OpenClaw v2026.5.12-beta.4 Release Summary

Key Changes

Plugin SDK & Workflow Orchestration

This release marks a major leap in the Plugin SDK, consolidating several workflow seams to empower plugin authors. The new infrastructure allows plugins to define typed session actions, send session-bound attachments through host-owned delivery, and schedule future session turns via cron-compatible plumbing. Additionally, a new finalize retry mechanism allows plugins to request bounded retries before a reply is finalized, enabling more sophisticated quality-control loops.

Agent & Session Management

Subagent coordination has been significantly hardened. A critical fix now preserves active background exec process references across session compaction, ensuring that long-running tasks are not orphaned when the context window is refreshed. Furthermore, the maxPingPongTurns ceiling for agent-to-agent communication has been raised from 5 to 20, facilitating more complex multi-agent coordination chains.

Model & Provider Enhancements

Support for self-hosted llama.cpp providers has been improved by implementing opportunistic /props discovery, allowing the system to accurately reflect the actual loaded runtime context window (n_ctx) rather than just the training ceiling. For OpenAI-compatible models, the system now correctly drains split provider stream frames, preventing session stalls during fragmented SSE or JSON responses.

Channel Integration Updates

• Slack: Added allowBots: "mentions" mode for parity with Discord, preventing infinite loops while allowing explicit bot-to-bot communication. Formatting hints for mrkdwn have also been added to ensure agents use Slack-specific syntax.
• Telegram: Fixed a critical issue where topic context from before a session start could be replayed into later turns. The /status and /think menus now correctly reflect resolved model thinking defaults.
• WhatsApp: Improved shutdown reliability by ensuring debounced inbound messages are fully drained before the socket closes.
• iMessage: Added support for tapback reactions as inbound agent events and provided a clear migration path from BlueBubbles to the native imsg path.

Infrastructure & Tooling

• Build System: The workspace has been upgraded to pnpm 11, and build utility paths now use Node's native --experimental-strip-types to eliminate deprecation warnings in Node 26.
• Gateway: Improved the robustness of session-kill HTTP paths by returning deterministic 400 invalid_request_error responses for malformed encoded keys.
• Memory Core: Introduced a bounded compaction step for MEMORY.md during dreaming promotions, preventing unbounded file growth that could lead to gateway freezes.

Impact

For Plugin Developers

Developers can now build more interactive and autonomous plugins. The ability to register sessionActions means plugins can trigger specific, validated workflows that the Gateway can dispatch. The new scheduling and attachment APIs remove the need for plugins to manage their own low-level delivery or timing logic.

For Power Users & Operators

• Reliability: Users running long-running subagent tasks will see a significant decrease in orphaned processes. Those using self-hosted models will see more accurate token budgeting in the Sessions table.
• Observability: Gateway logs now explicitly signal when embedded runs are undergoing auto-compaction, making it easier to distinguish between a healthy context refresh and a stalled agent.
• Diagnostics: openclaw doctor now provides a specific hint for GH_CONFIG_DIR mismatches, simplifying the setup of the GitHub skill in service environments where HOME is redirected.

For End Users

Users on Telegram and Slack will experience more consistent formatting and better session isolation. The addition of iMessage tapback support allows agents to react to user feedback more naturally.