: {results}')
-        return results
+        from open_webui.retrieval.web.main import SearchResult
+
+        search_results = []
+        for result in results[:count]:
+            url = get_firecrawl_result_url(result)
+            if not url:
+                continue
+
+            metadata = result.get('metadata') or {}
+            search_results.append(
+                SearchResult(
+                    link=url,
+                    title=result.get('title') or metadata.get('title'),
+                    snippet=result.get('description') or result.get('snippet') or metadata.get('description'),
+                )
+            )
+
+        log.info(f'FireCrawl search results: {search_results}')
+        return search_results
    except Exception as e:
        log.error(f'Error in FireCrawl search: {e}')
        return []
@@ -38,7 +38,7 @@ def search_searxng(
    """

    # Default values for optional parameters are provided as empty strings or None when not specified.
-    language = kwargs.get('language', 'all')
+    language = kwargs.get('language', 'all').strip().rstrip(',')
    safesearch = kwargs.get('safesearch', '1')
    time_range = kwargs.get('time_range', '')
    categories = ''.join(kwargs.get('categories', []))
@@ -5,6 +5,8 @@ import socket
 import ssl
 import urllib.parse
 import urllib.request
+
+import requests
 from datetime import datetime, time, timedelta
 from typing import (
    Any,
@@ -28,6 +30,7 @@ from langchain_core.documents import Document

 from open_webui.retrieval.loaders.tavily import TavilyLoader
 from open_webui.retrieval.loaders.external_web import ExternalWebLoader
+from open_webui.retrieval.web.firecrawl import scrape_firecrawl_url
 from open_webui.constants import ERROR_MESSAGES
 from open_webui.config import (
    ENABLE_RAG_LOCAL_WEB_FETCH,
@@ -216,39 +219,20 @@ class SafeFireCrawlLoader(BaseLoader, RateLimitMixin, URLProcessingMixin):

    def lazy_load(self) -> Iterator[Document]:
        try:
-            headers = {
-                'Content-Type': 'application/json',
-                'Authorization': f'Bearer {self.api_key}',
-            }
-
            for url in self.web_paths:
-                payload = {
-                    'url': url,
-                    'formats': ['markdown'],
-                    **self.params,
-                }
-                if self.timeout:
-                    payload['timeout'] = self.timeout * 1000
-
-                response = requests.post(
-                    f'{self.api_url}/v1/scrape',
-                    headers=headers,
-                    json=payload,
-                    timeout=self.timeout or 60,
-                    verify=self.verify_ssl,
-                )
-                response.raise_for_status()
-                data = response.json().get('data', {})
-                metadata = data.get('metadata', {})
-                source = metadata.get('url') or metadata.get('sourceURL') or url
-
-                yield Document(
-                    page_content=data.get('markdown', ''),
-                    metadata={'source': source},
+                doc = scrape_firecrawl_url(
+                    self.api_url,
+                    self.api_key,
+                    url,
+                    verify_ssl=self.verify_ssl,
+                    timeout=self.timeout,
+                    params=self.params,
                )
+                if doc is not None:
+                    yield doc
        except Exception as e:
            if self.continue_on_failure:
-                log.exception(f'Error extracting content from URLs: {e}')
+                log.warning(f'Error extracting content from URLs with Firecrawl: {e}')
            else:
                raise e

@@ -259,7 +243,7 @@ class SafeFireCrawlLoader(BaseLoader, RateLimitMixin, URLProcessingMixin):
                yield doc
        except Exception as e:
            if self.continue_on_failure:
-                log.exception(f'Error extracting content from URLs: {e}')
+                log.warning(f'Error extracting content from URLs with Firecrawl: {e}')
            else:
                raise e

@@ -20,6 +20,8 @@ log = logging.getLogger(__name__)


 def xml_element_contents_to_string(element: Element) -> str:
+    if element is None:
+        return ''
    buffer = [element.text if element.text else '']

    for child in element:
@@ -1,3 +1,5 @@
+import asyncio
+import io
 import hashlib
 import json
 import logging
@@ -127,6 +129,54 @@ def convert_audio_to_mp3(file_path):
        return None


+def transcode_audio_to_mp3(audio_data: bytes, content_type_header: str, output_path: str) -> bool:
+    """
+    Transcode audio bytes to MP3 if the Content-Type indicates a non-MP3 format.
+
+    Handles raw PCM audio (e.g. Gemini-TTS via OpenRouter/LiteLLM) by parsing
+    optional rate/channels from the Content-Type params, defaulting to 24kHz,
+    16-bit, mono. For other non-MP3 formats, uses pydub auto-detection.
+
+    Returns True if transcoding was performed, False if the data is already MP3.
+    Respects BYPASS_PYDUB_PREPROCESSING — when set, writes raw bytes and logs a warning.
+    """
+    mime_type = content_type_header.split(';')[0].strip().lower()
+
+    if mime_type in ('audio/mpeg', 'audio/mp3'):
+        return False
+
+    if BYPASS_PYDUB_PREPROCESSING:
+        log.warning(
+            f'TTS returned {mime_type} but BYPASS_PYDUB_PREPROCESSING is set; writing raw audio without transcoding'
+        )
+        return False
+
+    if mime_type in ('audio/pcm', 'audio/l16', 'audio/raw'):
+        # Parse optional rate/channels from Content-Type params,
+        # default: 24kHz, 16-bit, mono (standard for Gemini TTS).
+        ct_params = {}
+        for part in content_type_header.split(';')[1:]:
+            key_val = part.strip().split('=')
+            if len(key_val) == 2:
+                ct_params[key_val[0].strip().lower()] = key_val[1].strip()
+
+        sample_rate = int(ct_params.get('rate', 24000))
+        channels = int(ct_params.get('channels', 1))
+
+        audio_segment = AudioSegment.from_raw(
+            io.BytesIO(audio_data),
+            sample_width=2,
+            frame_rate=sample_rate,
+            channels=channels,
+        )
+    else:
+        audio_segment = AudioSegment.from_file(io.BytesIO(audio_data))
+
+    audio_segment.export(str(output_path), format='mp3')
+    log.info(f'Transcoded {mime_type} audio to MP3: {output_path}')
+    return True
+
+
 def set_faster_whisper_model(model: str, auto_update: bool = False):
    whisper_model = None
    if model:
@@ -178,6 +228,7 @@ class STTConfigForm(BaseModel):
    ENGINE: str
    MODEL: str
    SUPPORTED_CONTENT_TYPES: list[str] = []
+    ALLOWED_EXTENSIONS: list[str] = []
    WHISPER_MODEL: str
    DEEPGRAM_API_KEY: str
    AZURE_API_KEY: str
@@ -219,6 +270,7 @@ async def get_audio_config(request: Request, user=Depends(get_admin_user)):
            'ENGINE': request.app.state.config.STT_ENGINE,
            'MODEL': request.app.state.config.STT_MODEL,
            'SUPPORTED_CONTENT_TYPES': request.app.state.config.STT_SUPPORTED_CONTENT_TYPES,
+            'ALLOWED_EXTENSIONS': request.app.state.config.STT_ALLOWED_EXTENSIONS,
            'WHISPER_MODEL': request.app.state.config.WHISPER_MODEL,
            'DEEPGRAM_API_KEY': request.app.state.config.DEEPGRAM_API_KEY,
            'AZURE_API_KEY': request.app.state.config.AUDIO_STT_AZURE_API_KEY,
@@ -254,6 +306,7 @@ async def update_audio_config(request: Request, form_data: AudioConfigUpdateForm
    request.app.state.config.STT_ENGINE = form_data.stt.ENGINE
    request.app.state.config.STT_MODEL = form_data.stt.MODEL
    request.app.state.config.STT_SUPPORTED_CONTENT_TYPES = form_data.stt.SUPPORTED_CONTENT_TYPES
+    request.app.state.config.STT_ALLOWED_EXTENSIONS = form_data.stt.ALLOWED_EXTENSIONS

    request.app.state.config.WHISPER_MODEL = form_data.stt.WHISPER_MODEL
    request.app.state.config.DEEPGRAM_API_KEY = form_data.stt.DEEPGRAM_API_KEY
@@ -295,6 +348,7 @@ async def update_audio_config(request: Request, form_data: AudioConfigUpdateForm
            'ENGINE': request.app.state.config.STT_ENGINE,
            'MODEL': request.app.state.config.STT_MODEL,
            'SUPPORTED_CONTENT_TYPES': request.app.state.config.STT_SUPPORTED_CONTENT_TYPES,
+            'ALLOWED_EXTENSIONS': request.app.state.config.STT_ALLOWED_EXTENSIONS,
            'WHISPER_MODEL': request.app.state.config.WHISPER_MODEL,
            'DEEPGRAM_API_KEY': request.app.state.config.DEEPGRAM_API_KEY,
            'AZURE_API_KEY': request.app.state.config.AUDIO_STT_AZURE_API_KEY,
@@ -387,8 +441,12 @@ async def speech(request: Request, user=Depends(get_verified_user)):

                r.raise_for_status()

-                async with aiofiles.open(file_path, 'wb') as f:
-                    await f.write(await r.read())
+                audio_data = await r.read()
+                content_type_header = r.headers.get('Content-Type', 'audio/mpeg')
+
+                if not transcode_audio_to_mp3(audio_data, content_type_header, file_path):
+                    async with aiofiles.open(file_path, 'wb') as f:
+                        await f.write(audio_data)

                async with aiofiles.open(file_body_path, 'w') as f:
                    await f.write(json.dumps(payload))
@@ -576,7 +634,7 @@ async def speech(request: Request, user=Depends(get_verified_user)):
            async with aiohttp.ClientSession(timeout=timeout, trust_env=True) as session:
                mistral_payload = {
                    'input': payload.get('input', ''),
-                    'model': request.app.state.config.TTS_MODEL or 'mistral-tts-latest',
+                    'model': request.app.state.config.TTS_MODEL or 'voxtral-mini-tts-2603',
                    'voice_id': payload.get('voice', ''),
                    'response_format': 'mp3',
                }
@@ -1107,6 +1165,11 @@ def transcribe(request: Request, file_path: str, metadata: Optional[dict] = None
    else:
        if is_audio_conversion_required(file_path):
            file_path = convert_audio_to_mp3(file_path)
+            if not file_path:
+                raise HTTPException(
+                    status_code=status.HTTP_400_BAD_REQUEST,
+                    detail='Audio conversion failed. The audio file may be corrupted or empty.',
+                )

        try:
            file_path = compress_audio(file_path)
@@ -1126,7 +1189,12 @@ def transcribe(request: Request, file_path: str, metadata: Optional[dict] = None

    results = []
    try:
-        with ThreadPoolExecutor() as executor:
+        if getattr(request.app.state.config, 'STT_ENGINE', '') == '':
+            max_workers = 1
+        else:
+            max_workers = None
+
+        with ThreadPoolExecutor(max_workers=max_workers) as executor:
            # Submit tasks for each chunk_path
            futures = [
                executor.submit(transcription_handler, request, chunk_path, metadata, user)
@@ -1242,12 +1310,19 @@ async def transcription(

    try:
        safe_name = os.path.basename(file.filename) if file.filename else ''
-        ext = safe_name.rsplit('.', 1)[-1] if '.' in safe_name else ''
+        ext = safe_name.rsplit('.', 1)[-1].lower() if '.' in safe_name else ''
+
+        allowed_extensions = getattr(request.app.state.config, 'STT_ALLOWED_EXTENSIONS', [])
+        if allowed_extensions and ext not in allowed_extensions:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail='Invalid audio file extension',
+            )

        id = uuid.uuid4()

        filename = f'{id}.{ext}'
-        contents = file.file.read()
+        contents = await file.read()

        file_dir = os.path.join(CACHE_DIR, 'audio', 'transcriptions')
        os.makedirs(file_dir, exist_ok=True)
@@ -1266,7 +1341,7 @@ async def transcription(
            if language:
                metadata = {'language': language}

-            result = transcribe(request, file_path, metadata, user)
+            result = await asyncio.to_thread(transcribe, request, file_path, metadata, user)

            return {
                **result,
@@ -1345,7 +1420,7 @@ async def get_available_models(request: Request) -> list[dict]:
        except Exception as e:
            log.error(f'Error fetching models: {str(e)}')
    elif request.app.state.config.TTS_ENGINE == 'mistral':
-        available_models = [{'id': 'mistral-tts-latest'}]
+        available_models = [{'id': 'voxtral-mini-tts-2603'}]
    return available_models


@@ -1431,11 +1506,14 @@ async def get_available_voices(request) -> dict:
                        response.raise_for_status()
                        voices_data = await response.json()

-                        for voice in voices_data:
-                            voice_id = voice.get('voice_id', voice.get('id', ''))
-                            voice_name = voice.get('name', voice_id)
-                            if voice_id:
-                                available_voices[voice_id] = voice_name
+                        # Mistral returns a paginated response: {"items": [...], "page": ..., "total": ...}
+                        voices_list = voices_data.get('items', []) if isinstance(voices_data, dict) else voices_data
+                        for voice in voices_list:
+                            if isinstance(voice, dict):
+                                voice_id = voice.get('voice_id', voice.get('id', ''))
+                                voice_name = voice.get('name', voice_id)
+                                if voice_id:
+                                    available_voices[voice_id] = voice_name
            except Exception as e:
                log.error(f'Error fetching Mistral voices: {str(e)}')

@@ -172,10 +172,17 @@ async def get_session_user(
    user=Depends(get_current_user),
    db: AsyncSession = Depends(get_async_session),
 ):
+    token = None
    auth_header = request.headers.get('Authorization')
-    auth_token = get_http_authorization_cred(auth_header)
-    token = auth_token.credentials
-    data = decode_token(token)
+    if auth_header:
+        auth_token = get_http_authorization_cred(auth_header)
+        if auth_token is not None:
+            token = auth_token.credentials
+    if token is None:
+        token = request.cookies.get('token')
+    if token is None and getattr(request.state, 'token', None):
+        token = request.state.token.credentials
+    data = decode_token(token) if token else None

    expires_at = None

@@ -515,6 +522,18 @@ async def ldap_auth(
                        db=db,
                    )

+                    if request.app.state.config.WEBHOOK_URL:
+                        await post_webhook(
+                            request.app.state.WEBUI_NAME,
+                            request.app.state.config.WEBHOOK_URL,
+                            WEBHOOK_MESSAGES.USER_SIGNUP(user.name),
+                            {
+                                'action': 'signup',
+                                'message': WEBHOOK_MESSAGES.USER_SIGNUP(user.name),
+                                'user': user.model_dump_json(exclude_none=True),
+                            },
+                        )
+
                except HTTPException:
                    raise
                except Exception as err:
@@ -766,15 +785,16 @@ async def signup(
        raise HTTPException(500, detail='An internal error occurred during signup.')


-@router.get('/signout')
+@router.post('/signout')
 async def signout(request: Request, response: Response, db: AsyncSession = Depends(get_async_session)):
    # get auth token from headers or cookies
    token = None
    auth_header = request.headers.get('Authorization')
    if auth_header:
        auth_cred = get_http_authorization_cred(auth_header)
-        token = auth_cred.credentials
-    else:
+        if auth_cred is not None:
+            token = auth_cred.credentials
+    if token is None:
        token = request.cookies.get('token')

    if token:
@@ -853,6 +873,31 @@ async def signout(request: Request, response: Response, db: AsyncSession = Depen
    return JSONResponse(status_code=200, content={'status': True}, headers=response.headers)


+############################
+# OAuth Session Management
+############################
+
+
+@router.delete('/oauth/sessions/{provider:path}', response_model=bool)
+async def delete_oauth_session_by_provider(
+    provider: str,
+    user=Depends(get_verified_user),
+    db: AsyncSession = Depends(get_async_session),
+):
+    """
+    Disconnect the current user's OAuth session for a specific provider.
+    The provider string matches the 'provider' field in the oauth_session table
+    (e.g. 'mcp:server-id' for MCP connections).
+    """
+    result = await OAuthSessions.delete_sessions_by_user_id_and_provider(user.id, provider, db=db)
+    if not result:
+        raise HTTPException(
+            status_code=status.HTTP_404_NOT_FOUND,
+            detail='No OAuth session found for this provider',
+        )
+    return True
+
+
 ############################
 # AddUser
 ############################
@@ -8,6 +8,7 @@ from fastapi.responses import StreamingResponse


 from open_webui.utils.misc import get_message_list
+from open_webui.utils.middleware import serialize_output
 from open_webui.socket.main import get_event_emitter
 from open_webui.models.chats import (
    ChatForm,
@@ -676,11 +677,46 @@ async def get_user_pinned_chats(user=Depends(get_verified_user), db: AsyncSessio
 # GetChats
 ############################

+CHAT_EXPORT_BATCH_SIZE = 100

-@router.get('/all', response_model=list[ChatResponse])
-async def get_user_chats(user=Depends(get_verified_user), db: AsyncSession = Depends(get_async_session)):
-    result = await Chats.get_chats_by_user_id(user.id, db=db)
-    return [ChatResponse(**chat.model_dump()) for chat in result.items]
+
+async def generate_chat_export_ndjson(user_id: str):
+    """
+    Async generator that streams all user chats as NDJSON (one JSON object per line).
+
+    Uses short-lived DB sessions per batch to avoid holding locks for the
+    entire duration, which is critical for SQLite environments.
+    """
+    skip = 0
+
+    while True:
+        result = await Chats.get_chats_by_user_id(
+            user_id,
+            skip=skip,
+            limit=CHAT_EXPORT_BATCH_SIZE,
+            db=None,
+        )
+        if not result.items:
+            break
+
+        for chat in result.items:
+            try:
+                yield ChatResponse(**chat.model_dump()).model_dump_json() + '\n'
+            except Exception as e:
+                log.exception(f'Error serializing chat {chat.id}: {e}')
+
+        if len(result.items) < CHAT_EXPORT_BATCH_SIZE:
+            break
+
+        skip += CHAT_EXPORT_BATCH_SIZE
+
+
+@router.get('/all')
+async def get_user_chats(user=Depends(get_verified_user)):
+    return StreamingResponse(
+        generate_chat_export_ndjson(user.id),
+        media_type='application/x-ndjson',
+    )


 ############################
@@ -829,29 +865,31 @@ async def get_shared_chat_by_id(
    if user.role == 'pending':
        raise HTTPException(status_code=status.HTTP_401_UNAUTHORIZED, detail=ERROR_MESSAGES.NOT_FOUND)

-    if user.role == 'admin' and ENABLE_ADMIN_CHAT_ACCESS:
+    chat = await Chats.get_chat_by_share_id(share_id, db=db)
+
+    # Fallback: admins can also access any chat directly by chat ID
+    if not chat and user.role == 'admin' and ENABLE_ADMIN_CHAT_ACCESS:
        chat = await Chats.get_chat_by_id(share_id, db=db)
-    else:
-        chat = await Chats.get_chat_by_share_id(share_id, db=db)

    if not chat:
        raise HTTPException(status_code=status.HTTP_401_UNAUTHORIZED, detail=ERROR_MESSAGES.NOT_FOUND)

-    # Look up the original chat_id to check access grants
-    shared = await SharedChats.get_by_id(share_id, db=db)
-    if shared:
-        has_grant = await AccessGrants.has_access(
-            user_id=user.id,
-            resource_type='shared_chat',
-            resource_id=shared.chat_id,
-            permission='read',
-            db=db,
-        )
-        if not has_grant:
-            raise HTTPException(
-                status_code=status.HTTP_401_UNAUTHORIZED,
-                detail=ERROR_MESSAGES.ACCESS_PROHIBITED,
+    # Look up the original chat_id to check access grants (admins bypass)
+    if user.role != 'admin' or not ENABLE_ADMIN_CHAT_ACCESS:
+        shared = await SharedChats.get_by_id(share_id, db=db)
+        if shared:
+            has_grant = await AccessGrants.has_access(
+                user_id=user.id,
+                resource_type='shared_chat',
+                resource_id=shared.chat_id,
+                permission='read',
+                db=db,
            )
+            if not has_grant:
+                raise HTTPException(
+                    status_code=status.HTTP_401_UNAUTHORIZED,
+                    detail=ERROR_MESSAGES.ACCESS_PROHIBITED,
+                )

    return ChatResponse(**chat.model_dump())

@@ -930,6 +968,14 @@ async def update_chat_by_id(
    chat = await Chats.get_chat_by_id_and_user_id(id, user.id, db=db)
    if chat:
        updated_chat = {**chat.chat, **form_data.chat}
+
+        # Re-derive content from output for assistant messages so that
+        # frontend edits to output items are always reflected in content.
+        # serialize_output() is the single source of truth for this conversion.
+        for msg in updated_chat.get('history', {}).get('messages', {}).values():
+            if msg.get('role') == 'assistant' and msg.get('output'):
+                msg['content'] = serialize_output(msg['output'])
+
        chat = await Chats.update_chat_by_id(id, updated_chat, db=db)
        return ChatResponse(**chat.model_dump())
    else:
@@ -1183,10 +1229,11 @@ async def clone_chat_by_id(
 async def clone_shared_chat_by_id(
    id: str, user=Depends(get_verified_user), db: AsyncSession = Depends(get_async_session)
 ):
-    if user.role == 'admin':
+    chat = await Chats.get_chat_by_share_id(id, db=db)
+
+    # Fallback: admins can also access any chat directly by chat ID
+    if not chat and user.role == 'admin' and ENABLE_ADMIN_CHAT_ACCESS:
        chat = await Chats.get_chat_by_id(id, db=db)
-    else:
-        chat = await Chats.get_chat_by_share_id(id, db=db)

    if not chat:
        raise HTTPException(
@@ -8,6 +8,7 @@ from typing import Optional

 from open_webui.env import AIOHTTP_CLIENT_SESSION_SSL, AIOHTTP_CLIENT_TIMEOUT
 from open_webui.utils.auth import get_admin_user, get_verified_user
+from open_webui.utils.headers import get_custom_headers
 from open_webui.config import get_config, save_config, async_save_config
 from open_webui.config import BannerModel

@@ -49,8 +50,9 @@ class ImportConfigForm(BaseModel):


@router.post('/import', response_model=dict)
-async def import_config(form_data: ImportConfigForm, user=Depends(get_admin_user)):
+async def import_config(request: Request, form_data: ImportConfigForm, user=Depends(get_admin_user)):
    await async_save_config(form_data.config)
+    request.app.state.config._sync_to_redis()
    return get_config()


@@ -102,6 +104,7 @@ class OAuthClientRegistrationForm(BaseModel):
    client_id: str
    client_name: Optional[str] = None
    client_secret: Optional[str] = None
+    oauth_server_url: Optional[str] = None


@router.post('/oauth/clients/register')
@@ -116,18 +119,20 @@ async def register_oauth_client(
        if type:
            oauth_client_id = f'{type}:{form_data.client_id}'

+        oauth_server_url = form_data.oauth_server_url if form_data.oauth_server_url else form_data.url
+
        if form_data.client_secret:
            # Static credentials: skip dynamic registration, build from provided credentials
            oauth_client_info = await get_oauth_client_info_with_static_credentials(
                request,
                oauth_client_id,
-                form_data.url,
+                oauth_server_url,
                oauth_client_id=form_data.client_id,
                oauth_client_secret=form_data.client_secret,
            )
        else:
            oauth_client_info = await get_oauth_client_info_with_dynamic_client_registration(
-                request, oauth_client_id, form_data.url
+                request, oauth_client_id, oauth_server_url
            )
        return {
            'status': True,
@@ -154,6 +159,7 @@ class ToolServerConnection(BaseModel):
    headers: Optional[dict | str] = None
    key: Optional[str]
    config: Optional[dict]
+    info: Optional[dict] = None

    model_config = ConfigDict(extra='allow')

@@ -368,7 +374,12 @@ async def verify_tool_servers_config(request: Request, form_data: ToolServerConn
    try:
        if form_data.type == 'mcp':
            if form_data.auth_type in ('oauth_2.1', 'oauth_2.1_static'):
-                discovery_urls = await get_discovery_urls(form_data.url)
+                oauth_server_url = (
+                    form_data.info.get('oauth_server_url')
+                    if form_data.info and form_data.info.get('oauth_server_url')
+                    else form_data.url
+                )
+                discovery_urls = await get_discovery_urls(oauth_server_url)
                for discovery_url in discovery_urls:
                    log.debug(f'Trying to fetch OAuth 2.1 discovery document from {discovery_url}')
                    async with aiohttp.ClientSession(
@@ -427,7 +438,8 @@ async def verify_tool_servers_config(request: Request, form_data: ToolServerConn
                    if form_data.headers and isinstance(form_data.headers, dict):
                        if headers is None:
                            headers = {}
-                        headers.update(form_data.headers)
+                        custom_headers = get_custom_headers(form_data.headers, user)
+                        headers.update(custom_headers)

                    await client.connect(form_data.url, headers=headers)
                    specs = await client.list_tool_specs()
@@ -471,7 +483,8 @@ async def verify_tool_servers_config(request: Request, form_data: ToolServerConn
            if form_data.headers and isinstance(form_data.headers, dict):
                if headers is None:
                    headers = {}
-                headers.update(form_data.headers)
+                custom_headers = get_custom_headers(form_data.headers, user)
+                headers.update(custom_headers)

            url = get_tool_server_url(form_data.url, form_data.path)
            return await get_tool_server_data(url, headers=headers)
@@ -125,7 +125,13 @@ async def process_uploaded_file(

                if strict_match_mime_type(stt_supported_content_types, content_type):
                    file_path_processed = await asyncio.to_thread(Storage.get_file, file_path)
-                    result = transcribe(request, file_path_processed, file_metadata, user)
+                    result = await asyncio.to_thread(
+                        transcribe,
+                        request,
+                        file_path_processed,
+                        file_metadata,
+                        user,
+                    )

                    await process_file(
                        request,
@@ -4,6 +4,8 @@ import base64
 import json
 import asyncio
 import logging
+import posixpath
+from urllib.parse import unquote

 from open_webui.models.groups import Groups
 from open_webui.models.models import (
@@ -26,15 +28,16 @@ from fastapi import (
    Depends,
    HTTPException,
    Request,
-    status,
    Response,
+    status,
 )
-from fastapi.responses import FileResponse, StreamingResponse
+from fastapi.responses import RedirectResponse, StreamingResponse


 from open_webui.utils.auth import get_admin_user, get_verified_user
 from open_webui.utils.access_control import has_permission, filter_allowed_access_grants
-from open_webui.config import BYPASS_ADMIN_ACCESS_CONTROL, STATIC_DIR
+from open_webui.config import BYPASS_ADMIN_ACCESS_CONTROL
+from open_webui.env import ENABLE_PROFILE_IMAGE_URL_FORWARDING
 from open_webui.internal.db import get_async_session
 from sqlalchemy.ext.asyncio import AsyncSession

@@ -43,6 +46,34 @@ log = logging.getLogger(__name__)
 router = APIRouter()


+def _safe_static_redirect_path(url: str) -> Optional[str]:
+    """
+    If url is a same-origin static asset path, return a normalized path safe for
+    RedirectResponse Location. Otherwise None (caller should fall back to default).
+    Rejects traversal (..), encoded dots, query/fragment, and non-/static targets.
+    """
+    if not url or not isinstance(url, str):
+        return None
+    path = url.split('?', 1)[0].split('#', 1)[0].strip()
+    for _ in range(2):
+        decoded = unquote(path)
+        if decoded == path:
+            break
+        path = decoded
+    if '\x00' in path or '\\' in path:
+        return None
+    if not path.startswith('/'):
+        return None
+    normalized = posixpath.normpath(path)
+    if normalized in ('.', '/'):
+        return None
+    if not (normalized == '/static' or normalized.startswith('/static/')):
+        return None
+    if normalized == '/static':
+        return '/static/'
+    return normalized
+
+
 def is_valid_model_id(model_id: str) -> bool:
    return model_id and len(model_id) <= 256

@@ -108,18 +139,25 @@ async def get_models(
        db=db,
    )

-    return ModelAccessListResponse(
-        items=[
+    # Strip profile_image_url from meta — images are served via /model/profile/image.
+    items = []
+    for model in result.items:
+        data = model.model_dump()
+        if data.get('meta'):
+            data['meta'].pop('profile_image_url', None)
+        items.append(
            ModelAccessResponse(
-                **model.model_dump(),
+                **data,
                write_access=(
                    (user.role == 'admin' and BYPASS_ADMIN_ACCESS_CONTROL)
                    or user.id == model.user_id
                    or model.id in writable_model_ids
                ),
            )
-            for model in result.items
-        ],
+        )
+
+    return ModelAccessListResponse(
+        items=items,
        total=result.total,
    )

@@ -141,25 +179,12 @@ async def get_base_models(user=Depends(get_admin_user), db: AsyncSession = Depen

@router.get('/tags', response_model=list[str])
 async def get_model_tags(user=Depends(get_verified_user), db: AsyncSession = Depends(get_async_session)):
-    if user.role == 'admin' and BYPASS_ADMIN_ACCESS_CONTROL:
-        models = await Models.get_models(db=db)
-    else:
-        models = await Models.get_models_by_user_id(user.id, db=db)
-
-    tags_set = set()
-    for model in models:
-        if model.meta:
-            meta = model.meta.model_dump()
-            for tag in meta.get('tags', []):
-                try:
-                    name = tag.get('name') if isinstance(tag, dict) else str(tag)
-                    if name:
-                        tags_set.add(name)
-                except Exception:
-                    continue
-
-    tags = sorted(tags_set)
-    return tags
+    tags = await Models.get_all_tags(
+        user_id=user.id,
+        is_admin=(user.role == 'admin' and BYPASS_ADMIN_ACCESS_CONTROL),
+        db=db,
+    )
+    return sorted(tags)


 ############################
@@ -432,43 +457,72 @@ async def get_model_by_id(id: str, user=Depends(get_verified_user), db: AsyncSes

@router.get('/model/profile/image')
 async def get_model_profile_image(
+    request: Request,
    id: str,
    user=Depends(get_verified_user),
    db: AsyncSession = Depends(get_async_session),
 ):
-    model = await Models.get_model_by_id(id, db=db)
+    profile_image_url = None
+    updated_at = None

-    if model:
-        etag = f'"{model.updated_at}"' if model.updated_at else None
+    # First, check the database for regular models
+    model_meta = await Models.get_model_meta_by_id(id, db=db)
+    if model_meta:
+        meta, updated_at = model_meta
+        profile_image_url = (meta or {}).get('profile_image_url')

-        if model.meta.profile_image_url:
-            if model.meta.profile_image_url.startswith('http'):
+    # Fallback: check arena models stored in config (not in the DB)
+    if not profile_image_url:
+        arena_models = getattr(
+            getattr(request.app.state, 'config', None),
+            'EVALUATION_ARENA_MODELS',
+            [],
+        )
+        for arena_model in arena_models:
+            if arena_model.get('id') == id:
+                profile_image_url = arena_model.get('meta', {}).get('profile_image_url')
+                break
+
+    if profile_image_url:
+        if profile_image_url.startswith('http'):
+            if ENABLE_PROFILE_IMAGE_URL_FORWARDING:
                return Response(
                    status_code=status.HTTP_302_FOUND,
-                    headers={'Location': model.meta.profile_image_url},
+                    headers={'Location': profile_image_url},
                )
-            elif model.meta.profile_image_url.startswith('data:image'):
-                try:
-                    header, base64_data = model.meta.profile_image_url.split(',', 1)
-                    image_data = base64.b64decode(base64_data)
-                    image_buffer = io.BytesIO(image_data)
-                    media_type = header.split(';')[0].lstrip('data:')
+            # When forwarding is disabled, fall through to the
+            # default image to prevent client-side IP/UA/Referer
+            # leaks via 302 redirect to external origins.
+        elif profile_image_url.startswith('data:image'):
+            try:
+                header, base64_data = profile_image_url.split(',', 1)
+                image_data = base64.b64decode(base64_data)
+                image_buffer = io.BytesIO(image_data)
+                media_type = header.split(';')[0].lstrip('data:')

-                    headers = {'Content-Disposition': 'inline'}
-                    if etag:
-                        headers['ETag'] = etag
+                headers = {'Content-Disposition': 'inline'}
+                if updated_at:
+                    headers['ETag'] = f'"{updated_at}"'

-                    return StreamingResponse(
-                        image_buffer,
-                        media_type=media_type,
-                        headers=headers,
-                    )
-                except Exception as e:
-                    pass
+                return StreamingResponse(
+                    image_buffer,
+                    media_type=media_type,
+                    headers=headers,
+                )
+            except Exception:
+                pass
+        else:
+            safe_static = _safe_static_redirect_path(profile_image_url)
+            if safe_static:
+                return RedirectResponse(
+                    url=safe_static,
+                    status_code=status.HTTP_302_FOUND,
+                )

-        return FileResponse(f'{STATIC_DIR}/favicon.png')
-    else:
-        return FileResponse(f'{STATIC_DIR}/favicon.png')
+    return RedirectResponse(
+        url='/static/favicon.png',
+        status_code=status.HTTP_302_FOUND,
+    )


 ############################
@@ -92,10 +92,13 @@ async def get_notes(
    user_ids = list(set(note.user_id for note in notes))
    users = {user.id: user for user in await Users.get_users_by_user_ids(user_ids, db=db)}

+    pinned_note_ids = await Notes.get_pinned_note_ids(user.id, db=db)
+
    return [
        NoteUserResponse(
            **{
                **note.model_dump(),
+                'is_pinned': note.id in pinned_note_ids,
                'data': _truncate_note_data(note.data),
                'user': UserResponse(**users[note.user_id].model_dump()),
            }
@@ -135,6 +138,7 @@ async def get_pinned_notes(
        NoteUserResponse(
            **{
                **note.model_dump(),
+                'is_pinned': True,
                'data': _truncate_note_data(note.data),
                'user': UserResponse(**users[note.user_id].model_dump()),
            }
@@ -190,7 +194,9 @@ async def search_notes(
        filter['user_id'] = user.id

    result = await Notes.search_notes(user.id, filter, skip=skip, limit=limit, db=db)
+    pinned_note_ids = await Notes.get_pinned_note_ids(user.id, db=db)
    for note in result.items:
+        note.is_pinned = note.id in pinned_note_ids
        note.data = _truncate_note_data(note.data)
    return result

@@ -287,7 +293,8 @@ async def get_note_by_id(
        or has_public_write_access_grant(note.access_grants)
    )

-    return NoteResponse(**note.model_dump(), write_access=write_access)
+    pinned_note_ids = await Notes.get_pinned_note_ids(user.id, db=db)
+    return NoteResponse(**note.model_dump(), write_access=write_access, is_pinned=note.id in pinned_note_ids)


 ############################
@@ -338,6 +345,9 @@ async def update_note_by_id(

    try:
        note = await Notes.update_note_by_id(id, form_data, db=db)
+        pinned_note_ids = await Notes.get_pinned_note_ids(user.id, db=db)
+        note.is_pinned = note.id in pinned_note_ids
+
        await sio.emit(
            'note-events',
            note.model_dump(),
@@ -401,7 +411,10 @@ async def update_note_access_by_id(

    await AccessGrants.set_access_grants('note', id, form_data.access_grants, db=db)

-    return await Notes.get_note_by_id(id, db=db)
+    note = await Notes.get_note_by_id(id, db=db)
+    pinned_note_ids = await Notes.get_pinned_note_ids(user.id, db=db)
+    note.is_pinned = note.id in pinned_note_ids
+    return note


 ############################
@@ -440,7 +453,9 @@ async def pin_note_by_id(
    ):
        raise HTTPException(status_code=status.HTTP_403_FORBIDDEN, detail=ERROR_MESSAGES.DEFAULT())

-    note = await Notes.toggle_note_pinned_by_id(id, db=db)
+    note = await Notes.toggle_note_pinned_by_id(id, user.id, db=db)
+    pinned_note_ids = await Notes.get_pinned_note_ids(user.id, db=db)
+    note.is_pinned = note.id in pinned_note_ids
    return note


@@ -86,6 +86,17 @@ log = logging.getLogger(__name__)
 #
 ##########################################

+# Headers that become stale after aiohttp auto-decompresses the upstream
+# response body.  Forwarding them verbatim causes desktop / programmatic
+# clients to attempt decompression of an already-decoded payload, resulting
+# in ZlibError.  See https://github.com/aio-libs/aiohttp/issues/4462.
+_STRIP_PROXY_HEADERS = frozenset({'Content-Encoding', 'Content-Length', 'Transfer-Encoding'})
+
+
+def _clean_proxy_headers(raw_headers) -> dict:
+    """Return a copy of *raw_headers* with stale encoding headers removed."""
+    return {k: v for k, v in raw_headers.items() if k not in _STRIP_PROXY_HEADERS}
+

 async def send_get_request(url, key=None, user: UserModel = None):
    timeout = aiohttp.ClientTimeout(total=AIOHTTP_CLIENT_TIMEOUT_MODEL_LIST)
@@ -163,7 +174,7 @@ async def send_request(
        r.raise_for_status()

        if stream:
-            response_headers = dict(r.headers)
+            response_headers = _clean_proxy_headers(r.headers)
            if content_type:
                response_headers['Content-Type'] = content_type

@@ -1116,7 +1127,7 @@ async def generate_chat_completion(

            payload = apply_model_params_to_body_ollama(params, payload)
            if not bypass_system_prompt:
-                payload = apply_system_prompt_to_body(system, payload, metadata, user)
+                payload = await apply_system_prompt_to_body(system, payload, metadata, user)

        await check_model_access(user, model_info, bypass_filter)
    else:
@@ -1271,7 +1282,7 @@ async def generate_openai_chat_completion(
            system = params.pop('system', None)

            payload = apply_model_params_to_body_openai(params, payload)
-            payload = apply_system_prompt_to_body(system, payload, metadata, user)
+            payload = await apply_system_prompt_to_body(system, payload, metadata, user)

        await check_model_access(user, model_info)
    else:
@@ -62,7 +62,7 @@ from open_webui.utils.session_pool import (
 )

 from open_webui.utils.auth import get_admin_user, get_verified_user
-from open_webui.utils.headers import include_user_info_headers
+from open_webui.utils.headers import include_user_info_headers, get_custom_headers
 from open_webui.utils.anthropic import is_anthropic_url, get_anthropic_models

 log = logging.getLogger(__name__)
@@ -76,6 +76,17 @@ log = logging.getLogger(__name__)
 #
 ##########################################

+# Headers that become stale after aiohttp auto-decompresses the upstream
+# response body.  Forwarding them verbatim causes desktop / programmatic
+# clients to attempt decompression of an already-decoded payload, resulting
+# in ZlibError.  See https://github.com/aio-libs/aiohttp/issues/4462.
+_STRIP_PROXY_HEADERS = frozenset({'Content-Encoding', 'Content-Length', 'Transfer-Encoding'})
+
+
+def _clean_proxy_headers(raw_headers) -> dict:
+    """Return a copy of *raw_headers* with stale encoding headers removed."""
+    return {k: v for k, v in raw_headers.items() if k not in _STRIP_PROXY_HEADERS}
+

 async def send_get_request(
    request: Request = None,
@@ -204,7 +215,8 @@ async def get_headers_and_cookies(
        headers['Authorization'] = f'Bearer {token}'

    if config.get('headers') and isinstance(config.get('headers'), dict):
-        headers = {**headers, **config.get('headers')}
+        custom_headers = get_custom_headers(config.get('headers'), user, metadata)
+        headers.update(custom_headers)

    return headers, cookies

@@ -428,6 +440,7 @@ async def get_all_models_responses(request: Request, user: UserModel) -> list:
            connection_type = api_config.get('connection_type', 'external')
            prefix_id = api_config.get('prefix_id', None)
            tags = api_config.get('tags', [])
+            provider = api_config.get('provider', '')

            model_list = response if isinstance(response, list) else response.get('data', [])
            if not isinstance(model_list, list):
@@ -448,6 +461,9 @@ async def get_all_models_responses(request: Request, user: UserModel) -> list:
                if connection_type:
                    model['connection_type'] = connection_type

+                if provider:
+                    model['provider'] = provider
+
    log.debug(f'get_all_models:responses() {responses}')
    return responses

@@ -477,6 +493,39 @@ async def get_filtered_models(models, user, db=None):
    return filtered_models


+async def get_openai_loaded_models(request: Request, models: dict, api_base_urls: list):
+    """
+    Fetch loaded-model state from providers that expose it and annotate
+    each model dict with a ``loaded`` boolean.
+
+    Currently supports:
+      - **llama.cpp** – queries ``GET /slots`` and matches slot model IDs.
+    """
+    api_configs = request.app.state.config.OPENAI_API_CONFIGS
+    api_keys = request.app.state.config.OPENAI_API_KEYS
+
+    for idx, url in enumerate(api_base_urls):
+        api_config = api_configs.get(
+            str(idx),
+            api_configs.get(url, {}),
+        )
+        provider = api_config.get('provider', '')
+
+        if provider == 'llama.cpp':
+            try:
+                root_url = url.rstrip('/').removesuffix('/v1')
+                key = api_keys[idx] if idx < len(api_keys) else None
+                slots = await send_get_request(url=f'{root_url}/slots', key=key)
+                loaded_model_ids = (
+                    {s.get('model') for s in slots if s.get('model')} if isinstance(slots, list) else set()
+                )
+                for model_id, model in models.items():
+                    if model.get('urlIdx') == idx:
+                        model['loaded'] = model_id in loaded_model_ids
+            except Exception as e:
+                log.debug(f'Failed to fetch llama.cpp slots for idx {idx}: {e}')
+
+
@cached(
    ttl=MODELS_CACHE_TTL,
    key=lambda _, user: f'openai_all_models_{user.id}' if user else 'openai_all_models',
@@ -537,6 +586,7 @@ async def get_all_models(request: Request, user: UserModel) -> dict[str, list]:
                            'owned_by': 'openai',
                            'openai': model,
                            'connection_type': model.get('connection_type', 'external'),
+                            'provider': model.get('provider', ''),
                            'urlIdx': idx,
                        }

@@ -545,6 +595,9 @@ async def get_all_models(request: Request, user: UserModel) -> dict[str, list]:
    models = get_merged_models(map(extract_data, responses))
    log.debug(f'models: {models}')

+    # Fetch loaded state for providers that support it (e.g. llama.cpp /slots)
+    await get_openai_loaded_models(request, models, api_base_urls)
+
    request.app.state.OPENAI_MODELS = models
    return {'data': list(models.values())}

@@ -1066,7 +1119,7 @@ async def generate_chat_completion(

            payload = apply_model_params_to_body_openai(params, payload)
            if not bypass_system_prompt:
-                payload = apply_system_prompt_to_body(system, payload, metadata, user)
+                payload = await apply_system_prompt_to_body(system, payload, metadata, user)

        await check_model_access(user, model_info, bypass_filter)
    else:
@@ -1219,7 +1272,7 @@ async def generate_chat_completion(
            return StreamingResponse(
                stream_wrapper(r, content_handler=stream_chunks_handler),
                status_code=r.status,
-                headers=dict(r.headers),
+                headers=_clean_proxy_headers(r.headers),
            )
        else:
            try:
@@ -1304,7 +1357,7 @@ async def embeddings(request: Request, form_data: dict, user):
            return StreamingResponse(
                stream_wrapper(r),
                status_code=r.status,
-                headers=dict(r.headers),
+                headers=_clean_proxy_headers(r.headers),
            )
        else:
            try:
@@ -1425,7 +1478,7 @@ async def responses(
            return StreamingResponse(
                stream_wrapper(r),
                status_code=r.status,
-                headers=dict(r.headers),
+                headers=_clean_proxy_headers(r.headers),
            )
        else:
            try:
@@ -1542,7 +1595,7 @@ async def proxy(path: str, request: Request, user=Depends(get_verified_user)):
            return StreamingResponse(
                stream_wrapper(r),
                status_code=r.status,
-                headers=dict(r.headers),
+                headers=_clean_proxy_headers(r.headers),
            )
        else:
            try:
@@ -61,13 +61,7 @@ async def get_prompts(user=Depends(get_verified_user), db: AsyncSession = Depend
 async def get_prompt_tags(user=Depends(get_verified_user), db: AsyncSession = Depends(get_async_session)):
    if user.role == 'admin' and BYPASS_ADMIN_ACCESS_CONTROL:
        return await Prompts.get_tags(db=db)
-    else:
-        prompts = await Prompts.get_prompts_by_user_id(user.id, 'read', db=db)
-        tags = set()
-        for prompt in prompts:
-            if prompt.tags:
-                tags.update(prompt.tags)
-        return sorted(list(tags))
+    return await Prompts.get_tags_by_user_id(user.id, db=db)


@router.get('/list', response_model=PromptAccessListResponse)
@@ -57,6 +57,7 @@ from open_webui.retrieval.web.utils import get_web_loader
 from open_webui.retrieval.web.ollama import search_ollama_cloud
 from open_webui.retrieval.web.perplexity_search import search_perplexity_search
 from open_webui.retrieval.web.brave import search_brave
+from open_webui.retrieval.web.brave_llm_context import search_brave_llm_context
 from open_webui.retrieval.web.kagi import search_kagi
 from open_webui.retrieval.web.mojeek import search_mojeek
 from open_webui.retrieval.web.bocha import search_bocha
@@ -350,7 +351,7 @@ async def update_embedding_config(request: Request, form_data: EmbeddingModelUpd
    unload_embedding_model(request)
    try:
        request.app.state.config.RAG_EMBEDDING_ENGINE = form_data.RAG_EMBEDDING_ENGINE
-        request.app.state.config.RAG_EMBEDDING_MODEL = form_data.RAG_EMBEDDING_MODEL
+        request.app.state.config.RAG_EMBEDDING_MODEL = form_data.RAG_EMBEDDING_MODEL.strip()
        request.app.state.config.RAG_EMBEDDING_BATCH_SIZE = form_data.RAG_EMBEDDING_BATCH_SIZE
        request.app.state.config.ENABLE_ASYNC_EMBEDDING = form_data.ENABLE_ASYNC_EMBEDDING
        request.app.state.config.RAG_EMBEDDING_CONCURRENT_REQUESTS = form_data.RAG_EMBEDDING_CONCURRENT_REQUESTS
@@ -480,6 +481,8 @@ async def get_rag_config(request: Request, user=Depends(get_admin_user)):
        'DOCUMENT_INTELLIGENCE_MODEL': request.app.state.config.DOCUMENT_INTELLIGENCE_MODEL,
        'MISTRAL_OCR_API_BASE_URL': request.app.state.config.MISTRAL_OCR_API_BASE_URL,
        'MISTRAL_OCR_API_KEY': request.app.state.config.MISTRAL_OCR_API_KEY,
+        'PADDLEOCR_VL_BASE_URL': request.app.state.config.PADDLEOCR_VL_BASE_URL,
+        'PADDLEOCR_VL_TOKEN': request.app.state.config.PADDLEOCR_VL_TOKEN,
        # MinerU settings
        'MINERU_API_MODE': request.app.state.config.MINERU_API_MODE,
        'MINERU_API_URL': request.app.state.config.MINERU_API_URL,
@@ -529,6 +532,7 @@ async def get_rag_config(request: Request, user=Depends(get_admin_user)):
            'GOOGLE_PSE_API_KEY': request.app.state.config.GOOGLE_PSE_API_KEY,
            'GOOGLE_PSE_ENGINE_ID': request.app.state.config.GOOGLE_PSE_ENGINE_ID,
            'BRAVE_SEARCH_API_KEY': request.app.state.config.BRAVE_SEARCH_API_KEY,
+            'BRAVE_SEARCH_CONTEXT_TOKENS': request.app.state.config.BRAVE_SEARCH_CONTEXT_TOKENS,
            'KAGI_SEARCH_API_KEY': request.app.state.config.KAGI_SEARCH_API_KEY,
            'MOJEEK_SEARCH_API_KEY': request.app.state.config.MOJEEK_SEARCH_API_KEY,
            'BOCHA_SEARCH_API_KEY': request.app.state.config.BOCHA_SEARCH_API_KEY,
@@ -597,6 +601,7 @@ class WebConfig(BaseModel):
    GOOGLE_PSE_API_KEY: Optional[str] = None
    GOOGLE_PSE_ENGINE_ID: Optional[str] = None
    BRAVE_SEARCH_API_KEY: Optional[str] = None
+    BRAVE_SEARCH_CONTEXT_TOKENS: Optional[int] = None
    KAGI_SEARCH_API_KEY: Optional[str] = None
    MOJEEK_SEARCH_API_KEY: Optional[str] = None
    BOCHA_SEARCH_API_KEY: Optional[str] = None
@@ -686,6 +691,8 @@ class ConfigForm(BaseModel):
    DOCUMENT_INTELLIGENCE_MODEL: Optional[str] = None
    MISTRAL_OCR_API_BASE_URL: Optional[str] = None
    MISTRAL_OCR_API_KEY: Optional[str] = None
+    PADDLEOCR_VL_BASE_URL: Optional[str] = None
+    PADDLEOCR_VL_TOKEN: Optional[str] = None

    # MinerU settings
    MINERU_API_MODE: Optional[str] = None
@@ -887,6 +894,16 @@ async def update_rag_config(request: Request, form_data: ConfigForm, user=Depend
        if form_data.MISTRAL_OCR_API_KEY is not None
        else request.app.state.config.MISTRAL_OCR_API_KEY
    )
+    request.app.state.config.PADDLEOCR_VL_BASE_URL = (
+        form_data.PADDLEOCR_VL_BASE_URL
+        if form_data.PADDLEOCR_VL_BASE_URL is not None
+        else request.app.state.config.PADDLEOCR_VL_BASE_URL
+    )
+    request.app.state.config.PADDLEOCR_VL_TOKEN = (
+        form_data.PADDLEOCR_VL_TOKEN
+        if form_data.PADDLEOCR_VL_TOKEN is not None
+        else request.app.state.config.PADDLEOCR_VL_TOKEN
+    )

    # MinerU settings
    request.app.state.config.MINERU_API_MODE = (
@@ -1067,6 +1084,8 @@ async def update_rag_config(request: Request, form_data: ConfigForm, user=Depend
        request.app.state.config.GOOGLE_PSE_API_KEY = form_data.web.GOOGLE_PSE_API_KEY
        request.app.state.config.GOOGLE_PSE_ENGINE_ID = form_data.web.GOOGLE_PSE_ENGINE_ID
        request.app.state.config.BRAVE_SEARCH_API_KEY = form_data.web.BRAVE_SEARCH_API_KEY
+        if form_data.web.BRAVE_SEARCH_CONTEXT_TOKENS is not None:
+            request.app.state.config.BRAVE_SEARCH_CONTEXT_TOKENS = form_data.web.BRAVE_SEARCH_CONTEXT_TOKENS
        request.app.state.config.KAGI_SEARCH_API_KEY = form_data.web.KAGI_SEARCH_API_KEY
        request.app.state.config.MOJEEK_SEARCH_API_KEY = form_data.web.MOJEEK_SEARCH_API_KEY
        request.app.state.config.BOCHA_SEARCH_API_KEY = form_data.web.BOCHA_SEARCH_API_KEY
@@ -1152,6 +1171,8 @@ async def update_rag_config(request: Request, form_data: ConfigForm, user=Depend
        'DOCUMENT_INTELLIGENCE_MODEL': request.app.state.config.DOCUMENT_INTELLIGENCE_MODEL,
        'MISTRAL_OCR_API_BASE_URL': request.app.state.config.MISTRAL_OCR_API_BASE_URL,
        'MISTRAL_OCR_API_KEY': request.app.state.config.MISTRAL_OCR_API_KEY,
+        'PADDLEOCR_VL_BASE_URL': request.app.state.config.PADDLEOCR_VL_BASE_URL,
+        'PADDLEOCR_VL_TOKEN': request.app.state.config.PADDLEOCR_VL_TOKEN,
        # MinerU settings
        'MINERU_API_MODE': request.app.state.config.MINERU_API_MODE,
        'MINERU_API_URL': request.app.state.config.MINERU_API_URL,
@@ -1200,6 +1221,7 @@ async def update_rag_config(request: Request, form_data: ConfigForm, user=Depend
            'GOOGLE_PSE_API_KEY': request.app.state.config.GOOGLE_PSE_API_KEY,
            'GOOGLE_PSE_ENGINE_ID': request.app.state.config.GOOGLE_PSE_ENGINE_ID,
            'BRAVE_SEARCH_API_KEY': request.app.state.config.BRAVE_SEARCH_API_KEY,
+            'BRAVE_SEARCH_CONTEXT_TOKENS': request.app.state.config.BRAVE_SEARCH_CONTEXT_TOKENS,
            'KAGI_SEARCH_API_KEY': request.app.state.config.KAGI_SEARCH_API_KEY,
            'MOJEEK_SEARCH_API_KEY': request.app.state.config.MOJEEK_SEARCH_API_KEY,
            'BOCHA_SEARCH_API_KEY': request.app.state.config.BOCHA_SEARCH_API_KEY,
@@ -1947,6 +1969,17 @@ def search_web(request: Request, engine: str, query: str, user=None) -> list[Sea
            )
        else:
            raise Exception('No BRAVE_SEARCH_API_KEY found in environment variables')
+    elif engine == 'brave_llm_context':
+        if request.app.state.config.BRAVE_SEARCH_API_KEY:
+            return search_brave_llm_context(
+                request.app.state.config.BRAVE_SEARCH_API_KEY,
+                query,
+                request.app.state.config.WEB_SEARCH_RESULT_COUNT,
+                request.app.state.config.WEB_SEARCH_DOMAIN_FILTER_LIST,
+                request.app.state.config.BRAVE_SEARCH_CONTEXT_TOKENS,
+            )
+        else:
+            raise Exception('No BRAVE_SEARCH_API_KEY found in environment variables')
    elif engine == 'kagi':
        if request.app.state.config.KAGI_SEARCH_API_KEY:
            return search_kagi(
@@ -79,6 +79,7 @@ async def get_task_config(request: Request, user=Depends(get_verified_user)):
        'ENABLE_RETRIEVAL_QUERY_GENERATION': request.app.state.config.ENABLE_RETRIEVAL_QUERY_GENERATION,
        'QUERY_GENERATION_PROMPT_TEMPLATE': request.app.state.config.QUERY_GENERATION_PROMPT_TEMPLATE,
        'TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE': request.app.state.config.TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE,
+        'ENABLE_VOICE_MODE_PROMPT': request.app.state.config.ENABLE_VOICE_MODE_PROMPT,
        'VOICE_MODE_PROMPT_TEMPLATE': request.app.state.config.VOICE_MODE_PROMPT_TEMPLATE,
    }

@@ -99,6 +100,7 @@ class TaskConfigForm(BaseModel):
    ENABLE_RETRIEVAL_QUERY_GENERATION: bool
    QUERY_GENERATION_PROMPT_TEMPLATE: str
    TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE: str
+    ENABLE_VOICE_MODE_PROMPT: bool
    VOICE_MODE_PROMPT_TEMPLATE: Optional[str]


@@ -127,6 +129,7 @@ async def update_task_config(request: Request, form_data: TaskConfigForm, user=D
    request.app.state.config.QUERY_GENERATION_PROMPT_TEMPLATE = form_data.QUERY_GENERATION_PROMPT_TEMPLATE
    request.app.state.config.TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE = form_data.TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE

+    request.app.state.config.ENABLE_VOICE_MODE_PROMPT = form_data.ENABLE_VOICE_MODE_PROMPT
    request.app.state.config.VOICE_MODE_PROMPT_TEMPLATE = form_data.VOICE_MODE_PROMPT_TEMPLATE

    return {
@@ -145,6 +148,7 @@ async def update_task_config(request: Request, form_data: TaskConfigForm, user=D
        'ENABLE_RETRIEVAL_QUERY_GENERATION': request.app.state.config.ENABLE_RETRIEVAL_QUERY_GENERATION,
        'QUERY_GENERATION_PROMPT_TEMPLATE': request.app.state.config.QUERY_GENERATION_PROMPT_TEMPLATE,
        'TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE': request.app.state.config.TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE,
+        'ENABLE_VOICE_MODE_PROMPT': request.app.state.config.ENABLE_VOICE_MODE_PROMPT,
        'VOICE_MODE_PROMPT_TEMPLATE': request.app.state.config.VOICE_MODE_PROMPT_TEMPLATE,
    }

@@ -159,6 +163,7 @@ async def generate_title(request: Request, form_data: dict, user=Depends(get_ver

    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -187,7 +192,7 @@ async def generate_title(request: Request, form_data: dict, user=Depends(get_ver
    else:
        template = DEFAULT_TITLE_GENERATION_PROMPT_TEMPLATE

-    content = title_generation_template(template, form_data['messages'], user)
+    content = await title_generation_template(template, form_data['messages'], user)

    max_tokens = models[task_model_id].get('info', {}).get('params', {}).get('max_tokens', 1000)

@@ -236,6 +241,7 @@ async def generate_follow_ups(request: Request, form_data: dict, user=Depends(ge

    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -264,7 +270,7 @@ async def generate_follow_ups(request: Request, form_data: dict, user=Depends(ge
    else:
        template = DEFAULT_FOLLOW_UP_GENERATION_PROMPT_TEMPLATE

-    content = follow_up_generation_template(template, form_data['messages'], user)
+    content = await follow_up_generation_template(template, form_data['messages'], user)

    payload = {
        'model': task_model_id,
@@ -304,6 +310,7 @@ async def generate_chat_tags(request: Request, form_data: dict, user=Depends(get

    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -332,7 +339,7 @@ async def generate_chat_tags(request: Request, form_data: dict, user=Depends(get
    else:
        template = DEFAULT_TAGS_GENERATION_PROMPT_TEMPLATE

-    content = tags_generation_template(template, form_data['messages'], user)
+    content = await tags_generation_template(template, form_data['messages'], user)

    payload = {
        'model': task_model_id,
@@ -366,6 +373,7 @@ async def generate_chat_tags(request: Request, form_data: dict, user=Depends(get
 async def generate_image_prompt(request: Request, form_data: dict, user=Depends(get_verified_user)):
    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -394,7 +402,7 @@ async def generate_image_prompt(request: Request, form_data: dict, user=Depends(
    else:
        template = DEFAULT_IMAGE_PROMPT_GENERATION_PROMPT_TEMPLATE

-    content = image_prompt_generation_template(template, form_data['messages'], user)
+    content = await image_prompt_generation_template(template, form_data['messages'], user)

    payload = {
        'model': task_model_id,
@@ -446,6 +454,7 @@ async def generate_queries(request: Request, form_data: dict, user=Depends(get_v

    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -474,7 +483,7 @@ async def generate_queries(request: Request, form_data: dict, user=Depends(get_v
    else:
        template = DEFAULT_QUERY_GENERATION_PROMPT_TEMPLATE

-    content = query_generation_template(template, form_data['messages'], user)
+    content = await query_generation_template(template, form_data['messages'], user)

    payload = {
        'model': task_model_id,
@@ -524,6 +533,7 @@ async def generate_autocompletion(request: Request, form_data: dict, user=Depend

    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -552,7 +562,7 @@ async def generate_autocompletion(request: Request, form_data: dict, user=Depend
    else:
        template = DEFAULT_AUTOCOMPLETE_GENERATION_PROMPT_TEMPLATE

-    content = autocomplete_generation_template(template, prompt, messages, type, user)
+    content = await autocomplete_generation_template(template, prompt, messages, type, user)

    payload = {
        'model': task_model_id,
@@ -586,6 +596,7 @@ async def generate_autocompletion(request: Request, form_data: dict, user=Depend
 async def generate_emoji(request: Request, form_data: dict, user=Depends(get_verified_user)):
    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -611,7 +622,7 @@ async def generate_emoji(request: Request, form_data: dict, user=Depends(get_ver

    template = DEFAULT_EMOJI_GENERATION_PROMPT_TEMPLATE

-    content = emoji_generation_template(template, form_data['prompt'], user)
+    content = await emoji_generation_template(template, form_data['prompt'], user)

    payload = {
        'model': task_model_id,
@@ -651,6 +662,7 @@ async def generate_emoji(request: Request, form_data: dict, user=Depends(get_ver
 async def generate_moa_response(request: Request, form_data: dict, user=Depends(get_verified_user)):
    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -29,7 +29,7 @@ from open_webui.models.users import (
 )

 from open_webui.constants import ERROR_MESSAGES
-from open_webui.env import STATIC_DIR
+from open_webui.env import ENABLE_PROFILE_IMAGE_URL_FORWARDING, STATIC_DIR
 from open_webui.internal.db import get_async_session


@@ -193,6 +193,7 @@ class SharingPermissions(BaseModel):
    public_skills: bool = False
    notes: bool = False
    public_notes: bool = True
+    public_chats: bool = False


 class AccessGrantsPermissions(BaseModel):
@@ -477,12 +478,15 @@ async def get_user_profile_image_by_id(user_id: str, user=Depends(get_verified_u
    user = await Users.get_user_by_id(user_id)
    if user:
        if user.profile_image_url:
-            # check if it's url or base64
            if user.profile_image_url.startswith('http'):
-                return Response(
-                    status_code=status.HTTP_302_FOUND,
-                    headers={'Location': user.profile_image_url},
-                )
+                if ENABLE_PROFILE_IMAGE_URL_FORWARDING:
+                    return Response(
+                        status_code=status.HTTP_302_FOUND,
+                        headers={'Location': user.profile_image_url},
+                    )
+                # When forwarding is disabled, fall through to the
+                # default image to prevent client-side IP/UA/Referer
+                # leaks via 302 redirect to external origins.
            elif user.profile_image_url.startswith('data:image'):
                try:
                    header, base64_data = user.profile_image_url.split(',', 1)
@@ -550,6 +554,8 @@ async def update_user_by_id(
                        detail=ERROR_MESSAGES.ACTION_PROHIBITED,
                    )

+    except HTTPException:
+        raise
    except Exception as e:
        log.error(f'Error checking primary admin status: {e}')
        raise HTTPException(
@@ -631,6 +637,8 @@ async def delete_user_by_id(user_id: str, user=Depends(get_admin_user), db: Asyn
                status_code=status.HTTP_403_FORBIDDEN,
                detail=ERROR_MESSAGES.ACTION_PROHIBITED,
            )
+    except HTTPException:
+        raise
    except Exception as e:
        log.error(f'Error checking primary admin status: {e}')
        raise HTTPException(
@@ -888,13 +888,15 @@ async def get_event_emitter(request_info, update_db=True):
                )

            elif event_type == 'embeds':
-                message = await Chats.get_message_by_id_and_message_id(
-                    request_info['chat_id'],
-                    request_info['message_id'],
-                )
+                event_payload = event_data.get('data', {})
+                embeds = event_payload.get('embeds', [])

-                embeds = event_data.get('data', {}).get('embeds', [])
-                embeds.extend(message.get('embeds', []))
+                if not event_payload.get('replace', False):
+                    message = await Chats.get_message_by_id_and_message_id(
+                        request_info['chat_id'],
+                        request_info['message_id'],
+                    )
+                    embeds.extend(message.get('embeds', []))

                await Chats.upsert_message_to_chat_by_id_and_message_id(
                    request_info['chat_id'],
@@ -948,17 +950,27 @@ async def get_event_emitter(request_info, update_db=True):

 async def get_event_call(request_info):
    async def __event_caller__(event_data):
-        response = await sio.call(
-            'events',
-            {
-                'chat_id': request_info.get('chat_id', None),
-                'message_id': request_info.get('message_id', None),
-                'data': event_data,
-            },
-            to=request_info['session_id'],
-            timeout=WEBSOCKET_EVENT_CALLER_TIMEOUT,
-        )
-        return response
+        session_id = request_info['session_id']
+
+        # Fast-fail if the client has disconnected.
+        if session_id not in SESSION_POOL:
+            log.warning(f'Event caller: session {session_id} no longer connected')
+            return {'error': 'Client session disconnected.'}
+
+        try:
+            return await sio.call(
+                'events',
+                {
+                    'chat_id': request_info.get('chat_id', None),
+                    'message_id': request_info.get('message_id', None),
+                    'data': event_data,
+                },
+                to=session_id,
+                timeout=WEBSOCKET_EVENT_CALLER_TIMEOUT,
+            )
+        except TimeoutError:
+            log.warning(f'Event caller timed out for session {session_id}')
+            return {'error': 'Event call timed out. The browser tab may be inactive or closed.'}

    if 'session_id' in request_info and 'chat_id' in request_info and 'message_id' in request_info:
        return __event_caller__
@@ -238,9 +238,13 @@ async def fetch_url(
        content, _ = await asyncio.to_thread(get_content_from_url, __request__, url)

        # Truncate if configured (WEB_FETCH_MAX_CONTENT_LENGTH)
-        max_length = getattr(__request__.app.state.config, 'WEB_FETCH_MAX_CONTENT_LENGTH', None)
-        if max_length and max_length > 0 and len(content) > max_length:
-            content = content[:max_length] + '\n\n[Content truncated...]'
+        # Guard: content may be None if the web loader silently failed
+        if content is not None:
+            max_length = getattr(__request__.app.state.config, 'WEB_FETCH_MAX_CONTENT_LENGTH', None)
+            if max_length and max_length > 0 and len(content) > max_length:
+                content = content[:max_length] + '\n\n[Content truncated...]'
+        else:
+            content = ''

        return content
    except Exception as e:
@@ -467,9 +471,15 @@ async def execute_code(

            # Parse the output - pyodide returns dict with stdout, stderr, result
            if isinstance(output, dict):
-                stdout = output.get('stdout', '')
-                stderr = output.get('stderr', '')
-                result = output.get('result', '')
+                # Handle error responses from event_caller (e.g. session disconnected, timeout)
+                if output.get('error') and not output.get('stdout') and not output.get('result'):
+                    stderr = output['error']
+                    stdout = ''
+                    result = ''
+                else:
+                    stdout = output.get('stdout', '')
+                    stderr = output.get('stderr', '')
+                    result = output.get('result', '')
            else:
                stdout = ''
                stderr = ''
@@ -2568,8 +2578,11 @@ async def create_automation(
        if not user:
            return json.dumps({'error': 'User not found'})

-        # Always use the calling model for the automation
-        model_id = (__metadata__ or {}).get('model_id')
+        # Fall back to model dict ID since __metadata__ may predate model_id assignment
+        metadata = __metadata__ or {}
+        model_id = metadata.get('model_id') or (
+            metadata.get('model', {}).get('id') if isinstance(metadata.get('model'), dict) else None
+        )
        if not model_id:
            return json.dumps({'error': 'Could not detect current model'})

@@ -2671,7 +2684,7 @@ async def update_automation(
            is_active=automation.is_active,
        )

-        updated = await Automations.update(automation_id, form, next_run_ns(new_rrule, tz=tz))
+        updated = await Automations.update_by_id(automation_id, form, next_run_ns(new_rrule, tz=tz))

        return json.dumps(
            {
@@ -2888,6 +2901,9 @@ def _ns_to_dt(ns: int, tz) -> str:

 def _event_to_dict(event, tz) -> dict:
    """Convert a calendar event model to a human-friendly dict with local timestamps."""
+    alert_minutes = None
+    if event.meta and 'alert_minutes' in event.meta:
+        alert_minutes = event.meta['alert_minutes']
    return {
        'id': event.id,
        'calendar_id': event.calendar_id,
@@ -2897,6 +2913,7 @@ def _event_to_dict(event, tz) -> dict:
        'end': _ns_to_dt(event.end_at, tz) if event.end_at else None,
        'all_day': event.all_day,
        'location': event.location or '',
+        'reminder_minutes': alert_minutes if alert_minutes is not None else 10,
        'color': event.color,
        'is_cancelled': event.is_cancelled,
    }
@@ -2911,8 +2928,9 @@ async def search_calendar_events(
    __user__: dict = None,
 ) -> str:
    """
-    Search calendar events by text and/or date range.
-    Returns matching events across all accessible calendars.
+    Search calendar events, reminders, and scheduled items by text and/or date range.
+    Use this to check what's coming up, find a specific event or reminder, or list
+    the user's schedule for a time period.

    :param query: Search text to match against event title, description, or location (optional)
    :param start: Only return events starting at or after this datetime, e.g. "2026-04-20 00:00" (optional)
@@ -3003,20 +3021,24 @@ async def create_calendar_event(
    calendar_id: Optional[str] = None,
    all_day: bool = False,
    location: Optional[str] = None,
+    reminder_minutes: Optional[int] = None,
    __request__: Request = None,
    __user__: dict = None,
 ) -> str:
    """
-    Create a new calendar event. If no calendar_id is provided, the event is
-    added to the user's default calendar.
+    Create a calendar event, reminder, or alarm. Use this when the user wants to
+    schedule an event, set a reminder, create an alarm, or says things like
+    "remind me", "don't let me forget", "notify me at", or "add to my calendar".
+    For simple reminders, omit end/location/all_day and set reminder_minutes to 0.

-    :param title: Event title
-    :param start: Start datetime string in your local time (e.g. "2026-04-20 09:00" or "2026-04-20T09:00:00")
-    :param end: End datetime string in your local time (optional, omit for point-in-time events)
-    :param description: Event description (optional)
+    :param title: Event or reminder title (e.g. "Team standup", "Take medicine", "Call mom")
+    :param start: Start datetime in the user's local time (e.g. "2026-04-20 09:00")
+    :param end: End datetime in the user's local time (optional — omit for reminders or point-in-time events)
+    :param description: Event description or notes (optional)
    :param calendar_id: Target calendar ID (optional, uses default calendar if omitted)
    :param all_day: Whether this is an all-day event (default: false)
    :param location: Event location (optional)
+    :param reminder_minutes: Minutes before the event to send a notification (optional, default: 10). Use 0 for "at time of event", -1 for no notification.
    :return: JSON with the created event details including id
    """
    if __request__ is None:
@@ -3079,6 +3101,18 @@ async def create_calendar_event(
            # Default to 1 hour duration
            end_ns = start_ns + 3_600_000_000_000

+        # Build meta with reminder setting
+        meta = {}
+        if reminder_minutes is not None:
+            if isinstance(reminder_minutes, str):
+                try:
+                    reminder_minutes = int(reminder_minutes)
+                except ValueError:
+                    reminder_minutes = 10
+            meta['alert_minutes'] = reminder_minutes
+        else:
+            meta['alert_minutes'] = 10
+
        form = CalendarEventForm(
            calendar_id=calendar_id,
            title=title,
@@ -3087,6 +3121,7 @@ async def create_calendar_event(
            end_at=end_ns,
            all_day=all_day,
            location=location,
+            meta=meta,
        )

        event = await CalendarEvents.insert_new_event(user_id, form)
@@ -3114,6 +3149,7 @@ async def update_calendar_event(
    all_day: Optional[bool] = None,
    location: Optional[str] = None,
    is_cancelled: Optional[bool] = None,
+    reminder_minutes: Optional[int] = None,
    __request__: Request = None,
    __user__: dict = None,
 ) -> str:
@@ -3129,6 +3165,7 @@ async def update_calendar_event(
    :param all_day: Whether this is an all-day event (optional)
    :param location: New event location (optional)
    :param is_cancelled: Set to true to cancel the event (optional)
+    :param reminder_minutes: Minutes before the event to send a reminder notification (optional). Use 0 for "at time of event", -1 for no reminder. Accepts any positive integer for custom timing (e.g. 120 for 2 hours before).
    :return: JSON with the updated event details
    """
    if __request__ is None:
@@ -3149,8 +3186,10 @@ async def update_calendar_event(
            return json.dumps({'error': 'Event not found'})

        # Check write access to the event's calendar
-        cal = await Calendars.get_calendar_by_id(event.calendar_id)
-        if cal and cal.user_id != user_id and __user__.get('role') != 'admin':
+        if event.user_id != user_id and __user__.get('role') != 'admin':
+            cal = await Calendars.get_calendar_by_id(event.calendar_id)
+            if not cal:
+                return json.dumps({'error': 'Access denied'})
            user_group_ids = [g.id for g in await Groups.get_groups_by_member_id(user_id)]
            if not await AccessGrants.has_access(
                user_id=user_id,
@@ -3183,6 +3222,17 @@ async def update_calendar_event(
            except (ValueError, TypeError) as e:
                return json.dumps({'error': f'Invalid end datetime: {e}'})

+        # Build meta update with reminder setting if provided
+        meta = None
+        if reminder_minutes is not None:
+            if isinstance(reminder_minutes, str):
+                try:
+                    reminder_minutes = int(reminder_minutes)
+                except ValueError:
+                    reminder_minutes = None
+            if reminder_minutes is not None:
+                meta = {'alert_minutes': reminder_minutes}
+
        form = CalendarEventUpdateForm(
            title=title,
            description=description,
@@ -3191,6 +3241,7 @@ async def update_calendar_event(
            all_day=all_day,
            location=location,
            is_cancelled=is_cancelled,
+            meta=meta,
        )

        updated = await CalendarEvents.update_event_by_id(event_id, form)
@@ -3238,8 +3289,10 @@ async def delete_calendar_event(
            return json.dumps({'error': 'Event not found'})

        # Check write access
-        cal = await Calendars.get_calendar_by_id(event.calendar_id)
-        if cal and cal.user_id != user_id and __user__.get('role') != 'admin':
+        if event.user_id != user_id and __user__.get('role') != 'admin':
+            cal = await Calendars.get_calendar_by_id(event.calendar_id)
+            if not cal:
+                return json.dumps({'error': 'Access denied'})
            user_group_ids = [g.id for g in await Groups.get_groups_by_member_id(user_id)]
            if not await AccessGrants.has_access(
                user_id=user_id,
@@ -41,6 +41,7 @@ from starlette.datastructures import MutableHeaders
 from starlette.requests import Request
 from starlette.types import ASGIApp, Message, Receive, Scope, Send

+from open_webui.env import CUSTOM_API_KEY_HEADER
 from open_webui.internal.db import ScopedSession
 from open_webui.utils.auth import get_http_authorization_cred

@@ -87,6 +88,13 @@ class CommitSessionMiddleware:
            await self.app(scope, receive, send)
            return

+        path = scope.get('path', '')
+        # Keep health probes independent from sync session commit/remove
+        # so DB pressure cannot delay or fail probe responses.
+        if path in {'/health', '/ready', '/health/db'}:
+            await self.app(scope, receive, send)
+            return
+
        try:
            await self.app(scope, receive, send)
        except BaseException:
@@ -119,9 +127,16 @@ class CommitSessionMiddleware:


 class AuthTokenMiddleware:
-    """Extract the bearer/cookie/x-api-key credential and stash it on
+    """Extract the bearer/cookie/API-key credential and stash it on
    `request.state.token`.

+    The header used for API-key transport is controlled by the
+    ``CUSTOM_API_KEY_HEADER`` environment variable (default ``x-api-key``).
+    This is useful when Open WebUI sits behind a reverse proxy that
+    consumes the ``Authorization`` header for its own authentication —
+    set the env var to a unique header (e.g. ``X-OpenWebUI-Key``) so
+    the middleware checks that instead and avoids the 401 short-circuit.
+
    Routes that depend on `get_verified_user` etc. read this state.
    Also exposes `request.state.enable_api_keys` (snapshotted at request
    entry from runtime config) and stamps an `X-Process-Time` response
@@ -146,7 +161,7 @@ class AuthTokenMiddleware:
            if cookie_token:
                token = HTTPAuthorizationCredentials(scheme='Bearer', credentials=cookie_token)
        if token is None:
-            api_key = request.headers.get('x-api-key')
+            api_key = request.headers.get(CUSTOM_API_KEY_HEADER)
            if api_key:
                token = HTTPAuthorizationCredentials(scheme='Bearer', credentials=api_key)

@@ -360,7 +360,7 @@ async def execute_automation(app, automation: AutomationModel) -> None:
            await _record_run(automation.id, 'error', error='User not found')
            return

-        prompt = prompt_template(automation.data['prompt'], user)
+        prompt = await prompt_template(automation.data['prompt'], user)
        model_id = automation.data['model_id']
        terminal_config = automation.data.get('terminal')

@@ -73,6 +73,11 @@ async def generate_direct_chat_completion(
    request_id = str(uuid.uuid4())  # Generate a unique request ID

    event_caller = await get_event_call(metadata)
+    if event_caller is None:
+        raise Exception(
+            'Direct connection requires an active WebSocket session; '
+            'cannot generate completion in this context (e.g. background task).'
+        )

    channel = f'{user_id}:{session_id}:{request_id}'
    logging.info(f'WebSocket channel: {channel}')
@@ -180,10 +185,14 @@ async def generate_chat_completion(
            }

    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
+        # Merge the direct connection model into server models so that
+        # task functions (title, tags, etc.) can resolve a server-side
+        # task model while still having the direct model available.
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
-        log.debug(f'direct connection to model: {models}')
+        log.debug(f'direct connection to model: {request.state.model["id"]}')
    else:
        models = request.app.state.MODELS

@@ -193,7 +202,7 @@ async def generate_chat_completion(

    model = models[model_id]

-    if getattr(request.state, 'direct', False):
+    if getattr(request.state, 'direct', False) and model_id == getattr(request.state, 'model', {}).get('id'):
        return await generate_direct_chat_completion(request, form_data, user=user, models=models)
    else:
        # Check if user has access to the model
@@ -310,6 +319,7 @@ async def chat_completed(request: Request, form_data: dict, user: Any):

    if getattr(request.state, 'direct', False) and hasattr(request.state, 'model'):
        models = {
+            **request.app.state.MODELS,
            request.state.model['id']: request.state.model,
        }
    else:
@@ -14,7 +14,7 @@ async def get_function_module(request, function_id, load_from_db=True):
    """
    Get the function module by its ID.
    """
-    function_module, _, _ = await get_function_module_from_cache(request, function_id, load_from_db)
+    function_module, _, _ = await get_function_module_from_cache(request, function_id, load_from_db=load_from_db)
    return function_module


@@ -16,3 +16,26 @@ def include_user_info_headers(headers, user):
        FORWARD_USER_INFO_HEADER_USER_EMAIL: user.email,
        FORWARD_USER_INFO_HEADER_USER_ROLE: user.role,
    }
+
+
+def get_custom_headers(custom_headers: dict, user=None, metadata: dict = None) -> dict:
+    if not custom_headers or not isinstance(custom_headers, dict):
+        return {}
+
+    metadata = metadata or {}
+    template_vars = {
+        '{{CHAT_ID}}': metadata.get('chat_id', '') or '',
+        '{{MESSAGE_ID}}': metadata.get('message_id', '') or '',
+        '{{USER_ID}}': (user.id if user else '') or '',
+        '{{USER_NAME}}': (user.name if user else '') or '',
+    }
+
+    parsed_headers = {}
+    for key, value in custom_headers.items():
+        if not isinstance(value, str):
+            value = str(value)
+        for token, val in template_vars.items():
+            value = value.replace(token, val)
+        parsed_headers[key] = value
+
+    return parsed_headers
@@ -44,7 +44,7 @@ def create_httpx_client(headers=None, timeout=None, auth=None):
    return _build_httpx_client(headers=headers, timeout=timeout, auth=auth, verify=True)


-async def create_insecure_httpx_client(headers=None, timeout=None, auth=None):
+def create_insecure_httpx_client(headers=None, timeout=None, auth=None):
    return _build_httpx_client(headers=headers, timeout=timeout, auth=auth, verify=False)


@@ -155,20 +155,18 @@ class MCPClient:
        self.session = None

        try:
-            await asyncio.wait_for(
-                asyncio.shield(exit_stack.aclose()),
-                timeout=5.0,
-            )
-        except asyncio.TimeoutError:
+            # IMPORTANT: Do NOT use asyncio.shield() or asyncio.wait_for()
+            # because they create a new asyncio task, which violates the MCP SDK's
+            # requirement that its TaskGroup be exited in the exact same task.
+            # ALSO do NOT use anyio.CancelScope(shield=True) or anyio.fail_after(),
+            # because they push a new cancel scope onto the task, violating LIFO
+            # order when aclose() attempts to exit the inner TaskGroup.
+            # We simply call aclose() directly. If the task is cancelled, the
+            # sockets will eventually be cleaned up by garbage collection.
+            await exit_stack.aclose()
+        except TimeoutError:
            log.warning('MCPClient.disconnect() timed out after 5 s')
        except RuntimeError as exc:
-            # The MCP SDK's streamable_http transport uses anyio task
-            # groups and async generators internally.  When we close
-            # a session that was interrupted mid-flight these can
-            # raise RuntimeError ("aclose(): asynchronous generator is
-            # already running" or "Attempted to exit cancel scope in a
-            # different task").  Swallowing the error here prevents the
-            # orphaned coroutines from spinning at 100 % CPU.
            log.debug('MCPClient.disconnect() suppressed RuntimeError: %s', exc)
        except Exception as exc:
            log.debug('MCPClient.disconnect() error: %s', exc)
@@ -969,7 +969,7 @@ def get_source_context(sources: list, source_ids: dict = None, include_content:
    return context_string


-def apply_source_context_to_messages(
+async def apply_source_context_to_messages(
    request: Request,
    messages: list,
    sources: list,
@@ -995,13 +995,13 @@ def apply_source_context_to_messages(

    if RAG_SYSTEM_CONTEXT:
        return add_or_update_system_message(
-            rag_template(request.app.state.config.RAG_TEMPLATE, context, user_message),
+            await rag_template(request.app.state.config.RAG_TEMPLATE, context, user_message),
            messages,
            append=True,
        )
    else:
        return add_or_update_user_message(
-            rag_template(request.app.state.config.RAG_TEMPLATE, context, user_message),
+            await rag_template(request.app.state.config.RAG_TEMPLATE, context, user_message),
            messages,
            append=False,
        )
@@ -1164,6 +1164,15 @@ async def process_tool_result(
                                'url': file_url,
                            }
                        )
+                    elif item.get('type') == 'resource':
+                        resource = item.get('resource', {})
+                        text = resource.get('text', '')
+                        if isinstance(text, str) and text:
+                            try:
+                                text = json.loads(text)
+                            except json.JSONDecodeError:
+                                pass
+                            tool_response.append(text)
            tool_result = tool_response[0] if len(tool_response) == 1 else tool_response
        else:  # OpenAPI
            for item in tool_result:
@@ -1546,11 +1555,11 @@ async def chat_web_search_handler(request: Request, form_data: dict, extra_param

    except Exception as e:
        log.exception(e)
-        queries = [user_message]
+        queries = [user_message or '']

    # Check if generated queries are empty
    if len(queries) == 1 and queries[0].strip() == '':
-        queries = [user_message]
+        queries = [user_message or '']

    # Check if queries are not found
    if len(queries) == 0:
@@ -1991,7 +2000,7 @@ async def chat_completion_files_handler(
            )

        if len(queries) == 0:
-            queries = [get_last_user_message(body['messages'])]
+            queries = [get_last_user_message(body['messages']) or '']

        try:
            # Directly await async get_sources_from_items (no thread needed - fully async now)
@@ -2156,7 +2165,27 @@ async def load_messages_from_db(chat_id: str, message_id: str) -> Optional[list[
    return [{k: v for k, v in msg.items() if k in ('role', 'content', 'output', 'files')} for msg in db_messages]


-def process_messages_with_output(messages: list[dict]) -> list[dict]:
+def get_reasoning_format(model: dict) -> str | None:
+    """
+    Determine how reasoning should be included in reconstructed messages.
+
+    Returns:
+        'think_tags': Ollama expects <think> tags in content.
+        'reasoning_content': llama.cpp supports reasoning_content as a top-level field.
+        None: skip reasoning (safe default for strict providers).
+    """
+    provider = model.get('provider', '')
+    if provider == 'ollama':
+        return 'think_tags'
+    if provider == 'llama.cpp':
+        return 'reasoning_content'
+    return None
+
+
+def process_messages_with_output(
+    messages: list[dict],
+    reasoning_format: str | None = None,
+) -> list[dict]:
    """
    Process messages with OR-aligned output items for LLM consumption.

@@ -2168,7 +2197,11 @@ def process_messages_with_output(messages: list[dict]) -> list[dict]:
    for message in messages:
        if message.get('role') == 'assistant' and message.get('output'):
            # Use output items for clean OpenAI-format messages
-            output_messages = convert_output_to_messages(message['output'], raw=True)
+            output_messages = convert_output_to_messages(
+                message['output'],
+                raw=True,
+                reasoning_format=reasoning_format,
+            )
            if output_messages:
                processed.extend(output_messages)
                continue
@@ -2180,6 +2213,43 @@ def process_messages_with_output(messages: list[dict]) -> list[dict]:
    return processed


+SKILL_MENTION_RE = re.compile(r'<\$([^|>]+)\|?[^>]*>')
+
+
+def _get_text_parts(message: dict) -> list[str]:
+    """Return all text segments from a message's content."""
+    content = message.get('content')
+    if isinstance(content, str):
+        return [content]
+    if isinstance(content, list):
+        return [p.get('text', '') for p in content if isinstance(p, dict) and p.get('type') == 'text']
+    return []
+
+
+def extract_skill_ids_from_messages(messages: list[dict]) -> set[str]:
+    """Extract skill IDs from <$skillId|label> mention tags in messages."""
+    ids: set[str] = set()
+    for message in messages:
+        for text in _get_text_parts(message):
+            ids.update(m.group(1) for m in SKILL_MENTION_RE.finditer(text))
+    return ids
+
+
+def strip_skill_mentions(messages: list[dict]) -> None:
+    """Strip <$skillId|label> mention tags from message content in-place."""
+    strip_re = re.compile(r'<\$[^>]+>')
+    for message in messages:
+        content = message.get('content')
+        if isinstance(content, str) and strip_re.search(content):
+            message['content'] = strip_re.sub('', content).strip()
+        elif isinstance(content, list):
+            for part in content:
+                if isinstance(part, dict) and part.get('type') == 'text':
+                    text = part.get('text', '')
+                    if strip_re.search(text):
+                        part['text'] = strip_re.sub('', text).strip()
+
+
 async def process_chat_payload(request, form_data, user, metadata, model):
    # Pipeline Inlet -> Filter Inlet -> Chat Memory -> Chat Web Search -> Chat Image Generation
    # -> Chat Code Interpreter (Form Data Update) -> (Default) Chat Tools Function Calling
@@ -2217,6 +2287,9 @@ async def process_chat_payload(request, form_data, user, metadata, model):
    form_data = apply_params_to_form_data(form_data, model)
    log.debug(f'form_data: {form_data}')

+    # Guided regeneration: extract before it reaches the LLM provider
+    regeneration_prompt = form_data.pop('regeneration_prompt', None)
+
    # Load messages from DB when available — DB preserves structured 'output' items
    # which the frontend strips, causing tool calls to be merged into content.
    chat_id = metadata.get('chat_id')
@@ -2225,6 +2298,16 @@ async def process_chat_payload(request, form_data, user, metadata, model):
    if chat_id and user_message_id and not chat_id.startswith('local:'):
        db_messages = await load_messages_from_db(chat_id, user_message_id)
        if db_messages:
+            # Continue: frontend sends assistant_message_id when continuing
+            # an existing response. Load its content so the LLM sees prior output.
+            assistant_message_id = metadata.get('assistant_message_id')
+            if assistant_message_id:
+                assistant_message = await Chats.get_message_by_id_and_message_id(chat_id, assistant_message_id)
+                if assistant_message and (assistant_message.get('content') or assistant_message.get('output')):
+                    db_messages.append(
+                        {k: v for k, v in assistant_message.items() if k in ('role', 'content', 'output', 'files')}
+                    )
+
            system_message = get_system_message(form_data.get('messages', []))
            form_data['messages'] = [system_message, *db_messages] if system_message else db_messages

@@ -2252,13 +2335,19 @@ async def process_chat_payload(request, form_data, user, metadata, model):
                # Strip files field — it's been incorporated into content
                message.pop('files', None)

+    if regeneration_prompt:
+        form_data['messages'].append({'role': 'user', 'content': regeneration_prompt})
+
    # Process messages with OR-aligned output items for clean LLM messages
-    form_data['messages'] = process_messages_with_output(form_data.get('messages', []))
+    form_data['messages'] = process_messages_with_output(
+        form_data.get('messages', []),
+        reasoning_format=get_reasoning_format(model),
+    )

    system_message = get_system_message(form_data.get('messages', []))
    if system_message:  # Chat Controls/User Settings
        try:
-            form_data = apply_system_prompt_to_body(
+            form_data = await apply_system_prompt_to_body(
                system_message.get('content'), form_data, metadata, user, replace=True
            )  # Required to handle system prompt variables
        except Exception:
@@ -2316,7 +2405,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):

        if folder and folder.data:
            if 'system_prompt' in folder.data:
-                form_data = apply_system_prompt_to_body(folder.data['system_prompt'], form_data, metadata, user)
+                form_data = await apply_system_prompt_to_body(folder.data['system_prompt'], form_data, metadata, user)
            if 'files' in folder.data:
                if metadata.get('params', {}).get('function_calling') != 'native':
                    form_data['files'] = [
@@ -2371,6 +2460,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):
        form_data['files'] = files

    variables = form_data.pop('variables', None)
+    payload_tools = form_data.get('tools', None)  # snapshot before filters

    # Process the form_data through the pipeline
    try:
@@ -2396,8 +2486,8 @@ async def process_chat_payload(request, form_data, user, metadata, model):
    extra_params['__features__'] = features
    if features:
        if 'voice' in features and features['voice']:
-            if request.app.state.config.VOICE_MODE_PROMPT_TEMPLATE != None:
-                if request.app.state.config.VOICE_MODE_PROMPT_TEMPLATE != '':
+            if getattr(request.app.state.config, 'ENABLE_VOICE_MODE_PROMPT', True):
+                if request.app.state.config.VOICE_MODE_PROMPT_TEMPLATE:
                    template = request.app.state.config.VOICE_MODE_PROMPT_TEMPLATE
                else:
                    template = DEFAULT_VOICE_MODE_PROMPT_TEMPLATE
@@ -2461,12 +2551,14 @@ async def process_chat_payload(request, form_data, user, metadata, model):
    files = form_data.pop('files', None)
    form_data.pop('folder_id', None)

-    # Caller-provided OpenAI-style tools take precedence over server-side
-    # tool resolution (tool_ids, MCP servers, builtin tools).
-    payload_tools = form_data.get('tools', None)
+    # If the original caller provided tools, use them as-is (skip resolution).
+    # Otherwise, save any tools that filter inlets added for merging later.
+    inlet_filter_tools = None if payload_tools else form_data.get('tools', None)

-    # Skills
+    # Skills — extract IDs from message content (<$skillId|label> tags) so
+    # persisted chats work without relying on the frontend to send skill_ids.
    user_skill_ids = set(form_data.pop('skill_ids', None) or [])
+    user_skill_ids |= extract_skill_ids_from_messages(form_data.get('messages', []))
    model_skill_ids = set(model.get('info', {}).get('meta', {}).get('skillIds', []))

    all_skill_ids = user_skill_ids | model_skill_ids
@@ -2502,6 +2594,9 @@ async def process_chat_payload(request, form_data, user, metadata, model):
                append=True,
            )

+    # Strip <$skillId|label> mention tags so the model doesn't see raw markup.
+    strip_skill_mentions(form_data.get('messages', []))
+
    prompt = get_last_user_message(form_data['messages'])
    # TODO: re-enable URL extraction from prompt
    # urls = []
@@ -2767,6 +2862,8 @@ async def process_chat_payload(request, form_data, user, metadata, model):
                form_data['tools'] = [
                    {'type': 'function', 'function': tool.get('spec', {})} for tool in tools_dict.values()
                ]
+                if inlet_filter_tools:
+                    form_data['tools'].extend(inlet_filter_tools)
            else:
                # If the function calling is not native, then call the tools function calling handler
                try:
@@ -2797,7 +2894,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):

    # If context is not empty, insert it into the messages
    if sources and prompt:
-        form_data['messages'] = apply_source_context_to_messages(request, form_data['messages'], sources, prompt)
+        form_data['messages'] = await apply_source_context_to_messages(request, form_data['messages'], sources, prompt)

    # If there are citations, add them to the data_items
    sources = [
@@ -3894,6 +3991,12 @@ async def streaming_chat_response_handler(response, ctx):
                                            response_id = response_metadata.pop('response_id', None)
                                            if response_id:
                                                last_response_id = response_id
+
+                                        # Normalize and capture usage for DB persistence
+                                        if response_metadata.get('usage'):
+                                            response_metadata['usage'] = normalize_usage(response_metadata['usage'])
+                                            usage = response_metadata['usage']
+
                                        processed_data.update(response_metadata)
                                        processed_data.pop('done', None)

@@ -4269,6 +4372,8 @@ async def streaming_chat_response_handler(response, ctx):
                                            'data': data,
                                        }
                                    )
+                        except (asyncio.CancelledError, KeyboardInterrupt):
+                            raise
                        except Exception as e:
                            done = 'data: [DONE]' in line
                            if done:
@@ -4611,7 +4716,7 @@ async def streaming_chat_response_handler(response, ctx):
                            )
                            source_context = source_context.strip()
                            if source_context:
-                                rag_content = rag_template(
+                                rag_content = await rag_template(
                                    request.app.state.config.RAG_TEMPLATE,
                                    source_context,
                                    user_message,
@@ -4656,16 +4761,21 @@ async def streaming_chat_response_handler(response, ctx):
                            **form_data,
                            'model': model_id,
                            'stream': True,
+                            'metadata': metadata,
                        }

                        if ENABLE_RESPONSES_API_STATEFUL and last_response_id:
                            system_message = get_system_message(form_data['messages'])
                            new_form_data['messages'] = (
                                [system_message] if system_message else []
-                            ) + convert_output_to_messages(output, raw=True)
+                            ) + convert_output_to_messages(
+                                output, raw=True, reasoning_format=get_reasoning_format(model)
+                            )
                            new_form_data['previous_response_id'] = last_response_id
                        else:
-                            tool_messages = convert_output_to_messages(output, raw=True)
+                            tool_messages = convert_output_to_messages(
+                                output, raw=True, reasoning_format=get_reasoning_format(model)
+                            )

                            # Chat Completions providers don't support multimodal
                            # tool messages.  Extract images into a user message.
@@ -4762,8 +4872,7 @@ async def streaming_chat_response_handler(response, ctx):
                                code = sanitize_code(code)

                                if CODE_INTERPRETER_BLOCKED_MODULES:
-                                    blocking_code = textwrap.dedent(
-                                        f"""
+                                    blocking_code = textwrap.dedent(f"""
                                        import builtins
    
                                        BLOCKED_MODULES = {CODE_INTERPRETER_BLOCKED_MODULES}
@@ -4779,8 +4888,7 @@ async def streaming_chat_response_handler(response, ctx):
                                            return _real_import(name, globals, locals, fromlist, level)
    
                                        builtins.__import__ = restricted_import
-                                    """
-                                    )
+                                    """)
                                    code = blocking_code + '\n' + code

                                if request.app.state.config.CODE_INTERPRETER_ENGINE == 'pyodide':
@@ -4816,6 +4924,11 @@ async def streaming_chat_response_handler(response, ctx):

                                log.debug(f'Code interpreter output: {ci_output}')

+                                # Handle error responses from event_caller
+                                # (e.g. session disconnected, timeout)
+                                if isinstance(ci_output, dict) and ci_output.get('error'):
+                                    ci_output = {'stderr': ci_output['error']}
+
                                if isinstance(ci_output, dict):
                                    stdout = ci_output.get('stdout', '')

@@ -4879,9 +4992,12 @@ async def streaming_chat_response_handler(response, ctx):
                                **form_data,
                                'model': model_id,
                                'stream': True,
+                                'metadata': metadata,
                                'messages': [
                                    *form_data['messages'],
-                                    *convert_output_to_messages(output, raw=True),
+                                    *convert_output_to_messages(
+                                        output, raw=True, reasoning_format=get_reasoning_format(model)
+                                    ),
                                ],
                            }

@@ -4971,31 +5087,39 @@ async def streaming_chat_response_handler(response, ctx):
                await outlet_filter_handler(ctx)
            except asyncio.CancelledError:
                log.warning('Task was cancelled!')
-                try:
-                    await asyncio.shield(event_emitter({'type': 'chat:tasks:cancel'}))

+                # Close the response body iterator to trigger cleanup
+                # in stream_wrapper's finally block and release the
+                # upstream connection.  Without this, the async
+                # generator is orphaned and may spin in anyio internals.
+                if hasattr(response, 'body_iterator') and hasattr(response.body_iterator, 'aclose'):
+                    try:
+                        await asyncio.shield(response.body_iterator.aclose())
+                    except (asyncio.CancelledError, Exception):
+                        pass
+
+                async def save_cancelled_state():
+                    await event_emitter({'type': 'chat:tasks:cancel'})
                    if not ENABLE_REALTIME_CHAT_SAVE:
-                        # Save message in the database
-                        await asyncio.shield(
-                            Chats.upsert_message_to_chat_by_id_and_message_id(
-                                metadata['chat_id'],
-                                metadata['message_id'],
-                                {
-                                    'done': True,
-                                    'content': serialize_output(output),
-                                    'output': output,
-                                },
-                            )
+                        await Chats.upsert_message_to_chat_by_id_and_message_id(
+                            metadata['chat_id'],
+                            metadata['message_id'],
+                            {
+                                'done': True,
+                                'content': serialize_output(output),
+                                'output': output,
+                            },
                        )
                    else:
-                        await asyncio.shield(
-                            Chats.upsert_message_to_chat_by_id_and_message_id(
-                                metadata['chat_id'],
-                                metadata['message_id'],
-                                {'done': True},
-                            )
+                        await Chats.upsert_message_to_chat_by_id_and_message_id(
+                            metadata['chat_id'],
+                            metadata['message_id'],
+                            {'done': True},
                        )
-                except Exception:
+
+                try:
+                    await asyncio.shield(save_cancelled_state())
+                except (asyncio.CancelledError, Exception):
                    pass
                raise  # re-raise CancelledError for proper propagation

@@ -129,7 +129,11 @@ def get_content_from_message(message: dict) -> Optional[str]:
    return None


-def convert_output_to_messages(output: list, raw: bool = False) -> list[dict]:
+def convert_output_to_messages(
+    output: list,
+    raw: bool = False,
+    reasoning_format: str | None = None,
+) -> list[dict]:
    """
    Convert OR-aligned output items to OpenAI Chat Completion-format messages.

@@ -139,8 +143,14 @@ def convert_output_to_messages(output: list, raw: bool = False) -> list[dict]:

    Args:
        output: List of OR-aligned output items (Responses API format).
-        raw: If True, include reasoning blocks (with original tags) and code
-             interpreter blocks for LLM re-processing follow-ups.
+        raw: If True, include code interpreter blocks for LLM re-processing
+             follow-ups.
+        reasoning_format: How to include reasoning blocks in the output:
+            - None: skip reasoning (default, safe for strict providers).
+            - ``'think_tags'``: wrap in ``<think>`` tags inside content
+              (for Ollama, which expects reasoning as tagged content).
+            - ``'reasoning_content'``: set as ``reasoning_content`` top-level field
+              (for llama.cpp, which routes it via the chat template).
    """
    if not output or not isinstance(output, list):
        return []
@@ -148,19 +158,26 @@ def convert_output_to_messages(output: list, raw: bool = False) -> list[dict]:
    messages = []
    pending_tool_calls = []
    pending_content = []
+    pending_reasoning = []  # Only populated when reasoning_format == 'reasoning_content'

    def flush_pending():
-        nonlocal pending_content, pending_tool_calls
-        if pending_content or pending_tool_calls:
-            messages.append(
-                {
-                    'role': 'assistant',
-                    'content': '\n'.join(pending_content) if pending_content else '',
-                    **({'tool_calls': pending_tool_calls} if pending_tool_calls else {}),
-                }
-            )
-            pending_content = []
-            pending_tool_calls = []
+        nonlocal pending_content, pending_tool_calls, pending_reasoning
+        if not pending_content and not pending_tool_calls and not pending_reasoning:
+            return
+
+        message = {
+            'role': 'assistant',
+            'content': '\n'.join(pending_content) if pending_content else '',
+            **({'tool_calls': pending_tool_calls} if pending_tool_calls else {}),
+        }
+
+        if pending_reasoning:
+            message['reasoning_content'] = '\n'.join(pending_reasoning)
+
+        messages.append(message)
+        pending_content = []
+        pending_tool_calls = []
+        pending_reasoning = []

    for item in output:
        item_type = item.get('type', '')
@@ -231,27 +248,26 @@ def convert_output_to_messages(output: list, raw: bool = False) -> list[dict]:
                )

        elif item_type == 'reasoning':
-            if raw:
-                # Include reasoning with original tags for LLM re-processing
-                reasoning_text = ''
-                source_list = item.get('summary', []) or item.get('content', [])
-                for part in source_list:
-                    if part.get('type') == 'output_text':
-                        reasoning_text += part.get('text', '')
-                    elif 'text' in part:
-                        reasoning_text += part.get('text', '')
+            if not reasoning_format:
+                continue

-                if reasoning_text:
+            reasoning_text = ''
+            source_list = item.get('summary', []) or item.get('content', [])
+            for part in source_list:
+                if part.get('type') == 'output_text':
+                    reasoning_text += part.get('text', '')
+                elif 'text' in part:
+                    reasoning_text += part.get('text', '')
+
+            if reasoning_text:
+                if reasoning_format == 'think_tags':
+                    # Ollama: embed in content with the item's original tags
                    start_tag = item.get('start_tag', '<think>')
                    end_tag = item.get('end_tag', '</think>')
                    pending_content.append(f'{start_tag}{reasoning_text}{end_tag}')
-                    # NOTE: Some providers (e.g. Moonshot/Kimi K2.5) require
-                    # reasoning_content as a top-level field on assistant
-                    # messages.  This should be handled externally via a
-                    # pipeline filter or connection-level middleware, not
-                    # here — adding it universally breaks strict providers
-                    # (OpenAI, Vertex AI, Azure) that reject unknown fields.
-            # else: skip reasoning blocks for normal LLM messages
+                elif reasoning_format == 'reasoning_content':
+                    # llama.cpp: collect for reasoning_content field
+                    pending_reasoning.append(reasoning_text)

        elif item_type == 'open_webui:code_interpreter':
            # Always include code interpreter content so the LLM knows
@@ -597,6 +613,9 @@ def sanitize_text_for_db(text: str) -> str:
    """Remove null bytes and invalid UTF-8 surrogates from text for PostgreSQL storage."""
    if not isinstance(text, str):
        return text
+    # Fast path: skip work when there are no null bytes (the common case)
+    if '\x00' not in text:
+        return text
    # Remove null bytes
    text = text.replace('\x00', '').replace('\u0000', '')
    # Remove invalid UTF-8 surrogate characters that can cause encoding errors
@@ -608,17 +627,38 @@ def sanitize_text_for_db(text: str) -> str:
    return text


-def sanitize_data_for_db(obj):
-    """Recursively sanitize all strings in a data structure for database storage."""
+def _strip_null_bytes_deep(obj):
+    """Inner recursive walk — only called when null bytes are known to be present."""
    if isinstance(obj, str):
        return sanitize_text_for_db(obj)
    elif isinstance(obj, dict):
-        return {k: sanitize_data_for_db(v) for k, v in obj.items()}
+        return {k: _strip_null_bytes_deep(v) for k, v in obj.items()}
    elif isinstance(obj, list):
-        return [sanitize_data_for_db(v) for v in obj]
+        return [_strip_null_bytes_deep(v) for v in obj]
    return obj


+def sanitize_data_for_db(obj):
+    """Recursively sanitize all strings in a data structure for database storage.
+
+    Performs a fast pre-check: serializes the structure once and scans for
+    null bytes.  If none are found (the overwhelmingly common case), the
+    original object is returned immediately, skipping the expensive
+    recursive walk.
+    """
+    if isinstance(obj, str):
+        return sanitize_text_for_db(obj)
+    # Fast path: check for null bytes in the serialized form.
+    # json.dumps is implemented in C and much faster than a Python-level
+    # recursive walk over every leaf string.
+    try:
+        if '\x00' not in json.dumps(obj, ensure_ascii=False):
+            return obj
+    except (TypeError, ValueError):
+        pass
+    return _strip_null_bytes_deep(obj)
+
+
 def sanitize_metadata(metadata: dict) -> dict:
    """
    Return a JSON-safe copy of a metadata dict for database storage.
@@ -843,9 +883,9 @@ def throttle(interval: float = 10.0):
        last_calls = {}
        lock = threading.Lock()

-        def wrapper(*args, **kwargs):
+        async def wrapper(*args, **kwargs):
            if interval is None:
-                return func(*args, **kwargs)
+                return await func(*args, **kwargs)

            key = (args, freeze(kwargs))
            now = time.time()
@@ -855,7 +895,7 @@ def throttle(interval: float = 10.0):
                if now - last_calls.get(key, 0) < interval:
                    return None
                last_calls[key] = now
-            return func(*args, **kwargs)
+            return await func(*args, **kwargs)

        return wrapper

@@ -47,6 +47,7 @@ async def fetch_ollama_models(request: Request, user: UserModel = None):
            'created': int(time.time()),
            'owned_by': 'ollama',
            'ollama': model,
+            'loaded': 'expires_at' in model,
            'connection_type': model.get('connection_type', 'local'),
            'tags': model.get('tags', []),
        }
@@ -199,6 +200,8 @@ async def get_all_models(request, refresh: bool = False, user: UserModel = None)
                'connection_type': connection_type,
                'preset': True,
                **({'pipe': pipe} if pipe is not None else {}),
+                **({'provider': base_model.get('provider')} if base_model and base_model.get('provider') else {}),
+                **({'loaded': base_model.get('loaded')} if base_model and base_model.get('loaded') is not None else {}),
            }

            info = custom_model.model_dump()
@@ -287,9 +290,9 @@ async def get_all_models(request, refresh: bool = False, user: UserModel = None)
    # imported/custom model configs may reference tools or filters the user
    # hasn't installed, and trying to load those would cause persistent
    # "Failed to load function module" log spam on every model refresh.
-    for function_id in functions_by_id:
+    for function_id, function in functions_by_id.items():
        try:
-            await get_function_module_from_cache(request, function_id)
+            await get_function_module_from_cache(request, function_id, function=function)
        except Exception as e:
            log.debug(f'Failed to load function module for {function_id}: {e}')

@@ -1,4 +1,5 @@
 import base64
+from dataclasses import dataclass, field
 import copy
 import hashlib
 import logging
@@ -37,6 +38,7 @@ from open_webui.models.groups import Groups, GroupModel, GroupUpdateForm, GroupF
 from open_webui.config import (
    DEFAULT_USER_ROLE,
    ENABLE_OAUTH_SIGNUP,
+    OAUTH_CLIENT_TIMEOUT,
    OAUTH_REFRESH_TOKEN_INCLUDE_SCOPE,
    OAUTH_MERGE_ACCOUNTS_BY_EMAIL,
    OAUTH_PROVIDERS,
@@ -99,6 +101,7 @@ class OAuthClientMetadata(MCPOAuthClientMetadata):

 class OAuthClientInformationFull(OAuthClientMetadata):
    issuer: Optional[str] = None  # URL of the OAuth server that issued this client
+    resource: Optional[str] = None  # RFC 8707 resource indicator for JWT audience

    client_id: str
    client_secret: str | None = None
@@ -289,12 +292,34 @@ def get_parsed_and_base_url(server_url) -> tuple[urllib.parse.ParseResult, str]:
    return parsed, base_url


-async def get_authorization_server_discovery_urls(server_url: str) -> list[str]:
-    """
-    https://modelcontextprotocol.io/specification/2025-03-26/basic/authorization
-    """
+@dataclass
+class ProtectedResourceMetadata:
+    """RFC 9728 Protected Resource Metadata fields relevant to OAuth flows."""

+    resource: str | None = None
+    authorization_servers: list[str] = field(default_factory=list)
+
+    def get_discovery_urls(self, server_url: str) -> list[str]:
+        """Build all candidate OAuth discovery URLs from this metadata and the server URL."""
+        urls = []
+        for auth_server in self.authorization_servers:
+            urls.extend(_build_well_known_urls(auth_server.rstrip('/')))
+        urls.extend(_build_well_known_urls(server_url))
+        return urls
+
+
+async def get_protected_resource_metadata(server_url: str) -> ProtectedResourceMetadata:
+    """
+    Fetch RFC 9728 Protected Resource Metadata from an MCP server.
+
+    https://modelcontextprotocol.io/specification/2025-03-26/basic/authorization
+
+    Returns:
+        ProtectedResourceMetadata with the resource indicator (RFC 8707)
+        and authorization server URLs discovered from the metadata document.
+    """
    authorization_servers = []
+    resource = None
    try:
        async with aiohttp.ClientSession(trust_env=True) as session:
            async with session.post(
@@ -304,57 +329,66 @@ async def get_authorization_server_discovery_urls(server_url: str) -> list[str]:
                ssl=AIOHTTP_CLIENT_SESSION_SSL,
            ) as response:
                if response.status == 401:
+                    resource_metadata_urls = []
                    match = re.search(
                        r'resource_metadata=(?:"([^"]+)"|([^\s,]+))',
                        response.headers.get('WWW-Authenticate', ''),
                    )
                    if match:
-                        resource_metadata_url = match.group(1) or match.group(2)
-                        log.debug(f'Found resource_metadata URL: {resource_metadata_url}')
+                        resource_metadata_urls = [match.group(1) or match.group(2)]
+                        log.debug(f'Found resource_metadata URL: {resource_metadata_urls[0]}')
+                    else:
+                        # Fall back to well-known resource metadata URIs (RFC 9728 §4.2)
+                        parsed, base_url = get_parsed_and_base_url(server_url)
+                        if parsed.path and parsed.path != '/':
+                            path = parsed.path.rstrip('/')
+                            resource_metadata_urls.append(
+                                urllib.parse.urljoin(base_url, f'/.well-known/oauth-protected-resource{path}')
+                            )
+                        resource_metadata_urls.append(
+                            urllib.parse.urljoin(base_url, '/.well-known/oauth-protected-resource')
+                        )
+                        log.debug(f'No resource_metadata in header, trying well-known URIs: {resource_metadata_urls}')

-                        # Step 2: Fetch Protected Resource metadata
-                        async with session.get(
-                            resource_metadata_url, ssl=AIOHTTP_CLIENT_SESSION_SSL
-                        ) as resource_response:
-                            if resource_response.status == 200:
-                                resource_metadata = await resource_response.json()
+                    # Fetch Protected Resource metadata from candidate URLs
+                    for resource_metadata_url in resource_metadata_urls:
+                        try:
+                            async with session.get(
+                                resource_metadata_url, ssl=AIOHTTP_CLIENT_SESSION_SSL
+                            ) as resource_response:
+                                if resource_response.status == 200:
+                                    resource_metadata = await resource_response.json()

-                                # Step 3: Extract authorization_servers
-                                servers = resource_metadata.get('authorization_servers', [])
-                                if servers:
-                                    authorization_servers = servers
-                                    log.debug(f'Discovered authorization servers: {servers}')
+                                    resource = resource_metadata.get('resource') or None
+                                    if resource:
+                                        log.debug(f'Discovered resource indicator: {resource}')
+
+                                    servers = resource_metadata.get('authorization_servers', [])
+                                    if servers:
+                                        authorization_servers = servers
+                                        log.debug(f'Discovered authorization servers: {servers}')
+                                        break
+                        except Exception as e:
+                            log.debug(f'Failed to fetch resource metadata from {resource_metadata_url}: {e}')
+                            continue
    except Exception as e:
        log.debug(f'MCP Protected Resource discovery failed: {e}')

-    discovery_urls = []
-    for auth_server in authorization_servers:
-        auth_server = auth_server.rstrip('/')
-        discovery_urls.extend(
-            [
-                f'{auth_server}/.well-known/oauth-authorization-server',
-                f'{auth_server}/.well-known/openid-configuration',
-            ]
-        )
-
-    return discovery_urls
+    return ProtectedResourceMetadata(resource=resource, authorization_servers=authorization_servers)


-async def get_discovery_urls(server_url) -> list[str]:
-    urls = await get_authorization_server_discovery_urls(server_url)
+def _build_well_known_urls(server_url: str) -> list[str]:
+    """Build RFC 8414 / OIDC Discovery well-known URLs for a server URL."""
    parsed, base_url = get_parsed_and_base_url(server_url)
+    urls = []

    if parsed.path and parsed.path != '/':
-        # Generate discovery URLs based on https://modelcontextprotocol.io/specification/draft/basic/authorization#authorization-server-metadata-discovery
-        tenant = parsed.path.rstrip('/')
+        path = parsed.path.rstrip('/')
        urls.extend(
            [
-                urllib.parse.urljoin(
-                    base_url,
-                    f'/.well-known/oauth-authorization-server{tenant}',
-                ),
-                urllib.parse.urljoin(base_url, f'/.well-known/openid-configuration{tenant}'),
-                urllib.parse.urljoin(base_url, f'{tenant}/.well-known/openid-configuration'),
+                urllib.parse.urljoin(base_url, f'/.well-known/oauth-authorization-server{path}'),
+                urllib.parse.urljoin(base_url, f'/.well-known/openid-configuration{path}'),
+                urllib.parse.urljoin(base_url, f'{path}/.well-known/openid-configuration'),
            ]
        )

@@ -368,6 +402,12 @@ async def get_discovery_urls(server_url) -> list[str]:
    return urls


+async def get_discovery_urls(server_url) -> list[str]:
+    """Convenience: get all OAuth discovery URLs for a server URL."""
+    metadata = await get_protected_resource_metadata(server_url)
+    return metadata.get_discovery_urls(server_url)
+
+
 # TODO: Some OAuth providers require Initial Access Tokens (IATs) for dynamic client registration.
 # This is not currently supported.
 async def get_oauth_client_info_with_dynamic_client_registration(
@@ -390,7 +430,9 @@ async def get_oauth_client_info_with_dynamic_client_registration(
        )

        # Attempt to fetch OAuth server metadata to get registration endpoint & scopes
-        discovery_urls = await get_discovery_urls(oauth_server_url)
+        resource_metadata = await get_protected_resource_metadata(oauth_server_url)
+        resource = resource_metadata.resource
+        discovery_urls = resource_metadata.get_discovery_urls(oauth_server_url)
        for url in discovery_urls:
            async with aiohttp.ClientSession(trust_env=True) as session:
                async with session.get(url, ssl=AIOHTTP_CLIENT_SESSION_SSL) as oauth_server_metadata_response:
@@ -450,8 +492,9 @@ async def get_oauth_client_info_with_dynamic_client_registration(
                    oauth_client_info = OAuthClientInformationFull.model_validate(
                        {
                            **registration_response_json,
-                            **{'issuer': oauth_server_metadata_url},
-                            **{'server_metadata': oauth_server_metadata},
+                            'issuer': oauth_server_metadata_url,
+                            'server_metadata': oauth_server_metadata,
+                            'resource': resource,
                        }
                    )
                    log.info(
@@ -500,7 +543,9 @@ async def get_oauth_client_info_with_static_credentials(
        redirect_uri = f'{redirect_base_url}/oauth/clients/{client_id}/callback'

        # Discover server metadata (authorization endpoint, token endpoint, scopes, etc.)
-        discovery_urls = await get_discovery_urls(oauth_server_url)
+        resource_metadata = await get_protected_resource_metadata(oauth_server_url)
+        resource = resource_metadata.resource
+        discovery_urls = resource_metadata.get_discovery_urls(oauth_server_url)
        for url in discovery_urls:
            async with aiohttp.ClientSession(trust_env=True) as session:
                async with session.get(url, ssl=AIOHTTP_CLIENT_SESSION_SSL) as resp:
@@ -539,6 +584,7 @@ async def get_oauth_client_info_with_static_credentials(
            token_endpoint_auth_method=token_endpoint_auth_method,
            issuer=oauth_server_metadata_url,
            server_metadata=oauth_server_metadata,
+            resource=resource,
        )

        log.info(
@@ -581,6 +627,7 @@ class OAuthClientManager:
            'client_secret': oauth_client_info.client_secret,
            'client_kwargs': {
                'follow_redirects': True,
+                **({'timeout': int(OAUTH_CLIENT_TIMEOUT.value)} if OAUTH_CLIENT_TIMEOUT.value else {}),
                **({'scope': oauth_client_info.scope} if oauth_client_info.scope else {}),
                **(
                    {'token_endpoint_auth_method': oauth_client_info.token_endpoint_auth_method}
@@ -856,6 +903,11 @@ class OAuthClientManager:
                'refresh_token': token_data['refresh_token'],
                'client_id': client.client_id,
            }
+            # RFC 8707: include resource indicator so refreshed tokens retain correct audience
+            client_info = self.get_client_info(client_id)
+            if client_info and client_info.resource:
+                refresh_data['resource'] = client_info.resource
+
            if hasattr(client, 'client_secret') and client.client_secret:
                refresh_data['client_secret'] = client.client_secret

@@ -908,7 +960,11 @@ class OAuthClientManager:

        redirect_uri = client_info.redirect_uris[0] if client_info.redirect_uris else None
        redirect_uri_str = str(redirect_uri) if redirect_uri else None
-        return await client.authorize_redirect(request, redirect_uri_str)
+        # RFC 8707: pass resource indicator so the IdP sets the correct JWT audience
+        kwargs = {}
+        if client_info.resource:
+            kwargs['resource'] = client_info.resource
+        return await client.authorize_redirect(request, redirect_uri_str, **kwargs)

    async def handle_callback(self, request, client_id: str, user_id: str, response):
        client = self.get_client(client_id) or self.ensure_client_from_config(client_id)
@@ -923,7 +979,11 @@ class OAuthClientManager:
            # The Authlib client already has these configured during add_client().
            # Passing them again causes Authlib to concatenate them (e.g., "ID1,ID1"),
            # which results in 401 errors from the token endpoint. (Fix for #19823)
-            token = await client.authorize_access_token(request)
+            # RFC 8707: pass resource indicator for correct JWT audience on token exchange
+            token_kwargs = {}
+            if client_info and client_info.resource:
+                token_kwargs['resource'] = client_info.resource
+            token = await client.authorize_access_token(request, **token_kwargs)

            # Validate that we received a proper token response
            # If token exchange failed (e.g., 401), we may get an error response instead
@@ -13,7 +13,7 @@ import json
 # What goes out cannot be taken back. Let it be shaped
 # well before it leaves this place.
 # inplace function: form_data is modified
-def apply_system_prompt_to_body(
+async def apply_system_prompt_to_body(
    system: Optional[str],
    form_data: dict,
    metadata: Optional[dict] = None,
@@ -30,7 +30,7 @@ def apply_system_prompt_to_body(
            system = prompt_variables_template(system, variables)

    # Legacy (API Usage)
-    system = prompt_template(system, user)
+    system = await prompt_template(system, user)

    if replace:
        form_data['messages'] = replace_system_message_content(system, form_data.get('messages', []))
@@ -14,7 +14,7 @@ from open_webui.env import (
    OFFLINE_MODE,
    ENABLE_PIP_INSTALL_FRONTMATTER_REQUIREMENTS,
 )
-from open_webui.models.functions import Functions
+from open_webui.models.functions import FunctionModel, Functions
 from open_webui.models.tools import Tools

 log = logging.getLogger(__name__)
@@ -335,13 +335,16 @@ async def get_tool_module_from_cache(request, tool_id, load_from_db=True):
    return tool_module, frontmatter


-async def get_function_module_from_cache(request, function_id, load_from_db=True):
+async def get_function_module_from_cache(
+    request, function_id, function: FunctionModel | None = None, load_from_db=True
+):
    if load_from_db:
        # Always load from the database by default
        # This is useful for hooks like "inlet" or "outlet" where the content might change
        # and we want to ensure the latest content is used.

-        function = await Functions.get_function_by_id(function_id)
+        if function is None:
+            function = await Functions.get_function_by_id(function_id)
        if not function:
            raise Exception(f'Function not found: {function_id}')
        content = function.content
@@ -380,7 +383,11 @@ async def get_function_module_from_cache(request, function_id, load_from_db=True
    return function_module, function_type, frontmatter


+_installed_requirements = set()
+
+
 def install_frontmatter_requirements(requirements: str):
+    global _installed_requirements
    if not ENABLE_PIP_INSTALL_FRONTMATTER_REQUIREMENTS:
        log.info('ENABLE_PIP_INSTALL_FRONTMATTER_REQUIREMENTS is disabled, skipping installation of requirements.')
        return
@@ -392,12 +399,18 @@ def install_frontmatter_requirements(requirements: str):
    if requirements:
        try:
            req_list = [req.strip() for req in requirements.split(',')]
-            log.info(f'Installing requirements: {" ".join(req_list)}')
+            new_reqs = [req for req in req_list if req and req not in _installed_requirements]
+
+            if not new_reqs:
+                return
+
+            log.info(f'Installing requirements: {" ".join(new_reqs)}')
            subprocess.check_call(
-                [sys.executable, '-m', 'pip', 'install'] + PIP_OPTIONS + req_list + PIP_PACKAGE_INDEX_OPTIONS
+                [sys.executable, '-m', 'pip', 'install'] + PIP_OPTIONS + new_reqs + PIP_PACKAGE_INDEX_OPTIONS
            )
+            _installed_requirements.update(new_reqs)
        except Exception as e:
-            log.error(f'Error installing packages: {" ".join(req_list)}')
+            log.error(f'Error installing packages: {" ".join(new_reqs)}')
            raise e

    else:
@@ -21,6 +21,8 @@ from open_webui.env import (

 log = logging.getLogger(__name__)

+MAX_RETRY_COUNT = REDIS_SENTINEL_MAX_RETRY_COUNT
+

 # Let not our connections be timed out but deliver them from
 # partition. For the cache and the socket and the uptime
@@ -38,7 +40,7 @@ class SentinelRedisProxy:
    def _master(self):
        return self._sentinel.master_for(self._service, **self._kw)

-    async def __getattr__(self, item):
+    def __getattr__(self, item):
        master = self._master()
        orig_attr = getattr(master, item)

@@ -135,6 +135,9 @@ def convert_response_ollama_to_openai(ollama_response: dict) -> dict:

 async def convert_streaming_response_ollama_to_openai(ollama_streaming_response):
    has_tool_calls = False
+    # All chunks in a single completion must share the same id (OpenAI spec).
+    completion_id = f'chatcmpl-{str(uuid4())}'
+    first = True
    async for data in ollama_streaming_response.body_iterator:
        data = json.loads(data)

@@ -155,6 +158,12 @@ async def convert_streaming_response_ollama_to_openai(ollama_streaming_response)
            usage = convert_ollama_usage_to_openai(data)

        data = openai_chat_chunk_message_template(model, message_content, reasoning_content, openai_tool_calls, usage)
+        data['id'] = completion_id
+
+        # First chunk must carry delta.role (OpenAI spec).
+        if first:
+            data['choices'][0]['delta']['role'] = 'assistant'
+            first = False

        if done and has_tool_calls:
            data['choices'][0]['finish_reason'] = 'tool_calls'
@@ -35,7 +35,7 @@ def prompt_variables_template(template: str, variables: dict[str, str]) -> str:
    return template


-def prompt_template(template: str, user: Optional[Any] = None) -> str:
+async def prompt_template(template: str, user: Optional[Any] = None) -> str:
    USER_VARIABLES = {}

    if user:
@@ -58,6 +58,19 @@ def prompt_template(template: str, user: Optional[Any] = None) -> str:
                except Exception as e:
                    pass

+            # Resolve user groups from DB only when the template uses {{USER_GROUPS}}
+            groups = ''
+            if '{{USER_GROUPS}}' in template:
+                user_id = user.get('id')
+                if user_id:
+                    try:
+                        from open_webui.models.groups import Groups
+
+                        user_groups = await Groups.get_groups_by_member_id(user_id)
+                        groups = ', '.join(g.name for g in user_groups)
+                    except Exception:
+                        pass
+
            USER_VARIABLES = {
                'name': str(user.get('name')),
                'email': str(user.get('email')),
@@ -66,6 +79,7 @@ def prompt_template(template: str, user: Optional[Any] = None) -> str:
                'gender': str(user.get('gender')),
                'birth_date': str(birth_date),
                'age': str(age),
+                'groups': groups,
            }

    # Get the current date
@@ -88,6 +102,7 @@ def prompt_template(template: str, user: Optional[Any] = None) -> str:
    template = template.replace('{{USER_BIRTH_DATE}}', USER_VARIABLES.get('birth_date', 'Unknown'))
    template = template.replace('{{USER_AGE}}', str(USER_VARIABLES.get('age', 'Unknown')))
    template = template.replace('{{USER_LOCATION}}', USER_VARIABLES.get('location', 'Unknown'))
+    template = template.replace('{{USER_GROUPS}}', USER_VARIABLES.get('groups', ''))

    return template

@@ -243,11 +258,11 @@ def replace_messages_variable(template: str, messages: Optional[list[dict]] = No

 # Let the context given here not distort the question,
 # but illuminate it, so that the answer serves the one who asked.
-def rag_template(template: str, context: str, query: str):
+async def rag_template(template: str, context: str, query: str):
    if template.strip() == '':
        template = DEFAULT_RAG_TEMPLATE

-    template = prompt_template(template)
+    template = await prompt_template(template)

    if '[context]' not in template and '{{CONTEXT}}' not in template:
        log.debug("WARNING: The RAG template does not contain the '[context]' or '{{CONTEXT}}' placeholder.")
@@ -282,51 +297,51 @@ def rag_template(template: str, context: str, query: str):
    return template


-def title_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
+async def title_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
    prompt = get_last_user_message(messages)
    template = replace_prompt_variable(template, prompt)
    template = replace_messages_variable(template, messages)

-    template = prompt_template(template, user)
+    template = await prompt_template(template, user)

    return template


-def follow_up_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
+async def follow_up_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
    prompt = get_last_user_message(messages)
    template = replace_prompt_variable(template, prompt)
    template = replace_messages_variable(template, messages)

-    template = prompt_template(template, user)
+    template = await prompt_template(template, user)
    return template


-def tags_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
+async def tags_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
    prompt = get_last_user_message(messages)
    template = replace_prompt_variable(template, prompt)
    template = replace_messages_variable(template, messages)

-    template = prompt_template(template, user)
+    template = await prompt_template(template, user)
    return template


-def image_prompt_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
+async def image_prompt_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
    prompt = get_last_user_message(messages)
    template = replace_prompt_variable(template, prompt)
    template = replace_messages_variable(template, messages)

-    template = prompt_template(template, user)
+    template = await prompt_template(template, user)
    return template


-def emoji_generation_template(template: str, prompt: str, user: Optional[Any] = None) -> str:
+async def emoji_generation_template(template: str, prompt: str, user: Optional[Any] = None) -> str:
    template = replace_prompt_variable(template, prompt)
-    template = prompt_template(template, user)
+    template = await prompt_template(template, user)

    return template


-def autocomplete_generation_template(
+async def autocomplete_generation_template(
    template: str,
    prompt: str,
    messages: Optional[list[dict]] = None,
@@ -337,16 +352,16 @@ def autocomplete_generation_template(
    template = replace_prompt_variable(template, prompt)
    template = replace_messages_variable(template, messages)

-    template = prompt_template(template, user)
+    template = await prompt_template(template, user)
    return template


-def query_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
+async def query_generation_template(template: str, messages: list[dict], user: Optional[Any] = None) -> str:
    prompt = get_last_user_message(messages)
    template = replace_prompt_variable(template, prompt)
    template = replace_messages_variable(template, messages)

-    template = prompt_template(template, user)
+    template = await prompt_template(template, user)
    return template


@@ -17,8 +17,10 @@ high-cardinality label sets.

 from __future__ import annotations

+import datetime
+import logging
 import time
-from typing import Dict, List, Sequence, Any
+from typing import Dict, Iterable, List, Optional
 from base64 import b64encode

 from fastapi import FastAPI, Request
@@ -36,6 +38,8 @@ from opentelemetry.sdk.metrics.export import (
    PeriodicExportingMetricReader,
 )
 from opentelemetry.sdk.resources import Resource
+from sqlalchemy import Engine, func, select
+from sqlalchemy.orm import Session

 from open_webui.env import (
    OTEL_SERVICE_NAME,
@@ -46,7 +50,47 @@ from open_webui.env import (
    OTEL_METRICS_EXPORTER_OTLP_INSECURE,
    OTEL_METRICS_EXPORT_INTERVAL_MILLIS,
 )
-from open_webui.models.users import Users
+from open_webui.models.users import User
+
+logger = logging.getLogger(__name__)
+
+
+# ---------------------------------------------------------------------------
+# Sync DB helpers for OTel gauge callbacks
+#
+# The OTel Python SDK calls observable-instrument callbacks *synchronously*
+# from a background collection thread — async callbacks are NOT supported
+# (the SDK does not ``await`` the return value).
+#
+# Rather than bridging into the async event loop, we run plain synchronous
+# SQL queries using the sync engine that is already available at setup time.
+# This avoids any cross-thread / cross-loop concerns entirely.
+# ---------------------------------------------------------------------------
+
+
+def _count_total_users(db_engine: Engine) -> Optional[int]:
+    """Return the total number of registered users (sync)."""
+    with Session(db_engine) as session:
+        return session.execute(select(func.count()).select_from(User)).scalar()
+
+
+def _count_active_users(db_engine: Engine) -> Optional[int]:
+    """Return the number of users active within the last 3 minutes (sync)."""
+    three_minutes_ago = int(time.time()) - 180
+    with Session(db_engine) as session:
+        return session.execute(
+            select(func.count()).select_from(User).filter(User.last_active_at >= three_minutes_ago)
+        ).scalar()
+
+
+def _count_users_active_today(db_engine: Engine) -> Optional[int]:
+    """Return the number of users active since midnight today (sync)."""
+    now = int(datetime.datetime.now().timestamp())
+    today_midnight = now - (now % 86400)
+    with Session(db_engine) as session:
+        return session.execute(
+            select(func.count()).select_from(User).filter(User.last_active_at > today_midnight)
+        ).scalar()


 def _build_meter_provider(resource: Resource) -> MeterProvider:
@@ -106,7 +150,7 @@ def _build_meter_provider(resource: Resource) -> MeterProvider:
    return provider


-def setup_metrics(app: FastAPI, resource: Resource) -> None:
+def setup_metrics(app: FastAPI, resource: Resource, db_engine: Engine) -> None:
    """Attach OTel metrics middleware to *app* and initialise provider."""

    metrics.set_meter_provider(_build_meter_provider(resource))
@@ -124,32 +168,46 @@ def setup_metrics(app: FastAPI, resource: Resource) -> None:
        unit='ms',
    )

-    async def observe_active_users(
-        options: metrics.CallbackOptions,
-    ) -> Sequence[metrics.Observation]:
-        return [
-            metrics.Observation(
-                value=await Users.get_active_user_count(),
-            )
-        ]
+    # -- Observable gauge callbacks ----------------------------------------
+    # These are called synchronously by the OTel SDK from a background
+    # collection thread.  They use the sync DB engine directly — no async
+    # bridging required.

-    async def observe_total_registered_users(
+    def observe_total_users(
        options: metrics.CallbackOptions,
-    ) -> Sequence[metrics.Observation]:
-        # IMPORTANT: Use get_num_users() for efficient COUNT(*) query.
-        # Do NOT use len(get_users()["users"]) - it loads ALL user records into memory,
-        # causing connection pool exhaustion on high-latency databases (e.g., Aurora).
-        return [
-            metrics.Observation(
-                value=await Users.get_num_users() or 0,
-            )
-        ]
+    ) -> Iterable[metrics.Observation]:
+        try:
+            value = _count_total_users(db_engine)
+            if value is not None:
+                yield metrics.Observation(value=value)
+        except Exception:
+            logger.debug('Failed to observe total users', exc_info=True)
+
+    def observe_active_users(
+        options: metrics.CallbackOptions,
+    ) -> Iterable[metrics.Observation]:
+        try:
+            value = _count_active_users(db_engine)
+            if value is not None:
+                yield metrics.Observation(value=value)
+        except Exception:
+            logger.debug('Failed to observe active users', exc_info=True)
+
+    def observe_users_active_today(
+        options: metrics.CallbackOptions,
+    ) -> Iterable[metrics.Observation]:
+        try:
+            value = _count_users_active_today(db_engine)
+            if value is not None:
+                yield metrics.Observation(value=value)
+        except Exception:
+            logger.debug('Failed to observe users active today', exc_info=True)

    meter.create_observable_gauge(
        name='webui.users.total',
        description='Total number of registered users',
        unit='users',
-        callbacks=[observe_total_registered_users],
+        callbacks=[observe_total_users],
    )

    meter.create_observable_gauge(
@@ -159,11 +217,6 @@ def setup_metrics(app: FastAPI, resource: Resource) -> None:
        callbacks=[observe_active_users],
    )

-    async def observe_users_active_today(
-        options: metrics.CallbackOptions,
-    ) -> Sequence[metrics.Observation]:
-        return [metrics.Observation(value=await Users.get_num_users_active_today())]
-
    meter.create_observable_gauge(
        name='webui.users.active.today',
        description='Number of users active since midnight today',
@@ -55,4 +55,4 @@ def setup(app: FastAPI, db_engine: Engine):

    # set up metrics only if enabled
    if ENABLE_OTEL_METRICS:
-        setup_metrics(app, resource)
+        setup_metrics(app, resource, db_engine)
@@ -1,4 +1,5 @@
 import base64
+import copy
 import inspect
 import logging
 import re
@@ -7,6 +8,7 @@ import aiohttp
 import asyncio
 import yaml
 import json
+from urllib.parse import quote, urlencode

 from pydantic import BaseModel
 from pydantic.fields import FieldInfo
@@ -54,7 +56,7 @@ from open_webui.env import (
    FORWARD_SESSION_INFO_HEADER_MESSAGE_ID,
    REDIS_KEY_PREFIX,
 )
-from open_webui.utils.headers import include_user_info_headers
+from open_webui.utils.headers import include_user_info_headers, get_custom_headers
 from open_webui.tools.builtin import (
    search_web,
    fetch_url,
@@ -100,7 +102,6 @@ from open_webui.tools.builtin import (
    delete_calendar_event,
 )

-import copy
 from open_webui.utils.access_control import has_permission

 log = logging.getLogger(__name__)
@@ -189,10 +190,11 @@ async def get_tools(request: Request, tool_ids: list[str], user: UserModel, extr
                log.warning(f'Access denied to tool {tool_id} for user {user.id}')
                continue

-            module = request.app.state.TOOLS.get(tool_id, None)
-            if module is None:
-                module, _ = await load_tool_module_by_id(tool_id)
+            module = request.app.state.TOOLS.get(tool_id)
+            if module is None or request.app.state.TOOL_CONTENTS.get(tool_id) != tool.content:
+                module, _ = await load_tool_module_by_id(tool_id, content=tool.content)
                request.app.state.TOOLS[tool_id] = module
+                request.app.state.TOOL_CONTENTS[tool_id] = tool.content

            __user__ = {
                **extra_params['__user__'],
@@ -336,8 +338,9 @@ async def get_tools(request: Request, tool_ids: list[str], user: UserModel, extr

                        connection_headers = tool_server_connection.get('headers', None)
                        if connection_headers and isinstance(connection_headers, dict):
-                            for key, value in connection_headers.items():
-                                headers[key] = value
+                            metadata = extra_params.get('__metadata__', {})
+                            custom_headers = get_custom_headers(connection_headers, user, metadata)
+                            headers.update(custom_headers)

                        # Add user info headers if enabled
                        if ENABLE_FORWARD_USER_INFO_HEADERS and user:
@@ -726,7 +729,6 @@ def clean_properties(schema: dict):


 def clean_openai_tool_schema(spec: dict) -> dict:
-    import copy

    cleaned_spec = copy.deepcopy(spec)

@@ -759,6 +761,11 @@ def get_tool_specs(tool_module: object) -> list[dict]:
    return specs


+# Valid HTTP methods per OpenAPI 3.x – used to skip extension keys (x-*)
+# and non-operation path-item fields (summary, description, servers, parameters).
+OPENAPI_HTTP_METHODS = {'get', 'put', 'post', 'delete', 'options', 'head', 'patch', 'trace'}
+
+
 def resolve_schema(schema, components, resolved_schemas=None):
    """
    Recursively resolves a JSON schema using OpenAPI components.
@@ -795,6 +802,13 @@ def resolve_schema(schema, components, resolved_schemas=None):
    if 'items' in resolved_schema:
        resolved_schema['items'] = resolve_schema(resolved_schema['items'], components)

+    # Resolve composition keywords (oneOf, anyOf, allOf) which may contain $ref
+    for keyword in ('oneOf', 'anyOf', 'allOf'):
+        if keyword in resolved_schema and isinstance(resolved_schema[keyword], list):
+            resolved_schema[keyword] = [
+                resolve_schema(inner, components, resolved_schemas) for inner in resolved_schema[keyword]
+            ]
+
    return resolved_schema


@@ -811,7 +825,20 @@ def convert_openapi_to_tool_payload(openapi_spec):
    tool_payload = []

    for path, methods in openapi_spec.get('paths', {}).items():
+        if not isinstance(methods, dict):
+            continue
+
+        # Path-level parameters apply to all operations under this path
+        # unless overridden at the operation level (matched by name + in).
+        path_level_params = methods.get('parameters', [])
+        if not isinstance(path_level_params, list):
+            path_level_params = []
+
        for method, operation in methods.items():
+            if method not in OPENAPI_HTTP_METHODS:
+                continue
+            if not isinstance(operation, dict):
+                continue
            if operation.get('operationId'):
                tool = {
                    'name': operation.get('operationId'),
@@ -822,7 +849,21 @@ def convert_openapi_to_tool_payload(openapi_spec):
                    'parameters': {'type': 'object', 'properties': {}, 'required': []},
                }

-                for param in operation.get('parameters', []):
+                # Merge path-level and operation-level parameters.
+                # Operation-level params override path-level params with the
+                # same (name, in) pair per the OpenAPI spec.
+                op_params = operation.get('parameters', [])
+                if not isinstance(op_params, list):
+                    op_params = []
+                merged_params = {}
+                for param in path_level_params:
+                    if isinstance(param, dict) and param.get('name'):
+                        merged_params[(param['name'], param.get('in', ''))] = param
+                for param in op_params:
+                    if isinstance(param, dict) and param.get('name'):
+                        merged_params[(param['name'], param.get('in', ''))] = param
+
+                for param in merged_params.values():
                    param_name = param.get('name')
                    if not param_name:
                        continue
@@ -871,29 +912,40 @@ def convert_openapi_to_tool_payload(openapi_spec):


 async def set_tool_servers(request: Request):
-    request.app.state.TOOL_SERVERS = await get_tool_servers_data(request.app.state.config.TOOL_SERVER_CONNECTIONS)
+    try:
+        request.app.state.TOOL_SERVERS = await get_tool_servers_data(request.app.state.config.TOOL_SERVER_CONNECTIONS)
+    except Exception as e:
+        log.error(f'Error fetching tool server data: {e}')
+        request.app.state.TOOL_SERVERS = getattr(request.app.state, 'TOOL_SERVERS', None) or []

-    if request.app.state.redis is not None:
-        await request.app.state.redis.set(
-            f'{REDIS_KEY_PREFIX}:tool_servers', json.dumps(request.app.state.TOOL_SERVERS)
-        )
+    try:
+        if request.app.state.redis is not None:
+            await request.app.state.redis.set(
+                f'{REDIS_KEY_PREFIX}:tool_servers', json.dumps(request.app.state.TOOL_SERVERS)
+            )
+    except Exception as e:
+        log.error(f'Error caching tool_servers to Redis: {e}')

    return request.app.state.TOOL_SERVERS


 async def get_tool_servers(request: Request):
-    tool_servers = []
-    if request.app.state.redis is not None:
-        try:
-            tool_servers = json.loads(await request.app.state.redis.get(f'{REDIS_KEY_PREFIX}:tool_servers'))
-            request.app.state.TOOL_SERVERS = tool_servers
-        except Exception as e:
-            log.error(f'Error fetching tool_servers from Redis: {e}')
+    try:
+        tool_servers = []
+        if request.app.state.redis is not None:
+            try:
+                tool_servers = json.loads(await request.app.state.redis.get(f'{REDIS_KEY_PREFIX}:tool_servers'))
+                request.app.state.TOOL_SERVERS = tool_servers
+            except Exception as e:
+                log.error(f'Error fetching tool_servers from Redis: {e}')

-    if not tool_servers:
-        tool_servers = await set_tool_servers(request)
+        if not tool_servers:
+            tool_servers = await set_tool_servers(request)

-    return tool_servers
+        return tool_servers
+    except Exception as e:
+        log.error(f'Failed to load tool servers, skipping: {e}')
+        return getattr(request.app.state, 'TOOL_SERVERS', None) or []


 async def get_terminal_cwd(
@@ -1154,22 +1206,17 @@ async def get_tool_server_data(url: str, headers: Optional[dict]) -> Dict[str, A
                    error_body = await response.json()
                    raise Exception(error_body)

-                text_content = None
+                text_content = await response.text()

                # Check if URL ends with .yaml or .yml to determine format
                if url.lower().endswith(('.yaml', '.yml')):
-                    text_content = await response.text()
                    res = yaml.safe_load(text_content)
                else:
-                    text_content = await response.text()
-
-                try:
-                    res = json.loads(text_content)
-                except json.JSONDecodeError:
                    try:
+                        res = json.loads(text_content)
+                    except json.JSONDecodeError:
+                        # Fall back to YAML for non-.yml URLs that aren't valid JSON
                        res = yaml.safe_load(text_content)
-                    except Exception as e:
-                        raise e

    except Exception as err:
        log.exception(f'Could not fetch tool server spec from {url}')
@@ -1297,7 +1344,11 @@ async def execute_tool_server(

        matching_route = None
        for route_path, methods in paths.items():
+            if not isinstance(methods, dict):
+                continue
            for http_method, operation in methods.items():
+                if http_method not in OPENAPI_HTTP_METHODS:
+                    continue
                if isinstance(operation, dict) and operation.get('operationId') == name:
                    matching_route = (route_path, methods)
                    break
@@ -1311,6 +1362,10 @@ async def execute_tool_server(

        method_entry = None
        for http_method, operation in methods.items():
+            if http_method not in OPENAPI_HTTP_METHODS:
+                continue
+            if not isinstance(operation, dict):
+                continue
            if operation.get('operationId') == name:
                method_entry = (http_method.lower(), operation)
                break
@@ -1324,7 +1379,22 @@ async def execute_tool_server(
        query_params = {}
        body_params = {}

-        for param in operation.get('parameters', []):
+        # Merge path-level and operation-level parameters for execution.
+        path_level_params = methods.get('parameters', [])
+        if not isinstance(path_level_params, list):
+            path_level_params = []
+        op_params = operation.get('parameters', [])
+        if not isinstance(op_params, list):
+            op_params = []
+        merged_params = {}
+        for param in path_level_params:
+            if isinstance(param, dict) and param.get('name'):
+                merged_params[(param['name'], param.get('in', ''))] = param
+        for param in op_params:
+            if isinstance(param, dict) and param.get('name'):
+                merged_params[(param['name'], param.get('in', ''))] = param
+
+        for param in merged_params.values():
            param_name = param.get('name')
            if not param_name:
                continue
@@ -1342,11 +1412,10 @@ async def execute_tool_server(

        final_url = f'{url.rstrip("/")}{route_path}'
        for key, value in path_params.items():
-            final_url = final_url.replace(f'{{{key}}}', str(value))
+            final_url = final_url.replace(f'{{{key}}}', quote(str(value), safe=''))

        if query_params:
-            query_string = '&'.join(f'{k}={v}' for k, v in query_params.items())
-            final_url = f'{final_url}?{query_string}'
+            final_url = f'{final_url}?{urlencode(query_params)}'

        if operation.get('requestBody', {}).get('content'):
            if params:
@@ -22,12 +22,13 @@ aiocache
 aiofiles
 starlette-compress==1.7.0
 Brotli==1.2.0
+brotlicffi==1.2.0.1
 httpx[socks,http2,zstd,cli,brotli]==0.28.1
 starsessions[redis]==2.2.1

 sqlalchemy==2.0.48
 aiosqlite==0.21.0
-asyncpg==0.30.0
+psycopg[binary]==3.2.9
 alembic==1.18.4
 peewee==3.19.0
 peewee-migrate==1.14.3
@@ -19,13 +19,14 @@ aiocache==0.12.3
 aiofiles==25.1.0
 starlette-compress==1.7.0
 Brotli==1.2.0
+brotlicffi==1.2.0.1
 httpx[socks,http2,zstd,cli,brotli]==0.28.1
 starsessions[redis]==2.2.1
 python-mimeparse==2.0.0

-sqlalchemy==2.0.48
+sqlalchemy[asyncio]==2.0.48
 aiosqlite==0.21.0
-asyncpg==0.30.0
+psycopg[binary]==3.2.9
 alembic==1.18.4
 peewee==3.19.0
 peewee-migrate==1.14.3
@@ -144,9 +145,6 @@ pytest-docker~=3.2.5
 ## LDAP
 ldap3==2.9.1

-## Firecrawl
-firecrawl-py==4.18.0
-
 ## Trace
 opentelemetry-api==1.40.0
 opentelemetry-sdk==1.40.0
@@ -24,7 +24,7 @@ IF NOT "%WEBUI_SECRET_KEY_FILE%" == "" (

 IF "%PORT%"=="" SET PORT=8080
 IF "%HOST%"=="" SET HOST=0.0.0.0
-IF "%FORWARDED_ALLOW_IPS%"=="" SET "FORWARDED_ALLOW_IPS=*"
+IF "%FORWARDED_ALLOW_IPS%"=="" SET "FORWARDED_ALLOW_IPS='*'"
 SET "WEBUI_SECRET_KEY=%WEBUI_SECRET_KEY%"
 SET "WEBUI_JWT_SECRET_KEY=%WEBUI_JWT_SECRET_KEY%"

@@ -47,5 +47,5 @@ IF "%WEBUI_SECRET_KEY% %WEBUI_JWT_SECRET_KEY%" == " " (
 :: Execute uvicorn
 SET "WEBUI_SECRET_KEY=%WEBUI_SECRET_KEY%"
 IF "%UVICORN_WORKERS%"=="" SET UVICORN_WORKERS=1
-uvicorn open_webui.main:app --host "%HOST%" --port "%PORT%" --forwarded-allow-ips "%FORWARDED_ALLOW_IPS%" --workers %UVICORN_WORKERS% --ws auto
+uvicorn open_webui.main:app --host "%HOST%" --port "%PORT%" --forwarded-allow-ips %FORWARDED_ALLOW_IPS% --workers %UVICORN_WORKERS% --ws auto
 :: For ssl user uvicorn open_webui.main:app --host "%HOST%" --port "%PORT%" --forwarded-allow-ips '*' --ssl-keyfile "key.pem" --ssl-certfile "cert.pem" --ws auto
@@ -16,13 +16,26 @@ Based on a precedent of an unacceptable degree of spamming and unsolicited commu

 Any reports or solicitations arriving from sources other than our designated GitHub repository will be dismissed without consideration. We’ve seen how external engagements can dilute and compromise the integrity of community-driven projects, and we’re not here to gamble with the security and privacy of our user community.

+## Foreign CNAs and Vendor Disposition
+
+When a report is filed via GitHub Security Advisories and the maintainers close it as out-of-scope per this policy, that closure is the **vendor's disposition** of the issue. A CVE Numbering Authority (CNA) that mints a CVE for such an issue without reflecting that vendor disposition in the resulting record is acting against vendor disposition.
+
+We respond to such records by:
+
+1. Filing a **REJECT** request with the CVE Program (with **DISPUTED** as fallback);
+2. Cataloging the record publicly, naming the issuing CNA;
+3. Refusing to provide vendor statements, version mappings, fix references, or any other coordination that would lend authority to the record;
+4. Escalating repeated patterns from a single CNA to the CVE Program Root.
+
+**Channel compliance does not entitle a CNA to override vendor disposition.** Reporters who escalate a closed-as-out-of-scope GHSA report to a third-party CNA after vendor disposition has been issued are likewise considered to have acted against vendor disposition, and will be permanently barred from future GHSA submissions.
+
 ## Reporting a Vulnerability

 Reports not submitted through our designated GitHub repository will be disregarded, and we will categorically reject invitations to collaborate on external platforms. Our aggressive stance on this matter underscores our commitment to a secure, transparent, and open community where all operations are visible and contributors are accountable.

 We appreciate the community's interest in identifying potential vulnerabilities. However, effective immediately, we will **not** accept low-effort vulnerability reports. Ensure that **submissions are constructive, actionable, reproducible, well documented and adhere to the following guidelines**:

-1. **Report MUST be a vulnerability:** A security vulnerability is an exploitable weakness where the system behaves in an unintended way, allowing attackers to bypass security controls, gain unauthorized access, execute arbitrary code, or escalate privileges. Configuration options, missing features, and expected protocol behavior are **not vulnerabilities**.
+1. **Report MUST be a vulnerability:** A security vulnerability is an exploitable weakness where the system behaves in an unintended way, allowing attackers to bypass security controls, gain unauthorized access, execute arbitrary code, or escalate privileges. Configuration options, missing features, and expected protocol behavior are **not vulnerabilities**. A vulnerability must cross at least one of the security boundaries (Confidentiality, Integrity, Availability, Authenticity, Non-repudiation). **These boundaries are interpreted broadly; equivalent concepts in other security frameworks fall within them.**

 2. **No Vague Reports**: Submissions such as "I found a vulnerability" without any details will be treated as spam and will not be accepted.

@@ -33,7 +46,7 @@ We appreciate the community's interest in identifying potential vulnerabilities.
 > [!NOTE]
 > A PoC (Proof of Concept) is a **demonstration of exploitation of a vulnerability**. Your PoC must show:
 >
-> 1. Exactly what security boundary was crossed (Confidentiality, Integrity, Availability, Authenticity, Non-repudiation)
+> 1. Exactly what security boundary was crossed (Confidentiality, Integrity, Availability, Authenticity, Non-repudiation - These boundaries are interpreted broadly; equivalent concepts in other security frameworks fall within them)
 > 2. How this vulnerability is triggered/abused (inputs, endpoints, UI actions, etc.)
 > 3. What actions the attacker can now perform
 > 4. What data/action becomes possible that should not be possible
@@ -105,7 +118,14 @@ Your remediation guidance can include, for example:
 > - wrote comments with conflicting information
 > - used illogical and conflicting arguments

-**Non-compliant submissions will be closed, and repeat or extreme violators may be banned.** Our goal is to foster a constructive reporting environment where quality submissions promote better security for all users.
+12. **Self-Affecting Issues Are Not Vulnerabilities:** A vulnerability requires crossing a security boundary that affects **a party other than the reporter**. Crossing one of the five recognized security boundaries (Confidentiality, Integrity, Availability, Authenticity, Non-repudiation - These boundaries are interpreted broadly; equivalent concepts in other security frameworks fall within them) only against the reporter's own data, account, session, or environment is **not a vulnerability** - it is a bug, and belongs in the [Issue Tracker](https://github.com/open-webui/open-webui/issues), not in a security report.
+
+> [!NOTE]
+> This rule is about **who is harmed**, not about severity. A user modifying or deleting their own data, impairing their own session, observing their own configuration, or disabling security controls on their own account is out of scope under this rule, regardless of impact.
+>
+> If the same action also affects another user, the operator, the host system, or shared resources, identify that second party clearly in the PoC, and we want to hear about it.
+
+**Non-compliant submissions will be closed, and repeat or extreme violators may be banned from submitting reports.** Our goal is to foster a constructive reporting environment where quality submissions promote better security for all users.

 ## Where to report the vulnerability

@@ -114,10 +134,24 @@ If you feel like you are not able to follow ALL outlined requirements for vulner

 ## Expected Response Timeframe

-Due to the volume of incoming vulnerability reports, issues, discussions, pull requests, and general project maintenance — lately compounded by a large number of invalid AI-generated reports (see [AI report transparency](#ai-report-transparency)) — our capacity to respond is limited. Open WebUI is a community-driven project maintained by a small team, and security reports are handled alongside all other project responsibilities.
+Due to the very high volume of incoming vulnerability reports, issues, discussions, pull requests, and general project maintenance — lately compounded by an unbelievably high number of AI-generated reports (see [AI report transparency](#ai-report-transparency)) — our capacity to respond is limited. Open WebUI is a community-driven project maintained by a small team, and security reports are handled alongside all other project responsibilities.

 **Please expect several weeks** for your report to be triaged, investigated, fixed, and published. While we aim to respond to every report as quickly as possible, it is normal to experience periods of silence lasting up to several weeks. **This does not mean your report has been ignored** — it means we have not yet had the capacity to address it. The entire process can realistically take multiple weeks from initial submission to final publication. We appreciate your patience and understanding.

+## Report Handling
+
+If you report a valid vulnerability that somebody else reported before you, we will close your report as a duplicate. The earliest filing is the one we will handle going forward, and we will not publish multiple advisories for the same vulnerability.
+
+When multiple independent reporters describe the same vulnerability class but each demonstrates a **distinct and separate exploitation vector** — for example, the same missing authorization check reached through different endpoints — we will consolidate them into the earliest filing and credit every reporter who demonstrated a distinct path. Only one CVE will be issued for the consolidated advisory.
+
+### Why duplicate reports don't receive credit
+
+We credit only the earliest filer of a given vulnerability:
+
+1. **The first report did the work.** By the time a later report arrives, triage and fix are already in motion. Later reports don't change the outcome or timeline; crediting them would misrepresent what moved the fix.
+2. **Credit-for-duplicates incentivizes flooding.** If similar-but-later filings earn credit, the rational play is to skim open advisories and file variations. We already see this pressure — the first-filer rule is what limits it.
+3. **Co-discovery is different from duplication.** Multiple reporters **are credited** on one advisory **when each contributes a _distinct_ finding** — different vector, different affected component, different sub-path the earlier filing does not cover. That is the consolidation rule above. Filing a duplicate of an existing report is not co-discovery.
+
 ## Confidential Disclosure

 Vulnerability reports submitted through GitHub Security Advisories are **private and confidential**. Public disclosure of **ANY** details related to a submitted vulnerability report is **STRICTLY PROHIBITED** until the advisory has been **fully published** — not merely when a CVE ID has been assigned, but when the advisory itself is publicly visible.
@@ -157,4 +191,4 @@ For any other immediate concerns and questions, please create an issue in our [i

 ---

-_Last updated on **2026-03-20**._
+_Last updated on **2026-05-04**._
@@ -1,12 +1,12 @@
 {
 	"name": "open-webui",
-	"version": "0.9.0",
+	"version": "0.9.3",
 	"lockfileVersion": 3,
 	"requires": true,
 	"packages": {
 		"": {
 			"name": "open-webui",
-			"version": "0.9.0",
+			"version": "0.9.3",
 			"dependencies": {
 				"@azure/msal-browser": "^4.5.0",
 				"@codemirror/lang-javascript": "^6.2.2",
@@ -2284,9 +2284,9 @@
 			"license": "Apache-2.0"
 		},
 		"node_modules/@mermaid-js/parser": {
-			"version": "1.0.1",
-			"resolved": "https://registry.npmjs.org/@mermaid-js/parser/-/parser-1.0.1.tgz",
-			"integrity": "sha512-opmV19kN1JsK0T6HhhokHpcVkqKpF+x2pPDKKM2ThHtZAB5F4PROopk0amuVYK5qMrIA4erzpNm8gmPNJgMDxQ==",
+			"version": "1.1.0",
+			"resolved": "https://registry.npmjs.org/@mermaid-js/parser/-/parser-1.1.0.tgz",
+			"integrity": "sha512-gxK9ZX2+Fex5zu8LhRQoMeMPEHbc73UKZ0FQ54YrQtUxE1VVhMwzeNtKRPAu5aXks4FasbMe4xB4bWrmq6Jlxw==",
 			"license": "MIT",
 			"dependencies": {
 				"langium": "^4.0.0"
@@ -3582,9 +3582,9 @@
 			}
 		},
 		"node_modules/@sveltejs/kit": {
-			"version": "2.57.1",
-			"resolved": "https://registry.npmjs.org/@sveltejs/kit/-/kit-2.57.1.tgz",
-			"integrity": "sha512-VRdSbB96cI1EnRh09CqmnQqP/YJvET5buj8S6k7CxaJqBJD4bw4fRKDjcarAj/eX9k2eHifQfDH8NtOh+ZxxPw==",
+			"version": "2.59.1",
+			"resolved": "https://registry.npmjs.org/@sveltejs/kit/-/kit-2.59.1.tgz",
+			"integrity": "sha512-d8OON70AphLdDesuTIl//M2O6fRTIicX8aYv8vhCiYEhTTI2OboKqey0Hu1A4VFhqwgqtq0vKDmPFGkw8kKmgw==",
 			"license": "MIT",
 			"dependencies": {
 				"@standard-schema/spec": "^1.0.0",
@@ -5388,9 +5388,9 @@
 			"license": "MIT"
 		},
 		"node_modules/@xmldom/xmldom": {
-			"version": "0.8.12",
-			"resolved": "https://registry.npmjs.org/@xmldom/xmldom/-/xmldom-0.8.12.tgz",
-			"integrity": "sha512-9k/gHF6n/pAi/9tqr3m3aqkuiNosYTurLLUtc7xQ9sxB/wm7WPygCv8GYa6mS0fLJEHhqMC1ATYhz++U/lRHqg==",
+			"version": "0.8.13",
+			"resolved": "https://registry.npmjs.org/@xmldom/xmldom/-/xmldom-0.8.13.tgz",
+			"integrity": "sha512-KRYzxepc14G/CEpEGc3Yn+JKaAeT63smlDr+vjB8jRfgTBBI9wRj/nkQEO+ucV8p8I9bfKLWp37uHgFrbntPvw==",
 			"license": "MIT",
 			"engines": {
 				"node": ">=10.0.0"
@@ -10948,14 +10948,14 @@
 			}
 		},
 		"node_modules/mermaid": {
-			"version": "11.13.0",
-			"resolved": "https://registry.npmjs.org/mermaid/-/mermaid-11.13.0.tgz",
-			"integrity": "sha512-fEnci+Immw6lKMFI8sqzjlATTyjLkRa6axrEgLV2yHTfv8r+h1wjFbV6xeRtd4rUV1cS4EpR9rwp3Rci7TRWDw==",
+			"version": "11.14.0",
+			"resolved": "https://registry.npmjs.org/mermaid/-/mermaid-11.14.0.tgz",
+			"integrity": "sha512-GSGloRsBs+JINmmhl0JDwjpuezCsHB4WGI4NASHxL3fHo3o/BRXTxhDLKnln8/Q0lRFRyDdEjmk1/d5Sn1Xz8g==",
 			"license": "MIT",
 			"dependencies": {
 				"@braintree/sanitize-url": "^7.1.1",
 				"@iconify/utils": "^3.0.2",
-				"@mermaid-js/parser": "^1.0.1",
+				"@mermaid-js/parser": "^1.1.0",
 				"@types/d3": "^7.4.3",
 				"@upsetjs/venn.js": "^2.0.0",
 				"cytoscape": "^3.33.1",
@@ -10989,9 +10989,9 @@
 			}
 		},
 		"node_modules/mermaid/node_modules/uuid": {
-			"version": "11.1.0",
-			"resolved": "https://registry.npmjs.org/uuid/-/uuid-11.1.0.tgz",
-			"integrity": "sha512-0/A9rDy9P7cJ+8w1c9WD9V//9Wj15Ce2MPz8Ri6032usz+NfePxx5AcN3bN+r6ZL6jEo066/yNYB3tn4pQEx+A==",
+			"version": "11.1.1",
+			"resolved": "https://registry.npmjs.org/uuid/-/uuid-11.1.1.tgz",
+			"integrity": "sha512-vIYxrBCC/N/K+Js3qSN88go7kIfNPssr/hHCesKCQNAjmgvYS2oqr69kIufEG+O4+PfezOH4EbIeHCfFov8ZgQ==",
 			"funding": [
 				"https://github.com/sponsors/broofa",
 				"https://github.com/sponsors/ctavan"
@@ -11847,9 +11847,9 @@
 			}
 		},
 		"node_modules/postcss": {
-			"version": "8.5.8",
-			"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.8.tgz",
-			"integrity": "sha512-OW/rX8O/jXnm82Ey1k44pObPtdblfiuWnrd8X7GJ7emImCOstunGbXUpp7HdBrFQX6rJzn3sPT397Wp5aCwCHg==",
+			"version": "8.5.14",
+			"resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.14.tgz",
+			"integrity": "sha512-SoSL4+OSEtR99LHFZQiJLkT59C5B1amGO1NzTwj7TT1qCUgUO6hxOvzkOYxD+vMrXBM3XJIKzokoERdqQq/Zmg==",
 			"funding": [
 				{
 					"type": "opencollective",
@@ -1,6 +1,6 @@
 {
 	"name": "open-webui",
-	"version": "0.9.0",
+	"version": "0.9.3",
 	"private": true,
 	"scripts": {
 		"dev": "npm run pyodide:fetch && vite dev --host",
@@ -26,12 +26,15 @@ dependencies = [
    "aiocache==0.12.3",
    "aiofiles==25.1.0",
    "starlette-compress==1.7.0",
-    "Brotli==1.1.0",
+    "Brotli==1.2.0",
+    "brotlicffi==1.2.0.1",
    "httpx[socks,http2,zstd,cli,brotli]==0.28.1",
    "starsessions[redis]==2.2.1",
    "python-mimeparse==2.0.0",

-    "sqlalchemy==2.0.48",
+    "sqlalchemy[asyncio]==2.0.48",
+    "aiosqlite==0.21.0",
+    "psycopg[binary]==3.2.9",
    "alembic==1.18.4",
    "peewee==3.19.0",
    "peewee-migrate==1.14.3",
@@ -164,7 +167,6 @@ all = [
    "oracledb==3.4.2",
    "colbert-ai==0.2.22",

-    "firecrawl-py==4.18.0",
    "azure-search-documents==11.6.0",
    "unstructured==0.18.31",
 ]
@@ -22,7 +22,7 @@ const packages = [
 // static/pyodide/ so that the browser can install them offline via micropip.
 // Packages already provided by the Pyodide distribution (click, platformdirs,
 // typing_extensions, etc.) do NOT need to be listed here.
-const pypiPackages = ['black', 'pathspec', 'mypy_extensions'];
+const pypiPackages = ['black', 'pathspec', 'mypy_extensions', 'pytokens'];

 import { loadPyodide } from 'pyodide';
 import { setGlobalDispatcher, ProxyAgent } from 'undici';
@@ -71,6 +71,16 @@
 					metaThemeColorTag.setAttribute('content', '#171717');
 				}

+				const preloadHref = document.documentElement.classList.contains('dark')
+					? '/static/splash-dark.png'
+					: '/static/splash.png';
+				const preload = document.createElement('link');
+				preload.rel = 'preload';
+				preload.as = 'image';
+				preload.href = preloadHref;
+				preload.setAttribute('fetchpriority', 'high');
+				document.head.appendChild(preload);
+
 				window.matchMedia('(prefers-color-scheme: dark)').addListener((e) => {
 					if (localStorage.theme === 'system') {
 						if (e.matches) {
@@ -90,6 +100,8 @@
 				logo.id = 'logo';
 				logo.style =
 					'position: absolute; width: auto; height: 6rem; top: 44%; left: 50%; transform: translateX(-50%); display:block;';
+				logo.loading = 'eager';
+				logo.fetchPriority = 'high';
 				logo.src = isDarkMode ? '/static/splash-dark.png' : '/static/splash.png';

 				document.addEventListener('DOMContentLoaded', function () {
@@ -139,6 +151,8 @@
 					id="logo-her"
 					style="width: auto; height: 13rem"
 					src="/static/splash.png"
+					loading="eager"
+					fetchpriority="high"
 					class="animate-pulse-fast"
 				/>

@@ -328,7 +328,7 @@ export const userSignOut = async () => {
 	let error = null;

 	const res = await fetch(`${WEBUI_API_BASE_URL}/auths/signout`, {
-		method: 'GET',
+		method: 'POST',
 		headers: {
 			'Content-Type': 'application/json'
 		},
@@ -712,3 +712,33 @@ export const deleteAPIKey = async (token: string) => {
 	}
 	return res;
 };
+
+export const deleteOAuthSession = async (token: string, provider: string) => {
+	let error = null;
+
+	const res = await fetch(
+		`${WEBUI_API_BASE_URL}/auths/oauth/sessions/${encodeURIComponent(provider)}`,
+		{
+			method: 'DELETE',
+			headers: {
+				'Content-Type': 'application/json',
+				Authorization: `Bearer ${token}`
+			}
+		}
+	)
+		.then(async (res) => {
+			if (!res.ok) throw await res.json();
+			return res.json();
+		})
+		.catch((err) => {
+			console.error(err);
+			error = err.detail;
+			return null;
+		});
+
+	if (error) {
+		throw error;
+	}
+
+	return res;
+};
@@ -301,34 +301,53 @@ export const getSharedChatList = async (token: string = '', page: number = 1, fi
 };

 export const getAllChats = async (token: string) => {
-	let error = null;
-
 	const res = await fetch(`${WEBUI_API_BASE_URL}/chats/all`, {
 		method: 'GET',
 		headers: {
-			Accept: 'application/json',
-			'Content-Type': 'application/json',
+			Accept: 'application/x-ndjson',
 			...(token && { authorization: `Bearer ${token}` })
 		}
-	})
-		.then(async (res) => {
-			if (!res.ok) throw await res.json();
-			return res.json();
-		})
-		.then((json) => {
-			return json;
-		})
-		.catch((err) => {
-			error = err;
-			console.error(err);
-			return null;
-		});
+	});

-	if (error) {
-		throw error;
+	if (!res.ok) {
+		const err = await res.json();
+		console.error(err);
+		throw err;
 	}

-	return res;
+	const reader = res.body?.getReader();
+	if (!reader) {
+		throw new Error('Response body is not readable');
+	}
+
+	const decoder = new TextDecoder();
+	const chats: object[] = [];
+	let buffer = '';
+
+	while (true) {
+		const { done, value } = await reader.read();
+		if (done) break;
+
+		buffer += decoder.decode(value, { stream: true });
+		const lines = buffer.split('\n');
+		// Keep the last potentially incomplete line in the buffer
+		buffer = lines.pop() ?? '';
+
+		for (const line of lines) {
+			const trimmed = line.trim();
+			if (trimmed) {
+				chats.push(JSON.parse(trimmed));
+			}
+		}
+	}
+
+	// Process any remaining data in the buffer
+	const remaining = buffer.trim();
+	if (remaining) {
+		chats.push(JSON.parse(remaining));
+	}
+
+	return chats;
 };

 export const getChatListBySearchText = async (token: string, text: string, page: number = 1) => {
@@ -378,6 +378,7 @@ type RegisterOAuthClientForm = {
 	client_id: string;
 	client_name?: string;
 	client_secret?: string;
+	oauth_server_url?: string;
 };

 export const registerOAuthClient = async (
@@ -4,6 +4,19 @@ import { getOpenAIModelsDirect } from './openai';

 const TOOL_SERVER_FETCH_TIMEOUT = 10000;

+// Valid HTTP methods per OpenAPI 3.x – used to skip extension keys (x-*)
+// and non-operation path-item fields (summary, description, servers, parameters).
+const OPENAPI_HTTP_METHODS = new Set([
+	'get',
+	'put',
+	'post',
+	'delete',
+	'options',
+	'head',
+	'patch',
+	'trace'
+]);
+
 // Every request sent from here is a petition. May it reach
 // the one for whom it was intended, and return answered.
 export const getModels = async (
@@ -159,6 +172,39 @@ export const getModels = async (
 	return models;
 };

+export const unloadModel = async (token: string, model: string) => {
+	let error = null;
+
+	const res = await fetch(`${WEBUI_BASE_URL}/api/models/unload`, {
+		method: 'POST',
+		headers: {
+			Accept: 'application/json',
+			'Content-Type': 'application/json',
+			...(token && { authorization: `Bearer ${token}` })
+		},
+		body: JSON.stringify({ model })
+	})
+		.then(async (res) => {
+			if (!res.ok) throw await res.json();
+			return res.json();
+		})
+		.catch((err) => {
+			console.error(err);
+			if ('detail' in err) {
+				error = err.detail;
+			} else {
+				error = err;
+			}
+			return null;
+		});
+
+	if (error) {
+		throw error;
+	}
+
+	return res;
+};
+
 type ChatCompletedForm = {
 	model: string;
 	messages: Record<string, unknown>[];
@@ -495,9 +541,15 @@ export const executeToolServer = async (
 	let error = null;

 	try {
-		// Find the matching operationId in the OpenAPI spec
+		// Find the matching operationId in the OpenAPI spec (only valid HTTP methods)
 		const matchingRoute = Object.entries(serverData.openapi.paths).find(([_, methods]) =>
-			Object.entries(methods as any).some(([__, operation]: any) => operation.operationId === name)
+			Object.entries(methods as any).some(
+				([method, operation]: any) =>
+					OPENAPI_HTTP_METHODS.has(method) &&
+					operation &&
+					typeof operation === 'object' &&
+					operation.operationId === name
+			)
 		);

 		if (!matchingRoute) {
@@ -507,7 +559,11 @@ export const executeToolServer = async (
 		const [routePath, methods] = matchingRoute;

 		const methodEntry = Object.entries(methods as any).find(
-			([_, operation]: any) => operation.operationId === name
+			([method, operation]: any) =>
+				OPENAPI_HTTP_METHODS.has(method) &&
+				operation &&
+				typeof operation === 'object' &&
+				operation.operationId === name
 		);

 		if (!methodEntry) {
@@ -516,24 +572,36 @@ export const executeToolServer = async (

 		const [httpMethod, operation]: [string, any] = methodEntry;

+		// Merge path-level and operation-level parameters.
+		// Operation-level params override path-level params with the same (name, in).
+		const pathLevelParams: any[] = Array.isArray((methods as any).parameters)
+			? (methods as any).parameters
+			: [];
+		const opParams: any[] = Array.isArray(operation.parameters) ? operation.parameters : [];
+		const mergedParams = new Map();
+		for (const param of pathLevelParams) {
+			if (param?.name) mergedParams.set(`${param.name}:${param.in ?? ''}`, param);
+		}
+		for (const param of opParams) {
+			if (param?.name) mergedParams.set(`${param.name}:${param.in ?? ''}`, param);
+		}
+
 		// Split parameters by type
 		const pathParams: Record<string, any> = {};
 		const queryParams: Record<string, any> = {};
 		let bodyParams: any = {};

-		if (operation.parameters) {
-			operation.parameters.forEach((param: any) => {
-				const paramName = param?.name;
-				if (!paramName) return;
-				const paramIn = param?.in;
-				if (params.hasOwnProperty(paramName)) {
-					if (paramIn === 'path') {
-						pathParams[paramName] = params[paramName];
-					} else if (paramIn === 'query') {
-						queryParams[paramName] = params[paramName];
-					}
+		for (const param of mergedParams.values()) {
+			const paramName = param?.name;
+			if (!paramName) continue;
+			const paramIn = param?.in;
+			if (params.hasOwnProperty(paramName)) {
+				if (paramIn === 'path') {
+					pathParams[paramName] = params[paramName];
+				} else if (paramIn === 'query') {
+					queryParams[paramName] = params[paramName];
 				}
-			});
+			}
 		}

 		let finalUrl = `${url}${routePath}`;
@@ -152,6 +152,9 @@ export const getBaseModels = async (token: string = '') => {
 export const createNewModel = async (token: string, model: object) => {
 	let error = null;

+	const { id, base_model_id, name, meta, params, access_grants, is_active } = model as any;
+	const payload = { id, base_model_id, name, meta, params, access_grants, is_active };
+
 	const res = await fetch(`${WEBUI_API_BASE_URL}/models/create`, {
 		method: 'POST',
 		headers: {
@@ -159,7 +162,7 @@ export const createNewModel = async (token: string, model: object) => {
 			'Content-Type': 'application/json',
 			authorization: `Bearer ${token}`
 		},
-		body: JSON.stringify(model)
+		body: JSON.stringify(payload)
 	})
 		.then(async (res) => {
 			if (!res.ok) throw await res.json();
@@ -251,6 +254,9 @@ export const toggleModelById = async (token: string, id: string) => {
 export const updateModelById = async (token: string, id: string, model: object) => {
 	let error = null;

+	const { base_model_id, name, meta, params, access_grants, is_active } = model as any;
+	const payload = { id, base_model_id, name, meta, params, access_grants, is_active };
+
 	const res = await fetch(`${WEBUI_API_BASE_URL}/models/model/update`, {
 		method: 'POST',
 		headers: {
@@ -258,7 +264,7 @@ export const updateModelById = async (token: string, id: string, model: object)
 			'Content-Type': 'application/json',
 			authorization: `Bearer ${token}`
 		},
-		body: JSON.stringify({ ...model, id })
+		body: JSON.stringify(payload)
 	})
 		.then(async (res) => {
 			if (!res.ok) throw await res.json();
@@ -36,9 +36,12 @@
 	let auth_type = 'bearer';

 	let connectionType = 'external';
-	let azure = false;
+	let provider = '';
 	$: azure =
-		(url.includes('azure.') || url.includes('cognitive.microsoft.com')) && !direct ? true : false;
+		provider === 'azure' ||
+		((url.includes('azure.') || url.includes('cognitive.microsoft.com')) &&
+			!direct &&
+			provider === '');

 	let prefixId = '';
 	let enable = true;
@@ -98,7 +101,7 @@
 				key,
 				config: {
 					auth_type,
-					azure: azure,
+					...(provider ? { provider } : azure ? { azure: true } : {}),
 					api_version: apiVersion,
 					...(_headers ? { headers: _headers } : {})
 				}
@@ -186,7 +189,8 @@
 				connection_type: connectionType,
 				auth_type,
 				headers: headers ? JSON.parse(headers) : undefined,
-				...(!ollama && azure ? { azure: true, api_version: apiVersion } : {}),
+				...(provider ? { provider } : !ollama && azure ? { azure: true } : {}),
+				...(azure ? { api_version: apiVersion } : {}),
 				...(apiType ? { api_type: apiType } : {})
 			}
 		};
@@ -223,7 +227,7 @@
 				connectionType = connection.config?.connection_type ?? 'local';
 			} else {
 				connectionType = connection.config?.connection_type ?? 'external';
-				azure = connection.config?.azure ?? false;
+				provider = connection.config?.provider ?? (connection.config?.azure ? 'azure' : '');
 				apiVersion = connection.config?.api_version ?? '';
 				apiType = connection.config?.api_type ?? '';
 			}
@@ -491,22 +495,22 @@
 						{#if !ollama && !direct}
 							<div class="flex flex-row justify-between items-center w-full mt-2">
 								<label
-									for="prefix-id-input"
+									for="provider-select"
 									class={`mb-0.5 text-xs text-gray-500
 								${($settings?.highContrastMode ?? false) ? 'text-gray-800 dark:text-gray-100' : ''}`}
-									>{$i18n.t('Provider Type')}</label
+									>{$i18n.t('Provider')}</label
 								>

 								<div>
-									<button
-										on:click={() => {
-											azure = !azure;
-										}}
-										type="button"
-										class=" text-xs text-gray-700 dark:text-gray-300"
+									<select
+										id="provider-select"
+										bind:value={provider}
+										class="text-xs text-gray-700 dark:text-gray-300 bg-transparent outline-hidden"
 									>
-										{azure ? $i18n.t('Azure OpenAI') : $i18n.t('OpenAI')}
-									</button>
+										<option value="">{$i18n.t('Default')}</option>
+										<option value="azure">{$i18n.t('Azure OpenAI')}</option>
+										<option value="llama.cpp">{$i18n.t('llama.cpp')}</option>
+									</select>
 								</div>
 							</div>
 						{/if}
@@ -60,6 +60,7 @@

 	let oauthClientId = '';
 	let oauthClientSecret = '';
+	let oauthServerUrl = '';

 	let enable = true;
 	let loading = false;
@@ -86,10 +87,17 @@
 		// client_id is the tool server ID (used as the internal lookup key for both flows).
 		// For static, client_secret signals the backend to use the static credential path.
 		// The actual OAuth client_id/secret come from the connection info at save time.
-		const formData: { url: string; client_id: string; client_secret?: string } = {
+		const formData: {
+			url: string;
+			client_id: string;
+			client_secret?: string;
+			oauth_server_url?: string;
+		} = {
 			url: url,
 			client_id: id,
-			...(auth_type === 'oauth_2.1_static' ? { client_secret: oauthClientSecret } : {})
+			...(auth_type === 'oauth_2.1_static'
+				? { client_secret: oauthClientSecret, oauth_server_url: oauthServerUrl }
+				: {})
 		};

 		const res = await registerOAuthClient(localStorage.token, formData, 'mcp').catch((err) => {
@@ -336,7 +344,11 @@
 				description: description,
 				...(oauthClientInfo ? { oauth_client_info: oauthClientInfo } : {}),
 				...(auth_type === 'oauth_2.1_static'
-					? { oauth_client_id: oauthClientId, oauth_client_secret: oauthClientSecret }
+					? {
+							oauth_client_id: oauthClientId,
+							oauth_client_secret: oauthClientSecret,
+							oauth_server_url: oauthServerUrl
+						}
 					: {})
 			}
 		};
@@ -364,6 +376,7 @@
 		oauthClientInfo = null;
 		oauthClientId = '';
 		oauthClientSecret = '';
+		oauthServerUrl = '';

 		enable = true;
 		functionNameFilterList = '';
@@ -390,6 +403,7 @@
 			oauthClientInfo = connection.info?.oauth_client_info ?? null;
 			oauthClientId = connection.info?.oauth_client_id ?? '';
 			oauthClientSecret = connection.info?.oauth_client_secret ?? '';
+			oauthServerUrl = connection.info?.oauth_server_url ?? '';

 			enable = connection.config?.enable ?? true;
 			functionNameFilterList = connection.config?.function_name_filter_list ?? '';
@@ -730,6 +744,15 @@
 													placeholder={$i18n.t('Client Secret')}
 													required={false}
 												/>
+												<div class="flex flex-1 items-center">
+													<input
+														class={`w-full text-sm bg-transparent ${($settings?.highContrastMode ?? false) ? 'placeholder:text-gray-700 dark:placeholder:text-gray-100' : 'outline-hidden placeholder:text-gray-300 dark:placeholder:text-gray-700'}`}
+														type="text"
+														bind:value={oauthServerUrl}
+														placeholder={$i18n.t('OAuth Server URL')}
+														autocomplete="off"
+													/>
+												</div>
 											</div>
 										{/if}
 									</div>
@@ -194,10 +194,10 @@
 <div
 	class="pt-0.5 pb-1 gap-1 flex flex-row justify-between items-center sticky top-0 z-10 bg-white dark:bg-gray-900"
 >
-	<div class="text-lg font-medium px-0.5">
+	<div class="text-lg font-medium px-0.5 shrink-0">
 		{$i18n.t('Analytics')}
 	</div>
-	<div class="flex items-center gap-2">
+	<div class="flex items-center gap-2 flex-wrap justify-end min-w-0">
 		{#if groups.length > 0}
 			<select
 				bind:value={selectedGroupId}
@@ -525,7 +525,7 @@
 									TTS_MODEL = 'tts-1';
 								} else if (e.target?.value === 'mistral') {
 									TTS_VOICE = '';
-									TTS_MODEL = 'mistral-tts-latest';
+									TTS_MODEL = 'voxtral-mini-tts-2603';
 								} else {
 									TTS_VOICE = '';
 									TTS_MODEL = '';
@@ -6,7 +6,6 @@
 	import Switch from '$lib/components/common/Switch.svelte';
 	import SensitiveInput from '$lib/components/common/SensitiveInput.svelte';
 	import AddConnectionModal from '$lib/components/AddConnectionModal.svelte';
-	import ConfirmDialog from '$lib/components/common/ConfirmDialog.svelte';

 	import Cog6 from '$lib/components/icons/Cog6.svelte';
 	import Wrench from '$lib/components/icons/Wrench.svelte';
@@ -22,7 +21,6 @@

 	let showManageModal = false;
 	let showConfigModal = false;
-	let showDeleteConfirmDialog = false;
 </script>

 <AddConnectionModal
@@ -35,7 +33,8 @@
 		config: config
 	}}
 	onDelete={() => {
-		showDeleteConfirmDialog = true;
+		onDelete();
+		showConfigModal = false;
 	}}
 	onSubmit={(connection) => {
 		url = connection.url;
@@ -44,14 +43,6 @@
 	}}
 />

-<ConfirmDialog
-	bind:show={showDeleteConfirmDialog}
-	on:confirm={() => {
-		onDelete();
-		showConfigModal = false;
-	}}
-/>
-
 <ManageOllamaModal bind:show={showManageModal} urlIdx={idx} />

 <div class="flex gap-1.5">
@@ -7,7 +7,6 @@
 	import SensitiveInput from '$lib/components/common/SensitiveInput.svelte';
 	import Cog6 from '$lib/components/icons/Cog6.svelte';
 	import AddConnectionModal from '$lib/components/AddConnectionModal.svelte';
-	import ConfirmDialog from '$lib/components/common/ConfirmDialog.svelte';

 	import { connect } from 'socket.io-client';

@@ -21,16 +20,8 @@
 	export let config = {};

 	let showConfigModal = false;
-	let showDeleteConfirmDialog = false;
 </script>

-<ConfirmDialog
-	bind:show={showDeleteConfirmDialog}
-	on:confirm={() => {
-		onDelete();
-	}}
-/>
-
 <AddConnectionModal
 	edit
 	bind:show={showConfigModal}
@@ -40,7 +31,8 @@
 		config
 	}}
 	onDelete={() => {
-		showDeleteConfirmDialog = true;
+		onDelete();
+		showConfigModal = false;
 	}}
 	onSubmit={(connection) => {
 		url = connection.url;
@@ -184,6 +184,13 @@
 			toast.error($i18n.t('Mistral OCR API Key required.'));
 			return;
 		}
+		if (
+			RAGConfig.CONTENT_EXTRACTION_ENGINE === 'paddleocr_vl' &&
+			RAGConfig.PADDLEOCR_VL_BASE_URL === ''
+		) {
+			toast.error($i18n.t('PaddleOCR-vl API URL required.'));
+			return;
+		}

 		if (
 			RAGConfig.CONTENT_EXTRACTION_ENGINE === 'mineru' &&
@@ -356,6 +363,7 @@
 									<option value="datalab_marker">{$i18n.t('Datalab Marker API')}</option>
 									<option value="document_intelligence">{$i18n.t('Document Intelligence')}</option>
 									<option value="mistral_ocr">{$i18n.t('Mistral OCR')}</option>
+									<option value="paddleocr_vl">{$i18n.t('PaddleOCR-vl')}</option>
 									<option value="mineru">{$i18n.t('MinerU')}</option>
 								</select>
 							</div>
@@ -657,6 +665,19 @@
 									bind:value={RAGConfig.MISTRAL_OCR_API_KEY}
 								/>
 							</div>
+						{:else if RAGConfig.CONTENT_EXTRACTION_ENGINE === 'paddleocr_vl'}
+							<div class="my-0.5 flex gap-2 pr-2">
+								<input
+									class="flex-1 w-full text-sm bg-transparent outline-hidden"
+									placeholder={$i18n.t('Enter PaddleOCR-vl API Base URL')}
+									bind:value={RAGConfig.PADDLEOCR_VL_BASE_URL}
+								/>
+								<SensitiveInput
+									placeholder={$i18n.t('Enter PaddleOCR-vl API Token')}
+									bind:value={RAGConfig.PADDLEOCR_VL_TOKEN}
+									required={false}
+								/>
+							</div>
 						{:else if RAGConfig.CONTENT_EXTRACTION_ENGINE === 'mineru'}
 							<!-- API Mode Selection -->
 							<div class="flex w-full mt-2">
@@ -1349,6 +1370,14 @@
 									/>
 								</Tooltip>
 							</div>
+
+							{#if RAGConfig.RAG_TEMPLATE && (RAGConfig.RAG_TEMPLATE.match(/\[context\]/g) || []).length + (RAGConfig.RAG_TEMPLATE.match(/\{\{CONTEXT\}\}/g) || []).length > 1}
+								<div class="mt-1 text-xs text-gray-400 dark:text-gray-500">
+									{$i18n.t(
+										'This template contains multiple context placeholders ([context] or {{CONTEXT}}). Context will be injected at each occurrence.'
+									)}
+								</div>
+							{/if}
 						</div>
 					</div>
 				{/if}
@@ -22,7 +22,6 @@

 	import AddToolServerModal from '$lib/components/AddToolServerModal.svelte';
 	import AddTerminalServerModal from '$lib/components/AddTerminalServerModal.svelte';
-	import ConfirmDialog from '$lib/components/common/ConfirmDialog.svelte';

 	import {
 		getToolServerConnections,
@@ -40,8 +39,6 @@
 	let terminalConnections = [];
 	let showAddTerminalModal = false;
 	let editTerminalIdx: number | null = null;
-	let showDeleteTerminalConfirm = false;
-	let deleteTerminalIdx: number | null = null;

 	const addConnectionHandler = async (server) => {
 		servers = [...servers, server];
@@ -135,23 +132,12 @@
 	}}
 	onDelete={() => {
 		if (editTerminalIdx !== null) {
-			deleteTerminalIdx = editTerminalIdx;
-			showDeleteTerminalConfirm = true;
+			removeTerminalConnection(editTerminalIdx);
 			editTerminalIdx = null;
 		}
 	}}
 />

-<ConfirmDialog
-	bind:show={showDeleteTerminalConfirm}
-	on:confirm={() => {
-		if (deleteTerminalIdx !== null) {
-			removeTerminalConnection(deleteTerminalIdx);
-			deleteTerminalIdx = null;
-		}
-	}}
-/>
-
 <form
 	class="flex flex-col h-full justify-between text-sm"
 	on:submit|preventDefault={() => {
@@ -31,6 +31,7 @@
 		ENABLE_RETRIEVAL_QUERY_GENERATION: true,
 		QUERY_GENERATION_PROMPT_TEMPLATE: '',
 		TOOLS_FUNCTION_CALLING_PROMPT_TEMPLATE: '',
+		ENABLE_VOICE_MODE_PROMPT: true,
 		VOICE_MODE_PROMPT_TEMPLATE: ''
 	};

@@ -236,24 +237,15 @@

 				<div class="mb-2.5 flex w-full items-center justify-between">
 					<div class=" self-center text-xs font-medium">
-						{$i18n.t('Voice Mode Custom Prompt')}
+						{$i18n.t('Voice Mode Prompt')}
 					</div>

-					<Switch
-						state={taskConfig.VOICE_MODE_PROMPT_TEMPLATE != null}
-						on:change={(e) => {
-							if (e.detail) {
-								taskConfig.VOICE_MODE_PROMPT_TEMPLATE = '';
-							} else {
-								taskConfig.VOICE_MODE_PROMPT_TEMPLATE = null;
-							}
-						}}
-					/>
+					<Switch bind:state={taskConfig.ENABLE_VOICE_MODE_PROMPT} />
 				</div>

-				{#if taskConfig.VOICE_MODE_PROMPT_TEMPLATE != null}
+				{#if taskConfig.ENABLE_VOICE_MODE_PROMPT}
 					<div class="mb-2.5">
-						<div class=" mb-1 text-xs font-medium">{$i18n.t('Voice Mode Prompt')}</div>
+						<div class=" mb-1 text-xs font-medium">{$i18n.t('Prompt Template')}</div>

 						<Tooltip
 							content={$i18n.t('Leave empty to use the default prompt, or enter a custom prompt')}
@@ -20,6 +20,7 @@
 		'yacy',
 		'google_pse',
 		'brave',
+		'brave_llm_context',
 		'kagi',
 		'mojeek',
 		'bocha',
@@ -357,6 +358,39 @@
 									/>
 								</div>
 							</div>
+						{:else if webConfig.WEB_SEARCH_ENGINE === 'brave_llm_context'}
+							<div class="mb-2.5 flex w-full flex-col">
+								<div>
+									<div class=" self-center text-xs font-medium mb-1">
+										{$i18n.t('Brave Search API Key')}
+									</div>
+
+									<SensitiveInput
+										placeholder={$i18n.t('Enter Brave Search API Key')}
+										bind:value={webConfig.BRAVE_SEARCH_API_KEY}
+									/>
+								</div>
+								<div class="mt-1.5">
+									<div class=" self-center text-xs font-medium mb-1">
+										{$i18n.t('Context Tokens')}
+									</div>
+
+									<div class="flex w-full">
+										<div class="flex-1">
+											<input
+												class="w-full rounded-lg py-2 px-4 text-sm bg-gray-50 dark:text-gray-300 dark:bg-gray-850 outline-hidden"
+												type="number"
+												min="1024"
+												max="32768"
+												step="1024"
+												placeholder={$i18n.t('Max tokens to retrieve (1024-32768, default 8192)')}
+												bind:value={webConfig.BRAVE_SEARCH_CONTEXT_TOKENS}
+												autocomplete="off"
+											/>
+										</div>
+									</div>
+								</div>
+							</div>
 						{:else if webConfig.WEB_SEARCH_ENGINE === 'kagi'}
 							<div class="mb-2.5 flex w-full flex-col">
 								<div>
@@ -392,6 +392,24 @@
 				{/if}
 			</div>
 		{/if}
+
+		{#if permissions.chat.share}
+			<div class="flex flex-col w-full">
+				<div class="flex w-full justify-between my-1">
+					<div class=" self-center text-xs font-medium">
+						{$i18n.t('Chats Public Sharing')}
+					</div>
+					<Switch bind:state={permissions.sharing.public_chats} />
+				</div>
+				{#if defaultPermissions?.sharing?.public_chats && !permissions.sharing.public_chats}
+					<div>
+						<div class="text-xs text-gray-500">
+							{$i18n.t('This is a default user permission and will remain enabled.')}
+						</div>
+					</div>
+				{/if}
+			</div>
+		{/if}
 	</div>

 	<hr class=" border-gray-100/30 dark:border-gray-850/30" />
@@ -179,6 +179,21 @@
 			<div class="text-[11px] text-gray-400 dark:text-gray-500 uppercase tracking-wider">
 				{$i18n.t('Calendars')}
 			</div>
+			<button
+				class="p-0.5 rounded hover:bg-gray-100 dark:hover:bg-gray-800 transition"
+				title={$i18n.t('New calendar')}
+				on:click={onCreateCalendar}
+			>
+				<svg
+					xmlns="http://www.w3.org/2000/svg"
+					fill="none"
+					viewBox="0 0 24 24"
+					stroke-width="2"
+					stroke="currentColor"
+					class="size-3 text-gray-400 dark:text-gray-500"
+					><path stroke-linecap="round" stroke-linejoin="round" d="M12 4.5v15m7.5-7.5h-15" /></svg
+				>
+			</button>
 		</div>

 		{#each calendars as cal (cal.id)}
@@ -0,0 +1,148 @@
+<script lang="ts">
+	import { createEventDispatcher, getContext } from 'svelte';
+	import { toast } from 'svelte-sonner';
+
+	import Modal from '$lib/components/common/Modal.svelte';
+	import XMark from '$lib/components/icons/XMark.svelte';
+	import Spinner from '$lib/components/common/Spinner.svelte';
+
+	import { createCalendar } from '$lib/apis/calendar';
+
+	const i18n = getContext('i18n');
+	const dispatch = createEventDispatcher();
+
+	export let show = false;
+
+	let name = '';
+	let color = '#3b82f6';
+	let loading = false;
+
+	const PRESET_COLORS = [
+		'#3b82f6', // blue
+		'#ef4444', // red
+		'#22c55e', // green
+		'#f59e0b', // amber
+		'#8b5cf6', // violet
+		'#ec4899', // pink
+		'#06b6d4', // cyan
+		'#f97316' // orange
+	];
+
+	function reset() {
+		name = '';
+		color = '#3b82f6';
+		loading = false;
+	}
+
+	$: if (show) reset();
+
+	const submitHandler = async () => {
+		if (!name.trim()) {
+			toast.error($i18n.t('Name is required'));
+			return;
+		}
+
+		loading = true;
+		try {
+			const result = await createCalendar(localStorage.token, {
+				name: name.trim(),
+				color
+			});
+			if (result) {
+				toast.success($i18n.t('Calendar created'));
+				dispatch('created', result);
+				show = false;
+			}
+		} catch (err) {
+			toast.error(`${err}`);
+		} finally {
+			loading = false;
+		}
+	};
+</script>
+
+<Modal size="sm" bind:show>
+	<div>
+		<!-- Header -->
+		<div class="flex justify-between items-center dark:text-gray-100 px-5 pt-4 pb-2">
+			<h3 class="text-base font-medium">{$i18n.t('New Calendar')}</h3>
+			<button
+				class="self-center shrink-0 ml-2"
+				aria-label={$i18n.t('Close')}
+				on:click={() => (show = false)}
+			>
+				<XMark className="size-5" />
+			</button>
+		</div>
+
+		<!-- Form -->
+		<div class="px-5 pb-2 flex flex-col gap-3">
+			<!-- Name -->
+			<div>
+				<div class="mb-1 text-xs text-gray-500">{$i18n.t('Name')}</div>
+				<input
+					class="w-full text-sm bg-transparent outline-hidden font-primary placeholder:text-gray-300 dark:placeholder:text-gray-700"
+					type="text"
+					bind:value={name}
+					placeholder={$i18n.t('Calendar name')}
+					on:keydown={(e) => {
+						if (e.key === 'Enter') submitHandler();
+					}}
+				/>
+			</div>
+
+			<!-- Color -->
+			<div>
+				<div class="mb-1 text-xs text-gray-500">{$i18n.t('Color')}</div>
+				<div class="flex items-center gap-2 flex-wrap">
+					{#each PRESET_COLORS as c}
+						<button
+							class="size-6 rounded-full transition-all border-2 {color === c
+								? 'border-gray-800 dark:border-white scale-110'
+								: 'border-transparent hover:scale-110'}"
+							style="background-color: {c};"
+							on:click={() => (color = c)}
+							aria-label={c}
+						/>
+					{/each}
+
+					<label
+						class="size-6 rounded-full overflow-hidden cursor-pointer border-2 transition-all {!PRESET_COLORS.includes(
+							color
+						)
+							? 'border-gray-800 dark:border-white scale-110'
+							: 'border-transparent hover:scale-110'}"
+						style="background-color: {color};"
+						title={$i18n.t('Custom color')}
+					>
+						<input type="color" bind:value={color} class="opacity-0 w-0 h-0 absolute" />
+					</label>
+				</div>
+			</div>
+		</div>
+
+		<!-- Bottom toolbar -->
+		<div class="flex items-center justify-end px-4 pb-3.5 pt-2 gap-2">
+			<button
+				class="px-3 py-1 text-xs text-gray-500 hover:text-gray-700 dark:hover:text-gray-200 transition"
+				type="button"
+				on:click={() => (show = false)}
+			>
+				{$i18n.t('Cancel')}
+			</button>
+			<button
+				class="px-3.5 py-1.5 text-sm bg-black hover:bg-gray-900 text-white dark:bg-white dark:text-black dark:hover:bg-gray-100 transition rounded-full flex items-center gap-2 {loading
+					? 'cursor-not-allowed'
+					: ''}"
+				on:click={submitHandler}
+				type="button"
+				disabled={loading}
+			>
+				{$i18n.t('Create')}
+				{#if loading}
+					<span class="shrink-0"><Spinner /></span>
+				{/if}
+			</button>
+		</div>
+	</div>
+</Modal>
@@ -356,6 +356,7 @@
 							<Markdown
 								id={`${message.id}-reply-to`}
 								content={message?.reply_to_message?.content}
+								allowEmbeds={false}
 							/>
 						</div>
 					</button>
@@ -527,6 +528,7 @@
 									id={message.id}
 									content={message.content}
 									paragraphTag="span"
+									allowEmbeds={!!message?.meta?.model_id}
 								/>{#if message.created_at !== message.updated_at && (message?.meta?.model_id ?? null) === null}<span
 										class="text-gray-500 text-[10px] pl-1 self-center">({$i18n.t('edited')})</span
 									>{/if}
@@ -70,6 +70,7 @@
 	import {
 		archiveChatById,
 		createNewChat,
+		deleteChatById,
 		getAllTags,
 		getChatById,
 		getChatList,
@@ -101,6 +102,7 @@
 	import Navbar from '$lib/components/chat/Navbar.svelte';
 	import ChatControls from './ChatControls.svelte';
 	import EventConfirmDialog from '../common/ConfirmDialog.svelte';
+	import DeleteConfirmDialog from '../common/ConfirmDialog.svelte';
 	import Placeholder from './Placeholder.svelte';
 	import FilesOverlay from './MessageInput/FilesOverlay.svelte';
 	import NotificationToast from '../NotificationToast.svelte';
@@ -119,8 +121,10 @@
 	let controlPaneComponent: ChatControls | undefined;

 	let messageInput: MessageInput | undefined;
+	let messagesRef: Messages | undefined;

 	let autoScroll = true;
+	let isNearTop = true;
 	let processing = '';
 	let messagesContainerElement: HTMLDivElement;

@@ -180,6 +184,12 @@
 		navigateHandler();
 	}

+	let saveControlsTimer;
+	$: if (!loading && !$temporaryChatEnabled && $chatId && params && chatFiles) {
+		clearTimeout(saveControlsTimer);
+		saveControlsTimer = setTimeout(saveControls, 400);
+	}
+
 	const navigateHandler = async () => {
 		// Mark the outgoing chat as read before loading the new one.
 		// $chatId still holds the previous chat here — loadChat() updates it.
@@ -187,6 +197,8 @@
 			updateLastReadAt($chatId);
 		}

+		clearTimeout(saveControlsTimer);
+		await saveControls();
 		loading = true;

 		prompt = '';
@@ -287,7 +299,7 @@
 		oldSelectedModelIds = structuredClone(selectedModelIds);
 	};

-	const resetInput = () => {
+	const resetInput = async () => {
 		selectedToolIds = [];
 		selectedFilterIds = [];
 		pendingOAuthTools = [];
@@ -296,10 +308,18 @@
 		codeInterpreterEnabled = false;

 		if (selectedModelIds.filter((id) => id).length > 0) {
-			setDefaults();
+			await setDefaults();
 		}
 	};

+	/** Check whether a terminal ID references an available system or direct terminal. */
+	const isTerminalAvailable = (tid: string): boolean => {
+		return (
+			($terminalServers ?? []).some((t) => t.id && t.id === tid) ||
+			($settings?.terminalServers ?? []).some((s) => s.url === tid)
+		);
+	};
+
 	const setDefaults = async () => {
 		if (!$tools) {
 			tools.set(await getTools(localStorage.token));
@@ -378,9 +398,12 @@
 				}
 			}

-			// Set Default Terminal
+			// Set Default Terminal — only if the referenced terminal actually exists
 			if (model?.info?.meta?.terminalId) {
-				selectedTerminalId.set(model.info.meta.terminalId);
+				const tid = model.info.meta.terminalId;
+				if (isTerminalAvailable(tid)) {
+					selectedTerminalId.set(tid);
+				}
 			}
 		}
 	};
@@ -718,6 +741,11 @@
 			});
 		}

+		// Clear stale selectedTerminalId if the referenced terminal no longer exists
+		if ($selectedTerminalId && !isTerminalAvailable($selectedTerminalId)) {
+			selectedTerminalId.set(null);
+		}
+
 		const pageSubscribe = page.subscribe(async (p) => {
 			if (p.url.pathname === '/') {
 				await tick();
@@ -807,6 +835,8 @@

 		return () => {
 			try {
+				clearTimeout(saveControlsTimer);
+				saveControls();
 				if (chatIdProp && !$temporaryChatEnabled) {
 					updateLastReadAt(chatIdProp);
 				}
@@ -1190,7 +1220,7 @@

 		autoScroll = true;

-		resetInput();
+		await resetInput();
 		await chatId.set('');
 		await chatTitle.set('');

@@ -1349,6 +1379,30 @@
 						? chatContent.history
 						: convertMessagesToHistory(chatContent.messages);

+				// Sanitize history: repair orphaned references from failed regenerations (#24424)
+				for (const message of Object.values(history.messages)) {
+					if (message.childrenIds) {
+						message.childrenIds = message.childrenIds.filter(
+							(childId) => history.messages[childId]
+						);
+					}
+				}
+				if (history.currentId && !history.messages[history.currentId]) {
+					const messageIds = Object.keys(history.messages);
+					let lastMessageId = null;
+					for (const messageId of messageIds) {
+						const message = history.messages[messageId];
+						if (
+							(message.childrenIds ?? []).length === 0 &&
+							(!lastMessageId ||
+								(message.timestamp ?? 0) > (history.messages[lastMessageId].timestamp ?? 0))
+						) {
+							lastMessageId = messageId;
+						}
+					}
+					history.currentId = lastMessageId ?? messageIds[0] ?? null;
+				}
+
 				chatTitle.set(chatContent.title);

 				params = chatContent?.params ?? {};
@@ -1360,6 +1414,7 @@
 				autoScroll = true;
 				await tick();

+				// Mark all non-current assistant messages as done
 				if (history.currentId) {
 					for (const message of Object.values(history.messages)) {
 						if (
@@ -1373,23 +1428,23 @@
 					}
 				}

-				const taskRes = await getTaskIdsByChatId(localStorage.token, $chatId).catch((error) => {
-					return null;
-				});
-
-				if (taskRes) {
-					taskIds = taskRes.task_ids;
-				}
-
-				// If no active tasks and current message is incomplete, generation was interrupted
+				// Reconcile active tasks with message state:
+				// If the response is already done, remaining tasks are just background
+				// work (follow-ups, title gen) that shouldn't block the input.
+				const pendingTaskIds = await getTaskIdsByChatId(localStorage.token, $chatId)
+					.then((res) => res?.task_ids ?? [])
+					.catch(() => []);
 				const currentMessage = history.currentId ? history.messages[history.currentId] : null;
-				if (
-					currentMessage &&
-					currentMessage.role === 'assistant' &&
-					!currentMessage.done &&
-					(!taskIds || taskIds.length === 0)
-				) {
-					currentMessage.done = true;
+				const responseComplete = currentMessage?.role === 'assistant' && currentMessage?.done;
+
+				if (pendingTaskIds.length > 0 && !responseComplete) {
+					taskIds = pendingTaskIds;
+				} else {
+					taskIds = null;
+					// No active tasks and message incomplete → generation was interrupted
+					if (currentMessage?.role === 'assistant' && !currentMessage.done) {
+						currentMessage.done = true;
+					}
 				}

 				await tick();
@@ -1411,6 +1466,10 @@
 		}
 	};

+	const scrollToTop = async () => {
+		await messagesRef?.scrollToTop();
+	};
+
 	let scrollRAF = null;
 	let contentsRAF = null;
 	const scheduleScrollToBottom = () => {
@@ -1967,11 +2026,13 @@
 		{
 			messages = null,
 			modelId = null,
-			modelIdx = null
+			modelIdx = null,
+			regenerationPrompt = null
 		}: {
 			messages?: any[] | null;
 			modelId?: string | null;
 			modelIdx?: number | null;
+			regenerationPrompt?: string | null;
 		} = {}
 	) => {
 		if (autoScroll) {
@@ -2083,7 +2144,10 @@
 				_history,
 				primaryResponseMessageId,
 				_chatId,
-				selectedModelIds.length > 1 ? messageIdsMap : undefined
+				{
+					messageIdsMap: selectedModelIds.length > 1 ? messageIdsMap : undefined,
+					regenerationPrompt
+				}
 			);

 			if (chatEventEmitter) clearInterval(chatEventEmitter);
@@ -2148,7 +2212,15 @@
 		_history,
 		responseMessageId,
 		_chatId,
-		messageIdsMap?: Record<string, string>
+		{
+			messageIdsMap,
+			regenerationPrompt,
+			continueResponse = false
+		}: {
+			messageIdsMap?: Record<string, string>;
+			regenerationPrompt?: string | null;
+			continueResponse?: boolean;
+		} = {}
 	) => {
 		const responseMessage = _history.messages[responseMessageId];
 		const userMessage = _history.messages[responseMessage.parentId];
@@ -2167,7 +2239,7 @@
 		files.push(
 			...(userMessage?.files ?? []).filter(
 				(item) =>
-					['doc', 'text', 'note', 'chat', 'collection'].includes(item.type) ||
+					['doc', 'text', 'note', 'chat', 'collection', 'folder'].includes(item.type) ||
 					(item.type === 'file' && !(item?.content_type ?? '').startsWith('image/'))
 			)
 		);
@@ -2204,6 +2276,7 @@
 				? { role: 'system', content: `${params?.system ?? $settings?.system ?? ''}` }
 				: undefined
 		].filter(Boolean);
+
 		if ($temporaryChatEnabled) {
 			messages = [
 				...messages,
@@ -2347,6 +2420,8 @@
 				...(messageIdsMap ? { message_ids: messageIdsMap } : {}),
 				parent_id: userMessage?.parentId ?? null,
 				user_message: userMessage,
+				...(regenerationPrompt ? { regeneration_prompt: regenerationPrompt } : {}),
+				...(continueResponse ? { assistant_message_id: responseMessageId } : {}),

 				background_tasks: {
 					...(!$temporaryChatEnabled && !_chatId && (userMessage?.parentId ?? null) === null
@@ -2416,6 +2491,16 @@
 						window.history.replaceState(history.state, '', `/c/${res.chat_id}`);
 						currentChatPage.set(1);
 						await chats.set(await getChatList(localStorage.token, $currentChatPage));
+
+						// Persist chat-level params (system prompt, advanced
+						// params) that the backend doesn't receive in the
+						// chat completion request.  Files are now persisted
+						// by the backend at chat creation time.
+						if (Object.keys(params).length > 0) {
+							await updateChatById(localStorage.token, res.chat_id, {
+								params: params
+							});
+						}
 					}
 				}
 			}
@@ -2562,13 +2647,8 @@
 			await sendMessage(history, userMessage.id, {
 				...(suggestionPrompt
 					? {
-							messages: [
-								...createMessagesList(history, message.id),
-								{
-									role: 'user',
-									content: suggestionPrompt
-								}
-							]
+							messages: createMessagesList(history, message.id),
+							regenerationPrompt: suggestionPrompt
 						}
 					: {}),
 				...((userMessage?.models ?? [...selectedModels]).length > 1
@@ -2601,7 +2681,8 @@
 					createMessagesList(history, responseMessage.id),
 					history,
 					responseMessage.id,
-					_chatId
+					_chatId,
+					{ continueResponse: true }
 				);
 			}
 		}
@@ -2715,6 +2796,13 @@
 		}
 	};

+	const saveControls = async () => {
+		if (!$chatId || $temporaryChatEnabled) return;
+		await updateChatById(localStorage.token, $chatId, { params, files: chatFiles }).catch((err) =>
+			console.error('[controls autosave]', err)
+		);
+	};
+
 	const MAX_DRAFT_LENGTH = 5000;
 	let saveDraftTimeout: ReturnType<typeof setTimeout> | null = null;

@@ -2777,6 +2865,33 @@
 			toast.error($i18n.t('Failed to archive chat.'));
 		}
 	};
+
+	let showDeleteConfirm = false;
+
+	const deleteChatHandler = async (id: string) => {
+		showDeleteConfirm = true;
+	};
+
+	const confirmDeleteChat = async () => {
+		const id = $chatId;
+		if (!id) return;
+
+		try {
+			const res = await deleteChatById(localStorage.token, id);
+			if (res) {
+				currentChatPage.set(1);
+				initNewChat();
+				await goto('/');
+				chats.set(await getChatList(localStorage.token, $currentChatPage));
+				pinnedChats.set(await getPinnedChatList(localStorage.token));
+				allTags.set(await getAllTags(localStorage.token));
+				toast.success($i18n.t('Chat deleted.'));
+			}
+		} catch (error) {
+			console.error('Error deleting chat:', error);
+			toast.error(`${error}`);
+		}
+	};
 </script>

 <svelte:head>
@@ -2787,7 +2902,19 @@
 	</title>
 </svelte:head>

-<audio id="audioElement" src="" style="display: none;"></audio>
+<audio id="audioElement" style="display: none;"></audio>
+
+<DeleteConfirmDialog
+	bind:show={showDeleteConfirm}
+	title={$i18n.t('Delete chat?')}
+	on:confirm={() => {
+		confirmDeleteChat();
+	}}
+>
+	<div class=" text-sm text-gray-500 flex-1 line-clamp-3">
+		{$i18n.t('This will delete')} <span class="  font-semibold">{$chatTitle}</span>.
+	</div>
+</DeleteConfirmDialog>

 <EventConfirmDialog
 	bind:show={showEventConfirmation}
@@ -2859,7 +2986,9 @@
 						bind:selectedModels
 						shareEnabled={!!history.currentId}
 						{initNewChat}
+						scrollToTop={!isNearTop ? scrollToTop : null}
 						{archiveChatHandler}
+						{deleteChatHandler}
 						{moveChatHandler}
 						onSaveTempChat={async () => {
 							try {
@@ -2910,10 +3039,12 @@
 									autoScroll =
 										messagesContainerElement.scrollHeight - messagesContainerElement.scrollTop <=
 										messagesContainerElement.clientHeight + 5;
+									isNearTop = messagesContainerElement.scrollTop <= 100;
 								}}
 							>
 								<div class=" h-full w-full flex flex-col">
 									<Messages
+										bind:this={messagesRef}
 										chatId={$chatId}
 										bind:history
 										bind:autoScroll
@@ -1,27 +1,14 @@
 <script lang="ts">
-	import { toast } from 'svelte-sonner';
-
-	import DOMPurify from 'dompurify';
-	import { marked } from 'marked';
-
-	import { getContext, tick, onDestroy } from 'svelte';
+	import { getContext, tick } from 'svelte';
 	const i18n = getContext('i18n');

-	import { chatCompletion } from '$lib/apis/openai';
-
 	import ChatBubble from '$lib/components/icons/ChatBubble.svelte';
 	import LightBulb from '$lib/components/icons/LightBulb.svelte';
-	import Markdown from '../Messages/Markdown.svelte';
-	import Skeleton from '../Messages/Skeleton.svelte';
-	import { chatId, models, socket } from '$lib/stores';

 	export let id = '';
-	export let messageId = '';

-	export let model = null;
-	export let messages = [];
 	export let actions = [];
-	export let onAdd = (e) => {};
+	export let onSetInputText = (text) => {};

 	let floatingInput = false;
 	let selectedAction = null;
@@ -29,11 +16,6 @@
 	let selectedText = '';
 	let floatingInputValue = '';

-	let content = '';
-	let responseContent = null;
-	let responseDone = false;
-	let controller = null;
-
 	$: if (actions.length === 0) {
 		actions = DEFAULT_ACTIONS;
 	}
@@ -54,25 +36,7 @@
 		}
 	];

-	const autoScroll = async () => {
-		const responseContainer = document.getElementById('response-container');
-		if (responseContainer) {
-			// Scroll to bottom only if the scroll is at the bottom give 50px buffer
-			if (
-				responseContainer.scrollHeight - responseContainer.clientHeight <=
-				responseContainer.scrollTop + 50
-			) {
-				responseContainer.scrollTop = responseContainer.scrollHeight;
-			}
-		}
-	};
-
-	const actionHandler = async (actionId) => {
-		if (!model) {
-			toast.error($i18n.t('Model not selected'));
-			return;
-		}
-
+	const actionHandler = (actionId) => {
 		let selectedContent = selectedText
 			.split('\n')
 			.map((line) => `> ${line}`)
@@ -80,27 +44,20 @@

 		let selectedAction = actions.find((action) => action.id === actionId);
 		if (!selectedAction) {
-			toast.error($i18n.t('Action not found'));
 			return;
 		}

 		let prompt = selectedAction?.prompt ?? '';
-		let toolIds = [];

 		// Handle: {{variableId|tool:id="toolId"}} pattern
 		// This regex captures variableId and toolId from {{variableId|tool:id="toolId"}}
 		const varToolPattern = /\{\{(.*?)\|tool:id="([^"]+)"\}\}/g;
 		prompt = prompt.replace(varToolPattern, (match, variableId, toolId) => {
-			toolIds.push(toolId);
 			return variableId; // Replace with just variableId
 		});

 		// legacy {{TOOL:toolId}} pattern (for backward compatibility)
 		let toolIdPattern = /\{\{TOOL:([^\}]+)\}\}/g;
-		let match;
-		while ((match = toolIdPattern.exec(prompt)) !== null) {
-			toolIds.push(match[1]);
-		}

 		// Remove all TOOL placeholders from the prompt
 		prompt = prompt.replace(toolIdPattern, '');
@@ -113,133 +70,17 @@
 		prompt = prompt.replace('{{CONTENT}}', selectedText);
 		prompt = prompt.replace('{{SELECTED_CONTENT}}', selectedContent);

-		content = prompt;
-		responseContent = '';
-
-		let res;
-		[res, controller] = await chatCompletion(localStorage.token, {
-			model: model,
-			model_item: $models.find((m) => m.id === model),
-			session_id: $socket?.id,
-			chat_id: $chatId,
-			messages: [
-				...messages,
-				{
-					role: 'user',
-					content: content
-				}
-			].map((message) => ({
-				role: message.role,
-				content: message.content
-			})),
-			...(toolIds.length > 0
-				? {
-						tool_ids: toolIds
-						// params: {
-						// 	function_calling: 'native'
-						// }
-					}
-				: {}),
-
-			stream: true // Enable streaming
-		});
-
-		if (res && res.ok) {
-			const reader = res.body.getReader();
-			const decoder = new TextDecoder();
-
-			const processStream = async () => {
-				while (true) {
-					// Read data chunks from the response stream
-					const { done, value } = await reader.read();
-					if (done) {
-						break;
-					}
-
-					// Decode the received chunk
-					const chunk = decoder.decode(value, { stream: true });
-
-					// Process lines within the chunk
-					const lines = chunk.split('\n').filter((line) => line.trim() !== '');
-
-					for (const line of lines) {
-						if (line.startsWith('data: ')) {
-							if (line.startsWith('data: [DONE]')) {
-								responseDone = true;
-
-								await tick();
-								autoScroll();
-								continue;
-							} else {
-								// Parse the JSON chunk
-								try {
-									const data = JSON.parse(line.slice(6));
-
-									// Append the `content` field from the "choices" object
-									if (data.choices && data.choices[0]?.delta?.content) {
-										responseContent += data.choices[0].delta.content;
-
-										autoScroll();
-									}
-								} catch (e) {
-									console.error(e);
-								}
-							}
-						}
-					}
-				}
-			};
-
-			// Process the stream in the background
-			try {
-				await processStream();
-			} catch (e) {
-				if (e.name !== 'AbortError') {
-					console.error(e);
-				}
-			}
-		} else {
-			toast.error($i18n.t('An error occurred while fetching the explanation'));
-		}
-	};
-
-	const addHandler = async () => {
-		const messages = [
-			{
-				role: 'user',
-				content: content
-			},
-			{
-				role: 'assistant',
-				content: responseContent
-			}
-		];
-
-		onAdd({
-			modelId: model,
-			parentId: messageId,
-			messages: messages
-		});
+		// Prepopulate the main chat input instead of inline streaming
+		onSetInputText(prompt);
+		closeHandler();
 	};

 	export const closeHandler = () => {
-		if (controller) {
-			controller.abort();
-		}
-
 		selectedAction = null;
 		selectedText = '';
-		responseContent = null;
-		responseDone = false;
 		floatingInput = false;
 		floatingInputValue = '';
 	};
-
-	onDestroy(() => {
-		if (controller) {
-			controller.abort();
-		}
-	});
 </script>

 <div
@@ -247,120 +88,82 @@
 	class="absolute rounded-lg mt-1 text-xs z-9999"
 	style="display: none"
 >
-	{#if responseContent === null}
-		{#if !floatingInput}
-			<div
-				class="flex flex-row shrink-0 p-0.5 bg-white dark:bg-gray-850 dark:text-gray-100 text-medium rounded-xl shadow-xl border border-gray-100 dark:border-gray-800"
-			>
-				{#each actions as action}
-					<button
-						aria-label={action.label}
-						class="px-1.5 py-[1px] hover:bg-gray-50 dark:hover:bg-gray-800 rounded-xl flex items-center gap-1 min-w-fit transition"
-						on:click={async () => {
-							selectedText = window.getSelection().toString();
-							selectedAction = action;
+	{#if !floatingInput}
+		<div
+			class="flex flex-row shrink-0 p-0.5 bg-white dark:bg-gray-850 dark:text-gray-100 text-medium rounded-xl shadow-xl border border-gray-100 dark:border-gray-800"
+		>
+			{#each actions as action}
+				<button
+					aria-label={action.label}
+					class="px-1.5 py-[1px] hover:bg-gray-50 dark:hover:bg-gray-800 rounded-xl flex items-center gap-1 min-w-fit transition"
+					on:click={async () => {
+						selectedText = window.getSelection().toString();
+						selectedAction = action;

-							if (action.prompt.includes('{{INPUT_CONTENT}}')) {
-								floatingInput = true;
-								floatingInputValue = '';
+						if (action.prompt.includes('{{INPUT_CONTENT}}')) {
+							floatingInput = true;
+							floatingInputValue = '';

-								await tick();
-								setTimeout(() => {
-									const input = document.getElementById('floating-message-input');
-									if (input) {
-										input.focus();
-									}
-								}, 0);
-							} else {
-								actionHandler(action.id);
-							}
-						}}
-					>
-						{#if action.icon}
-							<svelte:component this={action.icon} className="size-3 shrink-0" />
-						{/if}
-						<div class="shrink-0">{action.label}</div>
-					</button>
-				{/each}
-			</div>
-		{:else}
-			<div
-				class="py-1 flex dark:text-gray-100 bg-white dark:bg-gray-850 border border-gray-100 dark:border-gray-800 w-72 rounded-full shadow-xl"
-			>
-				<input
-					type="text"
-					id="floating-message-input"
-					class="ml-5 bg-transparent outline-hidden w-full flex-1 text-sm"
-					placeholder={$i18n.t('Ask a question')}
-					aria-label={$i18n.t('Ask a question')}
-					bind:value={floatingInputValue}
-					on:keydown={(e) => {
-						if (e.key === 'Enter') {
-							actionHandler(selectedAction?.id);
+							await tick();
+							setTimeout(() => {
+								const input = document.getElementById('floating-message-input');
+								if (input) {
+									input.focus();
+								}
+							}, 0);
+						} else {
+							actionHandler(action.id);
 						}
 					}}
-				/>
-
-				<div class="ml-1 mr-1">
-					<button
-						aria-label={$i18n.t('Submit question')}
-						class="{floatingInputValue !== ''
-							? 'bg-black text-white hover:bg-gray-900 dark:bg-white dark:text-black dark:hover:bg-gray-100 '
-							: 'text-white bg-gray-200 dark:text-gray-900 dark:bg-gray-700 disabled'} transition rounded-full p-1.5 m-0.5 self-center"
-						on:click={() => {
-							actionHandler(selectedAction?.id);
-						}}
-					>
-						<svg
-							xmlns="http://www.w3.org/2000/svg"
-							viewBox="0 0 16 16"
-							fill="currentColor"
-							class="size-4"
-						>
-							<path
-								fill-rule="evenodd"
-								d="M8 14a.75.75 0 0 1-.75-.75V4.56L4.03 7.78a.75.75 0 0 1-1.06-1.06l4.5-4.5a.75.75 0 0 1 1.06 0l4.5 4.5a.75.75 0 0 1-1.06 1.06L8.75 4.56v8.69A.75.75 0 0 1 8 14Z"
-								clip-rule="evenodd"
-							/>
-						</svg>
-					</button>
-				</div>
-			</div>
-		{/if}
+				>
+					{#if action.icon}
+						<svelte:component this={action.icon} className="size-3 shrink-0" />
+					{/if}
+					<div class="shrink-0">{action.label}</div>
+				</button>
+			{/each}
+		</div>
 	{:else}
 		<div
-			class="bg-white dark:bg-gray-850 dark:text-gray-100 rounded-3xl shadow-xl w-80 max-w-full border border-gray-100 dark:border-gray-800"
+			class="py-1 flex dark:text-gray-100 bg-white dark:bg-gray-850 border border-gray-100 dark:border-gray-800 w-72 rounded-full shadow-xl"
 		>
-			<div
-				class="bg-white dark:bg-gray-850 dark:text-gray-100 text-medium rounded-3xl px-3.5 pt-3 w-full"
-			>
-				<div class="font-medium">
-					<Markdown id={`${id}-float-prompt`} {content} />
-				</div>
-			</div>
+			<input
+				type="text"
+				id="floating-message-input"
+				class="ml-5 bg-transparent outline-hidden w-full flex-1 text-sm"
+				placeholder={$i18n.t('Ask a question')}
+				aria-label={$i18n.t('Ask a question')}
+				bind:value={floatingInputValue}
+				on:keydown={(e) => {
+					if (e.key === 'Enter') {
+						actionHandler(selectedAction?.id);
+					}
+				}}
+			/>

-			<div class="bg-white dark:bg-gray-850 dark:text-gray-100 text-medium rounded-4xl w-full">
-				<div
-					class=" max-h-80 overflow-y-auto w-full markdown-prose-xs px-3.5 py-3"
-					id="response-container"
+			<div class="ml-1 mr-1">
+				<button
+					aria-label={$i18n.t('Submit question')}
+					class="{floatingInputValue !== ''
+						? 'bg-black text-white hover:bg-gray-900 dark:bg-white dark:text-black dark:hover:bg-gray-100 '
+						: 'text-white bg-gray-200 dark:text-gray-900 dark:bg-gray-700 disabled'} transition rounded-full p-1.5 m-0.5 self-center"
+					on:click={() => {
+						actionHandler(selectedAction?.id);
+					}}
 				>
-					{#if !responseContent || responseContent?.trim() === ''}
-						<Skeleton size="sm" />
-					{:else}
-						<Markdown id={`${id}-float-response`} content={responseContent} />
-					{/if}
-
-					{#if responseDone}
-						<div class="flex justify-end pt-3 text-sm font-medium">
-							<button
-								class="px-3.5 py-1.5 text-sm font-medium bg-black hover:bg-gray-900 text-white dark:bg-white dark:text-black dark:hover:bg-gray-100 transition rounded-full"
-								on:click={addHandler}
-							>
-								{$i18n.t('Add')}
-							</button>
-						</div>
-					{/if}
-				</div>
+					<svg
+						xmlns="http://www.w3.org/2000/svg"
+						viewBox="0 0 16 16"
+						fill="currentColor"
+						class="size-4"
+					>
+						<path
+							fill-rule="evenodd"
+							d="M8 14a.75.75 0 0 1-.75-.75V4.56L4.03 7.78a.75.75 0 0 1-1.06-1.06l4.5-4.5a.75.75 0 0 1 1.06 0l4.5 4.5a.75.75 0 0 1-1.06 1.06L8.75 4.56v8.69A.75.75 0 0 1 8 14Z"
+							clip-rule="evenodd"
+						/>
+					</svg>
+				</button>
 			</div>
 		</div>
 	{/if}
@@ -92,6 +92,35 @@
 	let loading = false;
 	let error: string | null = null;

+	// ── Sort state ──────────────────────────────────────────────────────
+	type SortMode = 'name' | 'date';
+	let sortBy: SortMode = 'name';
+	let sortAsc = true;
+
+	const sortEntries = (items: FileEntry[]): FileEntry[] => {
+		return [...items].sort((a, b) => {
+			// Directories always first
+			if (a.type !== b.type) return a.type === 'directory' ? -1 : 1;
+			if (sortBy === 'date') {
+				const aTime = a.modified ?? 0;
+				const bTime = b.modified ?? 0;
+				return sortAsc ? aTime - bTime : bTime - aTime;
+			}
+			const cmp = a.name.localeCompare(b.name);
+			return sortAsc ? cmp : -cmp;
+		});
+	};
+
+	const toggleSort = (mode: SortMode) => {
+		if (sortBy === mode) {
+			sortAsc = !sortAsc;
+		} else {
+			sortBy = mode;
+			sortAsc = mode === 'name'; // name defaults asc, date defaults asc (oldest first)
+		}
+		entries = sortEntries(entries);
+	};
+
 	// ── Navigation history ──────────────────────────────────────────────
 	type NavEntry = { path: string; file: string | null };
 	let navHistory: NavEntry[] = [];
@@ -341,10 +370,7 @@
 				'Failed to load directory. Check your Terminal connection in Settings → Integrations.';
 			entries = [];
 		} else {
-			entries = result.sort((a, b) => {
-				if (a.type !== b.type) return a.type === 'directory' ? -1 : 1;
-				return a.name.localeCompare(b.name);
-			});
+			entries = sortEntries(result);
 		}
 	};

@@ -429,7 +455,8 @@
 							selectedExcelSheet = excelSheetNames[0];
 							const { excelToTable } = await import('$lib/utils/excelToTable');
 							const result = await excelToTable(wb.Sheets[selectedExcelSheet]);
-							fileOfficeHtml = result.html;
+							const DOMPurify = (await import('dompurify')).default;
+							fileOfficeHtml = DOMPurify.sanitize(result.html);
 						}
 					} else if (ext === 'pptx') {
 						const { pptxToImages } = await import('$lib/utils/pptxToHtml');
@@ -918,6 +945,8 @@
 				{loading}
 				{canGoBack}
 				{canGoForward}
+				{sortBy}
+				{sortAsc}
 				onGoBack={goBack}
 				onGoForward={goForward}
 				onNavigate={loadDir}
@@ -934,6 +963,7 @@
 				onUploadFiles={handleUploadFiles}
 				onDownloadDir={() => downloadFile(currentPath)}
 				onMove={handleMove}
+				onSort={toggleSort}
 			>
 				{#if fileImageUrl !== null || (fileOfficeSlides !== null && fileOfficeSlides.length > 0)}
 					<Tooltip content={$i18n.t('Reset view')}>
@@ -1253,7 +1283,8 @@
 						selectedExcelSheet = sheet;
 						const { excelToTable } = await import('$lib/utils/excelToTable');
 						const result = await excelToTable(excelWorkbook.Sheets[sheet]);
-						fileOfficeHtml = result.html;
+						const DOMPurify = (await import('dompurify')).default;
+						fileOfficeHtml = DOMPurify.sanitize(result.html);
 					}}
 					baseUrl={selectedTerminal?.url ?? ''}
 					apiKey={selectedTerminal?.key ?? ''}
@@ -1355,6 +1386,7 @@
 									onRename={handleRename}
 									onSelect={handleSelect}
 									onLongPress={enterSelectionMode}
+									showDate={sortBy === 'date'}
 								/>
 							{/each}
 						</ul>
@@ -30,6 +30,17 @@
 	export let selectedPaths: Set<string> = new Set();
 	export let onSelect: (entry: FileEntry, event: MouseEvent) => void = () => {};
 	export let onLongPress: () => void = () => {};
+	export let showDate: boolean = false;
+
+	const formatRelativeTime = (epoch: number): string => {
+		const diff = Math.floor(Date.now() / 1000) - epoch;
+		if (diff < 60) return 'just now';
+		if (diff < 3600) return `${Math.floor(diff / 60)}m ago`;
+		if (diff < 86400) return `${Math.floor(diff / 3600)}h ago`;
+		if (diff < 2592000) return `${Math.floor(diff / 86400)}d ago`;
+		if (diff < 31536000) return `${Math.floor(diff / 2592000)}mo ago`;
+		return `${Math.floor(diff / 31536000)}y ago`;
+	};

 	let dragOverFolder = false;

@@ -271,7 +282,14 @@
 				</span>
 			{/if}
 			{#if entry.type === 'file' && entry.size !== undefined && !renaming}
+				{#if showDate && entry.modified}
+					<span class="text-[10px] text-gray-400 shrink-0"
+						>{formatRelativeTime(entry.modified)}</span
+					>
+				{/if}
 				<span class="text-xs text-gray-400 shrink-0">{formatFileSize(entry.size)}</span>
+			{:else if entry.type === 'directory' && showDate && entry.modified && !renaming}
+				<span class="text-[10px] text-gray-400 shrink-0">{formatRelativeTime(entry.modified)}</span>
 			{/if}
 		</button>

@@ -6,6 +6,7 @@
 	import FilePlusAlt from '../../icons/FilePlusAlt.svelte';
 	import Spinner from '../../common/Spinner.svelte';
 	import Tooltip from '../../common/Tooltip.svelte';
+	import Dropdown from '$lib/components/common/Dropdown.svelte';

 	const i18n = getContext('i18n');

@@ -21,6 +22,11 @@
 	export let onDownloadDir: () => void = () => {};
 	export let onMove: (source: string, destFolder: string) => void = () => {};

+	// Sort controls
+	export let sortBy: 'name' | 'date' = 'name';
+	export let sortAsc: boolean = true;
+	export let onSort: (mode: 'name' | 'date') => void = () => {};
+
 	// Back / forward navigation
 	export let canGoBack = false;
 	export let canGoForward = false;
@@ -161,6 +167,78 @@
 	</Tooltip>

 	{#if !selectedFile}
+		<Dropdown align="end" sideOffset={4}>
+			<Tooltip content={$i18n.t('Sort')}>
+				<button
+					class="shrink-0 p-1 rounded hover:bg-gray-100 dark:hover:bg-gray-800 transition text-gray-400 dark:text-gray-500 hover:text-gray-600 dark:hover:text-gray-400"
+					aria-label={$i18n.t('Sort')}
+				>
+					<svg
+						xmlns="http://www.w3.org/2000/svg"
+						viewBox="0 0 20 20"
+						fill="currentColor"
+						class="size-3.5"
+					>
+						<path
+							d="M2 3.75A.75.75 0 0 1 2.75 3h11.5a.75.75 0 0 1 0 1.5H2.75A.75.75 0 0 1 2 3.75ZM2 7.5a.75.75 0 0 1 .75-.75h7.508a.75.75 0 0 1 0 1.5H2.75A.75.75 0 0 1 2 7.5ZM14 7a.75.75 0 0 1 .75.75v6.69l1.72-1.72a.75.75 0 1 1 1.06 1.06l-3 3a.75.75 0 0 1-1.06 0l-3-3a.75.75 0 1 1 1.06-1.06l1.72 1.72V7.75A.75.75 0 0 1 14 7ZM2 11.25a.75.75 0 0 1 .75-.75h4.562a.75.75 0 0 1 0 1.5H2.75a.75.75 0 0 1-.75-.75Z"
+						/>
+					</svg>
+				</button>
+			</Tooltip>
+
+			<div slot="content">
+				<div
+					class="min-w-[150px] rounded-2xl p-1 z-[9999999] bg-white dark:bg-gray-850 dark:text-white shadow-lg border border-gray-100 dark:border-gray-800"
+				>
+					<button
+						type="button"
+						class="select-none flex rounded-xl py-1.5 px-3 w-full hover:bg-gray-50 dark:hover:bg-gray-800 transition items-center gap-2 text-sm"
+						on:click={() => onSort('name')}
+					>
+						<span class="flex-1 text-left">{$i18n.t('Name')}</span>
+						{#if sortBy === 'name'}
+							<svg
+								xmlns="http://www.w3.org/2000/svg"
+								viewBox="0 0 16 16"
+								fill="currentColor"
+								class="size-3 text-gray-500 dark:text-gray-400 transition-transform {sortAsc
+									? ''
+									: 'rotate-180'}"
+							>
+								<path
+									fill-rule="evenodd"
+									d="M11.78 9.78a.75.75 0 0 1-1.06 0L8 7.06 5.28 9.78a.75.75 0 0 1-1.06-1.06l3.25-3.25a.75.75 0 0 1 1.06 0l3.25 3.25a.75.75 0 0 1 0 1.06Z"
+									clip-rule="evenodd"
+								/>
+							</svg>
+						{/if}
+					</button>
+					<button
+						type="button"
+						class="select-none flex rounded-xl py-1.5 px-3 w-full hover:bg-gray-50 dark:hover:bg-gray-800 transition items-center gap-2 text-sm"
+						on:click={() => onSort('date')}
+					>
+						<span class="flex-1 text-left">{$i18n.t('Date Modified')}</span>
+						{#if sortBy === 'date'}
+							<svg
+								xmlns="http://www.w3.org/2000/svg"
+								viewBox="0 0 16 16"
+								fill="currentColor"
+								class="size-3 text-gray-500 dark:text-gray-400 transition-transform {sortAsc
+									? ''
+									: 'rotate-180'}"
+							>
+								<path
+									fill-rule="evenodd"
+									d="M11.78 9.78a.75.75 0 0 1-1.06 0L8 7.06 5.28 9.78a.75.75 0 0 1-1.06-1.06l3.25-3.25a.75.75 0 0 1 1.06 0l3.25 3.25a.75.75 0 0 1 0 1.06Z"
+									clip-rule="evenodd"
+								/>
+							</svg>
+						{/if}
+					</button>
+				</div>
+			</div>
+		</Dropdown>
 		<Tooltip content={$i18n.t('New Folder')}>
 			<button
 				class="shrink-0 p-1 rounded hover:bg-gray-100 dark:hover:bg-gray-800 transition text-gray-400 dark:text-gray-500 hover:text-gray-600 dark:hover:text-gray-400"
@@ -32,6 +32,7 @@
 	let confirmed = false;
 	let interrupted = false;
 	let assistantSpeaking = false;
+	let muted = false;

 	let emoji = null;
 	let camera = false;
@@ -63,7 +64,12 @@

 		console.log(videoInputDevices);
 		if (selectedVideoInputDeviceId === null && videoInputDevices.length > 0) {
-			selectedVideoInputDeviceId = videoInputDevices[0].deviceId;
+			const savedDeviceId = localStorage.getItem('selectedVideoInputDeviceId');
+			if (savedDeviceId && videoInputDevices.some((d) => d.deviceId === savedDeviceId)) {
+				selectedVideoInputDeviceId = savedDeviceId;
+			} else {
+				selectedVideoInputDeviceId = videoInputDevices[0].deviceId;
+			}
 		}
 	};

@@ -150,6 +156,10 @@

 	const transcribeHandler = async (audioBlob) => {
 		// Create a blob from the audio chunks
+		if (!audioBlob || audioBlob.size < 100) {
+			console.log('Audio blob too small or empty, skipping transcription');
+			return;
+		}

 		await tick();
 		const file = blobToFile(audioBlob, 'recording.wav');
@@ -231,6 +241,11 @@
 					}
 				});
 			}
+
+			if (audioStream) {
+				// hardware track muting disabled to prevent backend translation errors with malformed WebM files
+			}
+
 			mediaRecorder = new MediaRecorder(audioStream);

 			mediaRecorder.onstart = () => {
@@ -305,8 +320,8 @@
 					return;
 				}

-				if (assistantSpeaking && !($settings?.voiceInterruption ?? false)) {
-					// Mute the audio if the assistant is speaking
+				if (muted || (assistantSpeaking && !($settings?.voiceInterruption ?? false))) {
+					// Suppress mic input when muted or when assistant is speaking without interruption enabled
 					analyser.maxDecibels = 0;
 					analyser.minDecibels = -1;
 				} else {
@@ -320,6 +335,10 @@
 				// Calculate RMS level from time domain data
 				rmsLevel = calculateRMS(timeDomainData);

+				if (muted || (assistantSpeaking && !($settings?.voiceInterruption ?? false))) {
+					rmsLevel = 0;
+				}
+
 				// Check if initial speech/noise has started
 				const hasSound = domainData.some((value) => value > 0);
 				if (hasSound) {
@@ -622,6 +641,47 @@
 		chatStreaming = false;
 	};

+	const toggleMute = () => {
+		muted = !muted;
+		if (muted && hasStartedSpeaking) {
+			// Abort the ongoing recording so it doesn't accidentally send a partial sentence
+			hasStartedSpeaking = false;
+			confirmed = false;
+			audioChunks = [];
+			if (mediaRecorder && mediaRecorder.state === 'recording') {
+				mediaRecorder.stop();
+			}
+		}
+	};
+
+	let wasAssistantSpeaking = false;
+	$: {
+		if (assistantSpeaking && !wasAssistantSpeaking) {
+			wasAssistantSpeaking = true;
+		} else if (!assistantSpeaking && wasAssistantSpeaking) {
+			wasAssistantSpeaking = false;
+			// Auto unmute when AI finishes speaking
+			if (muted) {
+				muted = false;
+			}
+		}
+	}
+
+	const handleKeydown = (e: KeyboardEvent) => {
+		// Only handle M key when not typing in an input/textarea
+		if (e.key === 'm' || e.key === 'M') {
+			const target = e.target as HTMLElement;
+			if (
+				target.tagName !== 'INPUT' &&
+				target.tagName !== 'TEXTAREA' &&
+				!target.isContentEditable
+			) {
+				e.preventDefault();
+				toggleMute();
+			}
+		}
+	};
+
 	onMount(async () => {
 		const setWakeLock = async () => {
 			try {
@@ -659,6 +719,8 @@
 		eventTarget.addEventListener('chat', chatEventHandler);
 		eventTarget.addEventListener('chat:finish', chatFinishHandler);

+		document.addEventListener('keydown', handleKeydown);
+
 		return async () => {
 			await stopAllAudio();

@@ -668,6 +730,8 @@
 			eventTarget.removeEventListener('chat', chatEventHandler);
 			eventTarget.removeEventListener('chat:finish', chatFinishHandler);

+			document.removeEventListener('keydown', handleKeydown);
+
 			audioAbortController.abort();
 			await tick();

@@ -687,6 +751,9 @@
 		eventTarget.removeEventListener('chat:start', chatStartHandler);
 		eventTarget.removeEventListener('chat', chatEventHandler);
 		eventTarget.removeEventListener('chat:finish', chatFinishHandler);
+
+		document.removeEventListener('keydown', handleKeydown);
+
 		audioAbortController.abort();

 		await tick();
@@ -882,19 +949,42 @@
 			{/if}
 		</div>

-		<div class="flex justify-between items-center pb-2 w-full">
-			<div>
+		<div class="flex flex-col items-center gap-4 pb-4 w-full">
+			<button
+				type="button"
+				class="z-10"
+				on:click={() => {
+					if (assistantSpeaking) {
+						stopAllAudio();
+					}
+				}}
+			>
+				<div class="line-clamp-1 text-sm font-medium">
+					{#if loading}
+						{$i18n.t('Thinking...')}
+					{:else if muted}
+						{$i18n.t('Muted')}
+					{:else if assistantSpeaking}
+						{$i18n.t('Tap to interrupt')}
+					{:else}
+						{$i18n.t('Listening...')}
+					{/if}
+				</div>
+			</button>
+
+			<div class="flex items-center justify-center gap-4 z-10">
 				{#if camera}
 					<VideoInputMenu
 						devices={videoInputDevices}
 						on:change={async (e) => {
 							console.log(e.detail);
 							selectedVideoInputDeviceId = e.detail;
+							localStorage.setItem('selectedVideoInputDeviceId', e.detail);
 							await stopVideoStream();
 							await startVideoStream();
 						}}
 					>
-						<button class=" p-3 rounded-full bg-gray-50 dark:bg-gray-900" type="button">
+						<button class="p-3 rounded-full bg-gray-50 dark:bg-gray-900" type="button">
 							<svg
 								xmlns="http://www.w3.org/2000/svg"
 								viewBox="0 0 20 20"
@@ -912,7 +1002,7 @@
 				{:else}
 					<Tooltip content={$i18n.t('Camera')}>
 						<button
-							class=" p-3 rounded-full bg-gray-50 dark:bg-gray-900"
+							class="p-3 rounded-full bg-gray-50 dark:bg-gray-900"
 							type="button"
 							on:click={async () => {
 								await navigator.mediaDevices.getUserMedia({ video: true });
@@ -941,32 +1031,63 @@
 						</button>
 					</Tooltip>
 				{/if}
-			</div>

-			<div>
-				<button
-					type="button"
-					on:click={() => {
-						if (assistantSpeaking) {
-							stopAllAudio();
-						}
-					}}
-				>
-					<div class=" line-clamp-1 text-sm font-medium">
-						{#if loading}
-							{$i18n.t('Thinking...')}
-						{:else if assistantSpeaking}
-							{$i18n.t('Tap to interrupt')}
+				<Tooltip content={muted ? $i18n.t('Unmute') + ' (M)' : $i18n.t('Mute') + ' (M)'}>
+					<button
+						class="p-3 rounded-full transition-colors duration-200 {muted
+							? 'bg-red-500 text-white'
+							: 'bg-gray-50 dark:bg-gray-900'}"
+						type="button"
+						aria-label={muted ? $i18n.t('Unmute') : $i18n.t('Mute')}
+						on:click={toggleMute}
+					>
+						{#if muted}
+							<!-- Mic Off icon -->
+							<svg
+								xmlns="http://www.w3.org/2000/svg"
+								fill="none"
+								viewBox="0 0 24 24"
+								stroke-width="1.5"
+								stroke="currentColor"
+								class="size-5"
+							>
+								<path
+									stroke-linecap="round"
+									stroke-linejoin="round"
+									d="M12 18.75a6 6 0 0 0 6-6v-1.5m-6 7.5a6 6 0 0 1-6-6v-1.5m6 7.5v3.75m-3.75 0h7.5M12 15.75a3 3 0 0 1-3-3V4.5a3 3 0 1 1 6 0v8.25a3 3 0 0 1-3 3Z"
+								/>
+								<line
+									x1="3"
+									y1="3"
+									x2="21"
+									y2="21"
+									stroke="currentColor"
+									stroke-width="1.5"
+									stroke-linecap="round"
+								/>
+							</svg>
 						{:else}
-							{$i18n.t('Listening...')}
+							<!-- Mic On icon -->
+							<svg
+								xmlns="http://www.w3.org/2000/svg"
+								fill="none"
+								viewBox="0 0 24 24"
+								stroke-width="1.5"
+								stroke="currentColor"
+								class="size-5"
+							>
+								<path
+									stroke-linecap="round"
+									stroke-linejoin="round"
+									d="M12 18.75a6 6 0 0 0 6-6v-1.5m-6 7.5a6 6 0 0 1-6-6v-1.5m6 7.5v3.75m-3.75 0h7.5M12 15.75a3 3 0 0 1-3-3V4.5a3 3 0 1 1 6 0v8.25a3 3 0 0 1-3 3Z"
+								/>
+							</svg>
 						{/if}
-					</div>
-				</button>
-			</div>
+					</button>
+				</Tooltip>

-			<div>
 				<button
-					class=" p-3 rounded-full bg-gray-50 dark:bg-gray-900"
+					class="p-3 rounded-full bg-gray-50 dark:bg-gray-900"
 					on:click={async () => {
 						await stopAudioStream();
 						await stopVideoStream();
@@ -13,8 +13,11 @@
 	} from '$lib/stores';

 	import { getOAuthClientAuthorizationUrl } from '$lib/apis/configs';
+	import { deleteOAuthSession } from '$lib/apis/auths';
 	import { getTools } from '$lib/apis/tools';

+	import { toast } from 'svelte-sonner';
+
 	import Knobs from '$lib/components/icons/Knobs.svelte';
 	import Dropdown from '$lib/components/common/Dropdown.svelte';
 	import Tooltip from '$lib/components/common/Tooltip.svelte';
@@ -27,6 +30,7 @@
 	import Terminal from '$lib/components/icons/Terminal.svelte';
 	import ChevronRight from '$lib/components/icons/ChevronRight.svelte';
 	import ChevronLeft from '$lib/components/icons/ChevronLeft.svelte';
+	import LinkSlash from '$lib/components/icons/LinkSlash.svelte';

 	const i18n = getContext('i18n');

@@ -375,6 +379,39 @@
 								</div>
 							</div>

+							{#if (tools[toolId]?.authenticated ?? true) && toolId.startsWith('server:mcp:')}
+								<div class="shrink-0">
+									<Tooltip content={$i18n.t('Disconnect OAuth')}>
+										<button
+											class="self-center w-fit text-sm text-gray-600 dark:text-gray-400 hover:text-gray-700 dark:hover:text-gray-300 transition rounded-full"
+											type="button"
+											on:click={async (e) => {
+												e.stopPropagation();
+												e.preventDefault();
+
+												const parts = toolId.split(':');
+												const serverId = parts.at(-1) ?? toolId;
+												const provider = `mcp:${serverId}`;
+
+												try {
+													await deleteOAuthSession(localStorage.token, provider);
+													toast.success($i18n.t('OAuth session disconnected'));
+
+													// Refresh tools to update authenticated state
+													_tools.set(await getTools(localStorage.token));
+													selectedToolIds = selectedToolIds.filter((id) => id !== toolId);
+													await init();
+												} catch (err) {
+													toast.error(err ?? $i18n.t('Failed to disconnect'));
+												}
+											}}
+										>
+											<LinkSlash className="size-3.5" />
+										</button>
+									</Tooltip>
+								</div>
+							{/if}
+
 							{#if tools[toolId]?.has_user_valves && ($user?.role === 'admin' || ($user?.permissions?.chat?.valves ?? true))}
 								<div class=" shrink-0">
 									<Tooltip content={$i18n.t('Valves')}>
@@ -238,7 +238,13 @@
 			return;
 		}

-		const mineTypes = ['audio/webm; codecs=opus', 'audio/mp4'];
+		const mineTypes = [
+			'audio/webm; codecs=opus',
+			'audio/webm',
+			'audio/ogg; codecs=opus',
+			'audio/mp4',
+			'audio/wav'
+		];

 		mediaRecorder = new MediaRecorder(stream, {
 			mimeType: mineTypes.find((type) => MediaRecorder.isTypeSupported(type))
--- a/Show More
+++ b/Show More