Esta es una versión anterior. La versión actual es 2.3.4. LobeHub 2.3.4 · Fecha de publicación:
Archivos de instalación 2.3.0
| Arquitectura | Tipo | Ámbito | Tamaño del archivo | Descargar | SHA256 |
|---|---|---|---|---|---|
| 64 bits | EXE (NSIS) | Usuario actual | — | github.com | VirusTotal |
| 64 bits | EXE (NSIS) | Todos los usuarios | — | github.com | VirusTotal |
El enlace y el SHA256 proceden del manifiesto comunitario del Administrador de paquetes de Windows. Conviene comparar el hash del archivo descargado antes de ejecutarlo.
Notas de la versión
From PR #22056: https://github.com/lobehub/lobehub/pull/22056
🚀 LobeHub Release (v2.3.0)
Release Date: November 1, 2026
Since v2.2.21: 145 merged PRs · 7 contributors
Bots stop waiting to be @-mentioned. With Proactive Bot on by default, every bot reads its channels, steps in when it can help, and shows its owner every call it made. Channels, pages, memory, dashboards, tasks and goals round out the release.
🤖 Proactive Bot
Until now a bot answered only when someone @-mentioned it. Bot wake was previewed behind a rollout flag in v2.2.21. In v2.3.0 it is on for every bot: the bot judges each message in the channels it watches and decides for itself whether to speak up, stay quiet or wait.
- On by default, with an off switch — The rollout flag is gone, and every bot speaks up on its own unless its owner turns it off. The switch is the new Proactive participation block at the top of the channel settings' Wake policy tab and saves as soon as it changes. With it off, the Duties, Channels and Pace options are hidden, and only watch keywords still wake the bot. (#22024)
- Steps in when it can help — For a bot whose owner wrote duties, the bot no longer waits for a message that falls inside them. It also steps in when someone leaves an open question or problem that nobody has answered and the bot can settle it. A line right after a person goes to the model instead of being skipped, and the bot still stays out of replies addressed to someone else. This adds no extra model call. (#22038)
- Wake records — A new Wake records tab in bot settings lists every message the bot judged. It shows a 30-day summary (woke, skipped, deferred, proactive wakes) and, for each message, the outcome, the model score and a readable reason, with a link back to the message on Discord. Filter by outcome, proactive wakes only or unmarked only, and mark each decision ✓ right or ✗ wrong. (#21934)
- Clearer failures in IM — Replies explain fail-fast sandbox dispatch errors and Feishu error 41050 instead of a bare failure, the generic reply-failure message is clearer, and internal metadata stays out of public replies. (#21973, #21525, #21955, #21899)
✨ Highlights
- Real-time page collaboration — Several people can edit a page at once, page Agent edits run in a sandbox, and a page shows its stored content while the room syncs instead of an empty editor. (#20322, #21924, #21919)
- Channel conversations — Quote a message in a reply, follow Agent replies in their own threads nested under a Discord-style connector line, and write in a compact IM-style composer. Mentions list members and Agents separately. (#22025, #22042, #21994, #22026, #21885)
- Channels that open themselves — Every new Agent gets a direct channel and every new project a General channel. In a group with one Agent, every message goes straight to it. (#22018, #21996, #21927)
- Memory as stories — Legacy contexts are consolidated into contexts and dossiers and shown as stories on the contexts tab, and a stopped consolidation run resumes instead of starting over. (#21968, #22006, #22007)
- Dashboards — Widgets use a declarative card layout with presets and user parameters, dashboards get descriptions and a createDashboard tool, and widget approval shows its stages. (#22016, #22049, #21954, #21936)
- Tasks — Pick a task's model from the task tools or the CLI, keep private Agents on public tasks (make the task private or take it over), and use one "Follow the agent" switch for device and working directory. (#21852, #21937, #21932, #21848)
- Goals — Goals can be shared at workspace level, and you can rename them yourself or with smart rename. New goals no longer use the whole instruction as their title. (#22027, #21970)
- Persistent sandbox — The persistent sandbox is out of preview, and paid users default to their drive. (#22032)
- Docs, files and Favorites — Add topics to Favorites, link straight to library folders, pages and files by id, open Docs file links beside the chat, and share files by link. (#21950, #21923, #21913)
- Credentials — Workspace and private credential tabs, with sharing across multiple workspaces. (#21670)
🏗️ Core Agent & Architecture
- One member's run never interrupts another member's live run, and a queued follow-up is no longer resent once its message is written. (#22037, #21777)
- Agent management searches the MCP marketplace before installing a market plugin, and Agents can be created from the identity form. (#21703, #21987)
- A generic TTS model type with MiniMax t2a_v2 speech and a TTS spend label. (#21942)
- Image generation rejects unsupported parameters with a clear error instead of a ServerError. (#21945)
- Chinese vendors in the model bank are priced in international USD. (#21808)
- Verification writes generated criteria and reports in the account's language. (#22035)
- Model discovery and runtime configuration reach the Agent correctly. (#21895)
External Agents
- Claude Code project memory shows in the workspace panel, and Agent quota picks the Claude Code account per device. (#21234, #21984)
- Every background-task wake-up stays in the run's reply, and subagent threads stay linear across replicas and resumable after finalize. (#21939, #21917)
- The CLI status card shows the device the Agent actually runs on. (#21931, #21999)
📱 Platforms / Integrations
- Desktop: New Window in the File menu opens the current workspace. On macOS 27 the sidebar uses native AppKit glass, and the window background follows system theme switches. (#21928, #21854, #21921, #21894, #21920)
- White-label: distributions can set their own desktop app identity, and the brand is no longer hardcoded in LLM system prompts, tool descriptions, auth emails, the passkey prompt or the default assistant name. (#21890, #21930, #21902, #21882, #21946)
- CLI: lh topic send / lh message reply wake another conversation, lh topic send --new / lh message send start a new one, and lh workspace member update sets a member's title and responsibilities. (#21938, #21965, #21959)
- CLI: follows Agent runs over the multiplexed gateway socket, and detects system services with clearer upgrade restarts. (#22048, #21869)
- API / SDK: read the caller's credit quota through the API, SDK and lh. (#21933)
- Browser tool: key combos work, and setViewport is available. (#22011)
🖥️ User Experience
- Chat keeps its scroll position after you send and loads history again in 500-row pages. (#22030)
- Device pickers show running topics on hover, and device cards show disk usage next to CPU and memory. (#21915, #21940)
- Search the skills panel, filter topics to only your own per Agent, and unarchive from the bulk bar. (#21961, #21870, #21876)
- Project "+" buttons open a blank draft, and project Agents come first when creating a channel. (#21989, #22004)
- Manage canvas storage from a Works Canvas tab, and insert videos from the page editor's slash menu. (#21831, #21925)
- Acceptance hover cards open the acceptance in the browser, and a local report can accept or reject the delivery. (#21966, #21947)
- Eight Agent cover textures, with profile covers on badge cards. (#21868)
- The boot loading hint appears only on slow boots. (#21941)
🔒 Security & Reliability
- Security: the OIDC provider pins its request host to APP_URL. (#21985)
- Reliability: pages keep seeding and projecting on serverless builds, file proxy URLs resolve from the page origin, and a late linked-PR lookup no longer crashes the device view. (#21919, #21914, #21960)
- Reliability: quota limits a plan no longer reports are dropped. (#22000)
Database migrations (self-hosted)
This release adds six migrations (0193–0198): goal/file share and structured user memory tables, document history attribution, memory columns with subscription logins and device account bindings, indexes on the memory tables, a (message_id, member_id) unique index for channel jobs, and goal visibility. They run automatically on startup. (#21844, #21611, #21888, #21672, #21935, #21901, #22027)
Note
0196 builds indexes on the user_memories_* tables inside a transaction, which blocks writes to those tables while it runs. If they are large, create the same indexes with CREATE INDEX CONCURRENTLY IF NOT EXISTS before upgrading so the migration does nothing.
👥 Contributors
Huge thanks to 7 contributors who shipped 145 merged PRs this cycle.
@arvinxx · @AmAzing129 · @Innei · @rdmclin2 · @ONLY-yours · @MarioJames · @tjx666
Plus @lobehub-coder and @lobehubbot for maintenance.
Full Changelog: v2.2.21...v2.3.0
🤖 Generated with Claude Code