Lataa LobeHub 2.2.7

This release was automatically published from PR #16080.
Changes
See PR description: https://github.com/lobehub/lobehub/pull/16080
Commit Message
🚀 LobeHub Release (20260620)
Release Date: June 20, 2026
Since v2.2.6: 88 merged PRs · 7 contributors · 1 database migration
This week pairs a workspace-scoped database migration with broad runtime and UX work — native ASR transcription, audio attachments for audio-capable models, a smarter input-completion prompt, and a long list of reliability fixes.

🗄️ Database Migration (Action for Self-Hosted)
This release ships migration 0111_workspace_device_and_ai_infra_surrogate_pk (#16065). It is the headline change for operators — read this before upgrading.
What changes
- ai_providers / ai_models — adopt a surrogate _id primary key and workspace-scoped partial unique indexes, so provider/model identity is unique per workspace rather than per user.
- devices — the old (user_id, device_id) unique is replaced by two workspace-scoped partial uniques, letting personal and workspace-enrolled devices live in independent identity spaces.
- workspaces — adds the freeze trio frozen (boolean, default false), frozen_reason (text), frozen_at (timestamptz) backing workspace freeze risk control.
Operator impact
- Backwards-compatible and additive — every statement is guarded (IF EXISTS / IF NOT EXISTS / catalog checks). It is a NO-OP on cloud production (where the ai_infra side was already applied online) and a full rebuild on fresh / self-hosted databases.
- No manual SQL required — the migration runs automatically on application startup.
- Schedule the rollout in a low-traffic window and take a backup snapshot before deploying.
- If the migration fails, do not retry repeatedly — inspect migration logs and lock state first.
Rollback
Migration is additive only (new columns / indexes / surrogate PK). If your environment requires strict reversibility, follow your standard DB restore or migration-rollback policy from the pre-deploy snapshot.

✨ Highlights
- Native ASR transcription — new ASR (transcribe) runtime with OpenAI + Gemini-native backends, exposed via tRPC and the CLI (incl. remote audio URLs). (#15992, #15987)
- Audio in chat & agent mode — pass audio (mp3) attachments to audio-capable models like Gemini, with upload allowed in agent mode and a waveform player on render. (#15988, #16022)
- Input completion v1.2 — reworked completion prompt that predicts intent, including long-range continuations. (#15961)
- Market recommendations — task-template recommendations from Market, plus skill recommendations during agent creation. (#15558, #16016)
- Model intelligence — inject model knowledge cutoff into context and support GLM-5.2 reasoning effort. (#16048, #15972)
- Topic lifecycle — persist topic unread as a backend status and show a [Draft] hint on topics holding unsent input. (#16056, #15996)
- Workspace-scoped devices — distinguish local vs. gateway device targets, add a recent-directory button, and auto-activate a device for bot triggers on a local target. (#15921, #16040, #16032)

🏗️ Core Agent & Model Runtime
- Rework input completion prompt to v1.2 — predict intent, including long-range. (#15961)
- Inject model knowledge cutoff into context. (#16048)
- Support GLM-5.2 reasoning effort. (#15972)
- Derive OpenAI model routing from IDs. (#16046)
- Rename model type stt to standard asr (runtime migration, non-breaking). (#16002)
- Persist model config cache for faster startup. (#16047)
- Hetero-agent: assistant-anchored message-chain write spine. (#15930)
- Stop chat mode from auto-injecting local-system tools. (#15981)
- Centralize inbox agent meta fallback across read paths. (#15948)
- Expose missing usage diagnostics. (#15973)

🎙️ Audio & ASR
- Add ASR (transcribe) with OpenAI + Gemini-native backends, exposed via tRPC. (#15992)
- Pass audio (mp3) attachments to audio-capable models like Gemini. (#15988)
- Allow audio upload in agent mode and render audio with a waveform player. (#16022)
- CLI: support remote audio URLs in generate asr. (#15987)
- Remove the deprecated /webapi/stt route and frontend mic entry. (#15999)
- bot: recover MIME type for QQ c2c file attachments so audio reaches the model. (#16063)

📱 Devices, CLI & Integrations
- Distinguish local and gateway device targets. (#15921)
- Inject device-bound working directory into the server local-system tool chain. (#15887)
- Add a recent-directory button to the device detail panel. (#16040)
- Auto-activate a device for bot triggers on a local target. (#16032)
- Tolerate CRLF line endings when editing files & surface real edit errors. (#16061)
- CLI: stop connect daemon on logout. (#16038)
- Subscription: add contextual plans modal header i18n keys. (#16025)
- Linear: render get_issue result as a card instead of raw JSON. (#15953)

🖥️ Chat & User Experience
- Collapse FloatingChatPanel into a slim strip by default. (#15991)
- Recommend Market skills during agent creation. (#16016)
- Show a [Draft] hint on topics holding unsent input; persist topic unread as a backend status. (#15996, #16056)
- Polish chat-input attachment menu, agent mode & status hints; polish skill suggestion modal UX. (#16010, #16027)
- Lead collapsed tool-call summary with total call count. (#16003)
- Clarify skill installation flow. (#16034)
- Agent-documents: present documents as clickable links; harden explorer tree, titles, empty state & default tab. (#15989, #15998, #16005)
- Make working-directory clear button actually clear; hide Codex cloud config tab. (#16019, #16037)
- SPA: bootstrap app initialization; add use-switch-workspace hook. (#15937, #15979)
- Desktop: require re-auth for expired sessions; close DevTools with window shortcut. (#16014, #16023)
- Add a product glossary page. (#16001)

🔒 Reliability & Security
- Security: skip Market trusted token for synthetic eval userIds. (#15959)
- Treat parked runs as non-terminal in the client run-lifecycle. (#16072)
- Classify upstream-message fallbacks out of UpstreamHttpError; converge 2nd residue wave + unmap payload-too-large from ECW. (#16024, #16053)
- Preserve gateway error semantics and event body. (#16058, #16069)
- Hetero-agent: gate CC usage-limit guide on rejected status; don't treat an inline tool's task_notification as a post-task summary. (#16042, #16050)
- Load snapshot store via static import, not dynamic require. (#15982)
- Reject duplicate custom model ids; allow same-user generic edit-lock updates. (#15975, #15969)
- Avoid flashing unauthorized state in the tool auth alert; avoid task-template locale key flash. (#16068, #16071)
- Prevent blank screen when switching language; resolve resource header i18n namespace fallback. (#15977, #16066)
- Provide Market auth to SPA modals; suppress agent-mode notice for CLI agents. (#16045, #16044)
- Handle generated asset metadata dedup; cache & gate task-template recommendations. (#16057, #15993, #16013)
- userMemories: delete persona document when clearing all memories. (#15997)

🔧 Tooling & Maintenance
- Upgrade react-router to v8. (#16029)
- Use server config for business features; remove Fable campaign paths and client-side community publish/unpublish entries. (#16015, #15960, #15986)
- Drop dead markUserValidAction business slot. (#16018)
- Expose memory-user-memory prompts subpath exports; fix skill audit symlink inventory. (#15985, #16039)
- Translate non-English comments to English in scripts; clean up LOBE-XXX markers. (#16000, #15943, #16031)

👥 Contributors
Huge thanks to 7 contributors who shipped 88 merged PRs this cycle.
@hardy-one · @arvinxx · @tjx666 · @AmAzing129 · @rivertwilight · @Innei · @Rdmclin2
Plus @lobehubbot and renovate[bot] for maintenance.

Full Changelog: v2.2.6...release/weekly-20260620-recut