v2.2.16
🚀 Release v2.2.16
This release was automatically published from PR #19137.
Changes
See PR description: https://github.com/lobehub/lobehub/pull/19137
Commit Message
🚀 LobeHub Release (v2.2.16)
Release Date: September 3, 2026
Since v2.2.15: 246 merged PRs · 9 contributors
This release makes long-running Goals easier to plan, supervise, and accept; adds direct Agent sharing and a project-scoped Acceptance workspace; broadens local heterogeneous Agent support; and strengthens search, CLI, collaboration, performance, and gateway reliability.
✨ Highlights
- More heterogeneous Agents, with safer bindings — Factory Droid joins through ACP, TRAE and Grok Build gain provider binding support, and the deployment-provided LobeHub connection expands across Claude Code, Codex, Kimi Code, Pi, and Grok Build. (#18737, #18775, #18785, #18811, #18932)
- Full-text search gains an Elasticsearch path — A provider-neutral search layer, durable outbox, resumable indexing, operational recovery, and an opt-in Docker Compose profile make Elasticsearch practical without changing the default self-hosted setup. (#18807, #18902, #19028)
- The CLI covers more of the Agent workflow — Workspace selection, resilient Goal execution, operation inspection and replay, execution snapshots, and stronger daemon readiness checks make
lhmore useful for repeatable local and remote workflows. (#18797, #18874, #19000, #19015, #19020) - Workspace collaboration becomes more actionable — Document comment threads gain participant notifications and deep links, shared documents can receive likes, and Tasks add assignments, mentions, and member-aware tool actions. (#18750, #18999, #19035, #19039)
- Large workloads feel lighter — A CI-enforced first-screen graph cut, off-main-thread hashing, streaming upload paths, and larger PDF reads reduce startup and file-processing stalls. (#18945, #19033, #19065, #19073)
- Gateway and runtime failures recover more clearly — Device handshake deadlines, explicit device-read failures, recoverable offline-device states, bounded tool-loop handling, and synthetic tool-failure diagnosis prevent silent fallbacks and stuck execution. (#18965, #19021, #19022, #19069, #19116)
🏗️ Core Agent & Architecture
Goals, Tasks & Acceptance
- Goal planning now records exploration structure, schedules independent Tasks in parallel, and bounds Goal reads, lifecycle transitions, and deadlines. (#18942, #18944, #18970)
- Goal creation and execution stay non-blocking, while the Goal page adds contextual chat, liveness heartbeats, explicit process controls, structured acceptance criteria, and a final result-review view. (#18979, #19050, #19124)
- Acceptance lists support pagination and batch actions, and reports, decisions, and evidence are clearer during review and reruns. (#18843, #18844, #18896, #18897, #18901)
- Projects now host their own Acceptance workspace, including scoped queries, contextual empty states, and a notice when a shared viewer cannot perform review actions. (#19042, #19064)
Runtime, Context & Agent Work
- Agent Work verification no longer redispatches an already delivered Work item, and gateway tool-state updates are coalesced to reduce redundant traffic. (#18969, #19060)
- Repeated tool loops are bounded, synthetic tool failures retain their reason and tool identity, and network-success responses with empty completions can be retried safely. (#18965, #19083, #19116)
- Heterogeneous Agent lifecycle hooks persist across operations, while operation completion propagates through the hook chain. (#18408, #18946)
- Agent documents keep their folder relationships, and workspace resource management is less restrictive where ownership and permissions already allow the operation. (#18964, #19107)
📱 Platforms & Integrations
Heterogeneous Agents
- Factory Droid is available as a first-class ACP Agent with session resume, image input, cancellation, model configuration, and permission interventions. (#18775)
- TRAE supports provider binding through OpenAI Responses, including isolated profiles and preserved authentication across model bindings. (#18811, #18932)
- The deployment-provided LobeHub connection covers more local Agent drivers with operation-scoped credentials and capability checks, including Grok Build support. (#18737, #18785)
Desktop, Gateway & Messaging
- Device clients time out black-holed handshakes, distinguish gateway read failures from an empty device list, and expose offline-device failures as recoverable states instead of silently degrading execution. (#19021, #19022, #19069)
- Desktop renderer updates use the next OTA flow, update notifications are clearer, and listener and root cleanup reduces stale IPC state. (#18745, #18839, #18954)
- Bot topics can be filtered by source, background Task and evaluation runs avoid noisy completion pushes, and Telegram webhooks gain verification hardening. (#18788, #18866, #18930)
- Mobile Agent search aligns with the shared search contract, while chat image proxying and media references improve image delivery across integrations. (#18743, #18784, #18835)
🖥️ User Experience
- Direct Agent links provide a controlled visitor-chat surface with configurable grants, limits, and exposure settings. (#18998)
- The Goal experience adds an exploration map, a detail portal, compact running-process controls, and a dedicated result-review view. (#18928, #18942, #18921, #19124)
- Workspace browsing gains a more capable explorer, Agent Overview and Workspace are separated more clearly, and Agent Documents are distinguished from knowledge-base lookup. (#18845, #18892, #19085)
- Local files gain a focused preview toolbar, large PDFs can be read without the earlier cap, and
.vand.svfiles are recognized as attachments. (#18758, #19073, #19118) - Gemini 3.8 Flash is available with the supported thinking levels, and React Artifact previews boot reliably in Sandpack. (#19058)
🔧 Search, CLI & Tooling
- Full-text search supports PostgreSQL and Elasticsearch behind one repository contract, with permission-aware hydration, durable incremental sync, and resumable reindexing. (#18807)
- Elasticsearch deployments can recover interrupted indexing, while self-hosters can opt into a single-node Docker Compose profile without changing the default stack. (#18902, #19028)
lh workspacelists, inspects, and switches workspace scope;lh goal runbecomes more resilient; and daemon startup waits for actual readiness. (#19000, #19015, #19020)- Signed operation snapshots can be inspected and replayed across models or as a full trajectory, and CLI execution snapshots make runs easier to diagnose. (#18796, #18797, #18874)
- Evaluation cases can be captured from a conversation turn, and the Mesurer inspector improves local investigation of application behavior. (#18889, #18955)
🔒 Security, Reliability & Performance
- Security: Trusted authentication origins and apex Desktop redirects are handled explicitly, while direct Agent sharing keeps spend caps and capability grants under the creator's control. (#18924, #18925, #18998)
- Security: Signed trace snapshots and heterogeneous provider bindings scope credentials to the intended operation and Agent capability. (#18785, #18796)
- Reliability: Model and runtime failures preserve clearer classifications and reasoning details, and image tool results are no longer misreported as failed. (#18795, #19004, #19005, #19006)
- Reliability: Device connectivity failures no longer masquerade as an empty online set or leave clients indefinitely parked in a connecting state. (#19021, #19022)
- Performance: The first-screen static import graph is CI-gated and 41% smaller, large-file hashing moves off the UI thread, and uploads stream with bounded memory. (#18945, #19033, #19065)
- Observability: Model, router, search-cost, and time-series metrics add clearer attribution for runtime behavior and operational trends. (#19012, #19063, #19071)
👥 Contributors
Huge thanks to 9 contributors who shipped 246 merged PRs this cycle.
@arvinxx · @Innei · @tjx666 · @rdmclin2 · @AmAzing129 · @ONLY-yours · @sudongyuer · @nekomeowww · @cy948
Full Changelog: https://github.com/lobehub/lobehub/compare/v2.2.15...release/v2.2.16