Aquesta és una versió anterior. La versió actual és la 2.2.20. LobeHub 2.2.20 · Data de publicació:
Fitxers d’instal·lació 2.2.13
| Arquitectura | Tipus | Àmbit | Mida del fitxer | Baixa | SHA256 |
|---|---|---|---|---|---|
| 64 bits | EXE (NSIS) | Usuari actual | — | github.com | VirusTotal |
| 64 bits | EXE (NSIS) | Tots els usuaris | — | github.com | VirusTotal |
L’enllaç i l’SHA256 provenen del manifest comunitari del Gestor de paquets del Windows. Convé comparar el hash del fitxer baixat abans d’executar-lo.
Notes de la versió
This release was automatically published from PR #17807.
Changes
See PR description: https://github.com/lobehub/lobehub/pull/17807
Commit Message
🚀 LobeHub Release (20260801)
Release Date: August 1, 2026
Since v2.2.12: 86 merged PRs · 13 contributors
Home becomes an Agent dashboard, the desktop shell keeps every tab alive, and local Agents connect through one wizard — with a long tail of runtime, model, and Windows fixes behind it.
✨ Highlights
- Home as an Agent dashboard — Home is rebuilt around what your Agents are doing: activity, Topics, and a dashboard rail replace the old landing view. (#17680, #17756, #17798, #17735, #17781)
- One wizard for local Agents — Connecting a local Agent now runs through a single unified flow instead of per-connector setup paths. (#17794)
- Local MCP connectors on desktop — Install and run stdio and LAN-HTTP MCP connectors directly from the desktop app. (#17673)
- Per-tab keep-alive routers on desktop — The Electron shell routes each tab through its own keep-alive router, so switching tabs no longer resets in-progress views. (#17595, #17773)
- Cross-runtime local database — Local data now runs on one database layer shared across runtimes, replacing the runtime-specific paths. (#17682)
- Scoped tool allowlists in the Agent runtime — Tool access is enforced against the calling scope, so a sub-Agent can no longer reach tools outside its grant. (#17717)
- Faster Agent search — Agent filtering moved above the BM25 scan and scans are isolated per query, restoring real TopN ordering and true relevance scores. (#17755, #17720)
- Demand-aware SPA startup — The SPA loads work on demand instead of up front, cutting cold-start cost. (#17577)
🏗️ Core Agent & Runtime
Conversation correctness
- Sub-Agent final answers are preserved when steps are grouped. (#17725)
- Explicit model overrides survive in sub-Agent Topics. (#17766)
- Assistant step count now matches the actual number of LLM calls. (#17790)
- Assistant messages are reused on step retry instead of duplicated. (#17703)
- Duplicate assistant continuations are omitted. (#17726)
- Failed assistant placeholders are filtered, and the Claude prefill guard is hardened. (#17737)
- Thread conversation cancellation is isolated, so cancelling one thread leaves the others running. (#17719)
- Runtime todos are injected into the server Context. (#17671)
Data & workflows
- Local Agent Signal workflows are supported. (#17681)
- Agent documents accept backing IDs when reading. (#17763)
- GitHub Works are registered from heterogeneous shell gh runs. (#17666)
🖥️ Workspace & Home
- Workspace budget settings get their own tab and route slot. (#17775)
- Onboarding generates starter task recommendations, with the recommendation prompts centralized. (#17757, #17770)
- Agent Profile supports a Markdown source mode. (#17580)
- Messenger DM support, plus broader Workspace interaction polish. (#17771)
- Claude Code quota is sampled from bound devices and refreshed every two minutes. (#17797)
- Settings page headers are unified, the About layout is reworked, and Workspace and service-model pages are tidied. (#17667, #17696, #17698)
- Memory empty states, the prompt editor header, and Agent statistics are polished. (#17764)
- Conversation portal headers are refined. (#17751)
- Model Select surfaces stale-model state instead of a raw provider/model id, and the stale-model popup is compact. (#17748, #17754)
- Tab titles show the section and Agent name on Agent sub-pages. (#17711)
- The Home sidebar stays put on routes without a dedicated panel. (#17718)
- Sidebar task list cache persists, so the skeleton no longer replays on every visit. (#17706)
- The Inbox route stays valid during initialization. (#17759)
- Resolved briefs no longer reappear after navigation. (#17760)
- The prompt view-mode toggle no longer scrolls to the bottom. (#17746)
- Enter / Esc hotkeys are added to the AskUserQuestion intervention. (#17733)
- A copy-operation-id action is available in assistant message menus. (#17710)
- Copy and i18n polish across the OAuth provider card, model param controls, storage cleanup guidance, artifact deployment, and the zh-CN model rating labels. (#17736, #17749, #17740, #17722, #17712, #17714)
📱 Desktop & Platforms
- Native macOS context menus back the app-domain menus. (#17648)
- Oversized glob and spawn errors in the local system layer are guarded. (#17792)
- High-frequency Windows shell-execution issues are resolved. (#17576)
- Composio-backed connector data is supported. (#17677)
🧠 Models & Providers
- Kimi K3 reasoning-effort controls, plus a K3 Fast model description. (#17615, #17707)
- Moonshot supports structured output together with thinking. (#17729)
- OpenAI Responses reasoning items are persisted in full, and opaque reasoning state is scoped correctly. (#17705, #17694)
- Empty and image-only model completions are classified and captured instead of failing silently. (#17700, #17695)
- Builtin model lists can be hidden. (#17683)
- Params model config moves to a dedicated useParamsModelConfig hook. (#17616)
- RustFS init uses the S3 policy Id field. (#15603)
🔒 Security & Reliability
- Security: Tool allowlists are enforced per scope in the Agent runtime. (#17717)
- Security: globFiles / searchFiles scopes resolve against the user Workspace on Windows. (#17676)
- Security: Stale OIDC sessions are cleared on account switch. (#17734)
- Reliability: Cron schedules start by default. (#17709)
- Reliability: Library visibility syncs with its content, and phantom Agent-config saves are stopped. (#17715)
- Reliability: The Vite SPA watcher scope is narrowed. (#17693)
🔧 Tooling
- Authorized runtime DevDock. (#17652)
- Acceptance review: evidence-aware verification, aligned task-checklist grouping, restored single-page scrolling, and refined review interactions. (#17658, #17654, #17697, #17768)
- Todo tool rendering style is aligned. (#17787)
- Next.js updated to preview.10. (#17750)
- Scripts are excluded from the root type-check. (#17699)
- Legacy PostHog configuration is removed. (#17738)
👥 Contributors
Huge thanks to 13 contributors who shipped 86 merged PRs this cycle.
@sxjeru · @Coooolfan · @hardy-one · @northword · @qybaihe · @tjx666 · @arvinxx · @Innei · @rivertwilight · @Rdmclin2 · @nekomeowww · @cy948 · @AmAzing129
Full Changelog: v2.2.12...release/weekly-20260801