By devasher · Edited by Nominiclaw
openclaw v2026.6.8-beta.1 introduces rich text delivery for Telegram, native usage footer rendering, expanded model support including GLM-5.2 and Claude Haiku 4.5, and critical stability fixes for Gateway recovery and memory indexing.
openclaw v2026.6.8-beta.1 delivers enhanced channel delivery capabilities, native usage tracking visualization, expanded LLM provider support, and significant reliability improvements for agent recovery and state management.
Telegram delivery is now richer and more resilient, supporting structured rich text with tables, lists, and expandable blockquotes. The release also introduces a native CLI backend delivery path for Telegram and honors configured ACP bindings for WhatsApp delivery. Additionally, Slack outbound message_sent hooks are now emitted across all delivery paths, including streaming and TTS supplemental media.
Users can now render a full /usage full footer in-core using a declarative template DSL, removing the requirement for a renderer plugin. This system includes a built-in compact footer by default, with support for custom templates via messages.usageTemplate. The reply_payload_sending plugin hook now attaches a per-turn usageState snapshot, providing model, usage, cost, and timing metadata without requiring side calls.
Model handling has been expanded with the addition of GLM-5.2 (default for Z.AI Coding Plan endpoints) and Claude Haiku 4.5 catalog entries. Provider-prefix normalization has been tightened for OpenRouter and Google Vertex to prevent invalid API paths. Additionally, the system now supports /btw and /side commands in CLI-backed sessions, enabling ephemeral side-questions that do not pollute main chat history.
Recovery mechanisms are now sharper across account-scoped DM sends and generated media completions. A critical fix ensures that active main sessions are marked for restart recovery before a Gateway shutdown abort occurs. Subagent runs that yield while waiting on children are no longer falsely settled as cancelled when the yield-induced abort signal is tripped.
Memory indexing is more robust, with the ability to split OpenAI embedding batches that trigger 431 request_headers_too_large errors. QMD memory search now operates in a transient mode for one-shot agent turns to avoid unnecessary background lifecycle resources. To prevent state corruption on NFS volumes, SQLite now uses rollback journaling instead of WAL mode when an NFS mount is detected.
WebChat now preserves backscroll during streaming, preventing the view from forcing a scroll-to-bottom when a user has scrolled up. The Control UI now starts workspace files collapsed by default and restores sidebar session picker interactivity above the desktop workbench. On iOS, the app now forces reconnects for stale foreground gateways to ensure a responsive connection upon resume.
nominiclaw/nodes documentation now includes a consolidated openclaw.json example. The before_install hook is clarified as a plugin-runtime lifecycle hook, while security.installPolicy is the primary operator-owned install boundary.operator.admin privileges to prevent spoofing via requester headers.infer embedding create command now surfaces nested fetch failures (e.g., Undici error codes) instead of generic TypeError: fetch failed messages./model command confirmation now displays the resolved canonical model reference instead of the raw alias.This release adds rich text support for Telegram, native templated usage footers, support for GLM-5.2 and Claude Haiku 4.5, and improved recovery for Gateway restarts and subagent yields.
Yes, security changes now require operator.admin for HTTP session kills and model overrides, and non-admin HTTP requester-session kills are no longer accepted.
The provided source does not contain specific upgrade instructions or migration commands.